A physics-guided hybrid deep learning model for medium- and long-term photovoltaic power forecasting☆

Accurate medium to long-term power forecasting is key to ensuring the efficient integration of photovoltaics and the safe regulation of the power grid. To address the issues of insufficient physical consistency and the accumulation of long-term prediction errors in data-driven models, this paper proposes a physics-guided Convolutional Neural Networks-Bidirectional Long Short Memory Network-Attention hybrid model (PG-CNN-BiLSTM-Attention). The model constructs a differentiable physics-constrained regression layer, introducing physics-inspired consistency penalty terms based on first and second derivatives into the loss function, embedding domain knowledge of the photovoltaic system as physics-inspired regularizations into the deep learning training process. It uses CNN to extract spatial correlations of meteorological features, BiLSTM to capture temporal dependencies, and a self-attention mechanism to focus on critical time steps, with Bayesian optimization employed for hyperparameter tuning. Based on the actual operational data of a photovoltaic power station in Zhejiang Province, China, the model’s effectiveness was validated on three typical prediction tasks of 24 h, 72 h, and 144 h. The experimental results show that: 1) The PG-CNN-BiLSTM-Attention model achieved the best accuracy in all prediction tasks, with R 2 exceeding 0.85. In the 144-hour long-term prediction, R 2 improved from 0.289 of the CNN-BiLSTM-Attention model without physical guidance to 0.852; 2) The physical guidance mechanism significantly accelerates model convergence, with the loss decreasing substantially within approximately 50 epochs and stabilizing after around 150 epochs; 3) The model demonstrates strong noise robustness, maintaining R 2 above 0.8 under 20% Gaussian noise, outperforming Transformer and Informer. The results of this study provide a reliable technical solution for high-precision photovoltaic power prediction, which has significant practical implications for promoting the integration of renewable energy and the development of smart grids.