CN115271219A

CN115271219A - Short-term load prediction method and prediction system based on causal relationship analysis

Info

Publication number: CN115271219A
Application number: CN202210918368.0A
Authority: CN
Inventors: 周群星; 贾昆; 佟芳; 段振华; 李宏波; 张容福; 李海龙; 马国雷; ***; 李文明; 杨凡; 翟娟荟; 林鑫; 徐铁军; 王婷; 马文珍; 张衡; 高天
Original assignee: State Grid Qinghai Electric Power Co Ltd; Information and Telecommunication Branch of State Grid Qinghai Electric Power Co Ltd
Current assignee: State Grid Qinghai Electric Power Co Ltd; Information and Telecommunication Branch of State Grid Qinghai Electric Power Co Ltd
Priority date: 2022-08-01
Filing date: 2022-08-01
Publication date: 2022-11-01

Abstract

The application relates to a short-term load prediction method and a short-term load prediction system based on causal relationship analysis, which comprises the following steps: acquiring an original load sequence, and calling an OVMD algorithm to decompose the original load sequence into a seasonal wave sub-sequence, a long-term trend sub-sequence and a random fluctuation sub-sequence; matching influence factors for all subsequences by calling a Granger causal relationship analysis algorithm, and acquiring an influence factor sequence corresponding to each subsequence; outputting a prediction result corresponding to the seasonal fluctuation subsequence through an ARIMR model; respectively outputting a prediction result corresponding to the long-term trend subsequence and a prediction result corresponding to the random fluctuation subsequence through a Bi-LSTM model; and superposing the prediction results of the subsequences to generate a final prediction result, and displaying the final prediction result. The load forecasting method and the load forecasting device have the effect of improving the load forecasting accuracy.

Description

Short-term load prediction method and prediction system based on causal relationship analysis

Technical Field

The application relates to the technical field of load prediction, in particular to a short-term load prediction method and a short-term load prediction system based on causal relationship analysis.

Background

The short-term load prediction means that the load data values of the next days or hours are scientifically predicted by combining external factors such as meteorological data, electricity price data, economic data and the like according to the change rule of the historical load data values.

The traditional short-term load forecasting method mainly comprises a similar day method, a regression analysis method, a time series method, fuzzy forecasting, an expert system and the like, wherein the expert system is a computer system for learning expert knowledge constructed by compiling programs, and can learn knowledge and experience of experts, convert the knowledge and experience into rules, perform modeling forecasting and finally obtain a forecasting result. However, the method is easy to generate artificial errors during prediction, the conversion of the experience and knowledge of experts into rules is difficult, and the system is mainly applied to specific scenes and has poor portability.

In the process of implementing the application, the inventor finds that at least the following problems exist in the technology: for short-term load prediction, an expert system is easy to generate artificial errors during prediction, and further the problem of poor accuracy of a load prediction result is easy to generate.

Disclosure of Invention

In order to solve the technical problem that the short-term power load prediction accuracy is poor in the prior art, the application provides a short-term load prediction method and a short-term load prediction system based on causal relationship analysis.

In a first aspect, the application provides a short-term load prediction method based on causal relationship analysis, which adopts the following technical scheme:

acquiring an original load sequence, and calling an OVMD algorithm to decompose the original load sequence into a seasonal wave subsequence, a long-term trend subsequence and a random fluctuation subsequence;

calling a Granger causal relationship analysis algorithm to match influence factors for each subsequence, and acquiring an influence factor sequence corresponding to each subsequence;

inputting the influence factor sequence corresponding to the seasonal wave subsequence as an input variable into a pre-constructed ARIMR model, and outputting a prediction result corresponding to the seasonal wave subsequence through the ARIMR model;

respectively taking the influence factor sequence corresponding to the long-term trend subsequence as an input variable and taking the influence factor sequence corresponding to the random fluctuation subsequence as an input variable, inputting the input variables into a pre-constructed Bi-LSTM model, and respectively outputting a prediction result corresponding to the long-term trend subsequence and a prediction result corresponding to the random fluctuation subsequence through the Bi-LSTM model;

and superposing the prediction results of the subsequences to generate a final prediction result, and displaying the final prediction result.

By adopting the technical scheme, the load data is time sequence signal data, and can decompose the original load sequence, provide useful characteristics and further improve the prediction precision of the model, so that the OVMD (optimal modal decomposition method) is adopted to carry out modal decomposition on the original load sequence to obtain high-quality load subsequences, namely a seasonal basic load subsequence, a static trend subsequence and a random fluctuation subsequence; secondly, matching influence factors with strong correlation with corresponding subsequences for each subsequence by adopting a Granger causal relationship analysis algorithm, so that the influence of irrelevant factors or weak relevant influence factors on a prediction result is reduced while less data prediction dimensionality is ensured; finally, matching proper prediction models for the subsequences respectively by combining the fluctuation characteristics of each subsequence along with time and the number of corresponding influence factors; specifically, the ARIMR model is adopted to predict the seasonal basic load subsequence, the Bi-LSTM model is adopted to predict the static trend subsequence and the random fluctuation subsequence respectively, then all the subsequence prediction results are overlapped to obtain the final prediction result, and the advantages of all the prediction models are fully exerted by adopting a combined prediction mode, so that the accuracy of the final prediction result is improved.

Optionally, the obtaining the original load sequence includes:

controlling a preset acquisition device to acquire a load data value every other preset time, and receiving the load data value x (d, t), wherein x (d, t) is the load data value at the t moment on the day d;

combining all received load data values x (d, t) to generate an original load sequence x (t); wherein, x (t) is a data set formed by a plurality of x (d, t) according to the receiving time sequence;

if two adjacent load data values x (d, t 1) and x (d, t 2) exist in the original load sequence x (t), and the condition that | t1-t2| ≠ preset time length is met, determining the missing data quantity S between x (d, t 1) and x (d, t 2), and S = | t1-t2 |/preset time length; starting from v =1, calculate

Until v = S; wherein

And

all are preset weight values;

all the calculated x (d, t) are divided according to the sequence of v from small to large _1+v ) Added to the original load sequence x (t) between x (d, t 1) and x (d, t 2).

Along with the extensive application and popularization of smart electric meter, real-time load data acquisition's speed and accuracy all promote greatly, however at data acquisition's in-process, often can lead to the load data to be lacked because of all kinds of uncontrollable factors, if adopt above-mentioned missing data to carry out load characteristic analysis and load prediction, then can make the great error of load prediction result appearance, consequently, in order to avoid appearing great deviation, this application supplements the data of disappearance through adopting above-mentioned technical scheme, the corresponding processing is done to the data of the same day type of specifically adopting, with the influence that the above-mentioned problem of furthest reduces brought.

Optionally, after combining all received x (d, t) to generate the original load sequence x (t), the method further includes:

if the first load data value x (d, t) exists in the original load sequence x (t), the following formula is satisfied:

|x(d,t)-x(d,t-1)|>α(t)；

|x(d,t)-x(d,t+1)|>β(t)

modifying the first load data value x (d, t) to:

wherein x (d, t), x (d, t-1) and x (d, t + 1) respectively represent load data values at the t-th moment, the t-1-th moment and the t + 1-th moment of the d-th day, and alpha (t) and beta (t) are preset threshold values;

if a second load data value x (d, t) exists in the original load sequence x (t), satisfying: | x (d, t) -m (t) & gt>r (t), then the second load data value x (d, t) is modified to

Wherein r (t) is a preset threshold, and m (t) represents the average value of the load in y days before and after the receiving time corresponding to x (d, t).

