Under the background of “dual-carbon” strategy and transportation energy integration, deploying photovoltaic (PV) power generation systems in highway service areas has become an important direction of green infrastructure. It is of great significance to carry out day-ahead hourly power generation prediction to improve the local power consumption level and optimize the charging load scheduling strategy. Due to the stochastic and non-stationary nature of photovoltaic power generation, which is susceptible to meteorological factors, the traditional single model has obvious deficiencies in dealing with complex signals and constructing a robust prediction structure. In this paper, a variational modal decomposition-driven deep learning network (VMD-CNN-LSTM) is proposed to construct an intelligent prediction system including signal processing-feature extraction-timing prediction through the integration of physically-driven decomposition and dual-engine architecture. The variational modal decomposition is firstly introduced as a pre-optimization module to deconstruct the original non-smooth sequences into intrinsic modal components (IMFs) with clear physical meanings through frequency-domain orthogonalization. Subsequently, the cross-domain correlation analysis of meteorological features and modal components is realized by the constructed CNN-LSTM synergistic mechanism, which improves the coupled characterization ability of PV power generation and meteorological features. The experimental results show that compared with LSTM, CNN, VMD-LSTM, VMD-CNN, and CNN-LSTM models, the method in this paper improves more than 30% on average in the key metrics such as MAE and RMSE, and this result suggests that its anti-jamming mechanism and cross-scale learning capability provide a reliable technological path for high volatility PV power prediction.