Temporal Convolutional Network with a Physics-Informed Decoder for 15-Minute-Ahead Solar Power Forecasting

As renewable energy grows, photovoltaic power integration into power grids has increased, and high solar penetration creates challenges for power system operation. Accurate, near-real-time solar power forecasting over short time intervals is therefore important for supporting stable grid operation and efficient energy scheduling. This study presents a novel Enhanced Temporal Convolutional Network (TCN) model to predict solar power output 15 min in advance. The model uses dilated convolutions to capture temporal patterns, while multi-head attention and adaptive feature gating help identify important dependencies and relevant inputs. It also extracts long-term trends and short-term variations using seasonal decomposition. A physics-informed decoder builds the forecast by correcting a persistence-based reference forecast. The model was trained on a dataset comprising 5.7 years of real-world solar power generation records from Arizona. After bias correction, the model achieves a root mean squared error (RMSE) of 281.0 ± 4.6 Wh, a mean absolute error (MAE) of 108.8 ± 2.9 Wh, and a coefficient of determination (R2) of 0.928 ± 0.002 across three random seeds. Compared with the baseline models, the proposed Enhanced TCN reduces the RMSE by 7.0% to 18.8% over five deep learning models, by 8.3% over persistence and by 4.9% over smart persistence. Without bias correction, the model performs no better than smart persistence. The ablation study shows that the physics-informed decoder gives the main gain.