Under normal conditions, the load data should be continuous and smooth, if the difference between the load data of two adjacent points is too large, the load data may be a data abnormal point in the data acquisition process, and if the load characteristics are analyzed or predicted by using the abnormal data, not only the characteristics of the load performance are wrong, but also the load prediction result is greatly deviated.

Optionally, the invoking OVMD algorithm decomposes the original load sequence into a seasonal wave subsequence, a long-term trend subsequence, and a random wave subsequence, and includes:

calling a VMD decomposition method to decompose an original load sequence x (t), and setting an initial value of a parameter K carried in the VMD decomposition method to be 3, wherein K refers to the number of subsequences decomposed by the VMD;

calculating each subsequence after decomposition

Pearson's correlation coefficient with original load sequence x (t)

Calculating a residual component epsilon _K Pearson's correlation coefficient R between (t) and original payload sequence x (t) _K Wherein, in the step (A),

the number of the finger subsequence is J-th subsequence with K, K =3,4,5 \8230; j =1,2,3, \ 8230;, K;

if R is _K Satisfy the requirement of

Let K = K +1, based on the modified K value, call VMD decomposition method to decompose the original load sequence x (t) again, and calculate

And R _K ；

If it satisfies

And R is _K ≤ψ·R ₃ Determining the current K value as the optimal K value, wherein psi is a preset threshold value; and based on the determined optimal K value, calling an OVMD algorithm to decompose the original load sequence into a seasonal wave subsequence, a long-term trend subsequence and a random wave subsequence.

K is used as a required parameter in OVMD decomposition and is used for representing the number of decomposed subsequences, and when the value of K is too small, the decomposed subsequences are always subjected to aliasing phenomenon, so that the original intention of reducing the randomness fluctuation of the original load sequence is violated; when the value is too large, the phenomenon of excessive decomposition occurs, so that the complexity of sequence analysis and the difficulty of load prediction are increased; therefore, by adopting the technical scheme, before decomposition is carried out by using the OVMD method, the original load sequence is decomposed by adopting the VMD decomposition method one by one from K =3, and R is calculated by using a Pearson correlation coefficient calculation formula _K And

and circularly judging until the termination condition is met:

and R _K ≤ψ·R ₃ . Wherein the Pearson correlation coefficient is used to characterize the magnitude of the correlation, e.g.

Represents each subsequence U _KJ Magnitude of correlation, R, with original payload sequence x (t) _K Representing the residual component epsilon _K (t) the magnitude of the correlation between the original payload sequence x (t). End conditions

The sufficiency of VMD decomposition is ensured through the correlation between the original load sequence and each subsequence, and the occurrence of aliasing between the decomposed subsequences is effectively avoided; since in general R _K Will exhibit a decreasing tendency as K increases, and therefore, the termination condition R is set _K ≤ψ·R ₃ To determine the magnitude of the correlation coefficient between the original payload sequence and the residual components and to ensure a sufficient decomposition of the original payload sequence by setting a threshold psi. Finally, after the termination condition is met, the K value at the moment is the optimal K value determined by the method,

the setting of (2) also ensures that the target sequence does not show over-decomposition.

Optionally, the step of calling an OVMD algorithm to decompose the original load sequence into a seasonal wave subsequence, a long-term trend subsequence, and a random wave subsequence based on the determined optimal K value includes:

constructing a two-dimensional plane coordinate system, randomly setting a plurality of alpha values and a plurality of tau values, and generating a plurality of groups of (alpha, tau) coordinate information, wherein the alpha values and the tau values are self-contained parameters of the OVMD;

based on the determined optimal K value and a plurality of preset (alpha, tau) coordinate information, calling a VMD algorithm to decompose the original load sequence x (t) to obtain subsequences corresponding to decomposition positions of different combinations (alpha, tau)

Wherein the content of the first and second substances,

when the optimal K value is shown, the decomposed K-th subsequence corresponding to the g-th group (alpha, tau) is g =1,2,3 \8230, K =1,2 \8230;

combining the decomposed subsequences corresponding to each group (alpha, tau)

Calculating an adaptation coefficient E for each group (α, τ), wherein,

ε _g (t) when the optimal K value is obtained, residual components corresponding to the g-th group (alpha, tau) are obtained, and N is the sequence length of the original load sequence x (t);

determining coordinate information of a combination parameter (alpha, tau) corresponding to the minimum E value, and performing iterative updating on the position information of each parameter combination (alpha, tau) by utilizing a PSO algorithm;

if the minimum E value is larger than the preset minE and the iteration updating times do not reach the preset maximum iteration times, recalculating the adaptive coefficient E corresponding to each (alpha, tau) after updating, determining the coordinate information of the (alpha, tau) corresponding to the minimum E value, and calling a PSO algorithm to perform iteration updating on the position information of each parameter combination (alpha, tau);

if the minimum E value is less than or equal to a preset minE or the iteration updating times reach a preset maximum iteration time, determining the coordinate information of (alpha, tau) corresponding to the current minimum E value, and determining an optimal alpha value and an optimal tau value;

and decomposing the original load sequence into a seasonal wave sub-sequence, a long-term trend sub-sequence and a random wave sub-sequence by adopting an OVMD method based on the determined optimal K value, optimal alpha value and optimal tau value.

Alpha and tau are preset parameters required by OVMD decomposition, alpha refers to a penalty factor, tau refers to a guarantee coefficient, for the penalty factor alpha and the fidelity coefficient tau, different setting combinations of the penalty factor alpha and the fidelity coefficient tau can cause different degrees of residual errors (namely epsilon (t)) to appear in a final decomposition result, and large interference is generated on accuracy of load prediction.

Optionally, decomposing the original load sequence into a seasonal wave sub-sequence, a long-term trend sub-sequence and a stochastic wave sub-sequence by using an OVMD method based on the determined optimal K value, optimal α value and optimal τ value, includes:

decomposing the original load sequence by adopting an OVMD method to obtain a plurality of optimal subsequences based on the determined optimal K value, optimal alpha value and optimal tau value;

if the number of the optimal subsequences is more than 3, calculating the approximate entropy of each optimal subsequence, combining the two optimal subsequences with the minimum difference value of the approximate entropy to generate a new optimal subsequence until the number of the final optimal subsequences is 3, and determining the 3 optimal subsequences as a seasonal wave subsequence, a long-term trend subsequence and a random fluctuation subsequence;

and if the number of the optimal subsequences is 3, determining the 3 optimal subsequences as a seasonal wave subsequence, a long-term trend subsequence and a random wave subsequence according to the change rule of each optimal subsequence.

By adopting the technical scheme, the final prediction task amount of each subsequence is reduced, when the optimal K value is not 3, two optimal subsequences with the minimum approximate entropy difference in all subsequences corresponding to the optimal K value are merged to form a new optimal subsequence until the number of all subsequences is 3, so that the seasonal base load subsequence, the static trend subsequence and the random fluctuation subsequence are matched.

