CN114676176A

CN114676176A - Time series prediction method, device, equipment and program product

Info

Publication number: CN114676176A
Application number: CN202210303163.1A
Authority: CN
Inventors: 路杰程; 王硕佳; 叶碧荣
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2022-03-24
Filing date: 2022-03-24
Publication date: 2022-06-28
Anticipated expiration: 2042-03-24

Abstract

The application provides a time series prediction method, a device, equipment, a storage medium and a program product; the method comprises the following steps: acquiring a first historical time sequence and a plurality of second historical time sequences; carrying out causal test treatment on the first historical time sequence and each second historical time sequence; assigning each second historical time series to a set of significant sequences or a set of non-significant sequences; performing incremental verification processing on each second historical time sequence in the non-significant set based on the significant set, the non-significant set and the first historical time sequence; determining a second history time sequence of which the incremental test processing result meets the explanatory condition as a third history time sequence; and the third historical time sequence is used for the prediction processing of a plurality of base class prediction models to obtain a prediction time sequence corresponding to the first historical time sequence. By the method and the device, the interpretability of the screened historical time sequence can be effectively improved, and a decision reference is provided for prediction.

Description

Time series prediction method, device, equipment and program product

Technical Field

The present application relates to the field of artificial intelligence technologies, and in particular, to a time series prediction method, apparatus, device, storage medium, and program product.

Background

Artificial Intelligence (AI) is a theory, method, technique and application system that uses a digital computer or a machine controlled by a digital computer to simulate, extend and expand human Intelligence, perceive the environment, acquire knowledge and use the knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technique of computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a manner similar to human intelligence. Artificial intelligence is the research of the design principle and the realization method of various intelligent machines, so that the machines have the functions of perception, reasoning and decision making.

In the related art, the historical time series are usually screened manually by a priori knowledge, and the screened historical time series has poor interpretability due to the fact that the manual screening needs a large amount of accumulated a priori knowledge and the reserve of the a priori knowledge is insufficient.

The related art has no effective solution for how to effectively improve the interpretability of the screened historical time series.

Disclosure of Invention

The embodiment of the application provides a time sequence prediction method, a time sequence prediction device, a computer readable storage medium and a computer program product, which can effectively improve the interpretability of a screened historical time sequence and provide decision reference for prediction.

The technical scheme of the embodiment of the application is realized as follows:

the embodiment of the application provides a time series prediction method, which comprises the following steps:

acquiring a first historical time sequence and a plurality of second historical time sequences, wherein the second historical time sequences are covariates of the first historical time sequence, and stationarity indexes of the second historical time sequences are smaller than a stationarity threshold;

respectively carrying out causal test treatment on the first historical time sequence and each second historical time sequence;

assigning each of the second historical time series to a set of significant sequences or a set of non-significant sequences based on causal test processing results;

performing incremental inspection processing on each second historical time sequence in the non-significant set based on the significant set, the non-significant set and the first historical time sequence to obtain an incremental inspection processing result;

determining the second historical time sequence of which the incremental test processing result meets an explanatory condition as a third historical time sequence; and the third historical time sequence is used for prediction processing of a plurality of base class prediction models to obtain a prediction time sequence corresponding to the first historical time sequence.

The embodiment of the present application provides a time series prediction apparatus, including:

the device comprises an acquisition module, a processing module and a processing module, wherein the acquisition module is used for acquiring a first historical time sequence and a plurality of second historical time sequences, the second historical time sequences are covariates of the first historical time sequence, and stationarity indexes of the second historical time sequences are smaller than a stationarity threshold;

the causal test module is used for carrying out causal test treatment on the first historical time sequence and each second historical time sequence;

an assignment module to assign each of the second historical time series to a significant sequence set or a non-significant sequence set based on a causal test treatment result;

an increment inspection module, configured to perform increment inspection processing on each second historical time sequence in the non-significant set based on the significant set, the non-significant set, and the first historical time sequence to obtain an increment inspection processing result;

the determining module is used for determining the second historical time sequence of which the incremental test processing result meets the explanatory condition as a third historical time sequence; and the third historical time sequence is used for prediction processing of a plurality of base class prediction models to obtain a prediction time sequence corresponding to the first historical time sequence.

An embodiment of the present application provides an electronic device, including:

a memory for storing executable instructions;

and the processor is used for realizing the time series prediction method provided by the embodiment of the application when executing the executable instructions stored in the memory.

The embodiment of the present application provides a computer-readable storage medium, which stores executable instructions for causing a processor to implement the time series prediction method provided by the embodiment of the present application when executed.

Embodiments of the present application provide a computer program product or computer program comprising computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions to cause the computer device to execute the time series prediction method according to the embodiment of the present application.

The embodiment of the application has the following beneficial effects:

and setting interpretability conditions through causal inspection processing and incremental interpretation processing, and screening the acquired second historical time sequence, so that the reason that each acquired third historical time sequence is introduced into a subsequent base class prediction model is clear, the influence mechanism of each third historical time sequence on the first historical time sequence is reflected, the screened third historical time sequence has interpretability, the interpretability of the screened historical time sequence is effectively improved, and a decision reference is provided for prediction.

Drawings

Fig. 1 is a schematic structural diagram of a time-series prediction system architecture provided in an embodiment of the present application;

FIG. 2 is a schematic structural diagram of a prediction server of a time series provided by an embodiment of the present application;

fig. 3A to fig. 3G are schematic flow charts of a time series prediction method provided in an embodiment of the present application;

FIG. 4 is a schematic diagram illustrating a time-series prediction method according to an embodiment of the present disclosure;

fig. 5A to 5G are schematic diagrams illustrating a time series prediction method according to an embodiment of the present application.

Detailed Description

In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be described in further detail with reference to the attached drawings, the described embodiments should not be considered as limiting the present application, and all other embodiments obtained by a person of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

In the following description, reference is made to "some embodiments" which describe a subset of all possible embodiments, but it is understood that "some embodiments" may be the same subset or different subsets of all possible embodiments, and may be combined with each other without conflict.

In the following description, references to the terms "first \ second \ third" are only to distinguish similar objects and do not denote a particular order, but rather the terms "first \ second \ third" are used to interchange specific orders or sequences, where appropriate, so as to enable the embodiments of the application described herein to be practiced in other than the order shown or described herein.

Unless defined otherwise, all technical and scientific terms used in the examples of this application have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the embodiments of the present application is for the purpose of describing the embodiments of the present application only and is not intended to be limiting of the present application.

Before further detailed description of the embodiments of the present application, terms and expressions referred to in the embodiments of the present application will be described, and the terms and expressions referred to in the embodiments of the present application will be used for the following explanation.

(1) Covariate (Covariate): also known as an explanatory variable, is a variable that is linearly related to a dependent variable and is controlled by statistical techniques when discussing the relationship between independent and dependent variables. Commonly used covariates include premeasured scores for the dependent variables, demographic indicators, and characteristics that are significantly different from the dependent variables, among others. For example, when the dependent variable is a variable, the covariate may be a demographic variable, an economic variable, and the like.

(2) Time Series (Time Series): the time series (or called dynamic number series) refers to a number series formed by arranging the numerical values of the same statistical index according to the occurrence time sequence. The main purpose of time series prediction is to predict future time series based on the existing historical time series. Most of the economic data is given in time series. The time in the time series may be year, quarter, month or any other form of time depending on the time of observation.

(3) Dependent Variable (Dependent Variable): is a quantity that directly (depending on the purpose) causes a fluctuation due to a fluctuation in an independent variable, and is linearly related to a gradient variable.

(4) Granger cause and effect Test (Granger cause Test): in the time series case, the granger causal relationship between two variables X, Y is defined as: if the effect of predicting the variable Y under the condition including the past information of the variable X, Y is better than the effect of predicting Y only from the past information of Y alone, that is, if the variable X helps to explain the future change of the variable Y, the variable X is considered to be the cause of the guillain of the variable Y.

(5) Confidence (creditability): to the extent that it is trusted by a person or thing. Is the degree of confidence that a thing or thing is true empirically.

In the implementation process of the embodiment of the present application, the applicant finds that the following problems exist in the related art:

in the related art, for the screening of the historical time series, the historical time series is usually screened manually through priori knowledge, the screened historical time series lacks interpretability due to the fact that a large amount of priori knowledge is needed for manual screening and the priori knowledge is insufficient in reserve, and meanwhile, the statistical significance of the screened historical time series cannot be guaranteed due to the fact that the priori knowledge is different under different scenes and the uncertainty of the priori knowledge.

In the related art, the prediction of the historical time series is usually realized through a single prediction model, and because the single prediction model lacks multi-dimensional information input, and the model performance of the single prediction model is relatively single and has certain bias, the reliability and the credibility of the prediction result of the prediction model are lower.

According to the prediction method of the time series provided by the embodiment of the application, on one hand, the acquired second historical time series are screened by setting interpretability conditions through causal inspection processing and incremental interpretation processing, so that the reason that each acquired third historical time series is introduced into a subsequent base class prediction model is clear, the influence mechanism of each third historical time series on the first historical time series is reflected, the screened third historical time series has interpretability, and the interpretability of the screened historical time series is effectively improved. On the other hand, the prediction time sequence corresponding to the first historical time sequence is obtained through the prediction processing of the base class prediction models, so that the model performance of the base class prediction models is integrated, and the accuracy of the predicted prediction time sequence is effectively improved.

Embodiments of the present application provide a method, an apparatus, a device, a computer-readable storage medium, and a computer program product for predicting a time sequence, which can effectively improve interpretability of a filtered historical time sequence and provide a decision reference for prediction. An exemplary application of the time-series prediction apparatus provided in the embodiments of the present application is described below, and the apparatus provided in the embodiments of the present application may be implemented as various types of user terminals such as a notebook computer, a tablet computer, a desktop computer, a set-top box, a mobile device (e.g., a mobile phone, a portable music player, a personal digital assistant, a dedicated messaging device, and a portable game device), and may also be implemented as a server.

Referring to fig. 1, fig. 1 is a schematic diagram of an architecture of a prediction system 100 for time series provided in an embodiment of the present application, a terminal (an exemplary terminal 400 is shown) is connected to a server 200 through a network 300, and the network 300 may be a wide area network or a local area network, or a combination of both.

The terminal 400 is configured for use by a user of the client 410 for display on a graphical interface 410-1 (graphical interface 410-1 is illustratively shown). The terminal 400 and the server 200 are connected to each other through a wired or wireless network.

In some embodiments, the server 200 may be an independent physical server, may also be a server cluster or a distributed system formed by a plurality of physical servers, and may also be a cloud server providing basic cloud computing services such as a cloud service, a cloud database, cloud computing, a cloud function, cloud storage, a network service, cloud communication, a middleware service, a domain name service, a security service, a CDN, and a big data and artificial intelligence platform. The terminal 400 may be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, a smart voice interaction device, a smart home appliance, a vehicle-mounted terminal, etc., but is not limited thereto. The terminal and the server may be directly or indirectly connected through wired or wireless communication, and the embodiment of the present application is not limited.

