Enhanced xPatch for Short-Term Photovoltaic Power Forecasting: Supporting Sustainable and Resilient Energy Systems

Accurate short-term photovoltaic (PV) power forecasting is a cornerstone for enhancing grid stability and promoting the sustainable integration of renewable energy sources. However, the inherent volatility of PV power, driven by multi-scale temporal patterns and variable weather conditions, poses a significant challenge to existing forecasting methods. This paper proposes NNDecomp-AdaptivePatch-xPatch, an enhanced deep learning framework that extends the xPatch architecture with a neural network-based decomposition module and an adaptive patching mechanism. The neural network decomposition module separates input signals into trend and seasonal components for specialized processing, while adaptive patching dynamically adjusts temporal windows based on input characteristics. Experimental validation on five real-world PV datasets from Australia and China demonstrates significant performance improvements. The proposed method achieves superior accuracy across multiple prediction horizons, with substantial improvements in mean absolute error (MAE) compared to baseline methods. The enhanced framework effectively addresses the challenges of short-term PV prediction by leveraging adaptive multi-scale feature extraction, providing a practical and robust tool that contributes to the sustainable development of energy systems.