Optionally, the invoking of the Granger causal relationship analysis algorithm matches influence factors for each subsequence, and obtaining an influence factor sequence corresponding to each subsequence includes:

performing stationarity check on each subsequence, and converting each non-stationarity subsequence into a stable sequence;

based on a plurality of preset factors to be evaluated, acquiring each factor sequence C to be evaluated ⁱ (t)，C ⁱ (t)＝(C ⁱ 1,C ⁱ 2,C ⁱ 3,…,C ⁱ m) wherein C ⁱ (t) refers to a data set of the ith factor to be evaluated in a specified historical period, and m refers to the sequence length of the corresponding factor sequence to be evaluated;

calling Granger causal relationship analysis algorithm to determine each factor C to be evaluated ⁱ (t) and each subsequence U ^h (t) correlation value of, wherein U ^h (t)＝(U ^h 1,U ^h 2,U ^h 3,…,U ^h n)，U ^h (t) represents the h-th load subsequence, n refers to the number of sample points of the corresponding subsequence;

if the target C exists ⁱ (t) and target U ^h (t) so that the target C ⁱ (t) with the target U ^h (t) if the correlation value is higher than a preset correlation value, the target C is determined ⁱ (t) adding to the target U ^h (t) a predetermined sequence of influencing factors.

Under the power market environment, the influence of multiple uncertain factors can cause a lot of interference on the development of load prediction work, in terms of short-term load, factors such as electricity price, historical load, weather and the like are main influence factors of the load, and after the original load sequence is decomposed, the influence factor sequences with strong correlation are matched for each subsequence, so that the redundancy of variable selection is reduced, the model prediction difficulty is reduced, and the prediction precision is ensured; currently, a common method is to determine the magnitude of the correlation between the load subsequence and the variable sequence, and to screen the influence factors with higher correlation with the load subsequence in the variable sequence as the prediction variables; however, in actual use, the prediction model tends to fall into "pseudo regression" under all conditions in which the magnitude of the correlation between sequences is merely selected as the influence factor of the load subsequence. Therefore, in view of the unique advantages of the causal relationship analysis method in the aspect of feature extraction, the application proposes to match the relevant influencing factors of the load subsequence by using the Granger causal relationship analysis method so as to ensure that the influencing factors have a significant driving effect on the load.

In a second aspect, the present application provides a short-term load prediction system based on causal relationship analysis, which adopts the following technical solutions: the short-term load prediction system includes:

the original sequence decomposition module is used for acquiring an original load sequence and calling an OVMD algorithm to decompose the original load sequence into a seasonal wave subsequence, a long-term trend subsequence and a random wave subsequence;

the influence factor matching module is used for calling a Granger causal relationship analysis algorithm to match influence factors for each subsequence and acquiring an influence factor sequence corresponding to each subsequence;

the subsequence prediction module is used for inputting the influence factor sequence corresponding to the seasonal wave subsequence serving as an input variable into a pre-constructed ARIMR model and outputting a prediction result corresponding to the seasonal wave subsequence through the ARIMR model;

the subsequence prediction module is further used for inputting the influence factor sequence corresponding to the long-term trend subsequence as an input variable and the influence factor sequence corresponding to the random fluctuation subsequence as an input variable into a pre-constructed Bi-LSTM model, and respectively outputting the prediction result corresponding to the long-term trend subsequence and the prediction result corresponding to the random fluctuation subsequence through the Bi-LSTM model;

and the prediction result generation module is used for superposing the prediction results of the subsequences output by the subsequence prediction module to generate a final prediction result and displaying the final prediction result.

In a third aspect, the present application provides a short term load prediction platform comprising a memory and a processor, the memory having stored thereon a computer program that is loadable by the processor and operable to perform the method of any of the first aspects.

In a fourth aspect, the present application provides a computer readable storage medium storing a computer program that can be loaded by a processor and executed to perform the method as in any one of the first aspects.

In summary, the present application includes at least one of the following beneficial technical effects:

1. in the application, an OVMD method is adopted to decompose an original load subsequence to extract a subsequence which is convenient for analyzing influence factors, and then a Granger causal relationship analysis method is adopted to further determine the influence factors with strong correlation with the decomposed subsequence so as to improve the accuracy of subsequent prediction; finally, a combined prediction mode is adopted, the influence factors of each subsequence are used as input quantity, the prediction result corresponding to each subsequence is output, and finally the prediction results are combined to generate the final prediction result;

2. further, by calculating each subsequence corresponding to different K

Pearson correlation coefficient with original load sequence x (t)

And a corresponding residual component epsilon _K Pearson's correlation coefficient R between (t) and original payload sequence x (t) _K To do so by

And R is _K ≤ψ·R ₃ As a condition, the optimal K value is calibrated; and the optimal (alpha, tau) parameter combination is ratified through a PSO algorithm, so that the decomposition effect of the OVMD on the original load sequence is optimized.

Drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings needed to be used in the description of the embodiments will be briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without creative efforts.

Fig. 1 is a block flow diagram of a short-term load prediction method based on causal relationship analysis in an embodiment of the present application.

Fig. 2 is a schematic diagram illustrating steps of a short-term load prediction method based on causal relationship analysis in an embodiment of the present application.

FIG. 3 is a diagram illustrating the calibration procedure for embodying the value of K in the example.

FIG. 4 is a schematic diagram of the calibration step for embodying (α, τ) in the example.

Fig. 5 is a sequence diagram of a sub-sequence for embodying the original payload sequence to be decomposed in the embodiment.

Fig. 6 is a block diagram of a short-term load prediction method based on causal relationship analysis in the embodiment of the present application.

Description of reference numerals: 1. an original sequence decomposition module; 2. an influence factor matching module; 3. a subsequence prediction module; 4. and a prediction result generation module.

Detailed Description

The present application is described in further detail below with reference to the accompanying drawings.

The embodiment of the application discloses a short-term load prediction method based on causal relationship analysis, which is mainly used for analyzing and determining influence factors influencing the change of historical load data values in combination with the change rule of the historical load data values in a time dimension, and predicting the load data values in a period of time in the future by using the historical load data values and the corresponding influence factors.

The short-term load forecasting method is characterized in that an execution main body is a short-term load forecasting platform assisted by acquisition equipment; the short-term load forecasting platform comprises a memory and a processor, the collecting device is used for collecting historical load data values and can be an intelligent electric meter, and the short-term load forecasting platform is in communication connection with the collecting device and used for controlling the collecting device to collect the load data values and receiving the load data values collected by the collecting device.

With reference to fig. 1, the short-term load prediction method of the present application has the following main ideas: firstly, integrating historical load data values into an original load sequence in a time sequence; then, performing modal decomposition on the load sequence by adopting an optimal modal decomposition method (OVMD) to obtain a plurality of high-quality load subsequences; then, matching influence factors for each load subsequence by adopting a Granger causal relationship analysis algorithm so as to enable the matched influence factors and the corresponding subsequences to have strong correlation, obtain an optimal matching result and reduce the redundancy of data dimensions; selecting a specific prediction model for each subsequence according to the variation characteristics of each subsequence, predicting the corresponding subsequence by using the prediction model to obtain a prediction result, and finally overlapping the prediction results of all the subsequences to obtain a final prediction result; the accuracy of the prediction result is improved by using a method of combining and predicting a plurality of prediction models.

The following describes in detail a specific implementation method of the main operation flow of the short-term load prediction method with reference to fig. 2, which includes the following steps:

step 101, acquiring an original load sequence, and decomposing the original load sequence into a seasonal wave sub-sequence, a long-term trend sub-sequence and a random wave sub-sequence by using an OVMD method.