In some embodiments, the client of the terminal 400 obtains the first historical time series and the plurality of second historical time series, and sends the plurality of second historical time series to the server 200 through the network 300, and the server 200 performs screening and inference on the plurality of second historical time series to obtain a third historical time series, and sends the third historical time series to the terminal 400 for prediction processing to obtain a predicted time series corresponding to the first historical time series.

In other embodiments, the server 200 obtains the first historical time series and the plurality of second historical time series, performs screening and inference on the plurality of second historical time series to obtain a third historical time series, performs prediction processing based on the third historical time series to obtain a predicted time series corresponding to the first historical time series, and sends the predicted time series to the graphical interface 410-1 in the terminal 400 for display.

Referring to fig. 2, fig. 2 is a schematic structural diagram of a server 200 of a prediction method of a time series according to an embodiment of the present application, where the server 200 shown in fig. 2 includes: at least one processor 210, memory 250, at least one network interface 220. The various components in server 200 are coupled together by a bus system 240. It is understood that the bus system 240 is used to enable communications among the components. The bus system 240 includes a power bus, a control bus, and a status signal bus in addition to a data bus. For clarity of illustration, however, the various buses are labeled as bus system 240 in fig. 2.

The Processor 210 may be an integrated circuit chip having Signal processing capabilities, such as a general purpose Processor, a Digital Signal Processor (DSP), or other programmable logic device, discrete gate or transistor logic device, discrete hardware components, or the like, wherein the general purpose Processor may be a microprocessor or any conventional Processor, or the like.

The memory 250 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid state memory, hard disk drives, optical disk drives, and the like. Memory 250 optionally includes one or more storage devices physically located remote from processor 210.

The memory 250 includes volatile memory or nonvolatile memory, and can also include both volatile and nonvolatile memory. The nonvolatile Memory may be a Read Only Memory (ROM), and the volatile Memory may be a Random Access Memory (RAM). The memory 250 described in embodiments herein is intended to comprise any suitable type of memory.

In some embodiments, memory 250 is capable of storing data, examples of which include programs, modules, and data structures, or a subset or superset thereof, to support various operations, as exemplified below.

The operating system 251, which includes system programs for handling various basic system services and performing hardware-related tasks, such as a framework layer, a core library layer, a driver layer, etc., is used for implementing various basic services and for processing hardware-based tasks.

A network communication module 252 for communicating to other electronic devices via one or more (wired or wireless) network interfaces 220, exemplary network interfaces 220 including: bluetooth, wireless compatibility authentication (WiFi), and Universal Serial Bus (USB), among others.

In some embodiments, the prediction apparatus of the time series provided by the embodiments of the present application may be implemented in software, and fig. 2 shows a processing apparatus 255 of a three-dimensional face model stored in a memory 250, which may be software in the form of programs and plug-ins, and the like, and includes the following software modules: the acquisition module 2551, the cause and effect check module 2552, the assignment module 2553, the incremental check module 2554, the determination module 2555, which are logical and therefore can be arbitrarily combined or further split depending on the functionality implemented. The functions of the respective modules will be explained below.

The time series prediction method provided by the embodiment of the present application will be described in conjunction with an exemplary application and implementation of the server provided by the embodiment of the present application.

In some embodiments, fig. 4 is a schematic diagram illustrating a time-series prediction method provided by an embodiment of the present application. Referring to fig. 4, the first historical time series is subjected to causal test processing with each second historical time series, and each second historical time series is allocated to a significant sequence set or a non-significant sequence set based on the causal test processing result; based on the significant set, the non-significant set and the first historical time sequence, performing incremental inspection processing on each second historical time sequence in the non-significant set to obtain an incremental inspection processing result, determining the second historical time sequence of which the incremental inspection processing result meets an explanatory condition as a third historical time sequence, and performing causal inspection processing and incremental inspection processing to ensure that each third historical time sequence screened from the second historical time sequence can accurately determine the influence mechanism of the third historical time sequence on the first historical time sequence, thereby effectively improving the interpretability of the screened historical time sequence.

Referring to fig. 3A, fig. 3A is a schematic flowchart of a time-series prediction method provided in an embodiment of the present application, and will be described with reference to steps 101 to 105 shown in fig. 3A, where an execution subject of the following steps 101 to 105 may be the aforementioned server or terminal.

In step 101, a first historical time series and a plurality of second historical time series are obtained.

And the second historical time sequence is a covariate of the first historical time sequence, and the stationarity index of the second historical time sequence is smaller than the stationarity threshold.

As an example, in an application scenario of a popular investigation, the first historical time series may be a historical time series of an epidemic that characterizes a trend of the epidemic over time over a past period of time. The second historical time series may include a population flow historical time series that characterizes a trend of the population flow over time over a past period of time, an atmospheric temperature historical time series that characterizes a trend of the atmospheric temperature over time over a past period of time, and the like. It will be appreciated that both the population flow factor and the atmospheric temperature factor affect the prevalence of the epidemic, i.e., the second historical time series is linearly related to the first historical time series, i.e., the first historical time series is a dependent variable of the second historical time series and the second historical time series is a covariate of the first historical time series.

As an example, in an application scenario of revenue and expenditure prediction, the first historical time series may be a historical time series of revenue and expenditure, which characterizes a trend of revenue and expenditure over time of the service over a past period of time. The second historical time series may include a population flow historical time series that characterizes a trend of the population flow over time over a past period of time, an economic indicator historical time series that characterizes a trend of the economic indicator over time over a past period of time, and the like. It is understood that both population mobility factors and economic indicator factors can affect the changing trends of revenue and expenditure.

As an example, in an application scenario of automatic driving, the first historical time series may be a historical time series of track points during past driving of the vehicle, and the historical time series of track points represents a change trend of a driving track of the vehicle over a past period of time. The second historical time series may include a manipulation time series of the driver, and the like.

As an example, in an application scenario of disk usage prediction, the first historical time series may be a historical time series of disk usage that characterizes usage of the disk over a past period of time. The second historical time series may include a disk capacity historical time series, a resource quantity historical time series, and the like.

As an example, in an application scenario of log access, the first historical time series may be a historical access time series of the log that characterizes access to the log over a past period of time. The second historical time series may include visitor size historical time series, and the like.

As an example, in the application scenario of stock prediction, the first historical time series may be a historical time series of the rise and fall trend of a certain stock, and the historical time series of the rise and fall trend represents the change trend of the rise and fall situation of the stock over a past period of time. The second historical time series may include an enterprise business situation historical time series that characterizes a trend of population flow over time over a past period of time, an economic indicator historical time series that characterizes a trend of economic indicator over time over a past period of time, and the like. It can be understood that the enterprise business situation and the economic index factor both affect the change trend of the enterprise stocks.

In some embodiments, the stationarity indicator of the second historical time series may be determined by: the following processing is performed for any one of the second historical time series: determining each step difference historical time sequence of the second historical time sequence, performing stationarity check processing on the second historical time sequence and each step difference historical time sequence of the second historical time sequence to obtain first stationarity indexes respectively corresponding to the second historical time sequence and each step difference historical time sequence of the second historical time sequence, and performing fusion processing on the first stationarity indexes to obtain stationarity indexes of the second historical time sequence.

In some embodiments, the stationarity testing process includes ADF testing (Augmented Dickey-Fuller), PP testing (Phillips And Perron), KPSS testing (Kwiatkowski-Phillips-Schmidt-Shin), NP testing, ERS testing, And DF testing, among others.

Therefore, when a large number of second historical time sequences are obtained, the second historical time sequences with the stationarity indexes smaller than the stationarity threshold are determined to be the second historical time sequences of the subsequent causal test treatment through stationarity test treatment, so that the preliminary screening of the second historical time sequences is realized, the second historical time sequences with overlarge volatility (abnormity) are deleted, and the effectiveness of the second historical time sequences is effectively guaranteed.

In step 102, the first historical time series is subjected to a causal test with each second historical time series.

In some embodiments, referring to fig. 5B, fig. 5B is a schematic flowchart of a prediction method for time series according to an embodiment of the present application. The above step 102 can be implemented by: and performing causal test treatment on the first historical time sequence and each second historical time sequence with the stationarity index smaller than the stationarity threshold respectively to obtain a causal test treatment result of each second historical time sequence with the stationarity index smaller than the stationarity threshold.

In some embodiments, the causal test process is used to test the extent of causal association of the first historical time series and the second historical time series.

In some embodiments, referring to fig. 3B, fig. 3B is a flowchart illustrating a method for predicting a time series according to an embodiment of the present application, where the causal test process includes a variance ratio test process, and step 102 shown in fig. 3B may perform the following steps 1021 to 1023 on any one of the second historical time series to determine a causal test process result.

In step 1021, a first time value corresponding to the first historical time series is determined based on the sequence of hysteresis values of order p of the second historical time series and the sequence of hysteresis values of order q of the first historical time series.

The p-order hysteresis value sequence comprises values of all moments of the second historical time sequence within a range from the moment p to the current moment, the q-order hysteresis value sequence comprises values of all moments of the first historical time sequence within a range from the moment q to the current moment, the moment p and the moment q are any moments before the current moment, and the first moment value is any moment after the current moment.

In some embodiments, the range of values for p may be: p is more than or equal to 0 and less than t, and the value range of q can be as follows: q is more than or equal to 0 and less than t, wherein t represents the current moment.

As an example, when p is 3 seconds(s), q is 5s, the current time is 10s, and the first time value is 12s, the 3-order hysteresis value sequence of the second historical time sequence may be {1, 2, 3, 4, 7, 8, 9, 10}, where 1 represents a value of the second historical time sequence at the time of 3s, 2 represents a value of the second historical time sequence at the time of 4s, 3 represents a value of the second historical time sequence at the time of 5s, and 10 represents a value of the second historical time sequence at the time of 10 s. The sequence of hysteresis values of order 5 of the first historical time series may be {11, 12, 15, 19, 18, 16}, where 11 represents a value of the first historical time series at a time 5s, 12 represents a value of the first historical time series at a time 6s, and 16 represents a value of the first historical time series at a time 10 s.

As an example, the expression of the first time value corresponding to the first historical time series may be:

wherein, Y_mtCharacterizing a first time value, X, corresponding to a first historical time series_nt-iSequence of hysteresis values of order p, Y, characterizing a second historical time sequence_mt-jA sequence of hysteresis values of order q characterizing a first historical time sequence, e_mtCharacterizing a first tuning parameter, beta_m0Characterizing a second tuning parameter, beta_m1iCharacterizing a weight value, β, corresponding to the second historical time series_m2jAnd representing the weight value corresponding to the first historical time sequence.

In some embodiments, the step 1021 may be implemented as follows: carrying out weighted summation processing on each parameter in the p-order lag value sequence of the second historical time sequence to obtain a first weighted summation processing result; carrying out weighted summation processing on each parameter in the q-order lag value sequence of the first historical time sequence to obtain a second weighted summation processing result; and adding the first weighted sum processing result, the second adjusting parameter and the first adjusting parameter to obtain a first time value corresponding to the first historical time sequence.

