Short-term day-ahead power load demand load prediction method based on TPE-DART algorithm

Abstract

The invention discloses a short-term day-ahead power load demand load prediction method based on a TPE-DART algorithm, and belongs to the field of power system load prediction. The problem that an existing short-term load prediction method is low in prediction accuracy is solved. According to the method, a load prediction model is constructed by using a DART algorithm, optimization training is performed on the load prediction model by using a load-meteorological characteristic data set, and optimization constraint is performed on a hyper-parameter combination of the load prediction model by taking an evaluation index as a constraint condition in the optimization training process, so that an optimized load prediction model is obtained; and in the optimization training process, a TPE algorithm is adopted, and an optimal hyper-parameter combination is determined according to a five-fold cross check mode. And the optimal hyper-parameter combination mode is determined, and the influence of climate on load prediction is considered in the optimization training process, so that the prediction accuracy is improved. The method is mainly used for power load prediction.

Description

Technical Field

[0001] The invention belongs to the field of power system load forecasting. Background Art

[0002] Short-term load forecasting is a classic and important requirement of the power system to ensure the balance of electricity supply and demand, guarantee power supply security, and optimize power dispatch. It plays a key role in ensuring the safety and economic operation of regional power grids and improving the utilization efficiency of new energy. Accurate load forecasting is of great significance to the safe, stable and economic operation of the power system.

[0003] The existing short-term load forecasting methods include support vector machine (SVM), long short-term memory network (LSTM) and extreme gradient boosting tree (XGBoost), etc. However, they do not consider the impact of climate on load forecasting, resulting in poor prediction accuracy. The above problems need to be solved urgently. Summary of the invention

[0004] The purpose of the present invention is to solve the problem of low prediction accuracy of existing short-term load forecasting methods; the present invention provides a short-term day-ahead power load demand forecasting method based on the TPE-DART algorithm.

[0005] A short-term day-ahead power load demand forecasting method based on the TPE-DART algorithm comprises the following steps:

[0006] Obtaining a load-meteorological characteristic data set and preprocessing it, wherein the input data of each sample in the load-meteorological characteristic data set is the historical power load data and historical meteorological data of consecutive dates, and the output data is the actual power load data corresponding to the predicted date;

[0007] Set the initial hyperparameter combination of the DART algorithm and use the DART algorithm to build a load forecasting model;

[0008] The load-meteorological characteristic data set is used to optimize the load forecasting model. During the optimization training, the evaluation index is used as a constraint to optimize the hyperparameter combination of the load forecasting model to obtain the optimized load forecasting model. The TPE algorithm is used in the optimization training process, and the optimal hyperparameter combination is determined based on the five-fold cross-validation method.

[0009] The optimized load forecasting model is used to forecast the load on the date to be predicted.

[0010] Preferably, the data types included in the historical meteorological data include temperature data, air pressure data, precipitation data, radiation data and wind-related data.

[0011] Preferably, the temperature data include the temperature 2 m above the ground, the dew point 2 m above the ground and the perceived temperature; the air pressure data include the atmospheric pressure at sea level and the surface air pressure; the precipitation data include the rainfall and snowfall; the radiation data include the solar shortwave radiation, the horizontal solar direct radiation, the global oblique irradiance and the ground radiation; the wind-related data include the wind speed 10 m above the ground and the gust 10 m above the ground.

[0012] Preferably, the evaluation indicators include mean absolute error MAE and mean absolute percentage error MAPE;

[0013] The formula for calculating the mean absolute error is:

[0014] The mean absolute percentage error is calculated as:

[0015] Among them, y i and ihat They represent the actual value and predicted value of the i-th power load respectively, and n represents the total number of samples of power load.

[0016] Preferably, the load-meteorological characteristic data set is obtained by:

[0017] The load characteristic dataset and meteorological characteristic dataset of the forecast target area in the same historical time domain are obtained. After filtering the load characteristic dataset according to the importance of the characteristic value, the filtered load characteristic dataset is merged with the corresponding meteorological data in the meteorological characteristic dataset at the same time point according to the principle of the same time point to form a load-meteorological characteristic dataset.

