Accurate photovoltaic(PV) power forecasting in arid regions is challenged by dust accumulation and the nonlinear effects of rainfall, which jointly alter PV module surface states. However, most data-driven models treat meteorological inputs uniformly and fail to explicitly model how priorday precipitation modulates temperature-humidity dynamics that govern mud-film formation and rainfall-induced cleaning effects.To address this limitation, we propose a selective precipitation-fusion framework that introduces a gated-Hadamard interaction mechanism to selectively fuse precipitation categories with temperature and humidity-two variables physically linked to post-rainfall surface transitions. This design enhances sensitivity to rainfall-induced surface state changes while suppressing irrelevant feature interactions. The selectively fused features are subsequently processed by a temporal modeling backbone to capture multi-period PV generation patterns.Using one year of operational data from Alice Springs, Australia, the proposed framework achieves an R-squared (R2) of 0.936, significantly outperforming a comprehensive set of baseline models as well as variants without precipitation-aware fusion. Stratified evaluation across rainfall categories further reveals physically consistent behavior, with the highest accuracy under heavy rainfall due to the cleaning effect and the lowest under moderate rainfall due to mud-film formation.These results demonstrate that embedding physically guided and selectively designed meteorological interactions substantially improves both forecasting accuracy and physical interpretability for PV systems operating in arid regions.