In step 1022, a second time value corresponding to the first historical time series is determined based on the q-th order lag value series of the first historical time series.

And the second time value and the first time value have the same corresponding time.

In some embodiments, the expression of the second time value corresponding to the first historical time series may be:

wherein, Y_mt1A second time value, Y, corresponding to the first historical time sequence_mt-jA sequence of hysteresis values of order q characterizing a first historical time sequence, e_mtA first adjustment parameter is characterized in that,β_m0characterizing a second tuning parameter, beta_w2jThe weights are characterized.

In some embodiments, the above step 1022 may be implemented by: carrying out weighted summation processing on each parameter in the q-order lag value sequence of the first historical time sequence according to the respective corresponding weight to obtain a third weighted summation processing result; and adding the third weighted sum processing result, the first adjustment parameter and the second adjustment parameter to obtain a second time value corresponding to the first historical time sequence.

In step 1023, the first time value and the second time value are subjected to a variance ratio test, and the obtained variance ratio test result is determined as a causal test result.

In some embodiments, the variance ratio Test process comprises an F-Test process (F-Test), wherein the F-Test process is a Joint Hypothesis Test (Joint Hypothesis Test), which is a Test in which statistical values obey an F-distribution under a Null Hypothesis (H0). It is typically used to analyze statistical models that use more than one parameter to determine whether all or a portion of the parameters in the model are suitable for estimating the mother.

In some embodiments, the expression for the ratio of variance test result may be:

wherein S is₁Characterizing a first time value, S₂The second time value, F, characterizes the variance ratio test result.

In this way, the causal test processing is performed on the first historical time sequence and each second historical time sequence, so that the causal association degree between the first historical time sequence and the second historical time sequence can be accurately tested, and the second historical time sequence can be accurately classified according to the causal test processing result.

In step 103, each second historical time series is assigned to either a significant series set or a non-significant series set based on the causal test process results.

In some embodiments, referring to fig. 3B, step 103 illustrated in fig. 3B may assign each second historical time series to a significant series set or a non-significant series set by performing the following steps 1031 to 1032.

In step 1031, a second historical time instant sequence is assigned to the significant sequence set when the causal test process result is greater than or equal to the causal test threshold.

In some embodiments, the causal test threshold characterizes a critical value of whether a causal relationship of the second historical time-of-day sequence to the first historical time-of-day sequence is significant.

As an example, when the causal test threshold is 10.3 and the causal test process result is 10.3, the causal test process result is equal to the causal test threshold, and the second historical time sequence is assigned to the significant sequence set.

As an example, when the causal test threshold is 10.3 and the causal test treatment result is 10.5, the causal test treatment result is greater than the causal test threshold, and the second historical time sequence is assigned to the significant sequence set.

In step 1032, the second historical time instant sequence is assigned to the non-significant sequence set when the causal test treatment result is less than the causal test threshold.

As an example, when the causal test threshold is 10.3 and the causal test treatment result is 10.1, and the causal test treatment result is less than the causal test threshold, the second historical time sequence is assigned to the non-significant sequence set.

In this way, the causal test threshold and the causal test processing result are compared, so that the second historical time sequence is accurately allocated to the significant sequence set or the non-significant sequence set, and thus the second historical time sequence is accurately classified.

In some embodiments, referring to fig. 3C, fig. 3C is a flowchart illustrating a time series prediction method provided in an embodiment of the present application, and after step 103 shown in fig. 3C, the second historical time series in the significant sequence set and the non-significant sequence set may be grouped by performing steps 106 to 107.

In step 106, a second historical time series of the same type in the significant series set is assigned to the series group of the corresponding type.

As an example, referring to fig. 5C, the second historical time series of the same type in the set of significant sequences is assigned to variable group a through variable group N, where variable group a includes the second historical time series of the same type in the set of significant sequences, and variable group N includes the second historical time series of the same type in the set of significant sequences.

In step 107, a second historical time series of the same type in the set of non-significant sequences is assigned to the set of sequences of the corresponding type.

Wherein the sequence group includes a second historical time series in the set of significant sequences and a second historical time series in the set of non-significant sequences.

As an example, referring to fig. 5C, second historical time series of the same type in the set of insignificant sequences are assigned to variable group a through variable group N, where variable group a includes the second historical time series of the same type in the set of insignificant sequences and variable group N includes the second historical time series of the same type in the set of insignificant sequences.

In step 104, based on the significant set, the non-significant set and the first historical time series, performing incremental inspection processing on each second historical time series in the non-significant set to obtain an incremental inspection processing result.

In some embodiments, referring to fig. 3D, fig. 3D is a flowchart illustrating a prediction method of a time series provided in an embodiment of the present application, and step 104 illustrated in fig. 3D may perform the following steps 1041 to 1045 on any second historical time series in the non-significant set to determine an incremental verification processing result.

In step 1041, a rank of the first test matrix is determined based on the first historical time series, the significant set, and the second historical time series.

In some embodiments, referring to fig. 3E, fig. 3E is a flowchart illustrating a time series prediction method provided in an embodiment of the present application, and step 1041 illustrated in fig. 3E may be implemented by the following steps 10411 to 10415.

In step 10411, a sequence of i-th order lag values for the first historical time series, a sequence of j-th order lag values for each second historical time series in the significant set, and a sequence of k-th order lag values for the second historical time series are determined.

Wherein the i-order lag value comprises: the values of all the moments of the first historical time sequence within the range from the moment i to the current moment; the sequence of hysteresis values of order j includes: the values of all the moments of the second historical time sequence in the significant set from the moment j to the current moment are taken; the sequence of k-th order lag values includes: the second historical time series takes values from the time k to all times within the current time range.

As an example, when i is 2, j is 3, and k is 2, determining a sequence of hysteresis values of order 2 of the first historical time series, a sequence of hysteresis values of order 3 of each second historical time series in the significant set, and a sequence of hysteresis values of order 1 of the second historical time series, wherein the sequence of hysteresis values of order 2 of the first historical time series may be {10, 11}, 10 is a hysteresis value of order 2, and 11 is a hysteresis value of order 1; a 3 rd order lag value sequence {15, 28, 26} of any one second historical time sequence in the significant set, 15 being a 3 rd order lag value, 28 being a 2 nd order lag value, 26 being a 1 st order lag value; the second historical time series may have a 2 nd order lag value of 19, 21, 19 being a 2 nd order lag value, and 21 being a 1 st order lag value.

In step 10412, difference processing is performed on the values at any two different times in the i-order lag value sequence to obtain a first difference processing result.

As an example, when the i-order lag value sequence is {10, 11}, the 2-order lag value 10 and the 1-order lag value 11 are subjected to difference processing, and the resulting first difference processing result is 1.

In step 10413, the following is performed for any sequence of j-order lag values: and carrying out difference processing on the values of any two different moments in the j-order lag value sequence to obtain a second difference processing result.

As an example, when the j-order hysteresis value sequence is {15, 28, 26}, performing difference processing on values of any two different times in the j-order hysteresis value sequence, that is, performing difference processing on the 3-order hysteresis value 15 and the 2-order hysteresis value 28 to obtain a second difference processing result 13; carrying out difference processing on the 3-order lag value 15 and the 1-order lag value 26 to obtain a second difference processing result 11; the 2 nd order lag value 28 and the 1 st order lag value 26 are subjected to difference processing to obtain a second difference processing result 2.

In step 10414, the values at any two different times in the k-order lag value sequence are subjected to difference processing to obtain a third difference processing result.

As an example, when the k-order lag value sequence is {19, 21}, the values at any two different times in the k-order lag value sequence are subjected to difference processing, that is, the 2-order lag value 19 and the 1-order lag value 21 are subjected to difference processing, so as to obtain a second difference processing result 2.

In step 10415, a rank of the first test matrix is determined based on the first difference processing result, the second difference processing result, and the third difference processing result.

In some embodiments, the expression for the first check matrix may be:

therein, pi_iCharacterizing a first inspection matrix, x_t-iCharacterizing the first difference processing result, the second difference processing result and the third difference processing result, and belonging to_tCharacterizing a third parameter, c characterizing a first constant, x_tAnd representing the value of the current moment of the first historical time sequence.

By way of example, by means of a first check matrix Π_iAnd determining the rank of the first check matrix, wherein the rank of the first check matrix is the maximum order of the sub-formula which is not zero in the first check matrix.

In step 1042, a rank of the second test matrix is determined based on the first historical time series and the significant set.

In some embodiments, referring to fig. 3F, fig. 3F is a flowchart illustrating a time-series prediction method provided in an embodiment of the present application, and step 1042 illustrated in fig. 3F may be implemented by the following steps 10421 to 10425.

In step 10421, a sequence of nth order lag values for the first historical time series and a sequence of mth order lag values for each second historical time series in the significant set are determined.

Wherein the nth order lag value comprises: the values of all the moments of the first historical time sequence within the range from the moment n to the current moment; the m-order lag value sequence comprises: and the second historical time series in the significant set takes values from m time to all times within the range of the current time.

As an example, when n is 2 and m is 3, determining a 2 nd order hysteresis value sequence of the first historical time series, and a 3 rd order hysteresis value sequence of each second historical time series in the significant set, wherein the 2 nd order hysteresis value sequence of the first historical time series may be {10, 11}, 10 is a 2 nd order hysteresis value, and 11 is a 1 st order hysteresis value; any second historical time series in the significant set of 3 rd order lag values {15, 28, 26}, 15 being a 3 rd order lag value, 28 being a 2 nd order lag value, 26 being a 1 st order lag value.

In step 10422, the values at any two different times in the n-order lag value sequence are subjected to difference processing, so as to obtain a fourth difference processing result corresponding to the values at any two different times.

As an example, when the n-order lag value sequence is {10, 11}, the difference processing is performed on the 2-order lag value 10 and the 1-order lag value 11, and the resulting fourth difference processing result is 1.

In step 10423, the following is performed for any one sequence of m-order lag values: and carrying out difference processing on the values at any two different moments in the m-order lag value sequence to obtain a fifth difference processing result corresponding to the values at any two different moments.

As an example, when the m-order hysteresis value sequence is {15, 28, 26}, performing difference processing on values at any two different times in the m-order hysteresis value sequence, that is, performing difference processing on the 3-order hysteresis value 15 and the 2-order hysteresis value 28 to obtain a fifth difference processing result 13; carrying out difference processing on the 3-order lag value 15 and the 1-order lag value 26 to obtain a fifth difference processing result 11; the 2 nd order lag value 28 and the 1 st order lag value 26 are subjected to difference processing to obtain a fifth difference processing result 2.

In step 10424, the summation processing is performed on each fourth difference processing result to obtain a fourth summation processing result, and the summation processing is performed on each fifth difference processing result to obtain a fifth summation processing result.

