A Cross-Variable Time-Series Transformer Architecture Integrating Physical Features for Photovoltaic Energy Forecasting

Abstract

Accurate photovoltaic (PV) energy forecasting is vital for grid stability and the global low-carbon transition. However, existing data-driven and channel-independent PV energy forecasting models struggle to capture nonlinear meteorological couplings, heterogeneous physical scales across stations, and high-frequency non-stationary fluctuations. To address these limitations, this study proposes a Physics-Guided Cross-Variable Temporal Transformer architecture. Building upon a channel-independent foundation, we introduce a Cross-Variable Attention mechanism to explicitly reconstruct nonlinear photothermal couplings via dynamic attention weights. To resolve multi-station physical scale discrepancies, a Physical Feature-wise Linear Modulation network utilizes installed capacity as a static prior for adaptive cross-station scale alignment. During optimization, a Time Dynamics-Aware Perceiving Loss jointly penalizes absolute errors and first-order time differences, constraining the network’s tracking ability for transient ramping. Experiments demonstrate that the proposed architecture overcomes traditional channel-isolation limitations. The model achieves a 21.6% reduction in MSE compared to PatchTST, a 44.0% reduction compared to Autoformer, and a 2.7% improvement in R2 over Informer. This provides an accurate, generalizable, and physically interpretable solution for collaborative multi-station distributed PV energy forecasting.