A model training method and system for an electricity quantity prediction model

Abstract

The present invention discloses a method and system for model training of an electricity consumption prediction model. The method includes: analyzing and obtaining influencing factors associated with the output of a first prediction model, where the influencing factors include a first non-temporal external force influencing factor, a second temporal external force influencing factor, and a third internal cause influencing factor; dividing enterprises into regions based on the first non-temporal external force influencing factor, and separately training the electricity consumption prediction models for enterprises in different regions; dividing enterprises into categories based on the third internal cause influencing factor, and separately training the electricity consumption prediction models for different categories of enterprises within the same region; and training using the second temporal external force influencing factor as input data for the first prediction model. The present invention improves the scenario applicability of the electricity consumption prediction model and the accuracy of the prediction results in the corresponding scenario.

Description

Technical Field

[0001] The present invention relates to the technical field of power grid prediction, and particularly to a method and system for training a power consumption prediction model. Background Art

[0002] In the power system, power consumption prediction is of great significance. For example, through the prediction and analysis of power consumption, power enterprises can more reasonably determine the total quota of power sales, so as to make relevant decisions to guide the reasonable operation of power plants, transmission and distribution grids; accurate power consumption prediction can enable power enterprises to effectively control profit balance.

[0003] Accurate power consumption prediction can provide reliable guidance for the production and scheduling of electric energy. However, the existing power consumption prediction technologies often cannot simultaneously take into account the applicability to different power application scenarios and the accuracy of power consumption prediction. Summary of the Invention

[0004] In view of the problems existing in the above-mentioned prior art, the present invention provides a method and system for training a power consumption prediction model, which effectively improves the scene applicability of the power consumption prediction model and the accuracy of the prediction results in the corresponding scene. The technical solution is as follows:

[0005] In the first aspect, a method for training a power consumption prediction model is provided, and the method includes the following steps:

[0006] For a first prediction model with a first prediction duration, analyze and obtain influencing factors associated with the output of the first prediction model, where the influencing factors include a first non-temporal external influencing factor, a second temporal external influencing factor, and a third internal influencing factor;

[0007] Based on the first non-temporal external influencing factor, divide enterprises into regions, and respectively train the power consumption prediction models of enterprises in different regions;

[0008] Based on the third internal influencing factor, divide enterprises into categories, and respectively train the power consumption prediction models of different categories of enterprises in the same region;

[0009] Use the second temporal external influencing factor as the input data for training the first prediction model;

[0010] For a second prediction model with a second prediction duration, use the second temporal external influencing factor as the input data for training the second prediction model.

[0011] In some embodiments, the first non-temporal external force influencing factor characterizes the external force influencing factors suffered by an enterprise that is associated with the output result of the power consumption prediction model and does not change with time. The second temporal external force influencing factor characterizes the external force influencing factors suffered by an enterprise that is associated with the output result of the power consumption prediction model and changes with time. The third internal factor influencing factor characterizes the influencing factors related to the enterprise's own attributes. The first non-temporal external force influencing factor includes the geographical location area attribute of the enterprise. The second temporal external force influencing factor includes the weather temperature and weather category. The third internal factor influencing factor includes the characteristics of the enterprise's power consumption change curve.

[0012] In some embodiments, the classification of enterprises based on the third internal factor influencing factor includes:

[0013] Obtain the third internal factor influencing factor parameters of the enterprise. The third internal factor influencing factor parameters include the power consumption curve characteristic parameters and the power consumption cycle parameters of the enterprise.

[0014] Perform clustering analysis based on the third internal factor influencing factor parameters of the enterprise, and classify enterprises with similar third internal factor influencing factor parameters into one category.

[0015] In some embodiments, the model training method further includes:

[0016] Obtain the original data for the training process based on the determined model input data.

[0017] Optimize the original data based on a preset optimization strategy to obtain the direct input data for the model training process as the training sample data.

[0018] Perform the model training process based on the training sample data.

[0019] In some embodiments, the step of optimizing the original data based on a preset optimization strategy to obtain the direct input data for the model training process as the training sample data includes:

[0020] Perform distortion data analysis based on the original data.

[0021] For the distorted data, use the fitting curve point value data to fill or replace the distorted item to optimize the original data.

[0022] Amplify the original data to increase the training sample size.