As an example, the fourth difference processing result is 1, and the summation processing is performed to obtain a fourth summation processing result 1; and summing the fifth difference processing result 13, the fifth difference processing result 11 and the fifth difference processing result 2 to obtain a fifth summation processing result 26.

In step 10425, a rank of the second test matrix is determined based on the fourth summation processing result and the fifth summation processing result.

In some embodiments, the expression for the second check matrix may be:

wherein, Π x_t-1A second check matrix is characterized that is,

the result of the fourth summing process is characterized,

characterizing the result of the fifth summation process, e_tCharacterizing the third parameter, C characterizing the first constant, Δ x_tAnd representing the value error of the current moment of the first historical time sequence.

By way of example, by means of a second test matrix Π x_t-1And determining the rank of the second check matrix, which is the maximum order of the sub-formula in the second check matrix that is not zero.

In step 1043, the rank of the first check matrix and the rank of the second check matrix are subjected to difference processing to obtain a rank difference value.

As an example, when the rank of the first check matrix is 5 and the rank of the second check matrix is 6, the rank difference 1 is obtained by performing differential processing on the rank of the first check matrix and the rank of the second check matrix.

In step 1044, the first historical time sequence, the significant set, and the second historical time sequence are subjected to likelihood ratio test processing to obtain a likelihood ratio value and a test probability value.

In some embodiments, the Likelihood Ratio verification process (LR) is an indexing process that reflects authenticity, and belongs to a composite index that reflects both sensitivity and specificity. The likelihood ratio is defined as the ratio of the maximum value of the likelihood function with constraint condition to the maximum value of the likelihood function without constraint condition.

In step 1045, the rank difference, likelihood ratio, and test probability values are determined as the incremental test processing results.

In step 105, the second history time series in which the incremental test processing result satisfies the explanatory condition is determined as the third history time series.

And the third historical time sequence is used for the prediction processing of a plurality of base class prediction models to obtain a prediction time sequence corresponding to the first historical time sequence.

In some embodiments, the determining the second history time sequence in which the incremental verification processing result satisfies the explanatory condition as the third history time sequence in step 105 may be implemented by: and when the rank difference value is smaller than the difference threshold value, the likelihood ratio value is smaller than the likelihood ratio value threshold value and the test probability value is smaller than the test probability threshold value, determining the second historical time sequence as a third historical time sequence.

As an example, when the difference threshold is 10, the likelihood ratio threshold is 0.5, the test probability threshold is 0.8, the rank difference value of the second history time sequence is 8, the likelihood ratio value is 0.3, and the test probability value is 0.7, it is satisfied that the rank difference value 8 is smaller than the difference threshold 10, the likelihood ratio value 0.3 is smaller than the likelihood ratio threshold 0.5, and the test probability value 0.8 is smaller than the test probability threshold 0.8, and the second history time sequence is determined as the third history time sequence.

As an example, when the difference threshold is 10, the likelihood ratio threshold is 0.5, the test probability threshold is 0.8, the rank difference value of the second history time sequence is 11, the likelihood ratio value is 0.6, and the test probability value is 0.9, it is satisfied that the rank difference value 11 is greater than the difference threshold 10, the likelihood ratio value 0.6 is greater than the likelihood ratio threshold 0.5, and the test probability value 0.9 is greater than the test probability threshold 0.8, and the second history time sequence is not determined as the third history time sequence.

In this way, each second historical time sequence in the non-significant sequence set is screened through the difference threshold, the likelihood ratio threshold and the test probability threshold, and the second historical time sequence meeting the screening condition is determined as a third historical time sequence, so that the second historical time sequence in the non-significant sequence set is effectively screened.

In some embodiments, referring to fig. 3G, fig. 3G is a flowchart illustrating a time series prediction method provided in an embodiment of the present application, and after step 105 shown in fig. 3G, a predicted time series corresponding to the first historical time series may be determined by performing steps 108 to 109.

In step 108, based on the third historical time sequence, the significant sequence set, and the first historical time sequence, a plurality of base class prediction models are called to perform prediction processing, so as to obtain a base class prediction sequence corresponding to each base class prediction model.

In some embodiments, the base class prediction sequence is a prediction sequence of values subsequent to a current value of the first historical time sequence.

As an example, when the first historical time series is {10, 11, 16, 19}, the corresponding base class prediction series may be {11, 16, 28}, and the base class prediction series corresponds to a time later than the time of the first historical time series.

In some embodiments, the types of base class prediction models include univariate base class prediction models, multivariate base class prediction models, and industry experience models.

In some embodiments, referring to fig. 3G, fig. 3G is a flowchart illustrating a time-series prediction method provided in an embodiment of the present application, and step 108 illustrated in fig. 3G can be implemented by performing the following steps 1081 to 1083.

In step 1081, based on the first historical time sequence, a plurality of univariate base class prediction models are sequentially called to perform univariate prediction processing, so as to obtain univariate base class prediction sequences corresponding to the plurality of univariate base class prediction models respectively.

In some embodiments, the types of univariate base class prediction models include exponential smoothing models, differential autoregressive models, and sequence decomposition models; the above step 1081 can be implemented as follows: and based on the first historical time sequence, sequentially calling the index smoothing model, the differential autoregressive model and the sequence decomposition model to perform univariate prediction processing to obtain univariate base class prediction sequences respectively corresponding to the index smoothing model, the differential autoregressive model and the sequence decomposition model.

By way of example, referring to fig. 5B, fig. 5B is a schematic diagram illustrating a time-series prediction method provided by an embodiment of the present application. And based on the first historical time sequence, sequentially calling the index smoothing model, the differential autoregressive model and the sequence decomposition model to perform univariate prediction processing to obtain univariate base class prediction sequences respectively corresponding to the index smoothing model, the differential autoregressive model and the sequence decomposition model.

In some embodiments, the model is exponentially smoothed by modeling the first historical time series as three structural elements, namely, level, trend, and seasonality. The exponential representation exponential smoothing model may exponentially weight all past observations to determine a degree of confidence in the past values, the smoothing representation model being capable of smoothing fluctuations in the observations.

In some embodiments, the differential autoregressive model is created by modeling the time series as three parts, differential, autoregressive, and moving average. The d-order difference representation is modeled by d-order difference values to determine the stationarity of the time series. p representations are autoregressed using the p lag predictor building model for the current time period. q represents the moving average portion of the time series that introduces an error in period q. Let Y_tFor stationary time series, the differential autoregressive model ARIMA (p, d, q) can be expressed as:

wherein L is a hysteresis factor, ε_tIn order to be an error term, the error term,

and theta_iFor the model parameters, the model parameters may be estimated using a maximum likelihood method.

In some embodiments, the sequence decomposition model may be divided into two parts: an external circulation section and an internal circulation section. Wherein the outer loop portion is used to adjust the weight of the observed value for each time period, and the inner loop portion is used to update the seasonal and trend components.

Therefore, the univariate prediction processing is carried out through the exponential smoothing model, the differential autoregressive model and the sequence decomposition model, so that the advantages of different univariate base class prediction models are fully utilized, and the prediction accuracy is effectively ensured.

In step 1082, based on the first historical time sequence and the scene experience information, an industry experience model is called to perform experience processing, so as to obtain an experience base class prediction sequence corresponding to the industry experience model.

In some embodiments, in an application scenario of popular disease investigation, the scenario experience information may be a high-incidence population of epidemic diseases, and so on. The industry experience model is a prediction model constructed by constructing industry influence factors according to past experience according to an actual model application scene.

For example, referring to fig. 5B, based on the first historical time sequence and the scene experience information, an industry experience model is called to perform experience processing, so as to obtain an experience base class prediction sequence corresponding to the industry experience model.

In step 1083, based on the third historical time sequence, the significant sequence set, and the first historical time sequence, the multiple multivariate base class prediction models are sequentially called to perform multivariate prediction processing, so as to obtain multivariate base class prediction sequences respectively corresponding to the multiple multivariate base class prediction models.

In some embodiments, the types of multivariate base class prediction models include vector autoregressive models and long-short term memory networks, and step 1083 above can be implemented by: and sequentially calling the vector autoregressive model and the long-short term memory network to perform multivariate prediction processing based on the third historical time sequence, the significant sequence set and the first historical time sequence to obtain multivariate base class prediction sequences respectively corresponding to the vector autoregressive model and the long-short term memory network.

As an example, referring to fig. 5B, based on the third historical time series, the significant sequence set, and the first historical time series, the vector autoregressive model and the long-short term memory network are sequentially called to perform multivariate prediction processing, so as to obtain multivariate base class prediction sequences respectively corresponding to the vector autoregressive model and the long-short term memory network.

In some embodiments, the vector autoregressive model may infer from the historical time series that the extension of the historical time series corresponds to adding a moving average portion.

In some embodiments, the long-term and short-term memory network comprises an input gate, a forgetting gate and an output gate, and is used for achieving the goals of long-term memory and forgetting information.

In step 109, each base class prediction sequence is integrated to obtain a prediction time sequence corresponding to the first historical time sequence.

In some embodiments, the integration processing may be implemented by a Ridge Regression (Ridge Regression) model, and the integration processing is used to aggregate advantages of base class prediction models from different viewing angles, so that a prediction time sequence corresponding to a finally obtained first historical time sequence is not affected by an abnormal value of a certain base class prediction model with incomplete information, thereby achieving high reliability, interpretability, accuracy and target adaptability of the base class prediction model as far as possible under a limited condition, and effectively ensuring prediction accuracy of the prediction time sequence corresponding to the first historical time sequence.

In this way, interpretability conditions are set through causal inspection processing and incremental interpretation processing, and the obtained second historical time sequence is screened, so that the reason that each obtained third historical time sequence is introduced into a subsequent base class prediction model is clear, the influence mechanism of each third historical time sequence on the first historical time sequence is reflected, the third historical time sequence obtained through screening has interpretability, the interpretability of the screened historical time sequence is effectively improved, and a decision reference is provided for prediction.

In the following, an exemplary application of the embodiment of the present application in an actual application scenario of time series prediction will be described.

In medical field tasks such as medical insurance, epidemiological survey, and disease control, a scenario in which time-series form data is analyzed to support decision making and planning widely exists. In similar tasks such as medical insurance reimbursement prediction, epidemic trend prediction and the like, since the predicted target variables have a large directional influence on the overall strategic decision, it is required that covariates (e.g., economic variables, demographic variables) (explanatory variables) introduced into the prediction model can maximally improve the accuracy of the prediction model, and at the same time, the prediction model has strong interpretability, so that the reliability of the model as a whole is high, and reliable decisions can be made through the influence and relationship between the variables displayed by the model.

Meanwhile, since the related technologies usually face pain points with small historical data volume and insufficient sample volume, the sum of the numbers of usable scene-related variables and introduced covariates (explanatory variables) such as economic variables and demographic variables may exceed the number of samples, which requires that the embodiment of the present application builds a model by screening variables meeting the above requirements from a set of variables before predicting a time series, thereby achieving high reliability, interpretability, accuracy and target adaptability of the model as much as possible under limited conditions.

