A Lightweight Transfer Modeling Method for PV Power Forecasting in Large-Scale Transformer Areas Based on Spatiotemporal Feature Decoupling

Abstract

The high-precision forecasting of photovoltaic (PV) power in distribution transformer areas (DTAs) is essential for accurately tracking the power distribution characteristics within distribution networks. Existing methods struggle to balance the dual demands of modeling cost and forecasting accuracy. To address this, a lightweight transfer learning model for forecasting PV power in DTAs is proposed, focusing on the ultra-short-term timescale. Firstly, dynamic time warping (DTW) distance is introduced into the affinity propagation (AP) clustering algorithm to achieve sub-region division based on output consistency. To enhance the model's ability to learn spatiotemporal information, a spatiotemporal feature decoupling architecture is developed, combining Transformer encoder (TE) with the GraphSAGE graph neural network for joint modeling of all DTAs within each sub-region. Furthermore, a lightweight transfer learning framework based on shared sub-region features is designed to facilitate rapid cross-sub-region modeling. Finally, the proposed method is tested using data from two distinct climatic regions. The results show an average reduction of 10.95% in NRMSE for 4-hour-ahead forecasts and a 9-10 fold reduction in modeling time, confirming the superiority of the proposed approach.