In implementation, step 101 specifically includes the following sub-steps:

the short-term load forecasting platform controls the acquisition equipment to acquire the load data value of a certain area at intervals of preset time, wherein the preset time can be 1 hour, and the specifically acquired load data value can be an electric quantity value; after the short-term load forecasting platform receives the load data value, the load data value is stored in the following form according to the receiving time: x (d, t), wherein x (d, t) represents the load data value received at the tth time of day d of the short-term load forecasting platform. The time delay between the time when the load data value is acquired and the time when the short-term load prediction platform receives the load data value is ignored, so that x (d, t) can also be regarded as the power consumption at the t moment of the d day; and then, the short-term load forecasting platform sequentially arranges and collects the received load data values according to the sequence of the receiving time to generate an original load sequence x (t), wherein x (t) is a data set formed by a plurality of x (d, t).

In the data acquisition process, the acquired load data values are often missing or abnormal data exist due to various uncontrollable factors (the missing of the load data values specifically means the missing of the load data values at a certain moment, and the abnormal data means that the difference value of the load data values corresponding to two adjacent time points is greater than an ideal difference value), and the existence of the abnormal data easily causes the deviation of the results of the analysis of the load characteristics and the load prediction.

Therefore, in order to avoid the problem of missing load data values, step 101 specifically includes the following operations:

the short-term load forecasting platform calculates the time difference of the receiving time corresponding to any two adjacent load data values from the first load data value so as to judge whether the original load sequence has the problem of data value missing; specifically, if two adjacent load data values x (d, t 1) and x (d, t 2) exist, and | t1-t2| ≠ preset duration, it is indicated that a missing load data value exists between x (d, t 1) and x (d, t 2), at this time, the short-term load prediction platform determines the number S of the missing load data values between x (d, t 1) and x (d, t 2), and S = | t1-t2 |/preset duration; next, the short-term load prediction platform starts from v =1 and calculates x (d, t) _1+v ) Until v = S. Wherein

And

all are preset weight values; finally, the short-term load forecasting platform calculates the calculated x (d, t) according to the sequence of v from small to large _1+v ) And adding the data to the original load sequence x (t) between x (d, t 1) and x (d, t 2) to supplement the missing data.

In order to avoid the problem of load data value anomaly, step 101 specifically includes the following operations:

the short-term load prediction platform will start from the second load data value in x (t), and will bring all load data values one by one into the following two formulas:

| x (d, t) -x (d, t-1) | > α (t); formula 1-1

I x (d, t) -x (d, t + 1) | > beta (t), formula 1-2

If the target load data value x (d, t) simultaneously satisfies the formula 1-1 and the formula 1-2, the target load data value x (d, t) is an abnormal load data value, and at this time, the short-term load prediction platform modifies the target load data value x (d, t) to be:

wherein, x (d, t), x (d, t-1) and x (d, t + 1) respectively represent the load data values at the t-th time, the t-1-th time and the t + 1-th time of the d day, and alpha (t) and beta (t) are preset threshold values.

In addition, the short-term load prediction platform substitutes any load data value x (d, t) in the original load sequence x (t) into the following formula:

| x (d, t) -m (t) | > r (t); formulas 1 to 3

If x (d, t) satisfies:

where r (t) is a preset threshold, and m (t) represents an average value of the load data in y days before and after the receiving time corresponding to x (d, t), in this embodiment, y may be specifically 3, that is, an average value of the load data in 6 days in total in 3 days before t and 3 days after t.

Optionally, as can be seen from the foregoing, the original load sequence generated by the acquired load data is a time sequence itself, and the time sequence is a superposition of subsequences of the following four types of variations, that is: a long-term trend subsequence (the data value of the long-term trend subsequence shows a trend in a long period, for example, the data value changes all the time in a long period are always in an ascending trend), a seasonal fluctuation subsequence (the seasonal fluctuation subsequence generates regular changes due to the seasonal changes), a cyclic fluctuation subsequence (the cyclic wave subsequence has a period of several years and has irregular periodic continuous changes), and an irregular fluctuation subsequence (the irregular wave subsequence shows irregular changes due to the influence of a plurality of accidental factors); therefore, in order to analyze the change rule of the original load series more precisely and improve the prediction accuracy of the future short-term load, after the original load sequence x (t) is generated by the short-term load prediction platform and the original load sequence x (t) is checked for the abnormal problems, the short-term load prediction platform decomposes the original load sequence by using an OVMD method, and correspondingly, the superposition of the load subsequences obtained by subsequent decomposition is the original load sequence.

It should be noted that OVMD refers to an optimal modal decomposition method, which is a further optimization decomposition of a sequence proposed on the basis of VMD (variational modal decomposition); OVMD has four parameters: the number K of subsequences to be decomposed, a penalty factor alpha, a guarantee coefficient tau and a decomposition cut-off condition epsilon.

For the parameter epsilon, since the influence of the decomposition cut-off condition epsilon on the final decomposition result of the sequence is small, the parameter epsilon is often set according to a default value, and the specific value is the prior art, and is not described herein again.

For the parameter K, when the value of the parameter K is too small, aliasing phenomena among the decomposed subsequences are often caused, so that the original purpose of reducing the randomness fluctuation of the original load sequence is violated; when the value is too large, the over-decomposition phenomenon occurs, and the complexity of sub-sequence analysis and the difficulty of load prediction are increased. For the penalty factor α and the fidelity coefficient τ, different setting combinations of (α, τ) may cause different degrees of residual epsilon (t) to appear in the final decomposition result, and may cause great interference to the accuracy of the load prediction. For this reason, the present application provides the following way to rate the optimal value of the parameter K, the optimal value of the parameter α, and the optimal value of the parameter τ before decomposition with OVMD.

Specifically, with reference to fig. 3, the calibration procedure for the parameter K is as follows:

decomposing an original load sequence x (t) by adopting a VMD decomposition method, and setting an initial value of a parameter K carried in the VMD decomposition method to be 3, wherein K refers to the number of subsequences decomposed by the VMD;

calculating each subsequence after decomposition

Pearson's correlation coefficient with original load sequence x (t)

Calculating the residual component ε _K Pearson's correlation coefficient R between (t) and original payload sequence x (t) _K Wherein, in the step (A),

the number of the finger subsequence is J-th subsequence with K, K =3,4,5 \8230; j =1,2,3, \8230;, K;

if R is _K Satisfy the requirement of

Let K = K +1, based on the modified K value, the VMD decomposition method is used to decompose the original load sequence x (t) again, and calculate

And R _K ；

If it satisfies

And R is _K ≤ψ·R ₃ Then the current K value is determined as the optimal K value, where ψ is a preset threshold value.

In implementation, the short-term load prediction platform first adopts a traditional VMD decomposition method to decompose the original load sequence x (t), and sets an initial value of a parameter K carried in the VMD decomposition method to be 3, and then starts from K =3,the short-term load forecasting platform respectively calculates each decomposed subsequence

Pearson's correlation coefficient with original load sequence x (t)

(as in equations 1-4 below), the residual component ε is calculated _K Pearson's correlation coefficient R between (t) and historical loading sequence x (t) _K (formulae 1-5 and formulae 1-6, hereinafter)

R _K ＝p(x(t),ε _K (t)); formulas 1 to 6

Wherein, the first and the second end of the pipe are connected with each other,

denotes the J-th subsequence, ε, when the number of subsequences is K _K (t) the residual sequence obtained by subtracting all subsequences corresponding to the current K value from the original load sequence, wherein K =3,4,5 \8230; j =1,2,3, \8230; the Pearson correlation coefficient is used to measure the degree of correlation between two variables, and thus

Representing subsequences

Degree of correlation, R, with the original load sequence x (t) _K Representing the residual component epsilon _K (t) and the historical load sequence x (t), and the larger the absolute value of the pearson correlation coefficient, the greater the degree of correlation.

