CN112435077A - Sales forecasting method - Google Patents

Sales forecasting method Download PDF

Info

Publication number
CN112435077A
CN112435077A CN202011463674.7A CN202011463674A CN112435077A CN 112435077 A CN112435077 A CN 112435077A CN 202011463674 A CN202011463674 A CN 202011463674A CN 112435077 A CN112435077 A CN 112435077A
Authority
CN
China
Prior art keywords
algorithm
model
strategy
sales
search
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202011463674.7A
Other languages
Chinese (zh)
Inventor
赵冬
唐潮
唐永瑞
肖欣庭
池明辉
刘洋
孙永强
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Sichuan Changhong Electric Co Ltd
Original Assignee
Sichuan Changhong Electric Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Sichuan Changhong Electric Co Ltd filed Critical Sichuan Changhong Electric Co Ltd
Priority to CN202011463674.7A priority Critical patent/CN112435077A/en
Publication of CN112435077A publication Critical patent/CN112435077A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/02—Marketing; Price estimation or determination; Fundraising
    • G06Q30/0201—Market modelling; Market analysis; Collecting market data
    • G06Q30/0202—Market predictions or forecasting for commercial activities
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/02—Marketing; Price estimation or determination; Fundraising
    • G06Q30/0201—Market modelling; Market analysis; Collecting market data

Landscapes

  • Business, Economics & Management (AREA)
  • Strategic Management (AREA)
  • Engineering & Computer Science (AREA)
  • Accounting & Taxation (AREA)
  • Development Economics (AREA)
  • Finance (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Game Theory and Decision Science (AREA)
  • Data Mining & Analysis (AREA)
  • Economics (AREA)
  • Marketing (AREA)
  • Physics & Mathematics (AREA)
  • General Business, Economics & Management (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention relates to the field of sales prediction, in particular to a sales prediction method, which realizes automatic selection of corresponding algorithm strategy modeling for sales prediction and greatly improves the flexibility of algorithm modeling and the efficiency of algorithm modeling. The technical scheme is summarized as that data are collected and the data format is unified, and abstract modeling is carried out according to the data with the unified format; establishing an algorithm library for abstract modeling, and generating a corresponding parameter search range for an algorithm in the algorithm library; selecting a corresponding search strategy according to the search range; and selecting a corresponding model according to the corresponding search strategy to perform model fusion, and performing sales prediction according to the fusion model. The invention is applicable to sales forecasting.

