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

Abstract

The short-term day-ahead electric power load demand load prediction method based on TPE-DART algorithm belongs to the field of electric power system load prediction, and solves the problem of low prediction accuracy of existing short-term load prediction methods. The load prediction model is constructed by using the DART algorithm, the load prediction model is optimized and trained by using the load-weather feature data set, and in the optimization and training process, the evaluation index is used as the constraint condition to optimize and constrain the super parameter combination of the load prediction model, so that the optimized load prediction model is obtained; and in the optimization and training process, the TPE algorithm is adopted, and according to the five-fold cross-validation mode, the optimal super parameter combination is determined. The optimal super parameter combination mode is determined, and the influence of climate on load prediction is considered in the optimization and training process, so that the prediction accuracy is improved. The application is mainly used for electric power load prediction.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of power system load forecasting. BACKGROUND

[0002] Short-term load forecasting is a classic and important demand of power system to ensure power supply and demand balance, guarantee power supply safety and optimize power dispatching. It plays a key role in ensuring the safe 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 power systems.

[0003] The short-term load forecasting methods of the prior art are support vector machine (SVM), long short-term memory network (LSTM) and extreme gradient boosting tree (XGBoost), etc., but they do not consider the influence of climate on load forecasting, resulting in poor prediction accuracy. The above problems need to be solved. SUMMARY

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

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

[0006] Obtain a load-weather feature dataset, and preprocess it. The input data of each sample in the load-weather feature dataset is continuous date historical power load data and historical weather data, and the output data is the actual power load data corresponding to the prediction date;

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

[0008] Optimize and train the load forecasting model using the load-weather feature dataset, and optimize and constrain the hyperparameter combination of the load forecasting model in the optimization and training process using the evaluation index as the constraint condition to obtain the optimized load forecasting model. In the optimization and training process, the TPE algorithm is used, and the optimal hyperparameter combination is determined according to the five-fold cross-validation method;

[0009] Use the optimized load forecasting model to predict the load of the to-be-predicted date.

[0010] Preferably, the historical weather data includes temperature data, barometric pressure data, precipitation data, radiation data and wind-related data.

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

[0012] Preferably, the evaluation metrics include mean absolute error (MAE) and mean absolute percentage error (MAPE).

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

[0014] The formula for calculating the mean absolute percentage error is:

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

[0016] Preferably, the implementation method for obtaining the load-meteorological feature dataset is as follows:

[0017] The load feature dataset and meteorological feature dataset for the target area in the same historical time domain are obtained. After filtering the load feature dataset according to the importance of the feature values, the corresponding meteorological data in the load feature dataset and the meteorological feature dataset of the same time point are merged according to the principle of the same time point to form a load-meteorological feature dataset.

[0018] Preferably, the importance of the eigenvalues ​​is determined by the proportion of daily load importance within a week and the proportion of weekly load importance within a month in the statistical historical time domain, which determines the data corresponding to the sampling points to be retained in the sample.

[0019] Preferably, the preprocessing of the load-meteorological feature dataset is implemented as follows:

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

[0021] A short-term day-ahead power load demand forecasting device based on the TPE-DART algorithm includes a storage device, a processor, and a computer program stored in the storage device and executable on the processor. The processor executes the computer program to implement 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.

[0022] A computer-readable storage device stores a computer program that, 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 that, 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 of this invention are:

[0025] This invention utilizes the DART algorithm to construct a load forecasting model. The DART algorithm, based on extreme gradient boosting trees, introduces a Dart ensemble strategy based on drop-out and skip operations. The TPE algorithm is a Bayesian-based optimization algorithm. This invention uses a combination of the TPE algorithm and five-fold cross-validation to determine the optimal hyperparameter combination, thereby optimizing the load forecasting model and improving its predictive performance. Furthermore, the impact of climate on load forecasting is considered during the optimization training process, further enhancing the prediction accuracy.

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

[0027] This invention excels in forecasting tasks during seasonal variations and special holidays, demonstrating strong robustness and generalization ability. It provides an efficient and reliable solution for short-term power load forecasting, improving forecast accuracy and model applicability. Attached Figure Description

[0028] Figure 1 This is a flowchart of the short-term day-ahead power load demand forecasting method based on the TPE-DART algorithm described in this invention;

[0029] Figure 2 This is a line graph showing the electricity load of prefecture-level cities in the buffer zone between the Lesser Khingan Mountains and the Sanjiang Plain from 2022 to 2023.

