Dynamic adaptive load forecasting model combination method

Abstract

The application discloses a dynamic self-adaptive load prediction model combination method, and proposes a short-term load prediction framework. The framework can dynamically identify whether prediction result combination is needed based on the prediction accuracy of a plurality of prediction algorithms in the last period, and adjust the weight coefficient of the combined prediction model. The proposed framework fully utilizes the advantages of a plurality of single prediction models, and reduces the phenomenon of increasing prediction error due to improper combination. The method can dynamically determine the period according to the expected target, determine whether model combination is needed according to the prediction error in the judgment period of each model, and dynamically adjust the combination weight to obtain the final result. Through the above steps, the short-term and extremely short-term load prediction accuracy can be further improved, the accuracy and robustness of the prediction model in actual application are improved, and the construction of a new power system is promoted.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power systems, and particularly relates to power system load prediction. BACKGROUND

[0002] With global energy transformation and low-carbon development, more emerging distributed energy enters the power grid, bringing great challenges to the balance of power supply and demand. Power load, also known as electricity load, is the total power taken by the power equipment of the power user at a certain moment. Through the analysis of the historical load data of the user, the load situation for the user can be obtained to a certain extent. The power grid company needs accurate power load prediction information to ensure the balance of power supply and demand and improve the economic efficiency of operation. At the same time, a series of energy optimization problems related to the power system also need reliable load prediction as a guarantee, so higher requirements are put forward for improving the accuracy of load prediction.

[0003] Therefore, scholars at home and abroad have proposed a variety of load prediction schemes from different time scales. Popular prediction methods include multiple regression algorithms, artificial neural networks (ANN), support vector machines (SVM), and time series analysis. However, in the actual application of load prediction, a single model often has many shortcomings, such as difficulty in judging load mutations and possible overfitting of complex models. In order to overcome the shortcomings of single prediction algorithms in terms of prediction accuracy, a combination framework is considered to combine the prediction results of multiple single models by weighting. However, since the combination framework combines the prediction results of single models by determining the weight coefficients in multiple ways, it completely depends on the prediction errors of the previous single models. In some cases, the determination of the wrong weight coefficients may increase the prediction error of the combination framework.

[0004] To solve the above problems, a dynamic adjustment combination model is proposed in this paper, which identifies the relationship between the results of each prediction algorithm and the true value to determine whether the results of different algorithms need to be combined, avoiding the problem of large deviation between the prediction results and the true value due to blind combination in some periods, while obtaining higher prediction accuracy. SUMMARY

[0005] The application is used to solve the above technical problems, and the following technical solutions are adopted:

[0006] The application proposes a dynamic self-adaptive load prediction model combination method, which comprises the following steps:

[0007] Step (1): Obtain the load, weather, and other historical data of the user or region to be predicted; perform missing value interpolation completion and outlier cleaning operations on the obtained data, and perform feature engineering to form a data set.

[0008] Step (2): The data set after feature engineering is divided into training set and test set, and according to the expected short-term or ultra-short-term prediction result, the dynamic judgment time period is set to judge whether it is necessary to carry out multi-model combination.

[0009] Step (3): The training set set in (2) is trained by building deep neural network, LSTM, random forest and other models, and the prediction accuracy of each model is improved through appropriate parameter search method.

[0010] Step (4): When actually predicting, the virtual prediction results of different models obtained in step (3) are dynamically judged. If model combination is needed, the weight is obtained according to the error of dynamic judgment period and the final result is obtained by weighting; if combination is not needed, the result of the model with the highest accuracy in the last period is taken as the final result.

[0011] The dynamic judgment process of step (2) is:

[0012] If the expected result is the ultra-short-term load prediction result, the virtual prediction error values of multiple single models with 1 hour prediction time in advance are taken as the judgment basis; if the expected result is the short-term load prediction result, the virtual prediction error values of multiple single models at this time in the past 1 day are taken as the judgment basis.

[0013] In the load prediction combination model, the weight should not be negative or greater than 1. The model prediction results at the dynamic judgment time point are judged. If all single model virtual prediction results are greater than or less than the real value at the dynamic judgment time, the above error situation will occur when combining, so the multi-model combination method is not considered; if all single model virtual prediction results contain greater or less than the real value, select the prediction model result with the minimum error distributed on both sides of the real value, and obtain the final prediction result by weighting at the expected prediction time point.