Description

Sales forecasting method
Technical Field
The invention relates to the field of sales prediction, in particular to a sales prediction method.
Background
Generally, a sales prediction modeling process is to process raw data (sales data collected by a service) by a professional data processing engineer, model the data by a professional algorithm worker, and finally perform sales prediction by using an established model, aiming at a specific service problem. In the process, a professional algorithm person needs to investigate a prediction algorithm suitable for a specific service in the process of modeling data, specifically, a time series method and a machine learning method (including a deep learning method) are required, the input data formats corresponding to the time series method and the machine learning method are not necessarily identical, the time series method and the machine learning method which are currently available are various, the algorithm person consumes time in the process of manually selecting, the workload is high, and the algorithm modeling is extremely inflexible.
Disclosure of Invention
The invention aims to provide a sales forecasting method, which realizes automatic selection of corresponding algorithm strategy modeling for sales forecasting, and greatly improves the flexibility of algorithm modeling and the efficiency of algorithm modeling.
The invention adopts the following technical scheme to realize the purpose, and the sales prediction method comprises the following steps:
step (1), collecting data, unifying data formats, and performing abstract modeling according to the data with unified formats;
step (2), establishing an algorithm library for abstract modeling, and generating a corresponding parameter search range for an algorithm in the algorithm library;
step (3), selecting a corresponding search strategy according to the search range;
and (4) selecting a corresponding model according to the corresponding search strategy to perform model fusion, and performing sales prediction according to the fusion model.
Further, in step (1), the acquiring data includes: and collecting sales business data, national macro economic index data and competitive brand data.
Further, in step (2), the concrete method for establishing an algorithm library for abstract modeling includes: and adding a time series algorithm, a machine learning algorithm and a deep learning algorithm into the algorithm library.
Further, the time-series class algorithm includes: ARIMA algorithm, ARMA algorithm, ES algorithm, GARCH algorithm, HOLT-WINTERS algorithm, SARIMA algorithm, STL algorithm, TBATS algorithm, and PROPHET algorithm;
the machine learning class algorithm comprises: a CART algorithm, a CatBOost algorithm, a GBDT algorithm, a LightGBM algorithm, a KNN algorithm, a Linear regression algorithm, a RandomForest algorithm, an SVR algorithm and an XgBoost algorithm;
the deep learning algorithm comprises the following steps: NBEATS algorithm, TCN algorithm, WAVENET algorithm, and Seq2Seq algorithm.
Further, in step (2), the specific method for generating the corresponding parameter search range for the algorithm in the algorithm library includes: and generating parameter searching ranges for each algorithm through the permutation and combination of the parameters of the parameter configuration center algorithm.
Further, the specific method for generating the parameter search range for each algorithm through the permutation and combination of the parameter configuration center algorithm parameters includes: and (3) specifying algorithm parameters corresponding to the algorithm, setting a search value range for the algorithm parameters, and finally arranging and combining the parameters to generate a parameter search space of the algorithm.
Further, in step (3), the search strategy includes: a random search strategy and a grid search strategy;
the random search strategy comprises: randomly selecting a corresponding plurality of parameter combinations from the generated algorithm parameter search space for modeling, and selecting a topk model according to a loss function;
the grid search strategy comprises: and sequentially calculating each parameter combination in the generated algorithm parameter search space, and selecting a topk model according to the loss function.
Further, the method for obtaining the topk model comprises the following steps: generating corresponding algorithms according to the arranged and combined parameters, randomly selecting the algorithms from the algorithms, inputting data with a uniform format, and training each selected algorithm by taking RMSE or MSE as a loss function; then, a model of k in which RMSE or MSE is the smallest is selected to constitute a topk model as the topk model.
Further, in step (4), the method for model fusion includes: and carrying out model fusion according to an algorithm model integration strategy, wherein the algorithm model integration strategy comprises a top1 strategy, a random strategy, a simple average weighted fusion strategy and a loss-dependent weighted fusion strategy.
Further, the top1 strategy includes: selecting a model with the minimum RMSE or MSE from the topk models to perform actual prediction;