[0023] In some embodiments, the performing distortion data analysis based on the original data includes:

[0024] S621: Based on dataset A in the original data, randomly select a data a in A as the root node of the binary tree. Classify according to the size of other data in dataset A relative to a, and use them as elements in the two branches of the binary tree respectively, where a is between the minimum and maximum values in dataset A;

[0025] S622: For each branch, select a data b in the branch element set B as the main trunk, and classify based on the size of other data in the branch element set B relative to b to form new branches, where b is between the minimum and maximum values in dataset B;

[0026] S623: Repeat step S622 until all data in dataset A exist in the binary tree, and there is only one data in each branch of each binary tree;

[0027] S624: Traverse each data a in the finally formed binary tree i , calculate the number of branch times t required from the root element to the corresponding node of the data i , calculate data a i 's distortion parameter where j(x) = ln(x) + γ, and γ refers to the Euler constant;

[0028] S625: When s(a i ) ≥ 0.6, it can be determined that the node data is distorted; when s(a i [[ID=])) < 0.5, it is determined that the node data is normal data. When 0.5 < s(a i ) < 0.6, it is determined that the node data is an uncertain item.

[0029] In some embodiments, after the step of when 0.5 < s(a i ) < 0.6, it is determined that the node data is an uncertain item, it further includes: further judging the node data;

[0030] Generate a bell curve corresponding to dataset A according to dataset A. The number of data within the confidence interval (μ - 3σ, μ + 3σ) of the bell curve accounts for 0.9100 - 0.9973 of the total number of data in the dataset;

[0031] Based on the bell curve, the data with values within the confidence interval are regarded as reliable data, and the data with values outside the confidence interval are regarded as distorted data.

[0032] In some embodiments, the amplification of the original data to increase the training sample size includes:

[0033] S81: Perform cluster analysis based on the third internal influencing factor parameters of the enterprise, and classify the enterprises with similar third internal influencing factor parameters into one category;

[0034] S82: Generate virtual model input data of the said type by fusing model input data of one type based on the same category of enterprises as the model input data of the said type for the new virtual enterprise.

[0035] S83: Repeat step S82 to generate virtual model input data of each type for the new virtual enterprise.

[0036] In some embodiments, the training based on the second time-series external force influencing factors as the input data of the prediction model includes:

[0037] Analyze and obtain the influence coefficient of the second time-series external force influencing factors on the output of the prediction model;

[0038] Based on the influence coefficient and the second time-series external force influencing factor data as the input data for the training process of the prediction model.

[0039] In a second aspect, a model training system for an electric quantity prediction model is provided, and the system includes:

[0040] A model training associated data analysis unit, which is used to analyze and obtain the influencing factors associated with the output of the first prediction model for the first prediction model with a first prediction duration, and the influencing factors include the first non-time-series external force influencing factors, the second time-series external force influencing factors, and the third internal cause influencing factors;

[0041] A model classification first analysis unit, which is used to divide enterprises into regions based on the first non-time-series external force influencing factors, and respectively train the electricity consumption prediction models of enterprises in different regions;

[0042] A model classification second analysis unit, which is used to divide enterprises into categories based on the third internal cause influencing factors, and respectively train the electricity consumption prediction models of different categories of enterprises within the same region;

[0043] A first model training unit, which is used to train based on the second time-series external force influencing factors as the input data of the first prediction model;

[0044] A second model training unit, which is used to train the second prediction model with a second prediction duration based on the second time-series external force influencing factors as the input data of the second prediction model.

[0045] The model training method and system for an electric quantity prediction model of the present invention have the following beneficial effects:

[0046] For the influencing factors associated with the output of the power consumption prediction model, the present invention fully considers the types and degrees of influence. For different types of influencing factors, it determines whether the power consumption prediction model can commonly use the same prediction model parameters, the applicable scenarios where the same prediction model parameters can be commonly used, and the applicable scenarios where the same prediction model parameters cannot be used. It realizes training the power consumption prediction for different scenarios respectively with appropriate training data and obtaining the applicable power consumption prediction model for the corresponding scenario, improving the scenario applicability of the power consumption prediction model and the accuracy of the prediction results in the scenario. Description of the Drawings

[0047] Figure 1 is a flowchart of a model training method for a power consumption prediction model in an embodiment of the present application;

[0048] Figure 2 is a flowchart of an implementation manner of the training process of the prediction model based on the second time-series external force influencing factor in an embodiment of the present application;