If R is _K Satisfy the requirements of

And enabling K = K +1 by the short-term load prediction platform, decomposing the original load sequence x (t) by adopting the VMD decomposition method again based on the modified K value, and calculating

And R _K (ii) a Until it meets

And R is _K ≤ψ·R ₃ And then, the short-term load prediction platform determines the current K value as the optimal K value, and finally the optimal K value is calibrated.

Optionally, with reference to fig. 4, the calibration procedure for the parameter combination (α, τ) is as follows:

based on the determined optimal K value and a plurality of preset (alpha, tau) coordinate information, calling a VMD algorithm to decompose the original load sequence x (t) to obtain subsequences at decomposition positions corresponding to different combinations (alpha, tau)

Wherein the content of the first and second substances,

when the optimal K value is represented, the kth subsequence decomposed corresponding to the g group (alpha, tau) has g =1,2, 3\8230, K =1,2 \8230, K;

combining the decomposed subsequences corresponding to each group (alpha, tau)

Calculating an adaptation coefficient E for each group (α, τ), wherein,

determining coordinate information of a combination parameter (alpha, tau) corresponding to the minimum E value, and calling a PSO algorithm to iteratively update the position information of each parameter combination (alpha, tau);

if the minimum E value is larger than the preset minE and the iterative updating times do not reach the preset maximum iterative times, recalculating the adaptive coefficient E corresponding to each (alpha, tau) after updating, determining the coordinate information of (alpha, tau) corresponding to the minimum E value, and performing iterative updating on the position information of each parameter combination (alpha, tau) by utilizing a PSO algorithm;

if the minimum E value is less than or equal to the preset minE or the iteration updating times reach the preset maximum iteration times, determining the coordinate information of (alpha, tau) corresponding to the current minimum E value, and determining the optimal alpha value and the optimal tau value;

in implementation, the short-term load prediction platform firstly constructs a two-dimensional plane coordinate system, randomly combines and generates a plurality of (alpha, tau) coordinate information by taking the value of alpha as an abscissa and the value of tau as an ordinate, and marks all the set (alpha, tau) coordinate information in the constructed two-dimensional plane coordinate system. Then, based on the determined optimal K value and a plurality of preset (alpha, tau) coordinate information, the short-term load prediction platform calls a VMD algorithm to decompose the original load sequence x (t) to obtain decomposition subsequences corresponding to different combinations (alpha )

The short-term load prediction platform calculates an adaptive coefficient E of each combination (alpha, tau), the adaptive coefficient is a measuring standard used for measuring whether optimal solution is achieved in a PSO algorithm, the PSO algorithm (particle swarm optimization algorithm) is a random search algorithm based on swarm cooperation and developed by simulating bird swarm foraging behavior, all preset (alpha, tau) combination values are processed by the PSO algorithm to find the optimal (alpha, tau) combination, and the value of the adaptive coefficient E corresponding to the optimal (alpha, tau) combination is smaller than the preset minE.

After calculating and obtaining the adaptive coefficients E of all the groups (alpha, tau), the short-term load prediction platform determines the coordinate information of the (alpha, tau) corresponding to the minimum E value, and then calls a PSO algorithm to adjust the position information of each (alpha, tau) combination, so that all the (alpha, tau) combinations move from the initial coordinate position to the coordinate position close to the combination parameter (alpha, tau) corresponding to the optimal E value, and the iterative updating of the position information of each (alpha, tau) combination is realized.

When the position information of the (alpha, tau) combination is iteratively updated through the PSO each time, the short-term load prediction platform will count the number of iterative updates, and then, the short-term load prediction platform will determine whether the current minimum E value or the iterative update number satisfies a condition, where the condition to be satisfied is specifically: the minimum E value is smaller than the preset minE, or the iteration updating times reach the maximum iteration updating times.

If the minimum E value is larger than the preset minE and the current iteration updating times do not reach the maximum iteration times, the short-term load prediction platform calculates the adaptive coefficient E corresponding to each (alpha, tau) after updating, then iteratively updates the position information of each parameter combination (alpha, tau) according to the minimum E value and the coordinate information of (alpha, tau) corresponding to the minimum E value from the adaptive coefficient E, and calls a PSO algorithm; the above cycle is repeated until the aforementioned conditions are satisfied. If the condition is met, the short-term load prediction platform determines coordinate information of (alpha, tau) corresponding to the current minimum E value, determines an abscissa value corresponding to (alpha, tau) as an optimal alpha value, and determines an ordinate value corresponding to (alpha, tau) as an optimal tau value; in this embodiment, the optimal α value is 1526, and the optimal τ value is 0.371.

Optionally, after the optimal K value, the optimal α value, and the optimal τ value are calibrated, the short-term load prediction platform invokes an OVMD method to decompose the original load sequence, and combines the foregoing: the time series can be decomposed into a long-term trend subsequence, a seasonal fluctuation subsequence, a cyclic fluctuation subsequence, an irregular fluctuation subsequence, and combined with the load data predicted by the present application in a short term (days/months) in the future, so the present application combines the cyclic wave subsequence and the irregular fluctuation subsequence to be regarded as a random fluctuation subsequence, and limits the number of the finally decomposed subsequences to 3, so as to be convenient for matching the long-term trend subsequence, the seasonal fluctuation subsequence, and the random fluctuation subsequence. Considering that the decomposition number of the final optimal subsequence is determined by the optimal K value, if the optimal K is not equal to 3, the method adopts the following steps to combine the finally decomposed subsequence number into 3 subsequences;

In implementation, the approximate entropy is a random complexity, and can be used for describing the irregularity of the time series, and the more irregular the time series, the larger the corresponding approximate entropy is; therefore, the sub-sequences with similar entropy (namely the difference value of the approximate entropy is minimum, namely the complexity of two corresponding sub-sequences is similar) can be merged by calculating the approximate entropy of each optimal sub-sequence; the specific combination mode can be as follows: according to the position of the load data value in the sub-sequence, the load data values at the same position are weighted and summed to realize combination, for example, for two sub-sequences U needing to be combined ₁ (t) and U ₂ (t) mixing U ₁ (t) first load data value and U ₂ And (t) weighting and summing the first load data value to obtain the first load data value of the subsequences after combination.

As to how to determine the one-to-one correspondence relationship between the seasonal wave sub-sequence, the long-term trend sub-sequence, and the stochastic fluctuation sub-sequence and the optimal sub-sequence, the determination method may be obtained from the prior art, for example, by a manual determination method, the short-term load prediction platform may display a sequence diagram (as shown in fig. 5) of each optimal sub-sequence through a display device provided therein, the sequence diagram takes time (e.g., hours) as an abscissa and load values (unit of load value is megawatts) as an ordinate, and simultaneously displays a selection button for an operator to determine the one-to-one correspondence relationship between the seasonal wave sub-sequence, the long-term trend sub-sequence, and the stochastic fluctuation sub-sequence and the optimal sub-sequence according to experience; specifically, the subsequence with the largest fluctuation range is defined as a seasonal subsequence (such as IMF1 in fig. 5), the subsequence with the smallest fluctuation range is defined as a random fluctuation subsequence (such as IMF3 in fig. 5), and the remaining subsequence is defined as a long-term tendency subsequence (such as IMF2 in fig. 5).

And 102, calling a preset Granger causal relationship analysis algorithm to match influence factors for the subsequences, and acquiring an influence factor sequence corresponding to each subsequence.