the random strategy comprises the following steps: randomly selecting a model from the topk models for actual prediction;
the simple average weighted fusion strategy comprises the following steps: carrying out average weighting on the prediction results of the models, and then predicting;
the loss-dependent weighted fusion strategy comprises the following steps: and performing weighted fusion according to the weight of each model, and then performing prediction.
The method unifies the format of the collected data, avoids errors caused by inconsistent formats of subsequent input data, establishes an algorithm library, contains all needed algorithms in the algorithm library, does not need manual selection, generates a corresponding parameter search range for the algorithms in the algorithm library, selects a corresponding search strategy according to the search range, selects a corresponding model according to the corresponding search strategy for model fusion, and performs sales prediction according to the fusion model, realizes automatic selection of corresponding algorithm strategy modeling for sales prediction, and greatly improves the flexibility of algorithm modeling and the efficiency of algorithm modeling.
Drawings
FIG. 1 is a method flow diagram of a sales prediction method of the present invention.
Detailed Description
The sales forecasting method of the invention, the flow chart of which is shown in figure 1, comprises the following steps:
step 101, collecting data, unifying data formats, and performing abstract modeling according to the data with unified formats;
102, establishing an algorithm library for abstract modeling, and generating a corresponding parameter search range for an algorithm in the algorithm library;
103, selecting a corresponding search strategy according to the search range;
and 104, selecting a corresponding model according to the corresponding search strategy to perform model fusion, and performing sales prediction according to the fusion model.
In step 101, the acquiring data comprises: and collecting sales business data, national macro economic index data and competitive brand data.
In step 102, the concrete method for creating an algorithm library for abstract modeling includes: and adding a time series algorithm, a machine learning algorithm and a deep learning algorithm into the algorithm library.
The time series algorithm comprises the following steps: ARIMA algorithm, ARMA algorithm, ES algorithm, GARCH algorithm, HOLT-WINTERS algorithm, SARIMA algorithm, STL algorithm, TBATS algorithm, and PROPHET algorithm;
the machine learning algorithm comprises the following steps: a CART algorithm, a CatBOost algorithm, a GBDT algorithm, a LightGBM algorithm, a KNN algorithm, a Linear regression algorithm, a RandomForest algorithm, an SVR algorithm and an XgBoost algorithm;
the deep learning algorithm comprises the following steps: NBEATS algorithm, TCN algorithm, WAVENET algorithm, and Seq2Seq algorithm.
In step 102, the specific method for generating the corresponding parameter search range for the algorithm in the algorithm library includes: and generating parameter searching ranges for each algorithm through the permutation and combination of the parameters of the parameter configuration center algorithm.
The specific method for generating the parameter search range for each algorithm through the permutation and combination of the algorithm parameters of the parameter configuration center comprises the following steps: and (3) specifying algorithm parameters corresponding to the algorithm, setting a search value range for the algorithm parameters, and finally arranging and combining the parameters to generate a parameter search space of the algorithm.
In step 103, the search policy includes: a random search strategy and a grid search strategy;
the random search strategy comprises: randomly selecting a corresponding plurality of parameter combinations from the generated algorithm parameter search space for modeling, and selecting a topk model according to a loss function; the loss function is an RMSE or MSE function;
the grid search strategy comprises: and sequentially calculating each parameter combination in the generated algorithm parameter search space, and selecting a topk model according to the loss function.
the method for obtaining the topk model comprises the following steps: generating corresponding algorithms [ Alg1, Alg2, … and Algv ] according to the arranged and combined parameters, wherein Alg1, Alg2, … and Algv respectively represent different algorithms, then randomly selecting the algorithms from the algorithms, inputting data in a uniform format, and training each selected algorithm by taking RMSE or MSE as a loss function; then, a model of k with the minimum RMSE or MSE is selected to form a topk model, and the topk model is AlgTopk, which is then AlgTopk ═ Alg1, Alg2, …, Algv.
In step 104, the method for model fusion includes: and carrying out model fusion according to an algorithm model integration strategy, wherein the algorithm model integration strategy comprises a top1 strategy, a random strategy, a simple average weighted fusion strategy and a loss-dependent weighted fusion strategy.