[0018] Preferably, the importance of the characteristic value is to determine the data corresponding to the sampling points to be retained in the sample by statistically analyzing the importance ratio of the loads of each day of each week and the importance ratio of the loads of each week of each month in the historical time domain.

[0019] Preferably, the preprocessing of the load-meteorological characteristic data set is implemented as follows:

[0020] The load-meteorological characteristic data set is windowed to obtain multiple samples, and each sample is normalized to complete the preprocessing of the load-meteorological characteristic data set.

[0021] A short-term day-ahead power load demand load forecasting device based on the TPE-DART algorithm comprises a storage device, a processor, and a computer program stored in the storage device and executable on the processor, wherein the processor executes the computer program to implement a short-term day-ahead power load demand load forecasting method based on the TPE-DART algorithm as claimed in any one of claims 1 to 7.

[0022] A computer-readable storage device stores a computer program, which, when executed, implements the short-term day-ahead power load demand forecasting method based on the TPE-DART algorithm.

[0023] A computer program product includes a computer program, which, when executed by a processor, implements the short-term day-ahead power load demand forecasting method based on the TPE-DART algorithm as described above.

[0024] The beneficial effects brought by the present invention are:

[0025] The present invention uses the DART algorithm to construct a load forecasting model. The DART algorithm introduces a Dart integration strategy based on discarding and skipping operations on the basis of the extreme gradient boosting tree. The TPE algorithm is a Bayesian-based optimization algorithm. The present invention uses the TPE algorithm combined with a five-fold cross-validation to determine the optimal hyperparameter combination to optimize the load forecasting model, thereby improving the prediction performance of the model. In addition, the impact of climate on load forecasting is taken into account during the optimization training process, further improving the prediction accuracy.

[0026] The experimental results show that the load forecasting model of the present invention, namely, the TPE-DART model, is significantly superior to the traditional original ensemble learning and Prophet model in terms of mean absolute error (MAE) and 1-MAPE. Figure 7 and Figure 8 The performance improvements ranged from 9.6% to 48.6% and from 0.64% to 6.06% respectively.

[0027] The present invention performs particularly well in forecasting tasks during seasonal changes and special holidays, demonstrating strong robustness and generalization capabilities, providing an efficient and reliable solution for short-term power load forecasting, and improving forecast accuracy and model applicability. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a flow chart of the short-term day-ahead power load demand forecasting method based on the TPE-DART algorithm of the present invention;

[0029] Figure 2 It is a line chart of power load from 2022 to 2023 for prefecture-level cities in the buffer zone between Xiaoxing'anling and Sanjiang Plain;

[0030] Figure 3 yes Figure 2 Autocorrelation function (ACF) diagram of annual power load data;

[0031] Figure 4 yes Figure 2The partial autocorrelation function (PACF) diagram of annual power load data;

[0032] Figure 5 is a five-fold cross-validation evaluation diagram provided according to an embodiment of the present invention;

[0033] Figure 6 It is the load importance ratio diagram in different time periods; Figure 7 is the MAE value of annual load forecast of various types of forecast models; the horizontal axis is the name of each model, and the vertical axis is the MAE value;

[0034] Figure 8 It is a MAPE line chart of daily load forecast of various types of forecast models; the horizontal axis is the forecast days, the vertical axis is the model name, the color area is the value of 1-MAPE, the green area indicates excellent, the blue area indicates good, and the red area indicates poor. DETAILED DESCRIPTION

[0035] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0036] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0037] Specific implementation method 1. Figure 1 The present embodiment describes a method for predicting short-term day-ahead power load demand based on the TPE-DART algorithm, the method comprising the following steps:

[0038] S1. Obtain a load-meteorological characteristic data set and preprocess it, wherein the input data of each sample in the load-meteorological characteristic data set is the historical power load data and historical meteorological data of consecutive dates, and the output data is the actual power load data corresponding to the predicted date;

[0039] S2, setting the initial hyperparameter combination of the DART algorithm and building a load forecasting model using the DART algorithm;

[0040] The load-meteorological characteristic data set is used to optimize the load forecasting model. During the optimization training, the evaluation index is used as a constraint to optimize the hyperparameter combination of the load forecasting model to obtain the optimized load forecasting model. The TPE algorithm is used in the optimization training process, and the optimal hyperparameter combination is determined based on the five-fold cross-validation method.

