Explainable and physics-constrained PV power prediction via a hybrid framework Integrating secondary decomposition and improved Transformer-LSTM

Photovoltaic power generation (PVPG) is susceptible to meteorological conditions, exhibiting significant randomness and volatility. Therefore, accurate and reliable PVPG prediction is crucial for enhancing grid stability. However, existing data-driven prediction methods often overlook the system's inherent physical mechanism, which can lead to prediction results that violate actual operating laws. This study presents a physics-constrained hybrid model, integrating Transformer and Long Short-Term Memory (LSTM) networks with a secondary decomposition strategy, for the multi-step short-term forecasting of PVPG. Initially, a Seasonal and Trend Decomposition using Loess (STL) method is utilized to decompose the original dataset. Subsequently, variational mode decomposition (VMD), optimized by an improved Dream Optimization Algorithm (DOA), is utilized to decompose the residual term. Subsequently, the decomposed components and the screened features are fed into a hybrid Transformer-LSTM model, with its hyperparameter optimized by an improved Dream Optimization Algorithm, to complete the final power prediction. To ensure the predictions adhere to the physical principles of photovoltaic power generation, the model utilizes a designed physics-constrained loss function specifically. On the Australian dataset, the proposed model is evaluated and is observed to achieve better performance than other methods in both prediction accuracy and robustness. Specifically, on Site 1, the R-squared and RMSE for the overall prediction performance are 0.9423 and 0.2326, respectively, demonstrating superior prediction performance. Moreover, it also exhibits superior prediction capability across different datasets, seasons, and weather conditions. Finally, explainability analysis was conducted using SHAP method. This multi-step short-term PVPG prediction method has the potential to enhance grid stability and the stable regulation of energy.