Wherein the top1 strategy comprises: selecting the best model from the topk models for actual prediction; selecting a model with the minimum RMSE or MSE from AlgTopk to actually predict y; y is the prediction output;
the random strategy comprises the following steps: randomly selecting a model from the topk models to carry out actual prediction; namely randomly selecting a model from AlgTopk to actually predict y;
the simple average weighted fusion strategy comprises: the prediction results of the models are weighted averagely and then predicted, namely, y is (1/k) × (Algk1(X) + Algk2(X) + … + algkk (X));
the loss-dependent weighted fusion strategy comprises the following steps: performing weighted fusion according to the weight of each model, and enabling L1 to be RMSE (Algk1) to represent the loss of the Algk1 algorithm; l2 ═ RMSE (Algk2) indicating the loss of Algk2 algorithm; lk — rmse (Algkk), which represents the loss of Algkk algorithm; the actual prediction result is: y ═ 1.0/L1 × (Algk1(X) + (1.0/L2) × (Algk 2(X) + … + (1.0/Lk) × (X), and then prediction was performed.
In specific implementation, data services are collected, and data can be organized into the following format of table 1:
table 1 sales service direct correlation data
Figure BDA0002832390660000041
In the above table, product ID is product ID (pid1 indicates product ID is pid1, pidi indicates product ID is pidi, pidN indicates product ID is pidN, N indicates the number of products to be predicted in total is N), SalesTime is time period of sale (sti indicates time of sale with product ID being pidi, where i can range from 1 to N), salenumm is number of sale (sni indicates the number of sale of product ID is pidi at sti), salemoney is amount of sale (smii indicates the amount of sale of product ID is pidi at sti), salvprovision is province of sale, salecity is city of sale, productfe is product feature (productfeatfet 1 is 1P of product, productfeatfet 2 is 2 nd feature of product, ProductFeatureP is the second feature of product, and so on.
When national macro economic index data is collected, the data can be organized into the following format of table 2:
TABLE 2 national macroeconomic index data
Figure BDA0002832390660000042
Figure BDA0002832390660000051
In Table 2, date represents the time period and index represents the national macro economic indicator (where index1 represents the 1 st indicator, index2 represents the 2 nd indicator, and so on, and index J represents the J-th indicator; e.g., idxjk represents the value of index J in dtk this time period).
When competitive brand data is collected, the data may be organized into the following table 3 format:
TABLE 3 Competition data
date brand1 brand2 … brandM SalesProvinceC SalesCityC
dt1 brd11 brd21 … brdM1 spc1 scc1
dt2 brd12 brd22 … brdM2 spc2 scc2
… … … … … … …
dtk brd1k brd2k … brdMk spck scck
… … … … … … …
dtK brd1K brd2K … brdMK spcK sccK
In table 3, brand name of competitive brand (auction) (brdM indicates brand name brdM), date indicates time zone, salesprovidec indicates sales province, and salesictyc indicates sales city. For example, brdMK indicates brandBrand M, and during the dtK time period, sccK in spcK province is the sales of the city.
When other data related to the service is collected, the data may be organized into the following table 4 format:
table 4 other data related to the service
Figure BDA0002832390660000052
In table 4, date represents a time period, rindex represents a service-related index (rindex1 represents the 1 st related index, rindex2 represents the 2 nd related index, rindexQ represents the qth related index, for example, ridxqk represents the value of rindexQ in the dtk time period, and so on).
Finally, the collected data are processed and normalized into the following uniform format; the above 4 tables are combined (the combination conditions of tables 1 and 2, and 4 are that the time period of table 1 is equal to the time period fields of tables 2 and 4, and the combination conditions of tables 1 and 3 are that the time period of table 1, the province of sale, and the city of sale are equal to table 2).
Abstracting the data with the uniform format into a mathematical expression: y is [ y1, y2],
x ═ t, X11, X12, …, X1r, X21, X22, …, X2s, X31, X32, …, X3m, X41, X42, …, X4u ], y denotes the amount to be predicted; y1 denotes sales number; y2 denotes sales; x represents an available characteristic factor; t represents the time dimension; x11, x12, … and x1r represent characteristic factors directly related to the service in table 1; x21, x22, … and x2s represent the national macro economic indicators in table 2; x31, x32, … and x3m represent the competitive feature factors in table 3; x41, x42, …, x4u represent other service-related characterizing factors in table 4;
sales forecasting tasks, abstract modeling, are: and y is f (X), the input is X, the prediction output is y, and f is a specific algorithm model.
Generating parameter search ranges for each algorithm through permutation and combination of algorithm parameters of a parameter configuration center, wherein the RIMA algorithm parameters and the search ranges are set as follows in specific implementation:
Figure BDA0002832390660000061
Figure BDA0002832390660000062
the ARMA algorithm parameters and search range are set as follows:
Figure BDA0002832390660000063
Figure BDA0002832390660000064