[0049] Figure 3 is a structural diagram of a model training system for a power consumption prediction model in an embodiment of the present application. Detailed Embodiments

[0050] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

[0051] The embodiment of the present application provides a model training method for a power consumption prediction model. The method includes the following steps:

[0052] S1: For the first prediction model with a first prediction duration, analyze and obtain the influencing factors associated with the output of the first prediction model. The influencing factors include the first non-time-series external force influencing factor, the second time-series external force influencing factor, and the third internal cause influencing factor;

[0053] S2: Based on the first non-time-series external force influencing factor, divide enterprises into regions, and respectively train the power consumption prediction models of enterprises in different regions;

[0054] S3: Based on the third internal cause influencing factor, divide enterprises into categories, and respectively train the power consumption prediction models of different categories of enterprises within the same region;

[0055] S4: Use the second time-series external force influencing factor as the input data of the first prediction model for training;

[0056] S5: For the second prediction model with a second prediction duration, it is trained based on the second time-series external force influencing factors as the input data of the second prediction model.

[0057] In the embodiments of the present application, enterprises are classified based on the non-time-series external force influencing factors, i.e., the first non-time-series external force influencing factors, so as to jointly analyze the electricity consumption data of enterprises with common external force factors, fit the influence of the common external force factors on the electricity consumption of enterprises, and apply it to the process of predicting the future electricity consumption of enterprises. Considering that the production rhythms of different industries and the electricity consumption time cycles of different enterprises are different, in the present application, enterprises are further divided based on the internal cause influencing factors of enterprises, i.e., the third internal cause influencing factors. Enterprises with similar third internal cause influencing factor data are classified into one category, and the same prediction model parameters are commonly used for all enterprises of this type. On this basis, in the training process of each model, its model input data is determined based on the data of the second time-series external force influencing factors. Specifically, the model input data includes both the influencing factor parameters related to the electricity consumption of enterprises and the historical data of the electricity consumption of enterprises. Among them, the electricity consumption data of enterprises can have various types, such as electricity consumption values, electricity consumption peaks, average electricity consumption values in a preset time period, and other parameters used to characterize the electricity consumption characteristics of enterprises.

[0058] In the present application, for the influencing factors associated with the output of the electricity consumption prediction model, the influence types and degrees are fully considered. For different influencing factor types, it is determined whether the electricity consumption prediction model can commonly use the same prediction model parameters, the applicable scenarios where the same prediction model parameters can be commonly used, and the applicable scenarios where the same prediction model parameters cannot be used, realizing the training of the electricity consumption prediction in different scenarios using appropriate training data and obtaining the applicable electricity consumption prediction models in the corresponding scenarios, improving the scenario applicability of the electricity consumption prediction model and the accuracy of the prediction results in the corresponding scenarios.

[0059] In addition, in the present application, based on the different prediction durations of the electricity consumption prediction model, it is divided into a first prediction model and a second prediction model. Specifically, the first prediction duration is long-term, and the second prediction duration is short-term. In one implementation, the first prediction duration is 2 - 6 months, and the second prediction duration does not exceed 1 month. In the training of the first prediction model, the first non-time-series external force influencing factors, the second time-series external force influencing factors, and the third internal cause influencing factors are considered for model training. However, in the training of the second prediction model, considering that the influence of the first non-time-series external force influencing factors and the third internal cause influencing factors on the short-term electricity consumption prediction results is limited and can be ignored, the training of the second prediction model in the present application only considers the second time-series external force influencing factor parameters. The present application simultaneously realizes the training of the long-term prediction model and the short-term prediction model for electricity consumption prediction.

[0060] Further, the above-mentioned first non-temporal external force influencing factor characterizes the external force influencing factors suffered by an enterprise that are related to the output result of the electricity quantity prediction model and do not change with time. The second temporal external force influencing factor characterizes the external force influencing factors suffered by an enterprise that are related to the output result of the electricity quantity prediction model and change with time. The third internal factor influencing factor characterizes the influencing factors related to the enterprise's own attributes. The first non-temporal external force influencing factor includes the regional attribute of the geographical location where the enterprise is located. The second temporal external force influencing factor includes weather temperature and weather category. The third internal factor influencing factor includes the characteristics of the enterprise's electricity consumption change curve.

