Autonomous driving systems face critical challenges in adverse weather conditions, where dense haze significantly degrades both visual perception and 3D scene reconstruction accuracy. This paper proposes a novel framework DRL-TDR (Dual-scene Representation Learning for Traffic Dehazing and Reconstruction), to address the limitations of existing dehazing methods in autonomous driving scenarios. Traditional image-based approaches often neglect scene structure information in sequential data, while reconstruction-based methods struggle to disentangle geometric and atmospheric components without explicit supervision. DRL-TDR integrates three core modules: Hazy Scene Learning (HSL) for geometrically consistent coarse dehazing, Pseudo Haze Synthesis (PHS) to generate synthetic haze for closed-loop training, and a Color Correction Module (CCM) for fine-grained refinement. The framework leverages NeRF-based volumetric rendering to embed haze parameters, enabling joint optimization of scene reconstruction and dehazing. The pseudo-haze generation mechanism synthesizes realistic haze from historical clear scenes to enhance cross-domain alignment. Experiments on KITTI and KITTI360 datasets demonstrate improvements over state-of-the-art methods, achieving 11.3% higher PSNR and 10.8% better SSIM on KITTI, results confirm its effectiveness in enhancing 3D geometric accuracy and atmospheric condition recovery, which brings reliable perception in autonomous driving systems.