The embodiment of the application provides a time series prediction method, which comprises the steps of judging the multi-dimensional interrelation between multiple selectable sets of covariates (explanatory variables) (namely the second historical time series described above) and target variables (namely the first historical time series described above) through a series of time series descriptive inference models, selecting an optimal subset from the sets of the covariates (explanatory variables) through an interpretable method, and outputting a mode that a time series system formed in the sets influences the target variables; the time series prediction method provided by the embodiment of the application can be used for detecting the influence of the introduction of a certain variable on the whole system in a group, so that the complex relation between a single variable and a target variable is considered in variable screening, and the interpretability and the comprehensiveness of a model variable are improved.

On the basis of the screened covariate (credible interpretation variable) combination, the embodiment of the application predicts a base learner (the base class prediction model described above) as a base prediction model by constructing an integrated model and fitting three types of time sequences with single sequence, multiple sequences and industry prior, and completes the prediction task of the target variable by constructing an upper integrated prediction model.

By using a variable source grouping mode, screening and inference models in a group are constructed, so that the explanatory property is ensured, the characteristic quantity of a single model is reduced, and the problems that the model cannot be constructed due to insufficient freedom and excessive variables with still-acceptable explanatory performance need to be screened out are solved. The variable screening method comprehensively considers univariate prediction contribution information of the granger causal test output and multivariate time sequence system information of the increment synergistic integration test output to determine an optimal variable subset under certain conditions, so that the construction of a subsequent inference model and a prediction model has statistical credibility. The influence mechanism of each group of source variables is output by constructing a grouped VAR model, and the decision support information of variable change is provided.

The embodiment of the application integrates the advantages of a single-variable time series prediction model, a field experience model and a multi-sequence prediction model, the single-variable time series prediction model, the field experience model and the multi-sequence prediction model are used as base learners (base class prediction models described above) providing different visual angles, an upper-layer integrated model is built according to the base classes, all the variables screened in the previous part are input into each base learner, parameter search and upper-layer learning samples are built according to conditions, so that the output of each base learner is input into the integrated learner, the advantages of each model are integrated, and the model prediction accuracy is improved.

By way of example, referring to fig. 5A, fig. 5A is a schematic diagram illustrating a time-series prediction method provided by an embodiment of the present application. Under the scenes of medical insurance reimbursement decision, infectious disease control decision and the like, the method can automatically screen and output the optimal variables as independent variables of a prediction model according to the input data sets of few samples and multiple characteristic sources, and simultaneously output a mechanism for influencing target variables by the screened variables as auxiliary information for decision making. And automatically constructing an integrated prediction model based on the variable screening result and outputting the prediction values of the models, so that a user can make reliable planning and judgment by referring to the prediction view angle of each sub-model (the base class prediction model described above) and combining the prediction result of the integrated model.

By way of example, referring to fig. 5B, fig. 5B is a schematic diagram illustrating a time-series prediction method provided by an embodiment of the present application. The time series prediction method provided by the embodiment of the application mainly comprises a variable screening and model deducing part and an integrated prediction model part, and can output credible statistical deducing and predicting results through a series of statistical learning methods designed aiming at small sample multi-feature application scenes.

For example, referring to fig. 5C, fig. 5C is a schematic diagram illustrating a time-series prediction method provided by an embodiment of the present application.

The model of the embodiment of the application is applied to the scene with large feature quantity and small data set observation sample size. The embodiment of the present application assumes that covariates (explanatory variables) in a data set, that is, explanatory variables other than target variables, have been divided into a plurality of groups based on their different sources, for example, in medical insurance budget (target variable) prediction, the explanatory variables may be divided into medical insurance variable groups, macro economic variable groups, population structure variable groups, and the like. The embodiment of the application solves the problem that effective variable screening and inference cannot be carried out due to the fact that small samples have multiple characteristics and lack of freedom degrees because a Vector Autoregressive (VAR) model and a Vector Error Correction Model (VECM) cannot be fitted on all explanatory variables through grouping logic, obtains the influence of a single variable on a target variable (Glan's causal test), the influence of a homologous variable system on the target variable (grouping VAR model) and the influence of all variables on the target variable (integration model), obtains the interpretability of variable interaction relation from three dimensions, improves the reliability of the model, and provides reliable information from more angles for decision support and planning.

The initial time series data set structure of the embodiment of the present application is shown in the upper left of fig. 5C, and is composed of target variables and a plurality of sets of explanatory variables. And screening variables with the most interpretation capability on the target variables from each group of the explanatory variables, inputting the variables serving as credible covariates (explanatory variables) into a subsequent integrated prediction model, and simultaneously outputting the single variables and the influence mechanisms of the variables in the same group on the target variables. In the examples of the present application, the target variable is represented by Y, and any explanatory variable is represented by X.

Referring to fig. 5C, the process of unit root testing is described below. And (4) unit root detection, namely detecting the stationarity of the original time sequence. For example, in the medical insurance reimbursement scenario, the variables from sources such as macroscopic economic indicators and medical insurance introduced, such as number of people and payment, are often highly correlated with time trends, resulting in non-stationary factors. In order to select the appropriate difference order, except for variables that do not meet subsequent testing and modeling conditions, a stationarity test is required. In the embodiment of the application, the stationarity of the original time sequence is tested in a unit root test mode: ADF test (amplified dictionary-Fuller), PP test (Phillips and Perron), KPSS test (Kwiatkowski-Phillips-Schmidt-Shin), and the like.

For each variable in each group of explanatory variables, the embodiment of the application tests the original form of each difference to n-order differences, and the variable with the test result of each order difference being significant at a specified significance level is regarded as a stable variable and can be used for subsequent prediction. In this process, the embodiment of the present application selects the same-order difference result as low as possible, and removes variables that do not satisfy any test significance (the test significance characterizes that the P value by the ADF test, the PP test, and the KPSS test is greater than the P value threshold), while ensuring that as few explanatory variables as possible are removed.

Referring to fig. 5C, the process of the glange causal test is described below. The Granger causal test (Granger causal test) may search for the relationship between each explanatory variable X and the target variable Y and the potential contribution to its prediction and provide a way of influence of its relationship. The grangeje causal test uses a form of vector auto-regression within two variables to check whether the n-th order lag value of one variable contributes to predicting the other variable (accounting for the n-th order lag value of variable X's contribution to predicting the target variable Y). The present embodiment focuses on each pair of differential smooth variables of Y and X to examine the one-way grand causal relationship between them, as shown below, (one target variable Y and all explanatory variables X).

Wherein, Y_mtThe parameter beta is the value of Y at the time point t_mFor introducing an explanatory variable X_nP and q are the specified maximum hysteresis order, beta_wTo introduce no explanatory variable X_nFitting coefficient of regression equation in case, epsilon_mtIs random noise. For more than two Y_mtPerforming an F-test to determine the introduced variable X_nWhether the interpretability of the target variable Y prediction is improved. At a set significance level, the glargine causal test classifies each original set of explanatory variables into a glargine causal significant variable set and an insignificant set (the classification is based on the relationship between the test result of the F test and a significant threshold, the test result is greater than the significant threshold, the significant variable is a significant variable, and is less than the test threshold, the insignificant variable is a non-significant variable), and based on the significance result (including the significant variable set (i.e., the significant sequence set described above) and the insignificant variable set (i.e., the insignificant sequence set described above)), the embodiment of the present application outputs a final variable screening result using the following grouped covariance incremental test.

Referring to fig. 5C, the process of packet-coordinated incremental verification is described below. The granger causal relationship described above is only able to validate the pairing relationship between two variables, not the form of grouped multivariate predictive relationships that embodiments of the present application want to achieve. Relationship checking of single variables may ignore co-operative relationships, such as co-ordinates, that are generated in a system formed by multiple variables. When a linear combination of a certain subset of a series of time series variables shows stationarity, the embodiment of the present application judges that a synergistic effect exists in the time series. Due to the cyclostationarity of the linear combinations, long-term correlations may be observed in this subset, while short-term effects may not be detectable by a relationship test of the paired variables. Therefore, the embodiment of the application designs a grouping coordinated increment test to find a potential multivariate relation beneficial to final prediction. According to the variable selection results of the granger causal test described above, each insignificant variable will be tested along with the significant variables for the presence of a synergistic relationship and at a set significance level determine whether the insignificant variable, after addition to the significant variable set, is a beneficial extension of the variable set, thereby determining whether it is used in a subsequent predictive model.

For example, referring to fig. 5D, fig. 5D is a schematic diagram illustrating a time-series prediction method provided by an embodiment of the present application. The specific flow of the packet coordinated incremental check is shown in fig. 5D. The embodiment of the application uses a coordination test method proposed by Johansen (1988) as a basic model for coordination relationship judgment. In the Johansen method, a Vector Error Correction Model (VECM) is built on the lead-in variables to describe the long-term effects of the input variables, thereby testing multiple co-integrals at once.

For an input matrix X containing Y and X, the original Vector Autoregressive (VAR) model is as follows:

wherein x is_t-iIs x_tI order lag term of (a), p is the maximum lag order, and c is the intercept term.

On the basis of the original model, the error-corrected VECM model is as follows:

wherein, Δ x_t，Δx_t-iAnd a parameter phi_iTo merge all lag terms into x_t-1The latter difference variables and coefficients. The long-term effects in all t are absorbed in the matrix Π.

The embodiment of the application determines the linear combination with the coordination relation possibly contained in the matrix pi by checking the rank (rank) of the matrix pi. In the trace test proposed by Johansen, for model i, the hypothesis of matrix rank < n is examined to determine the range in which rank of the matrix lies, and the number rank (i) of the co-ordinates is obtained.

Based on the basic Johansen co-integration test, the grouped co-integration incremental test of the embodiment of the application aims to select variables from a series of variables to be selected, which contribute to the current model, to be added to the subsequent model. For each variable group, the present example performed Johansen Test and liklihood Ratio Test (LR Test) on the insignificant variables in each grand cause and effect Test along with all significant variables.

The specific operation is that firstly, a VECM model i is constructed for each group of selected (significant) variables (all significant variables) so as to obtain rank (i) determined by trace inspection under a certain significance level as a substrate rank; then, sequentially adding each non-significant variable j into a model variable to construct a VECM model ij to obtain rank (ij) of the model; and (4) carrying out difference delta rank on rank (i) and rank (ij) to determine whether the addition of the variable j under the same significance level improves the number of the coordination relations in the original model system.

Because the variables Δ rank are usually relatively close and cannot be compared more finely at the same level, the embodiment of the application constructs two VAR models for the original variable and each added variable, performs Likelihood Ratio test (Likelihood Ratio test) on the two models, and determines whether the added variable j enhances the model interpretation capability by judging the value of LRStat and the tested p value (determines whether the value of LRStat is less than the threshold of LRStat, determines whether the value of p is less than the threshold, determines whether the added variable j enhances the model interpretation capability, and determines that the added variable j enhances the model interpretation capability and retains the variable j when the value of LRStat is less than the threshold of LRStat and the value of p is less than the threshold of p).

