To address the issues of energy consumption and environmental instability caused by frequent switching of on-off control devices in artificial climate chambers, this paper proposes an intelligent control framework that integrates physical constraints with deep reinforcement learning. A Dual-Path Prediction Network (DPP) is constructed to build a high-precision simulation environment for the climate chamber, achieving comprehensive environmental prediction with a modeling accuracy of \(R^2 = 0.9979\) and \(MSE = 0.0016\) . Based on this, a discrete control strategy is designed using a Deep Q-Network (DQN), with a composite reward function that incorporates penalties for device switching. Experimental results show that, compared to traditional threshold-based control, the proposed method reduces the switching frequency of the heating plate and compressor by 78.4% and 72.5%, respectively, while maintaining temperature accuracy within ±0.5 \(^{\circ }\) C. This study provides an intelligent control paradigm for on-off control systems in climate chambers, offering practical value in reducing operational costs and enhancing environmental stability.