[0030] Figure 3 yes Figure 2 The annual electricity load data autocorrelation function (ACF) graph;

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

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

[0033] Figure 6 A graph showing the proportion of load importance in different time periods; Figure 7 These are the annual load forecast MAE values ​​for various forecasting models; where the horizontal axis represents the model name and the vertical axis represents the MAE value.

[0034] Figure 8 This is a daily load forecast MAPE line chart for various forecasting models; the horizontal axis represents the number of forecast days, the vertical axis represents the model name, and the color area represents the value of 1-MAPE, with green indicating excellent, blue indicating good, and red indicating poor. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0037] Detailed Implementation Method 1, see [link / reference] Figure 1 This embodiment describes a short-term day-ahead electricity demand forecasting method based on the TPE-DART algorithm, which includes the following steps:

[0038] S1. Obtain the load-meteorological feature dataset and preprocess it. The input data for each sample in the load-meteorological feature dataset are historical power load data and historical meteorological data for consecutive days, and the output data is the actual power load data corresponding to the predicted date.

[0039] S2. Set the initial hyperparameter combination for the DART algorithm and use the DART algorithm to build a load forecasting model;

[0040] The load forecasting model was optimized and trained using a load-meteorological feature dataset. During the optimization training process, evaluation indicators were used as constraints to optimize the hyperparameter combination of the load forecasting model, resulting in an optimized load forecasting model. Furthermore, the TPE algorithm was used during the optimization training process, and the optimal hyperparameter combination was determined based on the five-fold cross-validation method.

[0041] Both the TPE algorithm and the five-fold cross-validation are existing technologies. In specific applications, when determining the optimal hyperparameter combination, after determining the number of training rounds, initial validation times, number of candidates, and prior weights of the TPE algorithm, a prior search is performed within the search range of the hyperparameter combination. During 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 primarily evaluates the data selection, and its basic steps are as follows:

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

[0044] Training and validation:

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

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

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

[0048] Results Summary: Finally, the results of the five evaluations are averaged to obtain the final evaluation metric for the model. The advantage of this method is that it makes full use of the dataset, reduces the variance of the model evaluation, and yields more stable performance evaluation results.

[0049] The DART algorithm, or Dropouts Meet Multiple Additive Regression Trees, proposed by K.R. V. Ashmi, is an improved strategy to address the potential overfitting in ensemble learning strategies that use tree models as meta-learners. The DART algorithm improves the model's robustness in prediction by introducing randomness and dropout mechanisms, ultimately enhancing the model's generalization ability.

[0050] The DART algorithm introduces four new metrics as hyperparameter combinations: sample type, normalize type, rate drop, and skip drop.

[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 affects the data distribution during the training process of each meta-learner in ensemble learning. That is, different sampling methods will directly affect the model's generalization ability and overfitting risk.

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

[0053] Drop rate and skip rate, by controlling the proportion of nodes dropped and skipped in each tree, introduce more randomness into the fitting process, reducing the risk of overfitting while simplifying the model, similar to neural network pruning. While dropping and skipping can improve training speed at the node level, inappropriate values ​​risk affecting the model's robustness and generalization ability.

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

[0055] In this implementation, the optimal hyperparameter combination is determined by combining the TPE algorithm with five-fold cross-validation, thereby optimizing the load forecasting model and improving its predictive performance. Furthermore, the impact of climate on load forecasting is considered during the optimization training process, which further enhances the forecasting accuracy.

[0056] Furthermore, the historical meteorological data includes data types such as temperature data, air pressure data, precipitation data, radiation data, and wind-related data; specifically,

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

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

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

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

[0061] For specific methods of obtaining historical meteorological data, please refer to Table 1;

[0062] Table 1. Composition and Acquisition Methods of Historical Meteorological Data

[0063]

[0064]

[0065] Furthermore, the evaluation metrics for load forecasting models include mean absolute error (MAE) and mean absolute percentage error (MAPE).