[0061] In the embodiments of the present application, among them, the first non-temporal external force influencing factor considers the influence of regional development characteristics on electricity consumption. According to the characteristics such as the production place and business place of the enterprise, the enterprises are divided into regions, which is convenient for fitting the influence of common external force factors on the electricity consumption of the enterprise and acts on the prediction of the enterprise's future electricity quantity. The second temporal external force influencing factor considers that the weather temperature will affect the efficiency of the production interval and the use of air conditioners or blowers, which will increase the electricity consumption of the enterprise's production. If the weather condition is cloudy or rainy, the lighting situation is not very good, and the enterprise may need to turn on the production lighting in advance during production, which will increase the electricity consumption of the enterprise's production. The third internal factor influencing factor considers the enterprise's own electricity consumption characteristics. Based on the different production rhythms of different industries and the different electricity consumption time cycles of different enterprises, the enterprises are divided. For enterprises with similar third internal factor influencing factor data, the same prediction model parameters are used together.

[0062] Further, the above-mentioned classification of enterprises based on the third internal factor influencing factor includes:

[0063] Obtain the third internal factor influencing factor parameters of the enterprise. The third internal factor influencing factor parameters include the electricity quantity curve characteristic parameters and electricity consumption cycle parameters of the enterprise;

[0064] Perform clustering analysis based on the third internal factor influencing factor parameters of the enterprise, and divide the enterprises with similar third internal factor influencing factor parameters into one category.

[0065] In the embodiments of the present application, this clustering analysis process performs clustering by considering the correlation between the characteristics of different electricity quantities of the enterprise and the electricity consumption cycle. Specifically, during implementation, first, select several known typical enterprise types among all electricity-consuming enterprises to form known electricity consumption characteristic types. Clustering can be performed respectively based on the correlation between the average value of electricity quantity and the time cycle, and the correlation between variance and the time cycle. Compare the clustering results with the results of the known electricity consumption characteristic types, and select the clustering scheme whose classification results are closest to the results of the known electricity consumption characteristic types to perform clustering division on all electricity-consuming enterprises and divide them into the corresponding electricity consumption characteristic types.

[0066] A model training method for an electricity consumption prediction model according to an embodiment of the present application, based on the above steps S1 - S4, further includes:

[0067] S5: Obtain the original data for the training process based on the determined model input data;

[0068] S6: Optimize the original data based on a preset optimization strategy to obtain the direct input data for the model training process as the training sample data;

[0069] S7: Perform the model training process based on the training sample data.

[0070] In an embodiment of the present application, the original data collected is optimized based on a preset optimization strategy. Specifically, the original data includes different types of data under various influencing factors serving as model inputs, such as data of weather temperature categories, weather state categories, and usage feature data. Among them, the data of each category can be further subdivided into multiple sub - types of data. For example, the usage feature data can be parameters such as electricity consumption values, electricity consumption peaks, and average electricity consumption values in a preset time period for characterizing the electricity consumption characteristics of an enterprise.

[0071] A model training method for an electricity consumption prediction model according to an embodiment of the present application, based on the above steps S1 - S7, further includes:

[0072] Perform electricity consumption prediction for a future prediction duration based on the trained electricity consumption prediction model;

[0073] After the prediction duration actually occurs, based on the actual electricity consumption and predicted electricity consumption data during the first prediction duration or the second prediction duration, correct and optimize the model parameters of the trained electricity consumption prediction model to obtain an optimized electricity consumption prediction model.

[0074] In an embodiment of the present application, based on the already trained electricity consumption prediction model, during the actual use process, the electricity consumption prediction model will be continuously corrected and improved as new data occurs, making the electricity consumption prediction model more suitable for electricity consumption prediction in the current time period and improving the dynamic applicability of the electricity consumption prediction model.

[0075] In one implementation manner, in the above step S6, optimizing the original data based on a preset optimization strategy to obtain the direct input data for the model training process as the training sample data includes:

[0076] S61: Perform distortion data analysis based on the original data;

[0077] S62: For the distorted data, use the fitting curve point value data to fill or replace the distorted item to optimize the original data;

[0078] S63: Amplify the original data to increase the amount of training samples.

[0079] In the embodiments of the present application, the optimization process includes the analysis and processing of distorted data and the amplification of sample data. By analyzing and processing the distorted data, the accuracy of the sample data is improved, so as to improve the accuracy of the training model. By amplifying the sample data, overfitting in model training is avoided.

