Short-Term Distributed Photovoltaic Power Forecasting Based on Limited-Sample All-Sky Imagers

To address the challenges of insufficient spatiotemporal resolution in Numerical Weather Prediction (NWP) for capturing rapidly changing local cloud conditions and the limited sample size of high-resolution all-sky imager (ASI) data in ultra-short-term distributed photovoltaic (PV) power forecasting, this paper proposes a novel two-stage hybrid prediction method. The approach integrates long-term trend information from ECMWF NWP data with detailed local cloud features extracted from small-sample ASI data. First, a base model combining CrossFormer (for capturing cross-time and cross-variable dependencies) and XGBoost (for handling nonlinear relationships) is built using NWP meteorological variables and historical PV power. Second, a residual correction model based on XGBoost is developed. It utilizes cloud optical ratio (COR) features extracted from ASI images via an optimized tri-modal threshold segmentation strategy and enhanced by a Time-DP model (prototype module and assignment module) to overcome small-sample limitations and learn deeper cloud dynamics. This residual model compensates for power fluctuations caused by rapidly changing clouds that the base model fails to capture. Experimental results on two distributed PV plants (one with only ~1 month of ASI data) show significant improvements: the hybrid model substantially reduces RMSE and MAE, increases R2, and achieves over 98% Accuracy (ACC), demonstrating effectiveness in complex cloudy conditions and strong generalization with limited ASI samples.