Conceptual closed-loop design of automotive cooling systems leveraging Reinforcement Learning

The transition from conventional to battery electric vehicles has significantly altered cooling system requirements. Previously, the primary component to cool was the combustion engine, whose waste heat could also be used to heat the passenger compartment. In battery electric vehicles, the electric motor, inverter, and battery each operate optimally within different temperature ranges, with battery aging particularly affected by non-optimal temperatures. Consequently, the design of cooling systems for electric vehicles is a topic of high interest, requiring the comparison of various concepts to identify the best solution. Since the behaviour and performance of cooling system concepts largely depend on the control strategy employed, it is essential to consider this aspect for proper evaluation of their closed-loop performance. Reinforcement Learning (RL) offers a promising approach to rapidly design control strategies for thermal systems, as it learns these strategies based on a high-level objective function. Traditionally, training an RL controller involves considerable manual effort, such as hyperparameter tuning. This paper investigates whether RL can be applied to a cooling system to evaluate its optimal closed-loop performance with minimal manual tuning effort.