[0041] The TPE algorithm and the five-fold cross-validation are both existing technologies. In specific applications, when determining the optimal hyperparameter combination, after the number of training rounds, the number of initial verifications, the number of candidates, and the prior weights of the TPE algorithm are determined, a priori search is performed within the search range of the hyperparameter combination. In this process, the optimal parameters are determined based on the historical search results of the five-fold cross-validation.

[0042] The five-fold cross validation mainly evaluates data selection, and its basic steps are as follows:

[0043] Data splitting: The entire dataset is randomly divided into 5 subsets (folds) of equal size.

[0044] Training and validation:

[0045] (1) For each fold (from 1 to 5), use that fold as a validation set and the other 4 folds as training sets.

[0046] (2) Train the model and evaluate it on the validation set, recording the evaluation metrics (such as accuracy, F1-score, etc.).

[0047] Repeat the process: Repeat the above steps 5 times, each time choosing a different fold as the validation set.

[0048] Results summary: Finally, the five evaluation results are averaged to obtain the final evaluation index of the model. The advantage of this method is that it can make full use of the data set, reduce the variance of model evaluation, and obtain more stable performance evaluation results.

[0049] The DART algorithm, namely Dropouts meet Multiple AdditiveRegression Trees, was proposed by KVRashmi. It is an improvement strategy for the potential overfitting phenomenon in the ensemble learning strategy using tree models as meta-learners. The DART algorithm improves the robustness of the model in prediction by introducing randomness and dropout mechanisms, and ultimately improves the generalization ability of the model.

[0050] In the DART algorithm, four new indicators, sample type, normalization type, rate drop, and skip drop, are introduced as hyperparameter combinations.

[0051] The sample sampling class is a parameter that controls the sample sampling method in the DART algorithm, including uniform sampling and weighted sampling. The sampling method corresponding to this parameter will affect the data distribution during the training process of each meta-learner in ensemble learning, that is, different sampling methods will directly affect the generalization ability and overfitting risk of the model.

[0052] Normalization types include tree normalization (tree) and forest normalization (forest). The choice of different normalization types will affect the stability of the fitting process and the performance of the final model.

[0053] The drop rate and skip rate control the ratio of nodes in each tree to be dropped and skipped, respectively, to introduce more randomness into the fitting process, reduce the risk of overfitting, simplify the model, and complete an operation similar to neural network pruning. By dropping and skipping, the training speed can be increased at the node level, but inappropriate values ​​may affect the robustness and generalization ability of the model.

[0054] The DART algorithm is a specific implementation of the extreme gradient boosting tree algorithm. S3. Use the optimized load forecasting model to predict the load on the date to be predicted. Specifically, the historical load data of the historical continuous dates are used as the input of the optimized load forecasting model to predict the load on the date to be predicted.

[0055] In this implementation, the TPE algorithm is combined with a five-fold cross-validation to determine the optimal hyperparameter combination, thereby optimizing the load forecasting model and improving the model's forecasting performance. The impact of climate on load forecasting is also taken into account during the optimization training process, further improving the forecast accuracy.

[0056] Furthermore, the data types included in the historical meteorological data include temperature data, air pressure data, precipitation data, radiation data and wind-related data; specifically,

[0057] Temperature data include the temperature 2m above the ground, the dew point 2m above the ground and the perceived temperature;

[0058] Air pressure data include sea level atmospheric pressure and surface air pressure; precipitation data include rainfall and snowfall;

[0059] Radiation data include solar shortwave radiation, horizontal solar direct radiation, global oblique irradiance and ground radiation;

[0060] Wind-related data include wind speed at 10 m above the ground and wind gust at 10 m above the ground.

