Dual-Layer Drift Detection and Adaptive Retraining for Ultra-Short-Term Photovoltaic Power Forecasting

In operational environments, photovoltaic (PV) forecasting models progressively degrade due to drift events induced by changing environmental conditions, system aging, and surface soiling. Mainstream forecasting models are typically trained under static assumptions and rarely account for such non-stationary effects, resulting in biased predictions, reduced reliability, and suboptimal dispatch decisions. To address this limitation, we propose a drift-aware forecasting strategy that identifies drift points by jointly monitoring variations in data distributions and prediction errors, and dynamically adjusts model update timing to maintain forecasting accuracy. First, an integrated STL-EDDM strategy is developed to decompose input features, detect anomalous fluctuations, and track performance variations for robust drift identification. Subsequently, Neuralprophet is employed as the forecasting backbone to generate ultra-short-term PV power predictions, and the model is adaptively retrained when drift events are detected. Validation results from 10 PV power stations in China show that the proposed STL-EDDM strategy reduces the average MAE from 2.480 MW to 1.287 MW, corresponding to a 48.1% improvement over the non-adaptive baseline. Furthermore, comparative experiments on the PVOD dataset showed that the proposed framework outperformed the evaluated baseline models overall, reducing average MAE and RMSE by 15.2% and 16.5%, respectively, relative to the best-performing baseline. These findings demonstrate that the proposed drift-aware strategy enhances both interpretability and robustness in non-stationary PV forecasting scenarios. The framework provides operational insights into model degradation mechanisms and practical guidance for real-time model management, with potential applicability to a broader class of real-time forecasting tasks.