A Multi-Level Distributed Photovoltaic Power Forecasting Method Based on Knowledge Transfer

Conventional independently trained models for distributed photovoltaic (DPV) power forecasting at the feeder, main-transformer, and regional-grid levels do not explicitly reuse spatiotemporal representations across the distribution-network hierarchy. This paper proposes a knowledge-transfer-based multilevel method for short-term DPV power forecasting. First, a T-GCN encoding block integrating a temporal convolutional network and an adaptive graph convolutional network is developed at the distribution-transformer-area–feeder level to capture multiscale temporal dependencies and implicit spatial correlations. Second, the pretrained feeder-level encoder is transferred to the main-transformer task and frozen, while a Transformer-based cross-feeder aggregation network and a decoder are trained for the target task. At the grid level, a meteorology-driven weighting network fuses the representations of individual main transformers. In a regional case study using real-world operating data, the proposed method achieves equal-weighted mean relative reductions of approximately 20.0% in NRMSE and 17.5% in NMAE over an independently trained T-GCN baseline across the three forecasting levels. Identical-architecture comparisons at the main-transformer level further show that fine-tuning a pretrained encoder improves forecasting accuracy relative to training from random initialization, while freezing the pretrained encoder yields additional error reductions. These findings support the proposed pretraining-and-freezing strategy under the evaluated conditions.