In implementation, the influence factors make a change rule of the load sequence, the influence factors include internal factors and external factors, the internal factors refer to all historical load data values influencing future load changes in a historical time period, a set formed by all the historical load data values is the historical load sequence influencing the load changes, a next-hour load data value is influenced by a previous N-hour load data value, a sequence formed by the previous N-hour load data value is the historical load sequence, and N is the sequence length of the historical load sequence.

The external factor C refers to the influence of the external environment on the load data value, including temperature, electricity price, economy, humidity and the like, wherein C = (C) ¹ ,C ² ,…,C ⁱ )，C ⁱ Refers to the i influencing factors. In order to determine which influence factors influence each sufficient subsequence, the short-term load prediction platform takes each external factor as a factor to be evaluated, and the following steps are carried out:

based on a plurality of preset factors to be evaluated, each factor to be evaluated is obtainedSequence C of the hormone ⁱ (t)，C ⁱ (t)＝(C ⁱ 1,C ⁱ 2,C ⁱ 3,…,C ⁱ m) wherein, C ⁱ (t) refers to a data set of the ith factor to be evaluated in a specified historical period, and m refers to the sequence length of the corresponding factor sequence to be evaluated;

if there is a target C ⁱ (t) and target U ^h (t) so that the object C ⁱ (t) with the target U ^h (t) if the correlation value is higher than the preset correlation value, the target C is determined ⁱ (t) adding to the target U ^h (t) a predetermined sequence of influencing factors.

In the implementation, as the Granger causal analysis algorithm is used on the premise that the subsequences are required to be guaranteed to be stable sequences, the short-term load prediction platform firstly adopts an ADF (automatic document surface) verification method to perform stability verification on each subsequence, and if non-stable subsequences exist, the non-stable subsequences are converted into stable subsequences in a difference mode, and the difference can be performed for multiple times until the non-stable subsequences are converted into stable subsequence positions; then, the short-term load prediction platform can use a Granger causal analysis algorithm through preinstalled Eviews software, and meanwhile, the short-term load prediction platform newly establishes a new hypothesis H0, wherein the content of the hypothesis H0 is as follows: factor sequence C to be evaluated ⁱ (t) is the subsequence U ^h (t) influence factor.

The short-term load prediction platform verifies the confidence level of the hypothesis H0 (namely, the confidence level of the hypothesis H0) through the Granger causal analysis algorithm, and uses the confidence level to characterize the factor sequence C to be evaluated mentioned in the previous step ⁱ (t) and subsequence U ^h (t) correlation values between; wherein the higher the confidence level, the higher the corresponding correlation value. Accordingly, the predetermined correlation value may be characterized by a predetermined confidence level value (pre-prediction of the present application)The confidence level value is set to 0.99). The finally obtained U of each subsequence ^h (t) and C ⁱ The confidence level values for (t) are shown in the following table:

TABLE 1 Table 1 analysis of Granger's causal relationship between seasonal fluctuation subsequences and various extrinsic factors

Factor to be evaluated	Confidence level
		Temperature of	1.00000
Price of electricity	0.92132
		Wind speed	0.33247
Illuminance of light	0.21873
		Humidity	0.98745

TABLE 2 Granger causal relationship analysis of Long term trending subsequences with various extrinsic factors

Factor to be evaluated	Confidence level
		Temperature of	1.00000
Electricity price	1.00000
		Wind speed	0.93252
Illuminance of light	0.93714
		Humidity of air	0.99999

TABLE 3 Granger's causal relationship analysis of stochastic fluctuation subsequences and extrinsic factors

Factor to be evaluated	Confidence level
		Temperature of	1.00000
Electricity price	0.03257
		Wind speed	0.99999
Illuminance of light	0.93714
		Humidity of air	0.19256

According to the 3 tables, the temperature is the influence factor corresponding to the seasonal wave subsequences; the temperature, the electricity price and the humidity are influence factors corresponding to the long-term trend subsequence; the temperature and the wind speed are influence factors corresponding to the random fluctuation subsequence.

In addition, how to determine the effect U ^h (t) the sequence length N of the historical payload sequence can be determined by trial-by-trial method, N =1,2,3 \8230n, N is the subsequence U ^h (t) sequence length. The short-term load forecasting platform firstly obtains U from N = N ^h (t) the first N load data values form a historical load sequence, and then a short-term load forecasting platform creates an assumption H1 that the historical load sequence formed by the first N load data values is a subsequence U ^h (t) determining the confidence level value of the hypothesis H1 by calling a Granger causal analysis algorithm; the short-term load prediction platform determines historical load sequences and U corresponding to different sequence lengths N ^h (t) the confidence level of the hypothesis corresponding to the step (t), determining the N 'corresponding to the hypothesis with the highest confidence level, and finally adding the historical load sequence corresponding to the N' into the influence factor sequence C to finally obtain all the sub-sequences U ^h (t) sequence of influencing factors.

And 103, inputting the influence factor sequence corresponding to the seasonal fluctuation subsequence as an input variable into a pre-constructed ARIMR model, and outputting a prediction result corresponding to the seasonal fluctuation subsequence through the ARIMR model.

104, respectively inputting the influence factor sequence corresponding to the long-term trend subsequence as an input variable and the influence factor sequence corresponding to the random fluctuation subsequence as an input variable into a pre-constructed Bi-LSTM model, and respectively outputting a prediction result corresponding to the long-term trend subsequence and a prediction result corresponding to the random fluctuation subsequence through the Bi-LSTM model;

in implementation, the ARIMR model and the Bi-LSTM model are both existing models, and can be directly constructed and used through software (such as Eviews software), so the specific working principles of the two models are not described again here; it should be noted, however, that the present application sets the three parameter values (d, p, q) of the ARIMR model to (1, 4), and the following parameters of the Bi-LSTM model: the number of layers of a final model hidden layer, the number of nodes of each layer, an activation function, a learning rate, an iteration period and a batch size are correspondingly set to be (3, 96, tanh function, 0.005, 250 and 128), and other parameters are set according to default values, wherein the designated and set parameter values are values obtained through multiple experiments and training and used for improving the prediction accuracy of the prediction model.

In implementation, the short-term load prediction platform takes the influence factor sequence corresponding to the long-term trend subsequence as an input variable of the ARIMR model, outputs the prediction result of the long-term trend subsequence through the ARIMR model, takes the influence factor sequence corresponding to the random fluctuation subsequence as an input variable of the Bi-LSTM model, and outputs the prediction result of the random fluctuation subsequence through the Bi-LSTM model.

And 105, overlapping the prediction results of the subsequences to generate a final prediction result, and displaying the final prediction result.