[0080] In one implementation, the above-mentioned analysis of distorted data based on the original data includes the following steps:

[0081] S621: Based on the data set A in the original data, randomly select a data a in A as the root node of the binary tree, classify according to the size of other data in the data set A relative to a, and respectively serve as the elements in the two branches of the binary tree, where a is between the minimum value and the maximum value in the data set A;

[0082] S622: For each branch, select a data b in the branch element set B as the main trunk, and classify based on the size of other data in the branch element set B relative to b to form a new branch, where b is between the minimum value and the maximum value in the data set B;

[0083] S623: Repeat step S622 until all the data in the data set A exist in the binary tree, and there is only one data in each branch of each binary tree;

[0084] S624: Traverse each data a in the finally formed binary tree i , calculate the number of branch times t required from the root element to the corresponding node of the data i , calculate the data a i 's distortion parameter where j(x) = ln(x) + γ, and γ refers to the Euler constant;

[0085] S625: When s(a i ) ≥ 0.6, it can be determined that the node data is distorted; when s(a i ) < 0.5, it is determined that the node data is normal data, and when 0.5 < s(a i ) < 0.6, it is determined that the node data is an uncertain item.

[0086] In the embodiments of the present application, the distortion analysis is respectively performed on different data sets in the original data. This distortion analysis is realized by constructing a binary tree. Specifically, the distortion analysis is performed on the data on each node of the formed binary tree. Based on the number of branch times from the root node of the binary tree to the node data, the distortion parameter s(a i ) of the node data is judged. Further, based on the distortion parameter s(a i) Analyze the size to determine whether the node data is distorted. For the determined distorted data, in this application, the fitting curve point value data is used to fill in the missing values or replace the abnormal data to eliminate the abnormal data. Taking the electricity consumption data as an example, the source data within the statistical period of each enterprise is cut to obtain several stable electricity consumption time periods. Select the electricity consumption time period where the abnormal data exists. Without considering the abnormal data, a data fitting algorithm is used to obtain the fitting curve of this stable electricity consumption time period, and based on the time node corresponding to the abnormal data, the data value corresponding to the fitting curve is used to replace the abnormal data to correct the enterprise electricity source data.

[0087] In the above distortion parameter , D is the number of data in the data set A, and the original data can be divided into multiple data sets A.

[0088] Preferably, before analyzing the distortion of the original data, this embodiment further includes: performing normalization processing on the data. For the data set A = {a1, a2... a i ... a D-1 , a D}, calculate its normalized data set A' = {a'1, a'2... a' i ... a' D-1 , a' D}, where a′ i is the normalized value of the data a i ; a min is the minimum value of the elements in the data set A; a max is the maximum value of the elements in the data set A.

[0089] Based on this normalization processing, the above distortion data analysis process is performed on the data set A' = {a'1, a'2... a' i ... a' D-1 , a' D}.

[0090] Furthermore, in the above process of determining whether the node data is distorted based on the binary tree, it further includes:

[0091] When selecting a piece of data each time, that is, when randomly selecting a piece of data a from dataset A as the root node of the binary tree and selecting a piece of data b from the branch element set B as the main trunk, analyze the selected data, and judge the adaptability parameter value C of the data selection. If the adaptability parameter value C is larger, it indicates that the selected data is more beneficial for the binary tree to quickly detect distortion items. The calculation method of the adaptability parameter value C is as follows: Based on the selected data, mark all the data in the dataset or the branch element set that is greater than the selected data as the first dataset, and mark all the data in the dataset or the branch element set that is less than the selected data as the second dataset. Then, based on the first dataset and the second dataset, obtain the difference in data distribution between the first dataset and the second dataset, and at the same time obtain the aggregation degree of the data distribution in the first dataset and the second dataset. Determine the size of the adaptability parameter value C based on the ratio of the difference and the aggregation degree.

[0092] Further, in the above step S625, when 0.5 < s(a i ) < 0.6 and it is determined that the node data is an uncertain item, it further includes: further judging the node data;

[0093] According to dataset A, generate a bell-shaped curve corresponding to dataset A. The number of data within the confidence interval (μ - 3σ, μ + 3σ) of the bell-shaped curve accounts for between 0.9100 and 0.9973 of the total number of data in the dataset;

[0094] Based on the bell-shaped curve, take the data with values within the confidence interval as reliable data, and take the data with values outside the confidence interval as distorted data.