[0014] Firstly, the deviation matrix E(1) is obtained by the error e1, e2 of the two models with the lowest virtual prediction error:

[0015]

[0016] Next, the Lagrange multiplier method (2) is used to find the best weight ω1, ω2 between them, and the expected weight value should only be 1, that is: R = [1, 1] T .

[0017]

[0018] The final prediction result at time t is obtained by linear combination (3) of the weight and the actual prediction value of the single model ​

[0019]

[0020] The combination method adopts a combination framework composed of a model pre-training module, a dynamic judgment module and a weighted combination module, and systematically solves the combination judgment basis, weight determination and other problems in the combination of multiple models. The dynamic adjustment framework first obtains a single prediction result by calling multiple advanced load prediction models for prediction on the processed data. Then, by comparing the prediction result with the true value, it is judged whether there are prediction results distributed on both sides of the true value to determine whether multiple model combination is needed or the result of the single model with the highest prediction accuracy is selected as the final prediction result.

[0021] For short-term load prediction and very short-term load prediction, different dynamic judgment periods are set to ensure the generality of the overall framework when responding to different prediction demands.

[0022] Compared with the prior art, the technical scheme has the progress that:

[0023] The present application aims at the problem that the accuracy of a single model is difficult to further improve in short-term and very short-term load prediction, and realizes the full use of the advantages of multiple models to obtain better load prediction results through a dynamic self-adaptive load prediction combination framework.

[0024] Meanwhile, the present application considers the problem that the prediction error may increase due to the wrong weight determination in multiple model combination, and establishes different dynamic judgment and weight updating methods for short-term prediction and very short-term prediction respectively, which can adapt to the changes of prediction errors of multiple models and obtain more accurate prediction results.

[0025] The method of the present application avoids the problem that the prediction result deviates greatly from the true value due to blind combination in some period, and obtains higher prediction accuracy, so it has extremely important significance. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 is a dynamic combination framework diagram;

[0027] Figure 2 is a single model training diagram;

[0028] Figure 3 is a method flowchart of the present application;

[0029] Figure 4 is a result curve schematic diagram of region 1;

[0030] Figure 5 is a result curve schematic diagram of region 2;

[0031] Figure 6 is a schematic diagram of the result curve obtained by region 3. DETAILED DESCRIPTION

[0032] The application will be further described below with reference to the drawings. The following examples are only used to more clearly illustrate the technical solutions of the application, and cannot be used to limit the protection scope of the application.

[0033] Example 1: A dynamic adaptive load prediction model combination method, the method comprising the following steps:

[0034] Step (1): Obtain the historical data of the load, weather, etc. of the user or region to be predicted; perform missing value interpolation completion and outlier cleaning operation on the obtained data, and perform feature engineering to form a data set.

[0035] Step (2): Divide the data set after feature engineering into a training set and a test set, and set a dynamic judgment time period to judge whether multi-model combination is needed according to the expected short-term or ultra-short-term prediction result.

[0036] Step (3): Train the models such as deep neural network, LSTM, random forest, etc. on the training set set in (2), and improve the prediction accuracy of each model through appropriate parameter search method.

[0037] Step (4): When actually predicting, the virtual prediction results of different models obtained in step (3) are dynamically judged, if model combination is needed, the weights are obtained according to the error of the dynamic judgment period and the final result is obtained by weighting; if combination is not needed, the result of the model with the highest accuracy in the last period is taken as the final result.

[0038] The dynamic judgment process of step (2) is:

[0039] If the expected result is an ultra-short-term load prediction result, the virtual prediction error values of multiple single models with a prediction time of 1 hour in advance are taken as the judgment basis; if the expected result is a short-term load prediction result, the virtual prediction error values of multiple single models at this time point in the past 1 day are taken as the judgment basis.