In practice, the prediction result is in the form of a sequence, for example, the three prediction result sequences are: w is a group of ¹ (t)＝(W ¹ 1，W ¹ 2，W ¹ 3，，，，W ¹ Z)

W ² (t)＝(W ² 1，W ² 2，W ² 3，，，，W ² Z)

W ³ (t)＝(W ³ 1，W ³ 2，W ³ 3，，，，W ³ Z)

Wherein Z is the sequence length of the prediction result, namely the number of the predicted load values; the superposition manner of the prediction results of all the subsequences may be according to the method described above, that is, the load data prediction values at the same position in all the prediction result sequences are summed to form the load data prediction value in the final prediction result sequence, where, for example, the final prediction result sequence is:

W ^{general assembly} (t)＝(W ^{General assembly} 1，W ^{General assembly} 2，W ^{General (1)} 3，，，，W ^{General (1)} Z), wherein W ^{General (1)} Z＝W ¹ Z+W ² Z+W ³ Z。

In summary, the short-term load prediction method based on causal analysis disclosed by the application is mainly based on the OVMD, the original load sequence is subjected to optimal modal decomposition, and the final decomposition results are respectively named as a seasonal wave sub-sequence, a long-term trend sub-sequence and a random wave sub-sequence; in order to extract influence factors with high matching degree with each subsequence, a Granger causal relationship analysis method is adopted to serve the influence factors related to the matching of each subsequence as the input quantity of a prediction model; and finally, according to the result after causal relationship matching, combining the actual characteristics of the subsequences, considering the stable fluctuation characteristics of the seasonal base load components and the number of relevant influence factors, and selecting an autoregressive moving average model (ARIMR) for prediction. For the long-term trend subsequence and the random fluctuation subsequence, a Bi-LSTM model is selected for prediction, and finally the prediction results are superposed to obtain the final result.

The embodiment of the application also discloses a short-term load prediction system based on causal relationship analysis. Referring to fig. 6, a short term load prediction system includes.

The original sequence decomposition module 1 is used for acquiring an original load sequence, and calling an OVMD algorithm to decompose the original load sequence into a seasonal wave sub-sequence, a long-term trend sub-sequence and a random fluctuation sub-sequence;

the influence factor matching module 2 is used for calling a Granger causal relationship analysis algorithm to match influence factors for each subsequence, and acquiring an influence factor sequence corresponding to each subsequence;

the subsequence prediction module 3 is used for inputting the influence factor sequence corresponding to the seasonal wave subsequence serving as an input variable into a pre-constructed ARIMR model and outputting a prediction result corresponding to the seasonal wave subsequence through the ARIMR model;

the subsequence prediction module 3 is further configured to input the influence factor sequence corresponding to the long-term trend subsequence as an input variable and the influence factor sequence corresponding to the stochastic fluctuation subsequence as an input variable into a pre-constructed Bi-LSTM model, and output a prediction result corresponding to the long-term trend subsequence and a prediction result corresponding to the stochastic fluctuation subsequence through the Bi-LSTM model;

and the prediction result generation module 4 is used for superposing the prediction results of the subsequences output by the subsequence prediction module to generate a final prediction result and displaying the final prediction result.

Optionally, the original sequence decomposition module 1 is further configured to control a preset acquisition device to acquire a load data value every other preset time, and receive a load data value x (d, t), where x (d, t) is a load data value at the tth time of the day d; combining all received load data values x (d, t) to generate an original load sequence x (t); wherein, x (t) is a data set formed by a plurality of x (d, t) according to the receiving time sequence; the method is further used for determining the missing data quantity S between x (d, t 1) and x (d, t 2) if two adjacent load data values x (d, t 1) and x (d, t 2) exist in the original load sequence x (t), and the condition that the | t1-t2| ≠ preset time length is met, and S = | t1-t2 |/preset time length; also for the calculation starting from v =1

1,t _1+v ) Until v = S; wherein

And

are all preset weight values; all the calculated x (d, t) are arranged in the order of v from small to large _1+v ) Added between x (d, t 1) and x (d, t 2) in the original payload sequence x (t).

Optionally, the original sequence decomposition module 1 is further configured to have the second bit in the original payload sequence x (t)A load data value x (d, t) simultaneously satisfies the following equation: | x (d, t) -x (d, t-1) & gtnon phosphor>α(t)；|x(d,t)-x(d,t+1)|>β (t); then, the first load data value x (d, t) is modified to:

wherein x (d, t), x (d, t-1) and x (d, t + 1) respectively represent load data values at the t-th moment, the t-1-th moment and the t + 1-th moment of the d-th day, and alpha (t) and beta (t) are preset threshold values; and for the presence of a second load data value x (d, t) in the original load sequence x (t) satisfying: | x (d, t) -m (t) & gt>r (t), the second load data value x (d, t) is modified

Optionally, the original sequence decomposition module 1 is further configured to invoke a VMD decomposition method to decompose the original load sequence x (t), and set an initial value of a parameter K included in the VMD decomposition method to 3, where K denotes the number of subsequences decomposed by the VMD; and also for calculating each sub-sequence after decomposition

Pearson correlation coefficient with original load sequence x (t)

Calculating the residual component ε _K Pearson's correlation coefficient R between (t) and original payload sequence x (t) _K Wherein, in the process,

also for use in R _K Satisfy the requirement of

And then, enabling K = K +1, calling a VMD decomposition method to decompose the original load sequence x (t) again based on the modified K value, and calculating

And R _K (ii) a And also for satisfying

And R is _K ≤ψ·R ₃ When the current K value is the optimal K value, determining the current K value as the optimal K value, wherein psi is a preset threshold value; and based on the determined optimal K value, calling an OVMD algorithm to decompose the original load sequence into a seasonal wave sub-sequence, a long-term trend sub-sequence and a random wave sub-sequence.

Optionally, the original sequence decomposition module 1 is further configured to construct a two-dimensional plane coordinate system, randomly set a plurality of α values and a plurality of τ values, and generate a plurality of sets of (α, τ) coordinate information, where the α values and the τ values are self-contained parameters of the OVMD; and the VMD algorithm is invoked to decompose the original load sequence x (t) based on the determined optimal K value and a plurality of preset (alpha, tau) coordinate information to obtain subsequences at decomposition positions corresponding to different combinations (alpha, tau)

Wherein the content of the first and second substances,

when the optimal K value is shown, the decomposed K-th subsequence corresponding to the g-th group (alpha, tau) is g =1,2,3 \8230, K =1,2 \8230; and also for combining the decomposed subsequences corresponding to each group (. Alpha.,. Tau.)

Calculating an adaptation coefficient E for each group (α, τ), wherein,

ε _g (t) is the mostWhen the value of K is optimal, residual components corresponding to the g group (alpha, tau) are obtained, and N is the sequence length of the original load sequence x (t);

the original sequence decomposition module 1 is further configured to determine coordinate information of a combination parameter (α, τ) corresponding to the minimum E value, and iteratively update the position information of each parameter combination (α, τ) by using a PSO algorithm; the method is also used for recalculating the adaptive coefficient E corresponding to each updated (alpha, tau) when the minimum E value is greater than the preset minE and the iterative update times do not reach the preset maximum iterative times, determining the coordinate information of (alpha, tau) corresponding to the minimum E value, and calling a PSO algorithm to iteratively update the position information of each parameter combination (alpha, tau); otherwise, determining the coordinate information of (alpha, tau) corresponding to the current minimum E value, and determining an optimal alpha value and an optimal tau value; and decomposing the original load sequence into a seasonal wave sub-sequence, a long-term trend sub-sequence and a random wave sub-sequence by adopting an OVMD method based on the determined optimal K value, optimal alpha value and optimal tau value.

Optionally, the original sequence decomposition module 1 is further configured to decompose the original load sequence by using an OVMD method based on the determined optimal K value, optimal α value, and optimal τ value to obtain a plurality of optimal subsequences; the method is also used for calculating the approximate entropy of each optimal subsequence when the number of the optimal subsequences is more than 3, combining the two optimal subsequences with the minimum difference value of the approximate entropy to generate a new optimal subsequence until the final number of the optimal subsequences is 3, and determining the 3 optimal subsequences as a seasonal wave subsequence, a long-term trend subsequence and a random fluctuation subsequence; and determining the 3 optimal subsequences as a seasonal wave subsequence, a long-term trend subsequence and a random wave subsequence according to the change rule of each optimal subsequence when the number of the optimal subsequences is 3.

