Distributed-regional photovoltaic power generation prediction with limited data: A robust autoregressive transfer learning method

This paper proposes a distributed regional photovoltaic (PV) power generation prediction method to address scenarios with a very high percentage of missing data. The proposed approach enhances PV power generation information through two key strategies. Firstly, for the limited reference power stations with available PV generation data in the region, an interpretable prediction model based on the DSC-LightGBM algorithm is developed to enhance the accuracy of PV power generation forecasts. For the problem of less available meteorological data at these stations, physical modeling introduces solar characteristics such as solar altitude and solar time angles, and then the Shapley Additive exPlanations (SHAP) interpretable algorithm is used to analyze the importance of original and enhanced features. Secondly, to address a lack of data at large number non-reference power stations in the region, this paper proposes for the first time the autoregressive transfer learning method, which divides the power stations in the region into three levels of “source domain-virtual source domain-target domain” to increase the reference information source of power stations that lack data collection. The proposed algorithm is validated on three datasets with varying locations, capacity sizes, and meteorological characteristics. In high missing-data scenarios, the autoregressive transfer learning method reduces regional prediction errors by an average of 25.8 % to 50.3 % compared to other approaches. Additionally, for estimating the power output of individual non-reference power stations, this method achieves an average error reduction of 19.7 % to 50.8 %.