[0040] In the load prediction combination model, the weight should not be negative or greater than 1. The model prediction results at the dynamic judgment time point are judged, if all single model virtual prediction results are greater than or less than the real value at the dynamic judgment time, the above error situation will occur when combining, therefore the multi-model combination method is not considered; if all single model virtual prediction results contain both greater than or less than the real value, the prediction model result with the minimum error distributed on both sides of the real value is selected, and the final prediction result is obtained by weighting at the expected prediction time point.

[0041] First, the bias matrix E(1) is obtained by the virtual prediction errors e1, e2 of the two models with the lowest errors:

[0042]

[0043] Next, the optimal weights ω1, ω2 between the two are found using the Lagrange multiplier method (2), and the expected weight value should only be 1, that is: R = [1, 1] T .

[0044]

[0045] The final prediction result at time t is obtained by the linear combination (3) of the weights and the actual prediction values of the individual models

[0046]

[0047] The combination method adopts a combination framework composed of a model pre-training module, a dynamic judgment module, and a weighted combination module, systematically solving the combination judgment basis, weight determination, etc. The dynamic adjustment framework first calls multiple advanced load prediction models to obtain single prediction results by processing the data. Then, by comparing the prediction results with the true values, it is determined whether there are prediction results distributed on both sides of the true value to determine whether multiple model combination is needed or the result of the single model with the highest prediction accuracy is selected as the final prediction result.

[0048] Example 2: For short-term load prediction and very short-term load prediction, different dynamic judgment periods are set to ensure the generalization of the overall framework when responding to different prediction needs. The dynamic adaptive load prediction combination method designed by the present application is as shown in Figure 1 The single model training process is as shown in Figure 2 In combination with Figure 3 The specific embodiments of the dynamic adaptive load prediction combination method of the present application are described in detail as follows:

[0049] The load data and weather data collected from January 1, 2020 to June 30, 2022 in a certain area of Jiangsu Province are used as sample data for prediction, including 15-minute load data of three regions and temperature, pressure, precipitation, and humidity of the region every three hours. The load data is divided into three data sets by region, and the weather data is preprocessed by interpolation to correspond to the load data every 15 minutes. The mean absolute error percentage E MAPE (4) is selected as the evaluation index of the load prediction result.

[0050] ​

[0051] The single model selects three common but effective models for load prediction, including random forest, multilayer perceptron regression and long short-term memory network.

[0052] The random forest algorithm is to integrate multiple decision trees by using the idea of ensemble learning. The decision tree is a classifier, and for an input sample, N decision trees will have N classification results. Random forest collects all classification votes and assigns the class with the most votes as the final output.

[0053] The multilayer perceptron model is a forward structure artificial neural network, which includes an input layer, an output layer and several hidden layers. The output accuracy of the network is improved by updating the weights in the neural network, and the final output of the regression problem is obtained.

[0054] The long short-term memory neural network model is a special recurrent neural network, which mainly solves the problems of gradient disappearance and gradient explosion in the training process of long sequence model. The basic component unit includes three parts of forget gate, input gate and output gate.

[0055] The advantages of the three selected models are shown in Table 1, and the framework allows more advanced prediction models to be added to obtain more accurate prediction accuracy.

[0056] Table 1 Comparison of disadvantages of different prediction models

[0057] In this case, the three single models with the above parameter adjustment are used to train the processed data, and the final output is dynamically adjusted according to the obtained prediction results. In this case, the expected result is the regional date load prediction result, so the judgment of each time point prediction result is determined according to the error of the previous day virtual prediction.

[0058] If the outputs of the three prediction models at the time of the previous day are higher or lower than the true value, the output of the model with the highest accuracy at the time of the previous day is selected as the final output at the prediction time; if the outputs of the three prediction models at the time of the previous day are distributed on both sides of the true value, the two models with the highest accuracy are selected for weighted combination at the prediction time, and the weight coefficient is determined according to the error of the previous day.

[0059] The final result curve of the three regions is as shown in Figures 4-6 The average absolute error percentage results of the single model and the combined model are shown in Table 2:

[0060] Table 2 Comparison of average error percentage results of single model and combined model

[0061] Random Forest Multi-Layer Perceptron Long Short-Term Memory Neural Network Ensemble Model Region 1 Mean Absolute Error Percentage 4.0271143％ 5.0863662％ 5.0440221％ 2.9655145％ Region 2 Mean Absolute Error Percentage 4.1342760％ 7.1255418％ 9.3187109％ 2.2519953％ Region 3 Mean Absolute Error Percentage 3.1101842％ 2.4080004％ 1.9905607％ 1.8521834％

[0062] like Figure 4 As shown, the prediction results of the three single prediction models in region 1 are generally evenly distributed on both sides of the true value. Therefore, the prediction value of the weighted combined model is used as the output value.

