Research On Short-Term Power Load Forecasting Method Based On MA-GRU Model

The "double carbon" target has given a strong impetus to energy transformation,and the large amount of renewable energy sources on the grid has brought more uncertainties to the power system.Short-term power load forecasting is based on historical load sequences to determine future load demand.Accurate forecasting can provide more data support for power system scheduling and planning to ensure safe and stable operation of the system.With the development of smart grid,the quantity and dimensionality of power data have proliferated,and it is difficult for traditional forecasting methods to capture the nonlinear relationship of load.With the continuous development of computer science,machine learning methods can achieve the fitting of nonlinear relationships,but the prediction effect of a single model is still unsatisfactory for such complex variables as power load data.In order to improve the model’s ability to mine data features,this paper constructs a pure electricity load sequence from deep learning forecasting methods and establishes MA-GRU,a shortterm electricity load forecasting model based on a self-attentive mechanism fused with GRU,as follows:(1)Real load series from 2018 to 2021 for three Baltic countries are obtained from ENTSO-E platform,and after reasonable division according to experimental requirements,data exploration is performed on the training set based on statistical methods,and the sliding window size is determined using discrete Fourier transform with Pearson correlation coefficient to analyze the data for the whole year of 2018.To ensure data integrity,data outliers are detected and interpolated,and finally the training set data are processed into a load sequence suitable for conducting experiments using windowing methods.(2)S-RNN,LSTM,and GRU models were established and hyperparameter optimization was performed on the validation set using the TPE algorithm.The prediction results of the three models on the test set were analyzed to determine that the single recurrent neural network model has the problem of insufficient load sequence feature mining capability,and GRU was selected as the base model for subsequent research by integrating the prediction results and model parameter complexity.(3)Introducing the self-attentive mechanism,a short-term electric load forecasting model MAGRU based on the fusion of GRU with the self-attentive mechanism is established to enhance the feature extraction ability of a single model by using the self-attentive mechanism to enhance the data similarity relationship.After optimizing the model structure and model hyperparameters using the TPE algorithm,the model prediction capability is tested and compared with the single model of GRU before the change,which proves that the improved model effectively enhances the feature extraction capability of the single model.(4)The LR,SVR,RF,XGBoost,Cat Boost,MLP,S-RNN,LSTM,and GRU algorithms,which are commonly used in short-term load forecasting studies,were selected to build the forecasting models and compared with MA-GRU after adjusting the hyperparameters,proving that the present model has higher forecasting accuracy compared with the models in the same application area.The MA-GRU model is proved to have better generalization performance by conducting studies on the load series of Estonia and Latvia and the load series of Lithuania after COVID-19 pandemic.The experiments prove that the MA-GRU model established in this paper has MAE and RMSE values of 30.39 and 48.22 for the 2018-2019 load series in Lithuania,which are 25.64% and 20.33%lower than the single GRU network before the introduction of the self-attention mechanism,and 1-MAPE and R2 of 97.83% and 0.86,respectively,which improve the 0.7% and 0.1,which greatly improved the accuracy of the single model and received the highest rating compared to the single prediction model in the same domain.The problem of large prediction errors of the single model for data from sampling points near the peak load is effectively improved,and the errors of sampling points from 16:00 to 21:00 daily are reduced by 32.3%,44.6%,45.3%,54.4%,54.8%,and 39.7%,respectively.the MA-GRU model can still maintain good performance on the rest of the country datasets in the same region,and after encountering large contingency events,the model’s 1-MAPE value can reach 97.61%,which has good generalization ability.