In industrial manufacturing, peg-in-hole assembly relies on images captured by cameras equipped with high-intensity light sources to accurately identify hole positions. However, the extensive use of such bright lighting contradicts energy conservation efforts and green manufacturing principles. Executing assembly tasks in low-light environments can reduce energy consumption but presents challenges in accurately localizing the hole. In this work, a visual-based residual reinforcement learning (RL) assembly framework is introduced for assembly tasks in low-light conditions. Firstly, an unsupervised generative adversarial network, dubbed the low light assembly network (LLA-Net) model, is proposed for transferring the images captured in low light environments to common light environments. The shape of the assembly hole is integrated into the Network as prior information to supervise the generation of the hole images, improving the generation of local details of the hole region. Then, the transferred image is used to calculate the prior trajectory and input it to the RL-based policy to output the assembly action. The experiment shows that the proposed method can successfully perform assembly tasks under low-light conditions, contributing to achieving efficient energy conservation in the manufacturing industry.