Optionally, the influencing factor matching module 2 is further configured to invoke a Granger causal relationship analysis algorithm to determine each factor C to be evaluated ⁱ (t) and each subsequence U ^h (t) correlation value of, wherein U ^h (t)＝(U ^h 1,U ^h 2,U ^h 3,…,U ^h n)，U ^h (t) represents the h-th load subsequence, and n refers to the number of sample points of the corresponding subsequence; also for the presence of a target C ⁱ (t) and target U ^h (t) so that the object C ⁱ (t) with the target U ^h (t) when the correlation value is higher than the preset correlation value, the target C is set ⁱ (t) addition to target U ^h (t) a predetermined sequence of influencing factors.

The embodiment of the application also discloses a short-term load prediction platform, which comprises a memory and a processor, wherein the memory stores a computer program which can be loaded by the processor and executes the short-term load prediction method based on causal relationship analysis.

The embodiment of the present application further discloses a computer readable storage medium, which stores a computer program that can be loaded by a processor and execute the method for predicting short-term load based on causal relationship analysis as described above, and the computer readable storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk, an optical disk, or other various media capable of storing program codes.

The above examples are only used to illustrate the technical solutions of the present application, and do not limit the scope of the application. It is to be understood that the embodiments described are only some of the embodiments of the present application and not all of them. All other embodiments, which can be derived by a person skilled in the art from these embodiments without inventive step, are intended to be within the scope of the present application.

Claims

1. A short-term load prediction method based on causal relationship analysis is characterized by comprising the following steps:

acquiring an original load sequence, and calling an OVMD algorithm to decompose the original load sequence into a seasonal wave sub-sequence, a long-term trend sub-sequence and a random fluctuation sub-sequence;

2. The method of claim 1, wherein the obtaining of the original load sequence comprises:

if two adjacent load data values x (d, t 1) and x (d, t 2) exist in the original load sequence x (t), and the condition that | t1-t2| ≠ preset time length is met, determining the quantity S of missing data between x (d, t 1) and x (d, t 2), and S = | t1-t2 |/preset time length; starting from v =1, calculate

Until v = S; wherein

And

are all preset weight values;

all the calculations are performed in the order of v from small to largeDerived x (d, t) _1+v ) Added between x (d, t 1) and x (d, t 2) in the original payload sequence x (t).

3. The causal relationship analysis-based short-term load prediction method of claim 2, wherein after combining all received x (d, t) to generate an original load sequence x (t), further comprising:

|x(d，t)-x(d，t-1)|＞α(t)；

|x(d，t)-x(d，t+1)|＞β(0

modifying the first load data value x (d, t) to:

if a second load data value x (d, t) exists in the original load sequence x (t), satisfying: | x (d, t) -m (t) & lt>r (t), the second load data value x (d, t) is modified to

4. The causal analysis-based short-term load prediction method of claim 1, wherein said invoking the OVMD algorithm decomposes an original load sequence into a seasonal wave subsequence, a long-term trend subsequence, and a stochastic wave subsequence, comprising:

calculating each subsequence after decomposition

Pearson's correlation coefficient with original load sequence x (t)

refers to the J-th subsequence with the number of subsequences being K,

if R is _K Satisfy the requirement of

Then let K = K +1, based on the modified K value, call VMD decomposition method to decompose the original load sequence x (t) again, and calculate

And R _K ；

If it satisfies

And R is _K ≤ψ·R ₃ Determining the current K value as the optimal K value, wherein psi is a preset threshold value; and based on the determined optimal K value, calling an OVMD algorithm to decompose the original load sequence into a seasonal wave sub-sequence, a long-term trend sub-sequence and a random wave sub-sequence.

5. The causal relationship analysis-based short-term load prediction method of claim 4, wherein said invoking an OVMD algorithm to decompose an original load sequence into a seasonal wave subsequence, a long-term trend subsequence, and a stochastic wave subsequence based on the determined optimal K value comprises:

Wherein the content of the first and second substances,

combining the decomposed subsequences corresponding to each group (alpha, tau)

Calculating an adaptation coefficient E for each group (α, τ), wherein,

determining coordinate information of combination parameters (alpha, tau) corresponding to the minimum E value, and performing iterative updating on position information of each parameter combination (alpha, tau) by using a PSO algorithm;

if the minimum E value is larger than the preset minE and the iterative updating times do not reach the preset maximum iterative times, recalculating the adaptive coefficient E corresponding to each (alpha, tau) after updating, determining the coordinate information of (alpha, tau) corresponding to the minimum E value, and calling a PSO algorithm to iteratively update the position information of each parameter combination (alpha, tau);

if the minimum E value is less than or equal to a preset minE, or the iteration updating times reach a preset maximum iteration time, determining the coordinate information of (alpha, tau) corresponding to the current minimum E value, and determining an optimal alpha value and an optimal tau value;

6. The causal relationship analysis-based short-term load prediction method of claim 5, wherein said decomposing the original load sequence into a seasonal wave subsequence, a long-term trend subsequence, and a stochastic wave subsequence with the OVMD method based on the determined optimal K value, optimal alpha value, and optimal tau value comprises:

7. The causal relationship analysis-based short-term load prediction method of claim 1, wherein said invoking Granger causal relationship analysis algorithm matches influence factors for each subsequence, and obtaining a sequence of influence factors corresponding to each subsequence comprises:

calling a Granger causal relationship analysis algorithm to determine each factor C to be evaluated ⁱ (t) and each subsequence U ^h (t) correlation value of, wherein U ^h (t)＝(U ^h 1,U ^h 2,U ^h 3,…,U ^h n)，U ^h (t) represents the h-th load subsequence, n refers to the number of sample points of the corresponding subsequence;

8. A short term load prediction system based on causal relationship analysis, the short term load prediction system comprising: the original sequence decomposition module (1) is used for acquiring an original load sequence, and calling an OVMD algorithm to decompose the original load sequence into a seasonal wave sub-sequence, a long-term trend sub-sequence and a random fluctuation sub-sequence;

the influence factor matching module (2) is used for calling a Granger causal relationship analysis algorithm to match influence factors for each subsequence, and acquiring an influence factor sequence corresponding to each subsequence;

the subsequence prediction module (3) is used for inputting the influence factor sequence corresponding to the seasonal wave subsequence as an input variable into a pre-constructed ARIMR model, and outputting a prediction result corresponding to the seasonal wave subsequence through the ARIMR model;

the subsequence prediction module (3) is also used for inputting the influence factor sequence corresponding to the long-term trend subsequence as an input variable and the influence factor sequence corresponding to the random fluctuation subsequence as an input variable into a pre-constructed Bi-LSTM model, and respectively outputting the prediction result corresponding to the long-term trend subsequence and the prediction result corresponding to the random fluctuation subsequence through the Bi-LSTM model;

and the prediction result generation module (4) is used for superposing the prediction results of the subsequences output by the subsequence prediction module (3) to generate a final prediction result and displaying the final prediction result.

9. A short term load prediction platform comprising a memory and a processor, the memory having stored thereon a computer program which is loadable by the processor and adapted to perform the method of any of claims 1 to 7.

10. A computer-readable storage medium, in which a computer program is stored which can be loaded by a processor and which executes the method of any one of claims 1 to 7.