Interpretable prediction of multi-photovoltaic power stations via spatial-temporal multi-task learning with Transformer-XLSTM

Abstract

With numerous distributed photovoltaic (PV) power stations integrated into the electrical energy system, accurate power forecasting for multiple PV stations is essential to ensure grid stability. However, due to the intricate correlations and spatiotemporal dynamics among PV stations, prevailing methods struggle to capture and interpret cross-station interactions induced by meteorological variations and distribution. For this purpose, a novel interpretable prediction approach using spatial-temporal multi-task learning with Transformer-extended long short-term memory (XLSTM) for multi-PV power stations is proposed in this paper. First, a spatial feature extractor combining rotary position embedding, Transformer, dilated causal convolutional, and residual connection is presented to model PV power global-local features (meteorological distribution), and a temporal feature extractor based on the XLSTM with global attention mechanisms is designed to effectively capture long-term dependencies and critical temporal features (meteorological variations). Then, a multi-task prediction model based on spatial-temporal feature extractors is proposed for learning the coupling relationships for PV stations. In addition, to ensure training data quality, an isolation forest-based multi-task outlier detection model is incorporated into the forecasting approach. Finally, the Shapley additive explanations model is utilized to elucidate the relationships between the coupled features and multi-task outputs. Simulation experiments are conducted using operational data from PV stations in western China. Compared to 17 advanced benchmarks, the proposed approach achieves a root mean square error reduction in PV power prediction, with an average value of 91.63%. The experimental results also demonstrate that the proposed approach can interpret the importance of relevant features in multi-task outputs.