Abstract: The development of efficient Thermal Management Strategies (TMS) is critical for Hybrid Electric Drive Tracked Vehicles (HETVs) due to the severe thermal safety and energy consumption challenges encountered during complex operations. Conventional strategies struggle to balance high-precision temperature control with multi-objective collaborative optimization, while also requiring long development cycles and exhibiting weak generalization capabilities. To address these issues, we propose a hierarchical thermal management framework integrating a Gated Recurrent Unit Multi-Head Attention Twin Delayed Deep Deterministic Policy Gradient with Model Predictive Control (GMA-TD3-MPC). This framework dynamically integrates Reinforcement Learning (RL) and Model Predictive Control (MPC), utilizing a Gated Recurrent Unit with Multi-Head Attention (GRU-MHA) module to optimize energy consumption and temperature control precision under cyclic conditions; meanwhile, it implements a dynamic threshold triggering mechanism to seamlessly transfer control to the MPC controller when approaching thermal safety limits. Our simulation results demonstrate that compared to baseline strategies, the proposed method accelerates convergence by approximately 28% and 40% over DDPG and TD3, respectively. In a standard temperature environment (25 °C) under off-road conditions, compared to standalone MPC, the proposed strategy reduces temperature fluctuation ranges in high-temperature and low-temperature circuits by 44.19% and 6.45%, respectively, while also achieving a 5.53% reduction in total energy consumption and a 10.63% decrease in peak power demand. Furthermore, under high-temperature conditions (45 °C), the strategy reduces the total energy consumption by 13.41%.