the ES algorithm parameters and search range are set as follows:
Figure BDA0002832390660000065
Figure BDA0002832390660000066
the GARCH algorithm parameters and search ranges are set as follows:
Figure BDA0002832390660000067
Figure BDA0002832390660000068
the HOLT algorithm parameters and search range are set as follows:
Figure BDA0002832390660000069
Figure BDA00028323906600000610
the HOLT-WINTERS algorithm parameters and the search range are set as follows:
Figure BDA0002832390660000071
Figure BDA0002832390660000072
the SARIMA algorithm parameters and search range are set as follows:
Figure BDA0002832390660000073
Figure BDA0002832390660000074
the STL algorithm parameters and search range are set as follows:
Figure BDA0002832390660000075
Figure BDA0002832390660000076
the TBATS algorithm parameters and search range are set as follows:
Figure BDA0002832390660000077
Figure BDA0002832390660000078
the PROPHET algorithm parameters and the search range are set as follows:
Figure BDA0002832390660000079
Figure BDA00028323906600000710
Figure BDA0002832390660000081
the CART algorithm parameters and the search range are set as follows:
Figure BDA0002832390660000082
Figure BDA0002832390660000083
the parameters of the Catboost algorithm and the search range are set as follows:
Figure BDA0002832390660000084
Figure BDA0002832390660000085
the GBDT algorithm parameters and the search range are set as follows:
Figure BDA0002832390660000086
Figure BDA0002832390660000087
the LightGBM algorithm parameters and search range are set as follows:
Figure BDA0002832390660000091
Figure BDA0002832390660000092
the KNN algorithm parameters and the search range are set as follows:
Figure BDA0002832390660000093
Figure BDA0002832390660000094
the parameters of the linear regression algorithm and the search range are set as follows:
Figure BDA0002832390660000095
Figure BDA0002832390660000096
the RandomForest algorithm parameters and search range are set as follows:
Figure BDA0002832390660000097
Figure BDA0002832390660000098
the SVR algorithm parameters and the search range are set as follows:
Figure BDA0002832390660000099
Figure BDA00028323906600000910
the XgBoost algorithm parameters and search ranges are set as follows:
Figure BDA00028323906600000911
Figure BDA00028323906600000912
Figure BDA0002832390660000101
the NBEATS algorithm parameters and search range are set as follows:
Figure BDA0002832390660000102
Figure BDA0002832390660000103
the TCN algorithm parameters and search ranges are set as follows:
Figure BDA0002832390660000104
Figure BDA0002832390660000105
the WAVENET algorithm parameters and search ranges are set as follows:
Figure BDA0002832390660000106
Figure BDA0002832390660000107
the parameters and search ranges of the Seq2Seq algorithm are set as follows:
Figure BDA0002832390660000108
Figure BDA0002832390660000109
Figure BDA0002832390660000111
taking the ARIMA algorithm as an example, the parameter search range of the ARIMA algorithm is as follows:
Figure BDA0002832390660000112
the full-permutation combination is as follows:
Figure BDA0002832390660000113
the rest of the algorithms are analogized in turn to obtain SS ═ ARIMA _ P1, ARIMA _ P2, …, ARIMA _ P1296, ARMA _ P1, …, …, Seq2Seq _7, Seq2Seq _ 8.
When a strategy is specifically selected, generating corresponding algorithms [ Alg1, Alg2, …, Algv, … and Algv ] according to the generated algorithm parameter combinations, randomly selecting partial algorithms from the algorithms, and training each selected algorithm by using input data in a uniform format and taking RMSE or MSE as a loss function;
assuming that the randomly selected partial algorithm is [ ARIMA _ P1, ARMA _ P2, TCN _ P1, CART _ P3, GBDT _ P1], the corresponding RMSE is trained to be [161.1,121.3,120.2,159.0,100.0 ];
selecting a model of k with the minimum RMSE or MSE to form a topk model; the topk model was written as: AlgTopk ═ Algk1, Algk2, …, Algkk; assuming that k is 3, i.e. the top3 model is selected, AlgTop3 ═ GBDT _ P1, TCN _ P1, ARMA _ P2;
model fusion is carried out according to the topk model selected by the selection strategy and the algorithm model integration strategy, assuming that a grid search strategy is selected, the selected top3 model is as follows: AlgTop3 ═ GBDT _ P1, TCN _ P1, ARMA _ P2], then the corresponding RMSE is [100.0,120.2,121.3 ];
when the prediction is carried out through the fusion model, if the prediction result of each model is weighted averagely according to a simple average weighting fusion strategy, the prediction is carried out; that is, y ═ 1/k ═ Algk1(X) + Algk2(X) + … + algkk (X)), that is, y ═ 1/3 ═ GBDT _ P1(X) + TCN _ P1(X) + … + ARMA _ P2 (X));
if the weighted fusion strategy is based on the loss, the weighted fusion is carried out according to the weight of each model; and then predicting, namely:
y＝(1.0/L1)*Algk1(X)+(1.0/L2)*Algk2(X)+…+(1.0/Lk)*Algkk(X)
at this time: l1 ═ 100.0, L2 ═ 120.2, L3 ═ 121.3, Algk1 ═ GBDT _ P1, Algk2 ═ TCN _ P1, Algk3 ═ ARMA _ P2.
Therefore:
y＝(1.0/100.0)*GBDT_P1(X)+(1.0/120.2)*TCN_P1(X)+(1.0/121.3)*ARMA_P2(X)。
in conclusion, the method and the system greatly improve the flexibility of algorithm modeling, realize automatic selection of corresponding algorithm strategy modeling for sales prediction, and improve the efficiency of algorithm modeling.