[0061] The specific implementation method of obtaining historical meteorological data can be found in Table 1;

[0062] Table 1 Composition and acquisition methods of historical meteorological data

[0063]

[0064]

[0065] Furthermore, the evaluation indicators of the load forecasting model include mean absolute error MAE and mean absolute percentage error MAPE;

[0066] The formula for calculating the mean absolute error is:

[0067] The mean absolute percentage error is calculated as:

[0068] Among them, y i and ihat They represent the actual value and predicted value of the i-th power load respectively, and n represents the total number of samples of power load.

[0069] Furthermore, the load-meteorological characteristic data set is obtained as follows:

[0070] Obtain the load characteristic data set and meteorological characteristic data set of the prediction target area in the same historical time domain. After filtering the load characteristic data set according to the eigenvalue importance, merge the filtered load characteristic data set with the corresponding meteorological data in the meteorological characteristic data set at the same time point according to the principle of the same time point to form a load-meteorological characteristic data set. Specifically, the eigenvalue importance is to determine the data corresponding to the sampling points to be retained in the sample by counting the load importance ratio of each day of each week in the historical time domain and the load importance ratio of each week of each month. Feature Importance Value (FIV) selection is a method for evaluating the contribution of features to model prediction. It is often used in tree models (such as random forests, gradient boosting trees, etc.) to help understand which features are most important to the final prediction results. Figure 6 The load importance ratio diagram for different time periods is given. Figure 6 It can be seen that Figure A represents the FIV ratio of the 2nd and 3rd days in a week, and the FIV on the 2nd day is significantly higher than that on the 3rd day; Figure B represents the FIV ratio of each week from 1 to 4 weeks, and the data of the 1st week occupies a dominant position in the figure, reflecting its importance; Figure C reflects the FIV ratio of each day in the 1st week in Figure B, among which the data of the 2nd and 3rd days occupy a dominant position; in summary, the data of the 1st week is particularly important, and the 2nd and 3rd days have an absolute dominant advantage.

[0071] Furthermore, the implementation method of preprocessing the load-meteorological characteristic data set is as follows:

[0072] The load-meteorological characteristic data set is windowed to obtain multiple samples, and each sample is normalized to complete the preprocessing of the load-meteorological characteristic data set. Specifically, the autocorrelation function (ACF) and partial autocorrelation function (PACF) are introduced to determine the size of the sliding window. ACF estimates the autocorrelation coefficient by calculating the correlation between the two time periods before and after the sequence. For the autocorrelation coefficient of lag k, the following formula is used to calculate:

[0073]

[0074] In the formula, γ k is the autocorrelation coefficient of lag k, γ0 is the autocorrelation coefficient of lag 0, that is, the variance of the time series, that is, the variance of the time series, and the calculation formula is as follows:

[0075]

[0076] Where l(t) is a power load sequence containing N elements, with t as the time index, and l m is the average value of the power load series. t is the power load data value at time t, and μ is the data mean under the calculated data sequence.

[0077] The partial autocorrelation function (PACF) is a measure of correlation that excludes other lag terms. Its calculation method uses the Yule-Walker equation to solve as follows:

[0078]

[0079] In the formula, ρ k =ACF(l,k), which represents the ACF value of k lagged terms, and the calculation method is shown above. The PACF value of lagged k is the value in the YW matrix above.

[0080] Collected Figure 2 The power load data line chart shown in the figure collects the power load in 2022 and 2023. Figure 2 It can be seen that the data in this region shows obvious annual periodicity.

[0081] Depend on Figure 3 and Figure 4It can be seen that the ACF results show that the data has a strong periodic feature. Each point repeats at intervals of 24 hours, which is consistent with the periodicity of human activities. In addition, this periodicity gradually weakens over time. Further through the partial autocorrelation function, see Figure 4 It can be observed that there is a strong correlation between the 0-48 period and a certain correlation pattern between the 48-72 period.

