Spatiotemporal Multi-step PV Forecasting via LSTM-GAT and Flow-based Uncertainty

Photovoltaic power generation systems are an essential component of modern power systems, characterized by inherent instability and uncertainty. This poses a significant challenge for accurate prediction. This paper introduces a novel end-to-end architecture that integrates Long-Term Short-Term Memory (LSTM), Graph Attention Network (GAT), and Standardized Flow (Flow) models, known as the new LSTM-GAT-Flow model. This model leverages the LSTM module to capture the temporal dynamic features of photovoltaic data, the GAT module to shape spatial dependencies among multiple power stations using self-attention mechanisms, and the Flow module to implement probabilistic modeling of photovoltaic output to estimate uncertainties. Analysis on real datasets shows that the LSTM-GAT-Flow model not only performs better in point prediction but also provides a practical method for considering uncertainties. It demonstrates strong stability in actual power system operations and is highly valuable for scheduling and decision-making.