In the obtained result table, the embodiment of the present application first compares Δ rank (i.e., the above-described rank difference value), and when comparing with the Δ rank variable, first determines whether the LR p value (i.e., the above-described test probability value) is significant, and then compares the magnitude of LR Stat (i.e., the above-described likelihood ratio value), thereby ranking the contributions of the added variables. The embodiment of the application judges whether the variable result meets the requirement of adding the screening result variable by setting a certain threshold value.

And (4) carrying out vector autoregressive model analysis, and taking the proper variable subset required for constructing the prediction model as input characteristics after screening. In order to enhance the interpretability of variable selection, the embodiment of the application establishes the VAR models of each group according to the grouped variable selection results, and performs descriptive analysis according to the inferred results of the VAR models to output the system influence contained in the variable.

The impulse response function, one of the main results of the VAR model, will account for the effects that the dependent variable may have by applying a standard deviation to each of the explanatory variables.

An example of the output result of the impulse response analysis is shown in fig. 5E below, and fig. 5E is a schematic diagram illustrating the principle of the time series prediction method provided by the embodiment of the present application. In the four models constructed as shown in fig. 5E, the abscissa of each result graph is the time step, and the ordinate is the target variable change amount. Each curve describes the effect that applying a variance to the selected variable at the initial state will have on the target variable at each time step. The result of the impulse response function can show a mechanism that the variable exerts influence on the target variable in the system, help the embodiment of the application to make a decision and adjust the strategy in the time dimension, and predict in advance the impact mode that the adjustment may bring to the system. By grouping VAR model construction of homologous variables, homologous variables with higher possibility of association can be considered as a system inspection source internal influence mechanism, and the same group of associated information which is easy to analyze and control is provided for a decision maker while the embodiment of the application is allowed to use more variables for modeling to solve the defects that n is closer to p and n < p.

In the embodiment of the application, a series of prediction models are established according to the previous variable selection results, so that an accurate and robust set prediction model is established while a small amount of observation data is processed and internal and external influences of variables are balanced.

In the embodiment of the application, an integrated Model with five base learners is exemplarily constructed, including univariate time series prediction models (explicit smoothening, ARIMA, STL), Industrial Empirical models (Industrial Empirical Model), and multivariate time series models (VARMA, LSTM).

Next, the above-described basis learners will be described one by one.

Referring to fig. 5F, fig. 5F is a schematic diagram illustrating a principle of a prediction method of a time series according to an embodiment of the present application, namely, an Exponential Smoothing model (i.e., the Exponential Smoothing model described above) by modeling an original time series into three structural elements, namely, a level, a trend, and a seasonality. An exponential representation model in the name may exponentially weight all past observations to determine a degree of confidence in the past values, with a smooth representation model capable of smoothing fluctuations in observations. The prediction method of the model is as follows:

F_t+1＝(S_t+T_t)C_(t+1)-L (11)

S_t＝αx_t+(1-α)(S_t-1+T_t-1) (12)

T_t＝β(S_t+S_t-1)+(1-β)T_t-1 (13)

C_t＝C_(t+1)-L (14)

wherein, F_t+1Representing the predicted next-phase value, S_tWhere α denotes an exponential weighting operation, β denotes a smoothing coefficient for past values, C_tIndicating the periodic effect occurring every L period.

Referring to fig. 5F, fig. 5F is a schematic diagram illustrating a principle of a prediction method of time series, namely an ARIMA model (Autor), according to an embodiment of the present applicationAn iterative Integrated Moving Average) (i.e., the differential autoregressive model described above) by modeling the time series as three parts, differential, autoregressive, and Moving Average. The d-order difference representation is modeled by d-order difference values to determine the stationarity of the time series. p represents the autoregression using the model built with the p lag predictor for the current time period. q represents the moving average portion of the time series that introduces an error in period q. Let Y_tFor stationary time series, the ARIMA (p, d, q) model can be expressed as:

and theta_iFor the model parameters, the model parameters may be estimated using maximum likelihood.

Referring to fig. 5F, fig. 5F is a schematic diagram illustrating a principle of a prediction method of a time series according to an embodiment of the present application, and an STL model (i.e., the sequence decomposition model described above) may be divided into two parts: an external circulation and an internal circulation. Wherein the outer loop is used to adjust the weight of the observations at each time period and the inner loop is used to update the seasonal and trend components. For an input time series Y_t. Firstly, the weight of each observation point is initialized to 1, the initialization trend component is 0, and the observation points are divided into different subsequences according to the periodicity of the input time sequence. The first step of the external circulation is the internal circulation.

An internal loop comprising the steps of: first is detrending. Using Y_tSubtracting the trend component

To obtain a detrended sequence. The second step is sub-sequence smoothing, after de-trending, Loess smoothing is carried out on each sub-sequence, and the smoothed sub-sequences are combined to obtain a time sequence with the length of N +2 times the sequence period length

Next, two moving averages are performed on the subsequence line having a window size of the sequence period length using the subsequence low-pass filter, and then a moving average of a window size of 3 is performed. Sequences of length N

And performing Loess smoothing processing. Trending off seasonal factors

By

Minus

And (6) calculating. Y is_tSubtracting the trend component

A de-seasonal sequence can be obtained; new trend factor

Obtained by Loess smoothing of the de-seasonal sequence. Then check whether the trend factor converges, and if not, return to the first step. At the end of the inner circulation, the embodiment of the application obtains the final trend factor and the seasonal factor T_tAnd S_t. The residual can be calculated as R_t＝Y_t-T_t-S_t. The second step of the outer loop is to obtain a robustness weight for each data point, i.e.

h＝6median(|R_t|) (16)

Wherein R is_tCharacterizing residual errors, h characterizing weight, B (u) characterizing robustness results, and u is a robustness parameter.

For an observed value at a time point v, the robustness weight of the point is:

ρ_v＝B(|R_t|/h) (18)

this weight is used for Loess smoothing in the steps of the next inner loop.

The weight of the local value is adjusted by multiplying the weight. After several external and internal cycles as above, the STL will decompose the final trend and seasonal factors T_tAnd S_t. Next, ARIMA is used to predict future trends and seasonality. And adding the future trend predicted value and the seasonal predicted value to obtain a future prediction result of the STL.

Referring to fig. 5F, fig. 5F is a schematic diagram illustrating a principle of a time series prediction method provided in the embodiment of the present application, where an Industrial Empirical Model (Industrial Empirical Model) is a prediction Model constructed by constructing Industrial impact factors according to past experiences according to an actual Model application scenario.

The fund balance time series is a time series with quarterly property, monthly volatility and growth trend regularity, and the demonstration results and practices of operators prove that the fund balance time series is the time series with quarterly property, so that the overall fund balance forecasting method based on target time series characteristic factor fitting is used for building an industry experience model. The factors included in the industry experience model may be listed as: (1) initial year time difference (current _ year-start _ year), (2) secondary initial year time difference (current _ year-start _ year)²(3) a continuous month factor (current _ month), (4) a category month factor (category _ month), (4) a start year time difference and a continuous month factor (current _ year) are interacted, (4) a previous month value (y) and (5) a previous month value_{last_month}) (6) the average difference between the monthly and quarterly values of the previous period

(7) Last month value Start month time Difference interaction (y)_{last_month}Total _ month _ past), (8) initial month time difference index factor ((y) of difference between previous month value and initial month value_{initial_month}-y_{last_month})^{total_months_past}) The growth trend, volatility, quadratic factor, exponential factor, periodic effect and the like are covered. A factor model constructed based on the above factors as inputs may be fitted to the parameters of each factor using linear regression OLS estimation, and the fitted model is used to calculate the prediction results.

VARMA (input as target variables + post-screening model variables), in addition to the univariate models and empirical models constructed above for target variable prediction, the next two sections will also construct multivariate time series prediction models including VARMA and RNN (LSTM).

Referring to fig. 5F, fig. 5F is a schematic diagram of a time-series prediction method provided in an embodiment of the present application, and a VARMA model (i.e., the vector autoregressive model described above) is built on the variable set selected in the above variable screening section to infer an interpretation result of the descriptive analysis according to the variable relationship. The VARMA model can infer the extended added moving average portion from the historical time series, which is in the form of:

wherein e_t-iLag term of order i, phi, of error term_iThe addition of the error term part is a parameter matrix of the error term, and the moving average part is expanded for the original VAR model. q order selection and determination of whether to add constants and trend factors to the model will be done by grid search hyper-parameters. Rnn (lstm) (inputs are target variables + post-screening model variables). RNN is a neural network structure intended to handle sequential inputs, and has been successfully used for multiple tasks on time series in various scenarios. The RNN unit will traverse each input value. Each value X in the sequence_tGenerating a hidden layer H_t. Each hidden layer is connected with a weight w through a direction_hConnected to its subsequent layer H_t. The weights are shared between all hidden layers. When the RNN cell reaches the last value, the hidden layer of RNN generates the final output y for comparison with the actual value in the reverse learning step. The Long Short Term Memory (LSTM) used in the embodiments of the present application is oneA widely used version of RNN unit. The LSTM network (i.e., the long-short term memory network described above) includes an input gate, a forgetting gate, and an output gate to achieve the goals of long-term memory and forgetting information. The LSTM process can be simply written as:

H_t,C_t＝LSTM(X_t,H_t-1,C_t-1) (20)

wherein, C_tIs the cell state value during t.

Referring to fig. 5F, fig. 5F is a schematic diagram illustrating a time series prediction method and an integrated model provided in the embodiment of the present application, the embodiment of the present application aggregates advantages of different view models by constructing and estimating several types of time series prediction models, i.e., a univariate prediction model, an empirical model and a multivariate prediction model, as a basis learner, and constructing an integrated (Ensemble) model based on the basis of the time series prediction models, so that a final model result is not affected by an abnormal value predicted by a certain model with incomplete information. In order not to destroy the feature representation learned by the base learner, the embodiment of the present application constructs a linear model: the Ridge Regression (Ridge Regression) model is used as a high-level integrated model, and can be used for contracting parameters by a regularization method under the conditions that all features are fully used and the result of a certain base learner is not lost, so that the result of the base learner model is integrated, and the accuracy of the whole model is improved.

All the model hyper-parameters can determine the optimal parameters in a grid search mode, and the optimal parameters of each model are selected by calculating the time series prediction indexes of the model in a mode of dividing a training set and a test set, such as MSE, MAE, MAPE, SMAPE and the like.

When the base learners are different from each other greatly and the target value interpretation capability is weak, the embodiment of the application may consider adding the original features as input into a high-level model (such as a dotted line in the figure) to supplement the part not learned by the base learner, but since the high-level model is a simple linear model, the final selection needs to be made by balancing the overfitting risk of the method.