[0082] In summary, the size of the sliding window is determined according to the autocorrelation function (ACF) and the partial autocorrelation function (PACF). That is, according to the ACF, it can be determined that the width of the sliding window should be a multiple of 24, that is, the natural day is used as the basic window period, which is consistent with the objective rhythm of human life. Further, through the PACF, it is determined that the effective periodic cycle rhythm is 3 days, see Figure 3 and Figure 4 .

[0083] Specific implementation method 2: A short-term day-ahead power load demand load forecasting device based on the TPE-DART algorithm comprises a storage device, a processor, and a computer program stored in the storage device and executable on the processor, wherein the processor executes the computer program to implement a short-term day-ahead power load demand load forecasting method based on the TPE-DART algorithm as described in any one of claims 1 to 7.

[0084] Specific implementation method three: a computer-readable storage device, wherein the storage device stores a computer program, and when the computer program is executed, the short-term day-ahead power load demand forecasting method based on the TPE-DART algorithm is implemented.

[0085] Specific implementation method 4. A computer program product includes a computer program, which, when executed by a processor, implements the short-term day-ahead power load demand forecasting method based on the TPE-DART algorithm as described above.

[0086] Verification test:

[0087] The technical effect of the present invention is demonstrated by the following verification experiments.

[0088] The present invention collects Figure 2 The power load data line chart shown in the figure collects the power load in 2022 and 2023. Figure 2 It can be seen that the data in this region shows obvious annual periodicity.

[0089] Depend on Figure 3 , 4 It can be seen that after determining the width of the sliding window to 3, Figure 4We further analyzed the correlation in the time series data and found that the PACF value dropped to 0.1 at about 100 periods, indicating poor periodicity, so we would not use a period that was too long.

[0090] Depend on Figure 5 It can be seen that after a width of 3 days, using the prediction model error obtained in the cross-validation mentioned above, the model error at this time reaches the lowest point, which fully demonstrates that the prediction accuracy has been significantly improved within this time length; and at a width of 6 days, the error reaches the highest point, showing an overfitting or instability phenomenon. As the cycle increases, the error gradually decreases and tends to stabilize. Although the prediction error values ​​of 24-day widths and 27-day widths are close to those of 3-day widths, they are still higher than those of 3-day widths. Taking into account the characteristics of load forecasting, it is usually necessary to reserve one day's data for future forecasts, and the data of the first day cannot be obtained. In fact, only two days of data are available. Therefore, in actual application, the window width is set to 2 days.

[0091] Depend on Figure 7 It can be seen that only the MAE value of the Prophet model is close to the three ensemble learning models (original ensemble learning, DART and EPE-DART), but the performance is still not as good as the ensemble learning model. Compared with other models, ensemble learning has obvious advantages. The conclusion that the MAE value is relatively low shows that ensemble learning has higher accuracy in load forecasting for the region, and the prediction model of the present invention, that is, the EPE-DART model has the lowest MAE value, which is better than the original ensemble learning and DART.

[0092] Figure 8 The colors in the figure refer to the 1-MAPE values, where green is excellent, blue is good, and red is poor. Most models performed poorly before 150 days, while the three integrated learning models still performed well. After 200 days, since the temperature entered a stable state, the prediction effects of the general models were relatively accurate, with a wide range of excellent predictions. The main advantage of the TPE-DART model built in this article is reflected after 100 days, and it has a wider range of good predictions and excellent prediction ratios. The TPE-DART model is the load forecasting model of the present invention.

[0093] Although the present invention is described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the present invention. It should therefore be understood that many modifications may be made to the exemplary embodiments and that other arrangements may be devised without departing from the spirit and scope of the present invention as defined by the appended claims. It should be understood that the various dependent claims and features described herein may be combined in a manner different from that described in the original claims. It should also be understood that features described in conjunction with individual embodiments may be used in other described embodiments.

