Dynamic graph and multi-period temporal feature learning for predicting regional photovoltaic power generation

Accurate forecasting of photovoltaic (PV) power output has become increasingly important as large-scale PV installations continue to connect to modern power grids. In regional PV clusters, diurnal cycles and seasonal variations generate multi-period characteristics, while changing meteorological conditions introduce dynamic spatial correlations among plants. However, existing approaches often fail to distinguish subtle meteorological differences across PV plants and struggle to jointly capture multi-period temporal patterns, resulting in limited predictive performance. To address these challenges, this study develops a dynamic multi-period spatial–temporal network for regional PV power forecasting. Meteorological similarity between plants is first quantified to construct a weighted meteorological matrix, which is then combined with a geographic distance matrix to form a dynamic, weighted association graph that characterizes inter-plant relationships. A graph convolutional network is applied to capture the dynamic spatial dependencies embedded in this graph. In parallel, frequency-based analysis of PV output and meteorological conditions converts the original one-dimensional sequences into two-dimensional representations. A multi-scale convolutional network is subsequently designed to extract short-term intra-period features and long-term inter-period patterns from two-dimensional representation simultaneously. The extracted spatial–temporal representations are adaptively fused to generate future PV power predictions. Experiments on real-world regional PV system demonstrate that the proposed model yields substantially improved accuracy and generalization over state-of-the-art baselines. It achieves RMSE and MAE values of 11.65 MW and 5.60 MW, corresponding to reductions of 19% and 34%, respectively. Sensitivity analysis further reveals that multi-period temporal dependencies play a dominant role in ensuring reliable PV power forecasting, particularly for multi-step prediction horizons.