ABSTRACT Accurate multi‐station photovoltaic (PV) power forecasting is essential for mitigating the randomness and volatility of PV generation in large‐scale grid integration. Conventional approaches, such as summation and statistical upscaling, often overlook spatial correlations among stations, leading to suboptimal accuracy and increased model complexity. This paper proposes a graph convolutional network (GCN)‐based framework incorporating terrain features for enhanced multi‐station forecasting. A spatial feature extraction method is constructed using mutual information matrices fused with terrain features (distance, orientation and topography quantified via principal component analysis) to strengthen spatial dependency capturing ability of GCN. To further enrich the representation, a dual‐layer GCN architecture is proposed, where the first layer processes gridded numerical weather prediction (NWP) data in the main graph and the second layer facilitates inter‐station information propagation in the subgraph. A multi‐channel LSTM architecture is then employed, allowing each PV station to learn temporal dynamics independently while preserving inter‐station correlations captured by the GCN. Case studies on two datasets of real‐world PV stations demonstrate that the proposed method reduces forecasting errors by 0.5%–1.2% compared with benchmark models, achieving superior accuracy under different weather conditions. The results confirm that the proposed graph structures integrating terrain features with gridded NWP and multi‐channel learning provides significant improvements in forecasting performance.
Xie et al. (Thu,) studied this question.