Distributed photovoltaic (DPV) power forecasting is essential for grid stability but remains challenging due to strong intermittency and meteorological uncertainty. Existing data-driven models often lack physical interpretability and struggle to capture asymmetric dependence structures

leading to unreliable predictions during extreme weather. This paper proposes a copula-guided transformer (CGT) framework that integrates statistical dependence mining with physics-informed deep learning. Specifically

Gaussian copula is used for global feature screening

whereas Clayton and Gumbel copulas quantify asymmetric tail dependencies—revealing the conditional lower-tail inhibitory effect of rainfall and the upper-tail driving effect of irradiance on power output. These copula-derived parameters are embedded as physical priors into a guided encoder

where a temporal convolutional network (TCN) dynamically regulates attention weights to enhance physical consistency. Validated on real-world DPV data from China

the CGT model offers significant performance advantages. The results demonstrate superior robustness across clear-sky

rainy and high-volatility scenarios by effectively mitigating spurious overestimation and response lag.