Abstract: Taking the connecting ridge between Shackleton and de Gerlache craters as the research area, based on real-time illumination simulation data from November 1, 2026, to February 28, 2027, a dynamic illumination dataset with a spatial resolution of 20 m/pixel and a temporal resolution of 1 hour was constructed. A deep-learning framework is proposed to recognize regions with continuous 3-day illumination, in which an improved VGG network extracts illumination-friendly regions from each temporal frame, a bidirectional GRU network captures temporal illumination characteristics, and a consistent temporal-spatial attention mechanism highlights key spatiotemporal illumination features. An output head network integrates these features to generate target regions. Based on the extracted regions and an eight-direction rover mobility model, a Sun-synchronous A* path planning algorithm is further optimized to enable illumination-aware navigation. Simulation results demonstrate that the proposed method accurately recognizes 3-day consecutive illumination-friendly regions in the 20 m/pixel dynamic dataset and effectively supports efficient rover path planning in well-illuminated areas of the lunar south pole.