High precision prediction of time-varying photovoltaic power based on dynamic adjacency matrix and temporal spectral graph convolution network

Accurate photovoltaic (PV) power generation forecasts are crucial for renewable power scheduling and integration. However, the inherent intermittency and volatility renders the PV data time-variant, posing significant challenges for accurate prediction. To address this issue, we propose a novel hybrid forecasting model named adaptive temporal spectral graph convolutional network (ATS-GCN) to combine multi-domain information and dynamically extract truly relevant adjacency matrices, proving high precision PV power prediction under evolving and hyperdynamic scenarios. The ATS-GCN framework combines a continuous wavelet transform module for spectral feature extraction, an adaptive feature extractor based on curvilinear Mahalanobis distance for time-varying correlations calculating among PV data, a graph convolutional network and a gated recurrent unit for spatial–temporal features capturing. Furthermore, an optimized dropout strategy is incorporated to prevent overfitting. By fusing nonlinear features from multiple domains and dynamically capturing the time-varying correlations, ATS-GCN not only improves prediction accuracy but also enhances the cross-seasonal stability and interpretability of PV prediction. The proposed model is compared with baseline models on seven PV datasets with 3-step and 24-step forecasting horizon, achieving the smallest average error interquartile range span of 1.96 kW and the smallest average error median of 0.31 kW, proving the superiority of ATS-GCN. These results demonstrate the model’s strong predictive capability for practical PV power generation, rendering it a valuable tool for advancing intelligent energy management and supporting renewable energy integration into smart grids. • A novel ATS-GCN model for short-term PV power prediction is proposed. • Adaptive feature extractor based on curvilinear Mahalanobis distance captures nonlinear and time-varying PV data correlation. • Continuous wavelet transform extracting high and low spectral components of all PV data improves prediction accuracy. • An optimized dropout strategy designed for GCN-GRU framework improves prediction robustness.