Claims (10)

1. A sales prediction method, comprising:
step (1), collecting data, unifying data formats, and performing abstract modeling according to the data with unified formats;
step (2), establishing an algorithm library for abstract modeling, and generating a corresponding parameter search range for an algorithm in the algorithm library;
step (3), selecting a corresponding search strategy according to the search range;
and (4) selecting a corresponding model according to the corresponding search strategy to perform model fusion, and performing sales prediction according to the fusion model.
2. The sales prediction method of claim 1, wherein, in step (1), the collecting data comprises: and collecting sales business data, national macro economic index data and competitive brand data.
3. The sales forecasting method of claim 1, wherein in the step (2), the concrete method for establishing the algorithm library for the abstract modeling comprises: and adding a time series algorithm, a machine learning algorithm and a deep learning algorithm into the algorithm library.
4. The sales prediction method of claim 3, wherein the time series class algorithm comprises: ARIMA algorithm, ARMA algorithm, ES algorithm, GARCH algorithm, HOLT-WINTERS algorithm, SARIMA algorithm, STL algorithm, TBATS algorithm, and PROPHET algorithm;
the machine learning class algorithm comprises: a CART algorithm, a CatBOost algorithm, a GBDT algorithm, a LightGBM algorithm, a KNN algorithm, a Linear regression algorithm, a RandomForest algorithm, an SVR algorithm and an XgBoost algorithm;
the deep learning algorithm comprises the following steps: NBEATS algorithm, TCN algorithm, WAVENET algorithm, and Seq2Seq algorithm.
5. The sales forecasting method according to any one of claims 1 to 4, wherein in the step (2), the specific method for generating the corresponding parameter search range for the algorithm in the algorithm library includes: and generating parameter searching ranges for each algorithm through the permutation and combination of the parameters of the parameter configuration center algorithm.
6. The sales forecasting method according to claim 5, wherein the specific method of generating the parameter search range for each algorithm by the permutation and combination of the parameter configuration center algorithm parameters includes: and (3) specifying algorithm parameters corresponding to the algorithm, setting a search value range for the algorithm parameters, and finally arranging and combining the parameters to generate a parameter search space of the algorithm.
7. The sales prediction method of claim 6, wherein in step (3), the search strategy comprises: a random search strategy and a grid search strategy;
the random search strategy comprises: randomly selecting a corresponding plurality of parameter combinations from the generated algorithm parameter search space for modeling, and selecting a topk model according to a loss function;
the grid search strategy comprises: and sequentially calculating each parameter combination in the generated algorithm parameter search space, and selecting a topk model according to the loss function.
8. The sales prediction method according to claim 7, wherein the topk model acquisition method includes: generating corresponding algorithms according to the arranged and combined parameters, randomly selecting the algorithms from the algorithms, inputting data with a uniform format, and training each selected algorithm by taking RMSE or MSE as a loss function; then, a model of k in which RMSE or MSE is the smallest is selected to constitute a topk model as the topk model.
9. The sales prediction method according to claim 7 or 8, wherein in the step (4), the method of performing model fusion includes: and carrying out model fusion according to an algorithm model integration strategy, wherein the algorithm model integration strategy comprises a top1 strategy, a random strategy, a simple average weighted fusion strategy and a loss-dependent weighted fusion strategy.
10. The sales forecasting method of claim 9, wherein the top1 policy comprises: selecting a model with the minimum RMSE or MSE from the topk models to perform actual prediction;
the random strategy comprises the following steps: randomly selecting a model from the topk models for actual prediction;
the simple average weighted fusion strategy comprises the following steps: carrying out average weighting on the prediction results of the models, and then predicting;
the loss-dependent weighted fusion strategy comprises the following steps: and performing weighted fusion according to the weight of each model, and then performing prediction.
CN202011463674.7A 2020-12-11 2020-12-11 Sales forecasting method Pending CN112435077A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011463674.7A CN112435077A (en) 2020-12-11 2020-12-11 Sales forecasting method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011463674.7A CN112435077A (en) 2020-12-11 2020-12-11 Sales forecasting method

