Temperature rise is a salient characteristic of the braking device of heavy-duty vehicles (HDVs) during long downhill braking. The temperature-independent effect on braking performance is an inevitable challenge for the longitudinal controller, which also causes the overheating of mechanical or hydrodynamic braking devices and may bring about brake failures for heavy-duty vehicles. To this end, a universal Bi-level control framework combining a temperature hierarchy system performance prediction method and a deep reinforcement learning (DRL)-based controller is proposed for long downhill braking of heavy-duty vehicles. Firstly, the temperature-independent characteristic is clustered to predict the braking performance under various rotating speeds. Secondly, a data-driven model for temperature rising is built for the long-time braking thermal prediction and estimates the safety remaining braking time using the auxiliary braking. Thirdly, a temperature-hierarchy environmental perceptive control framework with Double Deep Q Network (DDQN) algorithm is exploited to achieve the target speed tracking accuracy. Thermal safety is ensured with the application of fast calculating for the thermal rising, along with the effective estimation of remaining braking performance on endurance braking. The proposed Bi-level longitudinal controller is compared with the average-temperature strategy and the classic PID strategy to validate its superiority in terms of speed-tracking accuracy on robust conditions. The simulation results show that the proposed strategy improves the speed tracking accuracy by 37.31% on constant slope conditions and 68.11% on varying slope conditions compared with the classic PID strategy. Furthermore, a processor-in-the-loop test experiment verifies its real-time application.