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

[0067] The formula for calculating the mean absolute percentage error is:

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

[0069] Furthermore, the method for obtaining the load-meteorological feature dataset is as follows:

[0070] The process involves acquiring load feature datasets and meteorological feature datasets for the target region within the same historical time domain. After filtering the load feature datasets based on feature value importance, the filtered load feature datasets are merged with their corresponding meteorological feature datasets at the same time point, forming a load-meteorological feature dataset. Specifically, feature value importance is determined by statistically analyzing the percentage of daily load importance within a week and the percentage of weekly load importance within a month to determine the data points to be retained in the sample. Feature Importance Value (FIV) is a method for evaluating the contribution of features to model prediction. It is commonly used in tree models (such as random forests and gradient boosting trees) to help understand which features are most important for the final prediction results. Figure 6 A load importance percentage chart for different time periods is provided, by Figure 6 As can be seen, Figure A represents the FIV percentage on the 2nd and 3rd days of the week, with the FIV on the 2nd day being significantly higher than that on the 3rd day; Figure B represents the weekly FIV percentage from week 1 to week 4, with the data from week 1 dominating, reflecting its importance; Figure C reflects the daily FIV percentage within week 1 in Figure B, with the data from days 2 and 3 dominating; in summary, the data from week 1 is particularly important, and the data from days 2 and 3 within it has an absolute dominant advantage.

[0071] Furthermore, the preprocessing method for the load-meteorological feature dataset is as follows:

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

[0073]

[0074] In the formula, γ k Let γ be the autocorrelation coefficient of lag k, and γ0 be the autocorrelation coefficient of lag 0, which is the variance of the time series. The formula for calculating the variance of the time series is as follows:

[0075]

[0076] In the formula, l(t) is a power load sequence containing N elements, with t as the time index, and l m It is the average value of the electricity load sequence. X t Let t be the power load data value at time t, and μ be the average value of the data in the calculated data series.

[0077] The partial autocorrelation function (PACF) is a measure of the degree of correlation after excluding other lagged terms. It is calculated using the Yule-Walker equation. as follows:

[0078]

[0079] In the formula, ρ k =ACF(l,k), representing the ACF value of k lags, calculated as described above. The PACF value with lag of k is the value in the YW matrix mentioned above.

[0080] Collected such as Figure 2 The line chart shown here contains electricity load data for 2022 and 2023. Figure 2 It can be seen that the data in this region exhibits a clear annual cycle.

[0081] Depend on Figure 3 and Figure 4As can be seen from the ACF results, the data exhibits a very strong periodicity, with each point repeating at 24-hour intervals, consistent with the periodicity of human activity. Furthermore, this periodicity gradually weakens over time. Further analysis using the partial autocorrelation function (see [link to documentation]) demonstrates this. Figure 4 A strong correlation was observed in the 0-48 period, and a certain correlation pattern was also observed in the 48-72 period.

[0082] In summary, the size of the sliding window is determined based on the autocorrelation function (ACF) and partial autocorrelation function (PACF). Specifically, the ACF indicates that the width of the sliding window should be a multiple of 24, using the calendar day as the basic window period, which aligns with the objective circadian rhythm of humans. Further analysis using PACF confirms that an effective periodic circadian rhythm is 3 days. (See...) Figure 3 and Figure 4 .

[0083] Specific Implementation Method 2: A short-term day-ahead power load demand forecasting device based on the TPE-DART algorithm, comprising 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 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.

[0084] Specific Implementation Method 3: A computer-readable storage device storing a computer program that, when executed, implements the short-term day-ahead power load demand forecasting method based on the TPE-DART algorithm.

[0085] Specific Implementation Method Four: A computer program product, comprising a computer program that, 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 experiment:

[0087] The technical effects of the present invention are demonstrated by the following verification tests.

[0088] This invention collects, for example, Figure 2 The line chart shown here contains electricity load data for 2022 and 2023. Figure 2 It can be seen that the data in this region exhibits a clear annual cycle.

[0089] Depend on Figure 3 , 4 It can be seen that after determining the width of the sliding window to be 3, Figure 4Further analysis of the correlation in the time series data revealed that the PACF value decreased to 0.1 at approximately 100 periods, indicating poor periodicity, thus avoiding the use of excessively long periods.

[0090] Depend on Figure 5 As can be seen, after a 3-day window, the prediction model error obtained using the cross-validation method mentioned above reaches its minimum, indicating a significant improvement in prediction accuracy within this timeframe. However, at a 6-day window, the error reaches its maximum, showing overfitting or instability. With increasing period, the error gradually decreases and tends to stabilize. Although the prediction error values ​​for 24-day and 27-day windows are close to those for a 3-day window, they are still higher. Considering the characteristics of load forecasting, one day's data is typically reserved for future predictions, but the first day's data is unavailable, leaving only two days' worth of data. Therefore, in practical applications, the window width is set to 2 days.

