A physics-constrained deep learning framework enhanced with signal decomposition for accurate short-term photovoltaic power generation forecasting

Accurate short-term forecasting of photovoltaic power generation is vital for maintaining the stability and efficiency of modern power systems . However, the variability and complexity of photovoltaic power, driven by meteorological factors , pose challenges for traditional models in achieving reliable forecasts. This study introduces a physics-constrained deep learning framework enhanced with signal decomposition to address these challenges. The framework employs complete ensemble empirical mode decomposition with adaptive noise to decompose photovoltaic power time series into intrinsic mode functions and a residual component, effectively extracting key dynamic features. These components are integrated with meteorological variables to construct a comprehensive feature matrix. A hybrid convolutional neural network-long short-term memory model captures spatial and temporal dependencies within the data. Furthermore, a customized photovoltaic power generation loss function, incorporating mean square error , regularization terms, and physical constraints, ensures the forecasts align with physical laws governing photovoltaic power generation. Evaluation results from extensive experiments demonstrate the framework's superior accuracy, robustness, and adherence to physical principles compared to baseline models. This work provides a novel and effective approach to enhancing photovoltaic power forecasting, supporting renewable energy integration into power grids, and improving overall system reliability. • The proposed framework of preprocessing, decomposition and modeling yields consistent forecasts for non-generation periods. • PVPG loss penalizes idle periods to avoid unrealistic forecasts, balance errors and improve accuracy and consistency. • With limited data the framework decomposition and feature splicing enable learning generalization and accurate forecasts.