Spatial‐Temporal Feature Fusion Network for Short‐Term Photovoltaic Forecasting Utilising Multi‐Source Heterogeneous Data

Accurate short‐term photovoltaic (PV) power forecasting plays a key role in mitigating challenges from PV's intermittent, stochastic output — for example, grid frequency fluctuations — thereby ensuring grid stability and optimising renewable energy scheduling. However, modelling multi‐source heterogeneous data and extracting spatial‐temporal features remains challenging. To address this, we proposed a novel deep hybrid model integrating self‐attention enhanced convolutional block attention modules (SCBAM), temporal convolutional networks (TCN) and U‐net (UNet)‐based feature fusion; this model is abbreviated as SCTU. It processes multi‐grid/single‐grid numerical weather prediction (NWP) data, alongside historical power data, via five parallel branches (tailored for spatial/temporal extraction). We evaluated SCTU on two Desert Knowledge Australia Solar Centre stations. Results show it outperforms benchmarks: normalised root mean square error (NRMSE) 3.985%/5.791% and coefficient of determination ( R 2 ) 0.9509/0.9397 on eco‐kinetics (EK‐PV) and service station (SS‐PV), respectively. Across sunny/overcast/rainy, SCTU maintains lower NRMSE (EK‐PV 3.432%–6.766%, SS‐PV 5.331%–7.246%) than benchmarks. Ablation studies confirm the importance of SCBAM, UNet and multi‐grid input in enhancing accuracy. SCTU exhibits strong robustness (rainy‐day NRMSE: EK‐PV 6.766%, SS‐PV 7.246%, lower than TCN, respectively) and generalisation (EK‐PV/SS‐PV datasets: R 2 0.9509/0.9397, NRMSE 3.985%/5.791%). Notably, it relies on high‐resolution NWP — delays or reduced extreme‐weather accuracy may lower performance — and its generalisation to high‐latitude regions needs validation.