A Novel Multi-Scale Feature Attention Network-Based Prediction Model for the Accurate Forecasting of Photovoltaic Power Generation

The global rise in energy demand has driven the large-scale integration of photovoltaic (PV) systems into the power grids, increasing the need for accurate forecasting models to manage solar energy variability and uncertainty. This study proposes a novel 24-hour-ahead PV power generation prediction model based on the Multi-Scale Feature Attention Network (MSFAN), which integrates convolutional neural networks (CNNs), attention mechanisms, and gated recurrent units within a non-sequential architecture to enhance prediction accuracy. The model also incorporates a two-step feature selection technique, Mutual Information-Interaction Gain (MI-IG), to extract the most relevant input features whilst minimizing redundancies. The proposed model was validated using real-world PV power generation data from the Arvand power plant in Mahan, Iran. The results indicate that the the model outperformed traditional methods, including LSTM, MLP, and Elman networks. It achieved a mean squared error (MSE) of 0.001, mean absolute error (MAE) of 0.02, mean absolute percentage error (MAPE) of 2.5%, and a coefficient of determination (R²) of 0.991—substantially surpassing the performance of the comparison models. These results underscore the model's potential to improve PV forecasting accuracy and support more stable and efficient power grid operations.