Energy Management Strategy of Fuel Cell Hybrid Electric Vehicle based on Dynamic Programming and Neural Networks

Increasing concern about climate change, air pollution and petroleum resource depletion has led regulators to impose more stringent standards in the automotive industry, which accounts for 22% of global CO2 emissions. Battery Electric Vehicles (BEVs) are the most popular among the different alternatives proposed, but Fuel Cell Hybrid Electric Vehicles (FCHEVs) are regaining attention after a setback during the last decade. HEVs have two or more power sources that propel the vehicle. Consequently, Energy Management Strategies (EMS) play a key role in the performance of such vehicles because they seek to optimize power split between those sources to minimize fuel consumption. Modern EMSs consider additional criteria, such as increasing lifecycle of components to minimize Well-to-Wheel (WTW) emissions, thus leading to multi-objective optimization problems. In this context, the present work aims at implementing a real time controller that enables to concurrently optimize fuel consumption, lifetime of the stack and energy utilization rate, while guaranteeing a Charge-Sustaining working mode of the vehicle. For this purpose, an offline analysis is first conducted applying Dynamic Programming (DP) to different drive cycles (among which WLTP and FTP75 are the most realistic ones), to obtain optimal power split policies by minimizing an adequate cost function while meeting constraints associated to the dynamics of the system. Specifically, DP is implemented in Matlab using the DPM function developed at ETH Zurich and a quasi-static model of the vehicle based on the architecture and components sizing of the Toyota Mirai 2021, the state of the art of FCHEVs. The vehicle is modelled through its load parameters and only considers the longitudinal dynamics. Powertrain components such as electric motor, battery and fuel cell stack are modelled using an efficiency map obtained through a data-driven approach. Then, the Matlab Deep Learning Toolbox is used to design and train an Artificial Neural Network (ANN) using the data obtained from the previous stage, to approximate the DP behavior and enable on-board implementation. For testing the effectiveness of the obtained NN controller, it was implemented on a more realistic HEV model developed in Simulink using Simscape libraries. Results show that the NN controller outperformed a simple PID in terms of overall cost (fuel consumption and battery electric power), computed as Gallon Equivalent. It was also observed a slightly different behavior between the higher fidelity model and the quasi-static one.