[0095] In the embodiment of the present application, for data with uncertain distortion, further determine whether the data is distorted data through the generated bell-shaped curve and confidence interval, and combine the above binary tree method to realize the distortion analysis process of all original data.

[0096] In the embodiment of the present application, in the above step S63, amplifying the original data to increase the training sample size includes:

[0097] S81: Perform clustering analysis based on the third internal influencing factor parameters of the enterprise, and divide the enterprises with similar third internal influencing factor parameters into one category;

[0098] S82: Based on the fusion of one type of model input data of the enterprises in the same category, generate virtual model input data of the said type as the model input data of the said type of the new virtual enterprise;

[0099] S83: Repeat step S82 to generate virtual model input data of each type for the new virtual enterprise.

[0100] In the embodiments of the present application, clustering analysis is utilized to generate virtual new data to supplement and enrich the training sample database. Specifically, with regard to the consideration of the third internal influencing factor, in this embodiment, based on the enterprise power consumption change curve analysis, the parameters of the third internal influencing factor are obtained. On this basis, step S81 includes: extracting power consumption change characteristic parameters based on the enterprise power consumption change curve, performing clustering analysis on all power consumption curves based on the extracted characteristic parameters, and classifying enterprises with similar third internal influencing factor parameters into one category through clustering analysis. Specifically, it is possible to perform clustering based on the change curvature of a specific period of the enterprise power consumption change curve to obtain several categories of power-consuming enterprises with similar power consumption change characteristics. Step S82 includes: analyzing parameters such as the distribution rule and change rule of the power consumption data of all enterprises in the same category based on the power consumption data of all enterprises in the same category, and virtual enterprise power consumption data that conforms to the characteristics of the power consumption data of all enterprises in this category can be generated, that is, virtual enterprise power consumption data is generated. Specifically, this generation method can be executed based on a virtual data generation model that has been pre-trained based on the generative adversarial network GAN. The virtual data generation model uses the same type of data of all enterprises in the same category as training sample data, and the sample data annotation performs a supervised training process until the loss function of the model meets the preset conditions, obtaining a trained virtual data generation model with performance indicators meeting the preset conditions, and executing the above virtual data generation process based on this virtual data generation model. Of course, the parameters of the third internal influencing factor of the enterprise not only include the enterprise power consumption change curve, but also other types of data. Correspondingly, when generating virtual data for the new virtual enterprise, it includes generating virtual data for each type of data.

[0101] Further, in the above steps S4 and S5, training based on the second temporal external influencing factor as the input data of the prediction model includes:

[0102] Analyzing and obtaining the influence coefficient of the second temporal external influencing factor on the output of the prediction model;

[0103] Performing the training process of the prediction model based on the influence coefficient and the second temporal external influencing factor data as the input data.

[0104] In the embodiments of the present application, for the training process of the power consumption prediction model, for the first prediction model or the second prediction model, in the input data of the model, the data of the second time-series external force influencing factors should be combined with the influence coefficients of the influencing factors as the input data of the model. The second prediction model performs data fitting on the influence coefficients and the data of the second time-series external force influencing factors. Through the training of a large number of training sample data, the trained second prediction model fully explores the fitting relationship between the data of the second time-series external force influencing factors of each category and the power consumption prediction data, that is, the output data of the model, under the influence coefficients of the second time-series external force influencing factors of each category. By adding the influence coefficients to the input data of the model, the accuracy of the power consumption prediction result of the power consumption prediction model of the present application is improved. Since the above-mentioned second time-series external force influencing factors include different categories of data, correspondingly, different categories of data have corresponding different influence coefficients.

[0105] The embodiments of the present application also provide a model training system for a power consumption prediction model, and the system includes:

[0106] The model training correlation data analysis unit is used to analyze and obtain the influencing factors associated with the output of the first prediction model for the first prediction model with a first prediction duration, and the influencing factors include the first non-time-series external force influencing factors, the second time-series external force influencing factors, and the third internal cause influencing factors;

[0107] The first model classification analysis unit is used to divide enterprises into regions based on the first non-time-series external force influencing factors, and perform model training on the power consumption prediction models of enterprises in different regions respectively;

[0108] The second model classification analysis unit is used to divide enterprises into categories based on the third internal cause influencing factors, and perform model training on the power consumption prediction models of different categories of enterprises in the same region respectively;

[0109] The first model training unit is used to perform training based on the second time-series external force influencing factors as the input data of the first prediction model;

[0110] The second model training unit is used to perform training on the second prediction model with a second prediction duration based on the second time-series external force influencing factors as the input data of the second prediction model.