Referring to fig. 5G, fig. 5G is a schematic diagram illustrating a time series prediction method provided in an embodiment of the present application, and fig. 5G shows time series prediction curves respectively corresponding to an exponential smoothing model, a differential regression model, a sequence decomposition model, an industry experience model, a vector autoregressive model, and a long-short term memory model, and the time series prediction indexes shown in table 1 below are obtained by comparing the time series prediction curve of each model with a sample curve.

The prediction indexes calculated by the examples of the present application are shown in table 1 below. Compared with single-sequence and multi-sequence bottom-layer predictors, the average absolute error (MAE), the average absolute percentage error (MAPE) and the symmetric average absolute percentage error (SMAPE) of the integrated prediction model are all reduced, and the accuracy and the reliability of the prediction method provided by the embodiment of the application are shown.

TABLE 1 time series prediction index Table

The beneficial effects of the embodiments of the present application can be summarized as the following aspects:

credibility of small sample multi-feature statistical learning: the statistical inference tools such as VAR models which are difficult to apply originally can be applied under the scene of small samples and multiple characteristics, the number of variables needing to be removed is reduced by means of a homologous grouping method, effective information is saved, an optimal variable subset is selected as far as possible to construct an inference model, and the credibility of the model is reserved to the greatest extent.

Interpretability: by using the Glangel causal test, the grouped increment collaborative test and the grouped VAR method, the reason of each variable introduced into the subsequent model is clear, the influence mechanism between the screened variable and a system composed of the target variable and other introduced variables can be output, and the interpretability of the subsequent prediction model is effectively ensured.

The accuracy is as follows: by means of the integrated model, a univariate, empirical model and a multivariate time series prediction model are introduced, advantages and defects of each model are comprehensively considered, model performance is further improved through parameter search and model adjustment according to scenes, and the comprehensive model is effectively guaranteed to be superior to a traditional single model.

Target adaptability: the method has the advantages that excessive prior knowledge is not needed, complex feature exploration and concept model construction are not needed, the screening inference and prediction model can be constructed in a self-adaptive mode only by selecting target variables and other variable groups in the data set, the difficulty of complex artificial feature processing under a multi-feature scene is reduced, the migration and popularization difficulty of the model is reduced, the configuration efficiency is improved, and the cost is saved.

When the grouping cooperative increment inspection is applied, in the embodiment of the application, besides the single unnoticed variable can be added on the basis of the grand causal significant variable for inspection, and then some variables are screened according to the limiting conditions, the screened variable group can also be added after the optimal variable is obtained by screening each time for increment inspection again, and iteration is continuously carried out until the remaining variables do not meet the screening conditions any more.

Under the condition that the feature quantity allows, the variable groups of similar sources can be combined as much as possible, so that more variables are introduced into the statistical inference results such as the impulse response analysis of the VAR, and more variable information is considered. In the case of sufficient samples, the number of sets of the variable screening part can be regarded as 1 to achieve the optimal screening and inference effects.

According to the practical use scene, more specially customized models can be added to the base learner of the integrated model, for example, the empirical model is subjected to scene adaptation change, more types of time series prediction models are added, the learning view angle of the model can be increased to a certain extent, and partial fitting capacity is improved.

It is understood that, in the embodiments of the present application, the data related to the historical time series and the like need to be approved or approved by the user when the embodiments of the present application are applied to specific products or technologies, and the collection, use and processing of the related data need to comply with the related laws and regulations and standards of the related countries and regions.

Continuing with the exemplary structure of the time-series prediction apparatus 255 provided by the embodiments of the present application as software modules, in some embodiments, as shown in fig. 2, the software modules stored in the time-series prediction apparatus 255 of the memory 250 may include: an obtaining module 2551, configured to obtain a first historical time sequence and a plurality of second historical time sequences, where the second historical time sequences are covariates of the first historical time sequence, and a stationarity index of the second historical time sequences is smaller than a stationarity threshold; a causal test module 2552, configured to perform a causal test on the first historical time sequence and each second historical time sequence; an assignment module 2553 for assigning each second historical time series to a significant sequence set or a non-significant sequence set based on the causal test treatment results; an increment inspection module 2554, configured to perform increment inspection processing on each second historical time sequence in the non-significant set based on the significant set, the non-significant set, and the first historical time sequence to obtain an increment inspection processing result; a determining module 2555, configured to determine the second history time sequence in which the incremental test processing result satisfies the explanatory condition as a third history time sequence; and the third historical time sequence is used for the prediction processing of a plurality of base class prediction models to obtain a prediction time sequence corresponding to the first historical time sequence.

In some embodiments, the cause and effect check module 2552 is further configured to perform the following for any one of the second historical time series: determining a first time value corresponding to the first historical time sequence based on the p-order lag value sequence of the second historical time sequence and the q-order lag value sequence of the first historical time sequence; the p-order hysteresis value sequence comprises values of all moments of the second historical time sequence from the moment p to the current moment, the q-order hysteresis value sequence comprises values of all moments of the first historical time sequence from the moment q to the current moment, the moment p and the moment q are any moments before the current moment, and the first moment value is any moment after the current moment; determining a second time value corresponding to the first historical time sequence based on the q-order hysteresis value sequence of the first historical time sequence, wherein the second time value is the same as the time corresponding to the first time value; and carrying out variance ratio test treatment on the first time value and the second time value, and determining the obtained variance ratio test result as a causal test treatment result.

In some embodiments, the assigning module 2553 is further configured to assign the second historical time sequence to the significant sequence set when the causal test treatment result is greater than or equal to the causal test threshold; and when the causal test processing result is less than the causal test threshold value, allocating the second historical time sequence to the non-significant sequence set.

In some embodiments, the time-series prediction device 255 further includes: the grouping module is used for distributing second historical time sequences of the same type in the obvious sequence set to sequence groups of corresponding types; allocating second historical time sequences of the same type in the non-significant sequence set to sequence groups of corresponding types; wherein the sequence group includes a second historical time series in the set of significant sequences and a second historical time series in the set of non-significant sequences.

In some embodiments, the delta check module 2554 is further configured to perform the following for any one of the second historical time series in the non-significant set: determining the rank of the first test matrix according to the first historical time sequence, the significant set and the second historical time sequence; determining the rank of the second test matrix according to the first historical time sequence and the significant set; carrying out differential processing on the rank of the first test matrix and the rank of the second test matrix to obtain a rank difference value; carrying out likelihood ratio test processing on the first historical time sequence, the significant set and the second historical time sequence to obtain a likelihood ratio value and a test probability value; and determining the rank difference value, the likelihood ratio value and the test probability value as the incremental test processing result.

In some embodiments, the determining module 2555 is further configured to determine the second historical time sequence as a third historical time sequence when the rank difference is less than the difference threshold, the likelihood ratio is less than the likelihood ratio threshold, and the test probability value is less than the test probability threshold.

In some embodiments, the delta check module 2554 is further configured to determine a sequence of i-th order lag values for the first historical time series, a sequence of j-th order lag values for each second historical time series in the significant set, and a sequence of k-th order lag values for the second historical time series; wherein the i-order lag value comprises: the values of all the moments of the first historical time sequence in the range from the moment i to the current moment; the sequence of hysteresis values of order j includes: the values of all the moments of the second historical time sequence in the significant set from the moment j to the current moment are taken; the sequence of k-th order lag values includes: the values of all the moments of the second historical time sequence from the moment k to the current moment; carrying out differential processing on values of any two different moments in the i-order lag value sequence to obtain a first differential processing result; the following is performed for any sequence of j-order lag values: carrying out differential processing on values of any two different moments in the j-order lag value sequence to obtain a second differential processing result; carrying out differential processing on values at any two different moments in the k-order lag value sequence to obtain a third differential processing result; determining a rank of the first test matrix based on the first, second, and third differential processing results.

In some embodiments, the delta check module 2554 is further configured to determine a sequence of n-th order lag values for the first historical time series, and a sequence of m-th order lag values for each second historical time series in the significant set; wherein the nth order lag value comprises: the values of all the moments of the first historical time sequence in the range from the moment n to the current moment; the m-order lag value sequence comprises: the values of all the moments of the second historical time sequence in the significant set from the moment m to the current moment are taken; carrying out difference processing on the values at any two different moments in the n-order lag value sequence to obtain a fourth difference processing result corresponding to the values at any two different moments; the following processing is performed for any one sequence of m-th order lag values: carrying out difference processing on the values at any two different moments in the m-order lag value sequence to obtain a fifth difference processing result corresponding to the values at any two different moments; summing each fourth difference processing result to obtain a fourth summation processing result, and summing each fifth difference processing result to obtain a fifth summation processing result; and determining the rank of the second test matrix based on the fourth summation processing result and the fifth summation processing result.

In some embodiments, the time-series prediction device 255 further includes: the prediction module is used for calling a plurality of base class prediction models to perform prediction processing based on the third historical time sequence, the significant sequence set and the first historical time sequence to obtain a base class prediction sequence corresponding to each base class prediction model; and the integration module is used for performing integration processing on each base class prediction sequence to obtain a prediction time sequence corresponding to the first historical time sequence.

In some embodiments, the types of base class prediction models include univariate base class prediction models, industry empirical models, and multivariate base class prediction models; the prediction module is further configured to sequentially call the multiple univariate base class prediction models to perform univariate prediction processing based on the first historical time sequence to obtain univariate base class prediction sequences corresponding to the multiple univariate base class prediction models respectively; calling an industry experience model to perform experience processing based on the first historical time sequence and scene experience information to obtain an experience base class prediction sequence corresponding to the industry experience model; and sequentially calling the multiple multivariate base class prediction models to perform multivariate prediction processing based on the third history time sequence, the significant sequence set and the first history time sequence to obtain the multivariate base class prediction sequences respectively corresponding to the multiple multivariate base class prediction models.

In some embodiments, the types of the univariate base class prediction model include an exponential smoothing model, a differential autoregressive model, and a sequence decomposition model; the prediction module is further configured to call the exponential smoothing model, the differential autoregressive model and the sequence decomposition model in sequence to perform univariate prediction processing based on the first historical time sequence, so as to obtain univariate base class prediction sequences respectively corresponding to the exponential smoothing model, the differential autoregressive model and the sequence decomposition model.

In some embodiments, the types of multivariate base class prediction models include vector autoregressive models and long-short term memory networks; the prediction module is further configured to sequentially call the vector autoregressive model and the long-short term memory network to perform multivariate prediction processing based on the third historical time sequence, the significant sequence set and the first historical time sequence, so as to obtain multivariate base class prediction sequences respectively corresponding to the vector autoregressive model and the long-short term memory network.

Embodiments of the present application provide a computer-readable storage medium storing executable instructions, which when executed by a processor, will cause the processor to perform a time-series prediction method provided by embodiments of the present application, for example, the time-series prediction method shown in fig. 3A.

In some embodiments, the computer-readable storage medium may be memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash, magnetic surface memory, optical disk, or CD-ROM; or may be various devices including one or any combination of the above memories.

