Design and development of a model-based predictive control system for automotive thermal management

Abstract

In automotive thermal management systems, an adaptive cooling operation is required as the engine’s heat rejection is constantly changing with the vehicle dynamics. Maintaining an optimal adaptive engine cooling can be achieved by delivering coolant flow at certain temperature states, as close as possible to actual needs. Currently, for the coolant temperature regulation, cooling fan control systems widely employ Proportional Integral Derivative (PID) controllers, as fast and light solutions in generating fan speed demands for the heat rejection process. Although interesting in terms of in-vehicle computational processing cost, the induction of disturbances from parameters not precisely estimated, in the system modelling, aggravate the low robustness in the regulation of the coolant temperature in the presence of uncertain thermal impacts. Aiming at improving the reduction of disturbances, a Model Predictive Control (MPC) strategy is proposed, as a method, algorithm, and strategy, applied on cooling fan control systems, for the generation of optimized fan speed demands in maintaining a predicted horizon of coolant temperatures at a set point configuration. Improvements on vehicle performance, fuel efficiency and emissions are potentially achieved with the application of machine learning strategies for the prediction and thermal optimization of the coolant temperature in a future control horizon, allowing a proposed Reinforcement Learning (RL) labeling model to perform searches for optimal fan speeds. The probabilistic strategy of the RL agent is improved in interacting and observing the coolant temperature response, from a thermal response model, with confidence from crosstime correlations with thermal impact variables, resulting in less deviance from configurable temperature set points when compared to classic feedback controllers. In addition, a human interpretable feature extraction process is proposed, using the Toeplitz Inverse Covariance-Based Clustering (TICC) method, in extracting accurate and interpretable structures in multivariate time series data, for addressing processing time concerns with the use of reliable and low dimensional feature representations. The results of an experimental physical evaluation demonstrate the effectiveness of the MPC solution in comparison to a classic controller, as it achieves the potential reductions of 1.53% and 0.61% in the consumption of fan power and fuel, respectively.