[0063] like Figure 5 As shown, in region 2, there are prolonged periods of significantly overestimating predicted values ​​between 7-9 AM and 11-20 AM. If the predicted results were linearly combined at this time, it would lead to a large error. Therefore, the dynamically adjusted combined model selects the predicted result closest to the true value as the actual predicted value, avoiding the problem of increased error caused by using the combined model for prediction.

[0064] like Figure 6 As shown, the predicted values ​​for region 3 fluctuate on one side of the actual values ​​at 3 o'clock, 10-16 o'clock, 18 o'clock, and 21 o'clock. Therefore, the dynamically adjusted combined model selects the model prediction result with the smallest error as the actual predicted value for that period.

[0065] The above analysis shows that when the values ​​of different prediction models are distributed around the true value, the dynamic adaptive combination model selects the prediction result obtained through a weighted linear combination. When the values ​​of all individual prediction algorithms are on one side of the true value, the dynamic adaptive combination model selects the prediction result closest to the true value as the actual prediction value.

[0066] Therefore, the method proposed in this invention achieves an effective combination of multiple single models in load forecasting.

[0067] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0068] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure Oneapparatuses that implement the functions specified in the flowchart or flowcharts and / or blocks. Figure One

[0069] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the flowchart or flowcharts and / or blocks. Figure One Figure One

[0070] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the flowchart or flowcharts and / or blocks. Figure One Figure One

[0071] The above description is merely that of the preferred embodiments of the application and modifications and alterations are possible without departing from the principles of the application as set forth in the claims.​​​​​

Claims

1. A method for combining dynamic adaptive load forecasting models, characterized in that, The method includes the following steps: Step (1): Obtain the load and historical weather data of the user or region to be predicted; perform missing value interpolation and outlier cleaning on the obtained data, and perform feature engineering to form a dataset; Step (2): Divide the dataset after feature engineering into training set and test set, and determine whether to combine multiple models based on the expected short-term (prediction before the day) or very short-term (prediction before a few hours) prediction results. Step (3): On the training set set in (2), build deep neural network, LSTM and random forest models for training, and improve the prediction accuracy of each model by using appropriate parameter search methods; Step (4): When making actual predictions, the virtual prediction results of different models obtained in step (3) are dynamically judged. If model combination is required, the weights are calculated based on the error of the dynamic judgment period and the final result is obtained by weighting. If combination is not required, the result of the model with the highest accuracy in the previous period is taken as the final result. The dynamic judgment process in step (2) is as follows: If the desired load forecast is for a very short-term period, the judgment is based on the virtual forecast error values ​​of multiple single models one hour in advance; if the desired load forecast is for a short-term period, the judgment is based on the virtual forecast error values ​​of multiple single models one day in advance. First, the errors of the two models with the lowest virtual prediction errors are compared. Obtain the deviation matrix (1): （1） Next, the optimal weights between the two are found using the Lagrange multiplier method (2). The expected sum of the weight values ​​should be 1. ; （2） By weights and the actual predicted values ​​of individual models The linear combination (3) yields the final prediction result at time t. : （3）。 2. The dynamic adaptive load forecasting model combination method according to claim 1, characterized in that, The combined framework used in the above combination method consists of a model pre-training module, a dynamic judgment module, and a weighted combination module.

3. The dynamic adaptive load forecasting model combination method according to claim 1, characterized in that, Different dynamic judgment periods are set for short-term load forecasting and very short-term load forecasting to ensure the generalizability of the overall framework in response to different forecasting needs.

4. The dynamic adaptive load forecasting model combination method as described in claim 3, characterized in that: If the method of selecting a model combination is determined to obtain the final result, the combination weight coefficients are dynamically adjusted based on the model error of the previous period.

Images