This paper constructs an optimization framework that integrates thermodynamics and machine learning, and proposes an improved deep reinforcement learning thermal management algorithm. By establishing a heat generation correction formula with SOC influence, a 3-D thermal resistance network model, etc., combined with COMSOL simulation (temperature error <3∘C) and 18650 battery pack experiments, the algorithm advantages are verified: the response lag is reduced to within 0.5 second, the control deviation is ±1∘C, the temperature difference is reduced by 56.9% compared with the traditional PID, and the energy consumption is reduced by 18%, providing a new path for safe and efficient operation of batteries.