Claims

1. A short-term day-ahead power load demand forecasting method based on the TPE-DART algorithm is characterized by: The method comprises the following steps: Obtaining a load-meteorological characteristic data set and preprocessing it, wherein the input data of each sample in the load-meteorological characteristic data set is the historical power load data and historical meteorological data of consecutive dates, and the output data is the actual power load data corresponding to the predicted date; Set the initial hyperparameter combination of the DART algorithm and use the DART algorithm to build a load forecasting model; The load-meteorological characteristic data set is used to optimize the load forecasting model. During the optimization training, the evaluation index is used as a constraint to optimize the hyperparameter combination of the load forecasting model to obtain the optimized load forecasting model. The TPE algorithm is used in the optimization training process, and the optimal hyperparameter combination is determined based on the five-fold cross-validation method. The optimized load forecasting model is used to forecast the load on the date to be predicted.

2. The short-term day-ahead power load demand forecasting method based on the TPE-DART algorithm according to claim 1 is characterized in that: The data types included in historical meteorological data include temperature data, air pressure data, precipitation data, radiation data and wind-related data.

3. The short-term day-ahead power load demand forecasting method based on the TPE-DART algorithm according to claim 2 is characterized in that: Temperature data include temperature 2m above the ground, dew point 2m above the ground and perceived temperature; air pressure data include sea level atmospheric pressure and surface air pressure; precipitation data include rainfall and snowfall; radiation data include solar shortwave radiation, horizontal solar direct radiation, global oblique irradiance and ground radiation; wind-related data include wind speed 10m above the ground and gusts 10m above the ground.

4. The short-term day-ahead power load demand forecasting method based on the TPE-DART algorithm according to claim 1 is characterized in that: Evaluation indicators include mean absolute error MAE and mean absolute percentage error MAPE; The formula for calculating the mean absolute error is: The mean absolute percentage error is calculated as: Among them, y i and ihat They represent the actual value and predicted value of the i-th power load respectively, and n represents the total number of samples of power load.

5. The short-term day-ahead power load demand forecasting method based on the TPE-DART algorithm according to claim 1 is characterized in that: The implementation method of obtaining the load-meteorological characteristic data set is as follows: The load characteristic dataset and meteorological characteristic dataset of the forecast target area in the same historical time domain are obtained. After filtering the load characteristic dataset according to the importance of the characteristic value, the filtered load characteristic dataset is merged with the corresponding meteorological data in the meteorological characteristic dataset at the same time point according to the principle of the same time point to form a load-meteorological characteristic dataset.

6. The short-term day-ahead power load demand forecasting method based on the TPE-DART algorithm according to claim 5 is characterized in that: The importance of the eigenvalue is to determine the data corresponding to the sampling points to be retained in the sample by statistically analyzing the importance ratio of the load of each day of each week and the importance ratio of the load of each week of each month in the historical time domain.

7. The short-term day-ahead power load demand forecasting method based on the TPE-DART algorithm according to claim 1 is characterized in that: The implementation method of preprocessing the load-meteorological characteristic data set is as follows: The load-meteorological characteristic data set is windowed to obtain multiple samples, and each sample is normalized to complete the preprocessing of the load-meteorological characteristic data set.

8. A short-term day-ahead power load demand forecasting device based on a TPE-DART algorithm, comprising a storage device, a processor, and a computer program stored in the storage device and executable on the processor, characterized in that: The processor executes a computer program to implement a short-term day-ahead power load demand forecasting method based on the TPE-DART algorithm as described in any one of claims 1 to 7.

9. A computer-readable storage device storing a computer program, characterized in that: When the computer program is executed, the short-term day-ahead power load demand forecasting method based on the TPE-DART algorithm as described in any one of claims 1 to 7 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the short-term day-ahead power load demand forecasting method based on the TPE-DART algorithm as described in any one of claims 1 to 7 is implemented.

Images