[0111] For the specific limitations of the model training system for the power prediction model, reference can be made to the limitations of the model training method for the power prediction model in the foregoing text, which will not be elaborated herein. Each unit in the above-mentioned model training system for the power prediction model can be implemented in whole or in part by software, hardware, and their combination. The above-mentioned units can be embedded in the processor of the computer device in hardware form or be independent of it, or can be stored in the memory of the computer device in software form, so as to facilitate the processor to call and execute the operations corresponding to each of the above units.

[0112] The present invention is not limited to the above specific embodiments. Various changes made by those of ordinary skill in the art starting from the above conceptions without creative efforts fall within the protection scope of the present invention.

Claims

1. A method for training a model of an electricity quantity prediction model, characterized in that, Including: For a first prediction model with a first prediction duration, analyze and obtain influencing factors associated with the output of the first prediction model. The influencing factors include a first non-temporal external force influencing factor, a second temporal external force influencing factor, and a third internal factor influencing factor. The first non-temporal external force influencing factor represents the external force influencing factor received by an enterprise that is associated with the output result of the power consumption prediction model and does not change with time. The second temporal external force influencing factor represents the external force influencing factor received by an enterprise that is associated with the output result of the power consumption prediction model and changes with time. The third internal factor influencing factor represents the influencing factor related to the enterprise's own attributes. The first non-temporal external force influencing factor includes the regional attribute of the geographical location where the enterprise is located. The second temporal external force influencing factor includes the weather temperature and weather category. The third internal factor influencing factor includes the characteristics of the enterprise's electricity consumption change curve; Based on the first non-temporal external force influencing factor, conduct regional division of enterprises, and separately train the electricity consumption prediction models for enterprises in different regions; Based on the third internal factor influencing factor, conduct category division of enterprises, and separately train the electricity consumption prediction models for different categories of enterprises within the same region; Based on the second temporal external force influencing factor, use it as the input data for the first prediction model for training; For a second prediction model with a second prediction duration, based on the second temporal external force influencing factor, use it as the input data for the second prediction model for training; The model training method further includes: obtaining the original data for the training process based on the determined model input data; optimizing the original data based on a preset optimization strategy to obtain the direct input data for the model training process as the training sample data; and performing the model training process based on the training sample data; The optimizing the original data based on a preset optimization strategy to obtain the direct input data for the model training process as the training sample data includes: performing distortion data analysis on the original data; for the distorted data, using the fitting curve point value data to fill or replace the distorted item to optimize the original data; and amplifying the original data to increase the training sample size; The distortion data analysis based on the original data includes: S621: Based on the data set A in the original data, randomly select a data a in A as the root node of the binary tree, classify according to the size of other data in the data set A relative to a, and use them as elements in the two branches of the binary tree respectively, where a is between the minimum value and the maximum value in the data set A; S622: For each branch, select a data b in the branch element set B as the main trunk, and classify according to the size of other data in the branch element set B relative to b to form new branches, where b is between the minimum value and the maximum value in the data set B; S623: Repeat step S622 until all the data in the data set A exist in the binary tree, and there is only one data in each branch of each binary tree; S624: Traverse each data a in the finally formed binary tree i , calculate the number of branches t required from the root element to the corresponding node of the data i , calculate the data a i 's distortion parameter where j(x) = ln(x) + γ, γ refers to the Euler constant; D is the number of data in the data set A; S625: When s(a i ) ≥ 0.6, it can be determined that the node data is distorted; when s(a i ) < 0.5, it is determined that the node data is normal data, and when 0.5 < s(a i ) < 0.6, it is determined that the node data is an uncertain item.

2. The model training method of an electricity quantity prediction model according to claim 1, wherein, The conducting category division of enterprises based on the third internal factor influencing factor includes: Obtaining the third internal factor influencing factor parameters of the enterprise. The third internal factor influencing factor parameters include the power consumption curve characteristic parameters and the electricity consumption cycle parameters of the enterprise; Based on the third internal factor influencing factor parameters of the enterprise, conduct cluster analysis, and divide enterprises with similar third internal factor influencing factor parameters into one category.

