Integrating Recurrent Neural Networks into Model Predictive Control for Thermal Torque Derating of Electric Machines

Abstract

For cost and space reasons, the electric machine (EM) and its cooling system in electrified powertrains for road vehicles are usually designed in such a way that the maximum power cannot be called up continuously. For this purpose, control systems are used that derate the power of the EM depending on its temperature. In this context, this paper presents the integration of a Recurrent Neural Network (RNN) model into a nonlinear model predictive control (MPC) to efficiently control the driving performance of an electric vehicle (EV) on a race track and reduce thermal derating effects. The discrete black-box RNN model incorporating a Long Short-Term Memory (LSTM) layer describing the thermal dynamics of the electric machine is combined with the continuous-time one-dimensional vehicle dynamics to form the hybrid MPC. For the selected example, the RNN-MPC outperforms a standard linear derating strategy by 6.23% in terms of energy consumption while achieving equal lap speed and maximum machine temperatures. Compared to an existing grey-box thermal network model, the RNN model significantly reduces the temperature prediction error by 28%. The trajectory tracking problem is formulated using acados and deployed on a dSPACE SCALEXIO embedded system to meet real-time requirements.