Short-Term Load Forecasting for A Power Supplying District Based on CEEMDAN-WPE-LSTM-Stacking Methods

To ensure the reliability of flexible control of low-voltage substation power supply, improve the stability and economic benefits of the distribution network, a short-term substation net load forecasting method based on CEEMDAN-WPE decomposition and LSTM-Stacking is proposed. First, after data preprocessing, in order to reduce the non-stationarity of the sequence and limit the model scale, the substation net load is decomposed into multiple components through complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), and k-means is used to reorganize the components according to the weighted permutation entropy (WPE) value. Then, according to the characteristics of the components, the low-complexity recombination components and the high-complexity recombination components are divided, and the corresponding feature subsets are input into long short-term memory (LSTM) network and Stacking for prediction respectively. Finally, the prediction results of each recombinant component are superimposed to obtain the final prediction result. The real data of a photovoltaic substation in Hebei is used for example analysis and compared with RF, XGBoost, LSTM and Stacking. The prediction results show that the CEEMDAN-WPE-LSTM-Stacking model proposed in this paper has higher prediction accuracy. The proposed decomposition method can effectively reduce the non-stationarity of the time series and improve the prediction accuracy of the model.