In some embodiments, executable instructions may be written in any form of programming language (including compiled or interpreted languages), in the form of programs, software modules, scripts or code, and may be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

By way of example, executable instructions may be deployed to be executed on one computing device or on multiple computing devices at one site or distributed across multiple sites and interconnected by a communication network.

In summary, the embodiment of the present application has the following beneficial effects:

(1) and setting interpretability conditions through causal inspection processing and incremental interpretation processing, and screening the acquired second historical time sequence, so that the reason that each acquired third historical time sequence is introduced into a subsequent base class prediction model is clear, the influence mechanism of each third historical time sequence on the first historical time sequence is reflected, the screened third historical time sequence has interpretability, the interpretability of the screened historical time sequence is effectively improved, and a decision reference is provided for prediction.

(2) When a large number of second historical time sequences are obtained, the second historical time sequences with the stationarity indexes smaller than the stationarity threshold are determined to be the second historical time sequences of the subsequent causal test treatment through stationarity test treatment, so that preliminary screening of the second historical time sequences is achieved, the second historical time sequences with overlarge volatility (abnormity) are deleted, and effectiveness of the second historical time sequences is effectively guaranteed.

(3) The causal test processing is carried out on the first historical time sequence and each second historical time sequence, so that the causal association degree between the first historical time sequence and the second historical time sequence can be accurately tested, and the second historical time sequence can be accurately classified according to the causal test processing result.

(4) And comparing the causal test threshold with the causal test processing result, so that the second historical time sequence is accurately allocated to the significant sequence set or the non-significant sequence set, and the accurate classification of the second historical time sequence is realized.

(5) And screening each second historical time sequence in the non-significant sequence set through a difference threshold, a likelihood ratio threshold and a test probability threshold, and determining the second historical time sequence meeting the screening condition as a third historical time sequence, thereby realizing the effective screening of the second historical time sequence in the non-significant sequence set.

(6) The univariate prediction processing is carried out through the exponential smoothing model, the difference autoregressive model and the sequence decomposition model, so that the advantages of different univariate base class prediction models are fully utilized, and the prediction accuracy is effectively ensured.

(7) The advantages of the base class prediction models used for aggregating different visual angles are processed in an integrated mode, so that the finally obtained prediction time sequence corresponding to the first historical time sequence is not influenced by an abnormal value of a certain base class prediction model with incomplete information, high reliability, interpretability, accuracy and target adaptability of the base class prediction model are achieved as far as possible under the limited condition, and the prediction accuracy of the prediction time sequence corresponding to the first historical time sequence is effectively guaranteed.

The above description is only an example of the present application, and is not intended to limit the scope of the present application. Any modification, equivalent replacement, and improvement made within the spirit and scope of the present application are included in the protection scope of the present application.

Claims

1. A method for predicting a time series, the method comprising:

determining the second history time sequence of which the incremental test processing result meets an explanatory condition as a third history time sequence; and the third historical time sequence is used for prediction processing of a plurality of base class prediction models to obtain a prediction time sequence corresponding to the first historical time sequence.

2. The method of claim 1, wherein said causally testing said first historical time series with each of said second historical time series comprises:

performing the following processing for any one of the second historical time series:

determining a first time value corresponding to the first historical time sequence based on the p-order lag value sequence of the second historical time sequence and the q-order lag value sequence of the first historical time sequence;

the p-order hysteresis value sequence comprises values of all moments of the second historical time sequence within a range from a moment p to a current moment, the q-order hysteresis value sequence comprises values of all moments of the first historical time sequence within a range from a moment q to the current moment, the moment p and the moment q are any moments before the current moment, and the first moment value is any moment after the current moment;

determining a second time value corresponding to the first historical time sequence based on a q-order lag value sequence of the first historical time sequence, wherein the second time value is the same as the time corresponding to the first time value;

and carrying out variance ratio test processing on the first time value and the second time value, and determining the obtained variance ratio test result as a causal test processing result.

3. The method of claim 2, wherein said assigning each of said second historical time series to either a significant series set or a non-significant series set based on causal test treatment results comprises:

assigning the second historical time-of-day sequence to the set of significant sequences when the causal test treatment outcome is greater than or equal to a causal test threshold;

assigning the second historical time-of-day sequence to the set of non-significant sequences when the causal test treatment outcome is less than the causal test threshold.

4. The method of claim 1, wherein after assigning each of the second historical time series to either a significant sequence set or a non-significant sequence set based on causal test treatment results, the method further comprises:

allocating the second historical time series of the same type in the significant sequence set to a sequence group corresponding to the type;

allocating the second historical time sequence of the same type in the non-significant sequence set to a sequence group corresponding to the type;

wherein the sequence group includes the second historical time series in the set of significant sequences and the second historical time series in the set of non-significant sequences.

5. The method of claim 1, wherein performing an incremental check process on each of the second historical time series in the insignificant set based on the significant set, the insignificant set, and the first historical time series to obtain an incremental check process result comprises:

performing the following for any one of the second historical time series in the non-significant set:

determining a rank of a first test matrix from the first historical time series, the significant set, and the second historical time series;

determining a rank of a second test matrix according to the first historical time series and the significant set;

carrying out differential processing on the rank of the first test matrix and the rank of the second test matrix to obtain a rank difference value;

carrying out likelihood ratio test processing on the first historical time sequence, the significant set and the second historical time sequence to obtain a likelihood ratio value and a test probability value;

and determining the rank difference value, the likelihood ratio value and the test probability value as an incremental test processing result.

6. The method of claim 5, wherein determining the second historical time series as the third historical time series when the incremental test processing result satisfies the explanatory condition comprises:

and when the rank difference value is smaller than a difference threshold value, the likelihood ratio value is smaller than a likelihood ratio threshold value, and the test probability value is smaller than a test probability threshold value, determining the second history time sequence as a third history time sequence.

7. The method of claim 5, wherein determining the rank of a first test matrix from the first historical time series, the significant set, and the second historical time series comprises:

determining a sequence of i-th order lag values for the first historical time series, a sequence of j-th order lag values for each of the second historical time series in the significant set, and a sequence of k-th order lag values for the second historical time series;

wherein the i-order lag value comprises: the values of all the moments of the first historical time sequence within the range from the moment i to the current moment; the j-order hysteresis value sequence comprises: the second historical time sequence in the significant set takes values from the time j to all times within the range of the current time; the sequence of k-th order lag values comprises: the values of the second historical time sequence from the k moment to all moments in the current moment range;

carrying out difference processing on values of any two different moments in the i-order lag value sequence to obtain a first difference processing result;

performing the following for any one of the sequence of j-order lag values: carrying out differential processing on the values of any two different moments in the j-order lag value sequence to obtain a second differential processing result;

carrying out differential processing on values of any two different moments in the k-order lag value sequence to obtain a third differential processing result;

determining a rank of the first test matrix based on the first, second, and third differential processing results.

8. The method of claim 5, wherein determining a rank of a second test matrix based on the first historical time series and the significant set comprises:

determining a sequence of nth order lag values for the first historical time series and a sequence of mth order lag values for each of the second historical time series in the significant set;

wherein the nth order lag value comprises: the values of all the moments of the first historical time sequence within the range from the moment n to the current moment; the sequence of m-th order lag values includes: the second historical time series in the significant set take values from m time to all times within the range of the current time;

carrying out difference processing on the values of any two different moments in the n-order lag value sequence to obtain a fourth difference processing result corresponding to the values of any two different moments;

performing the following for any one of the sequences of m-order lag values: carrying out difference processing on the values at any two different moments in the m-order lag value sequence to obtain a fifth difference processing result corresponding to the values at any two different moments;

summing each fourth difference processing result to obtain a fourth summation processing result, and summing each fifth difference processing result to obtain a fifth summation processing result;

determining a rank of the second test matrix based on the fourth summation processing result and the fifth summation processing result.

9. The method of claim 1, further comprising:

calling the plurality of base class prediction models to perform prediction processing based on the third historical time sequence, the significant sequence set and the first historical time sequence to obtain a base class prediction sequence corresponding to each base class prediction model;

and performing integrated processing on each base class prediction sequence to obtain a prediction time sequence corresponding to the first historical time sequence.

10. The method of claim 9, wherein the types of base class prediction models include univariate base class prediction models, industry empirical models, and multivariate base class prediction models;

the step of calling a plurality of base class prediction models to perform prediction processing based on the third historical time sequence, the significant sequence set and the first historical time sequence to obtain a base class prediction sequence corresponding to each base class prediction model includes:

based on the first historical time sequence, sequentially calling a plurality of univariate base class prediction models to perform univariate prediction processing to obtain univariate base class prediction sequences respectively corresponding to the plurality of univariate base class prediction models;

calling the industry experience model to perform experience processing based on the first historical time sequence and scene experience information to obtain an experience base class prediction sequence corresponding to the industry experience model;

and sequentially calling a plurality of multivariate base class prediction models to perform multivariate prediction processing based on the third historical time sequence, the significant sequence set and the first historical time sequence to obtain multivariate base class prediction sequences respectively corresponding to the multivariate base class prediction models.

11. The method of claim 10, wherein the types of univariate base class prediction models comprise an exponential smoothing model, a differential autoregressive model, and a sequence decomposition model;

the method for obtaining the univariate base class prediction sequences respectively corresponding to the univariate base class prediction models by sequentially calling the univariate base class prediction models to perform univariate prediction processing based on the first historical time sequence comprises the following steps:

and sequentially calling the exponential smoothing model, the differential autoregressive model and the sequence decomposition model to perform univariate prediction processing based on the first historical time sequence to obtain univariate base class prediction sequences respectively corresponding to the exponential smoothing model, the differential autoregressive model and the sequence decomposition model.

12. The method of claim 10, wherein the types of multivariate base class prediction models comprise a vector autoregressive model and a long-short term memory network;

the step of sequentially calling a plurality of multivariate base class prediction models to perform multivariate prediction processing based on the third historical time sequence, the significant sequence set and the first historical time sequence to obtain multivariate base class prediction sequences respectively corresponding to the multivariate base class prediction models comprises:

and sequentially calling the vector autoregressive model and the long-short term memory network to perform multivariate prediction processing based on the third historical time sequence, the significant sequence set and the first historical time sequence to obtain multivariate base class prediction sequences respectively corresponding to the vector autoregressive model and the long-short term memory network.

13. A time series prediction apparatus, characterized in that the apparatus comprises:

an assignment module to assign each of the second historical time series to a significant sequence set or a non-significant sequence set based on causal test processing results;

14. An electronic device, characterized in that the electronic device comprises:

a memory for storing executable instructions;

a processor for implementing the method of predicting a time series according to any one of claims 1 to 12 when executing executable instructions or a computer program stored in the memory.

15. A computer-readable storage medium storing executable instructions or a computer program, wherein the executable instructions, when executed by a processor, implement the time series prediction method of any one of claims 1 to 12.

16. A computer program product comprising a computer program or instructions, characterized in that the computer program or instructions, when executed by a processor, implement the method of predicting a time series according to any one of claims 1 to 12.