Load forecasting for virtual power plants using VMDLSTM-transformer with SHAP explainability analysis

As a novel power system architecture that integrates various distributed energy resources, virtual power plants (VPPs) exhibit high volatility and uncertainty in their load characteristics. This study proposes a hybrid method combining variational mode decomposition (VMD), long short-term memory (LSTM) networks, and the Transformer architecture to improve the accuracy of short-term load forecasting in VPPs. The proposed VMD-LSTM-Transformer model decomposes complex VPP load time series into intrinsic mode functions, processes temporal dependencies using LSTM networks, and captures long-range dependencies through the attention mechanism of the Transformer. Experimental results from a real VPP load dataset demonstrate that our method outperforms mainstream forecasting models. Additionally, SHAP values are used to expl ain the prediction results, identifying key features and revealing that humidity and wind speed have minimal impact on the accuracy of VPP load forecasting. Removing these low-impact features maintains the prediction performance.