Data-driven reinforcement learning-based parametrization of a thermal model in induction traction motors

Monitoring the temperature of induction traction motors is crucial for the safe and efficient operation of railway propulsion systems. Several thermal models were developed to capture the thermal behaviour of the induction motors. With proper calibrating of the thermal model parameters, they can be used to predict the motor’s temperature. Moreover, calibrated thermal models can be used in simulation to evaluate the motor’s performance under different operating conditions and find the optimal control strategies. Parameterization of the thermal model is usually performed in dedicated labs where the induction motor is operated under predefined operating conditions and calibrating algorithms are then used to find the model’s parameters. With the development of digital tools, including smart sensors, Internet of Things (IoT) devices, software applications, and various data collection platforms, operational data can be collected and used later to calibrate the parameters of the thermal model. Nevertheless, calibrating the model’s parameters from operational data collected from different driving cycles is challenging as the model has to capture the thermal behaviour from all driving cycles’ data. In this paper, a data-driven reinforcement learning-based parametrization method is proposed to calibrate a thermal model in induction traction motors. First, the thermal behaviour of the induction motor is modelled as a thermal equivalent network. Second, a reinforcement learning (RL) agent is designed and trained to calibrate the model parameters using the data collected from multiple driving cycles. The proposed method is validated by numerical simulation results. The results showed that the trained RL agent came up with a policy that adeptly handles diverse driving cycles with different performance characteristics.