3. The model training method of an electricity quantity prediction model according to claim 1, characterized in that When 0.5 < s(a i ) < 0.6, after determining that the node data is an uncertain item, it further includes: further judging the node data; According to dataset A, generate a bell-shaped curve corresponding to dataset A. The number of data within the confidence interval (μ - 3σ, μ + 3σ) of the bell-shaped curve accounts for between 0.9100 and 0.9973 of the total amount of data in the dataset; Based on the bell-shaped curve, use the data with values within the confidence interval as reliable data, and use the data with values outside the confidence interval as distorted data.

4. The model training method of an electricity quantity prediction model according to claim 1, wherein The amplifying the original data to increase the training sample size includes: S81: Based on the third internal factor influencing factor parameters of the enterprise, conduct cluster analysis, and divide enterprises with similar third internal factor influencing factor parameters into one category; S82: Generate virtual model input data of the said type by fusing model input data of one type based on the same category of enterprises as the model input data of the said type for the new virtual enterprise; S83: Repeat step S82 to generate virtual model input data of each type for the new virtual enterprise.

5. The model training method of an electricity quantity prediction model according to claim 1, characterized in that The training based on the second time-series external influence factor as the input data of the prediction model includes: Analyze and obtain the influence coefficient of the second time-series external influence factor on the output of the prediction model; Based on the influence coefficient and the second time-series external influence factor data as the input data, conduct the training process of the prediction model.

6. A model training system for an electricity quantity prediction model, characterized in that, It includes: The model training correlation data analysis unit is used to analyze and obtain the influence factors associated with the output of the first prediction model with a first prediction duration for the first prediction model. The influence factors include the first non-time-series external influence factor, the second time-series external influence factor, and the third internal influence factor. The first non-time-series external influence factor represents the external influence factor of the enterprise that is associated with the output result of the electricity consumption prediction model and does not change with time. The second time-series external influence factor represents the external influence factor of the enterprise that is associated with the output result of the electricity consumption prediction model and changes with time. The third internal influence factor represents the influence factor related to the enterprise's own attributes. The first non-time-series external influence factor includes the regional attribute of the enterprise's location. The second time-series external influence factor includes the weather temperature and weather category. The third internal influence factor includes the characteristics of the enterprise's electricity consumption change curve; The first model classification analysis unit is used to divide enterprises into regions based on the first non-time-series external influence factor and conduct model training for the electricity consumption prediction models of enterprises in different regions respectively; The second model classification analysis unit is used to divide enterprises into categories based on the third internal influence factor and conduct model training for the electricity consumption prediction models of different categories of enterprises within the same region respectively; The first model training unit is used to train based on the second time-series external influence factor as the input data of the first prediction model; The second model training unit is used to train the second prediction model with a second prediction duration based on the second time-series external influence factor as the input data of the second prediction model; The model training system also executes the following steps: Obtain the original data for the training process based on the determined model input data; optimize the original data based on a preset optimization strategy to obtain the direct input data for the model training process as the training sample data; Conduct the model training process based on the training sample data; The optimization process of optimizing the original data based on a preset optimization strategy to obtain the direct input data for the model training process as the training sample data includes: conducting distortion data analysis based on the original data; for the distorted data, use the fitting curve point value data to fill or replace the distorted item to optimize the original data; amplify the original data to increase the training sample size; The distortion data analysis based on the original data includes: S621: Based on the dataset A in the original data, randomly select a data a in A as the root node of the binary tree, classify according to the size of other data in the dataset A relative to a, and use them as elements in the two branches of the binary tree respectively, where a is between the minimum value and the maximum value in the dataset A; S622: For each branch, select a data b in the branch element set B as the main trunk, and classify according to the size of other data in the branch element set B relative to b to form a new branch, where b is between the minimum value and the maximum value in the dataset B; S623: Repeat step S622 until all the data in the dataset A exist in the binary tree and there is only one data in each branch of each binary tree; S624: Traverse each data a in the finally formed binary tree i , calculate the number of branches t required from the root element to the corresponding node of the data i , calculate the data a i 's distortion parameter where j(x) = ln(x) + γ, γ refers to the Euler constant; D is the number of data in the dataset A; S625: When s(a i ) ≥ 0.6, it can be determined that the node data is distorted; when s(a i ) < 0.5, it is determined that the node data is normal data, and when 0.5 < s(a i ) < 0.6, it is determined that the node data is an uncertain item.

Images