Embedded differentiable predictive thermal management for integrated cabin–battery systems in electric vehicles

Nonlinear model predictive control (NMPC) is effective for integrated cabin and battery thermal management of electric vehicles, but its iterative online optimization is difficult to deploy on resource-constrained automotive microcontrollers. This paper proposes an embedded-oriented differentiable predictive control framework that replaces online NMPC solving with offline predictive policy learning and lightweight online residual correction. A differentiable thermal model is unrolled over the prediction horizon, allowing the policy network to optimize a predictive-control-inspired objective without expert NMPC labels. Battery-temperature boundary information is incorporated into policy training using an augmented Lagrangian method (ALM) with adaptive penalty updates. During deployment, a predictive correction safety (PSC) layer uses the predicted peak battery temperature to reduce residual boundary deviations. In simulations, the proposed controller achieves NMPC-comparable thermal regulation and up to 11.8% energy savings over an industrial rule-based baseline. Stress tests and ablation studies show that ALM improves policy-level constraint handling, while PSC eliminates residual violations under the tested severe conditions. Desktop profiling shows about two-orders-of-magnitude lower average computation time than solver-based NMPC. CAN-based HiL experiments are further conducted on an Infineon TC397 automotive-grade microcontroller using a thermal plant model calibrated and independently validated with production-EV field-test data, verifying real-time execution within the assigned embedded task time budget.