A Short-Term Prediction Method for Photovoltaic Power Based on Spatio-Temporal-Sample Correlation Pattern Mining

Accurate short-term photovoltaic (PV) power prediction is crucial for ensuring grid stability, especially under the ‘dual-carbon’ goal and energy transition. However, limited historical data from new or under-maintenance PV plants leads to model underfitting and low prediction accuracy in deep learning-based approaches. To address this, this paper proposes a short-term PV power prediction model based on spatio-temporal-sample correlation mining using the Transformer architecture. The model incorporates a spatio-temporal attention module, leveraging spatial and temporal convolutional attention mechanisms to capture correlations between PV power and meteorological variables (e.g., irradiance, temperature, humidity). A fusion module integrates these features, while a batch sample correlation learning module uses multiple attention mechanisms to model inter-sample relationships. Simulation results demonstrate that the proposed method outperforms other models in small-sample scenarios by effectively mining spatio-temporal and inter-sample correlations.