[0091] Depend on Figure 7 It can be seen that only the Prophet model's MAE value is close to that of the three ensemble learning models (original ensemble learning, DART, and EPE-DART), but its performance is still inferior to that of the ensemble learning models. Compared with other models, ensemble learning has a clear advantage. The conclusion that the MAE value is relatively low indicates that ensemble learning has higher accuracy in load forecasting for this region, and the forecasting model of this invention, namely the EPE-DART model, has the lowest MAE value, performing better than the original ensemble learning and DART models.

[0092] Figure 8 The colors in the diagram represent 1-MAPE values, with green indicating excellent performance, blue indicating good performance, and red indicating poor performance. Most models performed poorly before 150 days, while the three ensemble learning models maintained good performance. After 200 days, as temperatures stabilized, the predictions of most models became more accurate, with a wider range of excellent predictions. The TPE-DART model described in this paper has a significant advantage after 100 days, exhibiting a larger proportion of good and excellent predictions. The TPE-DART model is the load forecasting model of this invention.

[0093] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.

Claims

1. A short-term day-ahead power load demand load forecasting method based on TPE-DART algorithm, characterized in that, The method comprises the following steps: obtaining a load-meteorological feature dataset, and preprocessing the load-meteorological feature dataset, wherein input data of each sample in the load-meteorological feature dataset is historical power load data and historical meteorological data of consecutive dates, and output data is actual power load data corresponding to a prediction date; setting an initial hyperparameter combination of a DART algorithm, and constructing a load prediction model by using the DART algorithm; optimizing and training the load prediction model by using the load-meteorological feature dataset, and optimizing and constraining the hyperparameter combination of the load prediction model by using an evaluation index as a constraint condition in the process of optimizing and training, to obtain an optimized load prediction model; and in the process of optimizing and training, a TPE algorithm is used, and an optimal hyperparameter combination is determined according to a five-fold cross-validation method; predicting a load of a to-be-predicted date by using the optimized load prediction model.

2. The TPE-DART based short term day-ahead electric power load demand load forecasting method according to claim 1, characterized in that, The historical meteorological data comprises temperature data, atmospheric pressure data, precipitation data, radiation data and wind-related data.

3. The TPE-DART based short term day-ahead electric power load demand load forecasting method according to claim 2, characterized in that, The temperature data comprises temperature 2m above the ground, dew point 2m above the ground and body temperature; the atmospheric pressure data comprises sea level atmospheric pressure and surface atmospheric pressure; the precipitation data comprises rainfall and snowfall; the radiation data comprises solar shortwave radiation, horizontal solar direct radiation, global tilt irradiance and ground radiation; and the wind-related data comprises wind speed 10m above the ground and gust 10m above the ground.

4. The TPE-DART based short term day-ahead electric power load demand load forecasting method according to claim 1, wherein, The evaluation index comprises mean absolute error (MAE) and mean absolute percentage error (MAPE). The formula for calculating the mean absolute error is: The formula for calculating the mean absolute percentage error is: where y i and y ihat represent the actual and predicted values of the i-th power load, respectively, and n represents the total number of samples of the power load.

5. The TPE-DART based short term day-ahead electric power load demand load forecasting method according to claim 1, wherein, The implementation manner of obtaining the load-meteorological feature dataset is as follows: obtaining a load feature dataset and a meteorological feature dataset of a prediction target region in a same historical time domain, screening the load feature dataset according to feature value importance, and then merging the screened load feature dataset and corresponding meteorological data in the meteorological feature dataset of the same time point to which the load feature dataset corresponds, to form the load-meteorological feature dataset.

6. The TPE-DART based short term day-ahead electric power load demand load forecasting method according to claim 5, characterized in that, The feature value importance is determined by statistical load importance proportions of each day in each week and load importance proportions of each week in each month in the historical time domain, to determine data corresponding to sampling points to be retained in the sample.

7. The TPE-DART based short term day-ahead electric power load demand load forecasting method according to claim 1, wherein, The implementation manner of preprocessing the load-meteorological feature dataset is as follows: windowing the load-meteorological feature dataset to obtain a plurality of samples, and performing normalization processing on each sample, to complete the preprocessing of the load-meteorological feature dataset.

8. A short-term day-ahead electric power load demand load 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 the short-term day-ahead power load demand load prediction method based on the TPE-DART algorithm.

9. A computer-readable storage device storing a computer program, wherein the computer program comprises instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1-8. The computer program is executed to implement the short-term day-ahead power load demand load prediction method based on the TPE-DART algorithm.

10. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the short-term day-ahead power load demand load prediction method based on the TPE-DART algorithm.

Images