Publications (1)

Publication Number Publication Date
CN112435077A true CN112435077A (en) 2021-03-02

Family

ID=74690986

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011463674.7A Pending CN112435077A (en) 2020-12-11 2020-12-11 Sales forecasting method

Country Status (1)

Country Link
CN (1) CN112435077A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113837487A (en) * 2021-10-13 2021-12-24 国网湖南省电力有限公司 Power system load prediction method based on combined model
CN114299633A (en) * 2021-12-28 2022-04-08 中国电信股份有限公司 Automobile driving accident prediction method and device, electronic equipment and storage medium

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113837487A (en) * 2021-10-13 2021-12-24 国网湖南省电力有限公司 Power system load prediction method based on combined model
CN114299633A (en) * 2021-12-28 2022-04-08 中国电信股份有限公司 Automobile driving accident prediction method and device, electronic equipment and storage medium

Similar Documents

Publication Publication Date Title
Qu et al. Research and application of ensemble forecasting based on a novel multi-objective optimization algorithm for wind-speed forecasting
Li et al. Combining probability density forecasts for power electrical loads
Wang et al. A new ARMAX model based on evolutionary algorithm and particle swarm optimization for short-term load forecasting
CN107067341B (en) RBFNN power distribution automation system state operation evaluation method based on multi-level entropy weight
CN110070228B (en) BP neural network wind speed prediction method for neuron branch evolution
CN112435077A (en) Sales forecasting method
CN113344192B (en) Enterprise-level motor system energy-saving optimization automatic control method and system
Draginčić et al. AHP-based group decision making approach to supplier selection of irrigation equipment
Zhang et al. A collaborative service group-based fuzzy QoS-aware manufacturing service composition using an extended flower pollination algorithm
CN109754122A (en) A kind of Numerical Predicting Method of the BP neural network based on random forest feature extraction
Tavares et al. Comparison of PV power generation forecasting in a residential building using ANN and DNN
CN114091776A (en) K-means-based multi-branch AGCNN short-term power load prediction method
CN117634678A (en) Low-carbon park carbon emission prediction method based on actual operation scene
Siriweera et al. Constraint-driven dynamic workflow for automation of big data analytics based on graphplan
CN111539558A (en) Power load prediction method adopting optimized extreme learning machine
CN116578858A (en) Air compressor fault prediction and health degree evaluation method and system based on graphic neural network
Zeng et al. You Only Train Once: A highly generalizable reinforcement learning method for dynamic job shop scheduling problem
Chang et al. A fuzzy neural network for the flow time estimation in a semiconductor manufacturing factory
CN116027741A (en) Edge cloud collaborative artificial intelligence framework for complex manufacturing scene
Shi et al. Load forecasting for regional integrated energy system based on two-phase decomposition and mixture prediction model
Kalinowski et al. The concept of ant colony algorithm for scheduling of flexible manufacturing systems
CN107222328A (en) Predict the method and device of saturation flex point
Bharati et al. A performance prediction of worm-worm wheel and its multi objective design parameter optimization based on genetic algorithm
Owda et al. Using artificial neural network techniques for prediction of electric energy consumption
Saffarian et al. Generator maintenance scheduling using discrete firefly algorithm

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication

Application publication date: 20210302

RJ01 Rejection of invention patent application after publication