Optimal energy management for multistack fuel cell vehicles based on hybrid quantum reinforcement learning

Abstract

This article proposes a driving condition recognition (DCR)-based hybrid quantum deep deterministic policy gradient (HQDDPG) method for energy management in multistack fuel cell vehicle hybrid power systems (MFCV HPSs) and its quantum simulation setup on digital signal processors (DSPs). Driving conditions are initially segmented into microtrips and clustered into three types. The DCR method, using a learning vector quantization neural network (LVQNN), is then developed, thus accurately and efficiently identifying driving condition types. Subsequently, quantum reinforcement learning (RL) is proposed to achieve optimal energy management of MFCV HPSs, i.e., power allocation among the multiple fuel cells to minimize the economic metrics based on the DCR results. Compared to classical large-scale neural networks, quantum RL reduces parameters by combining a parameterized quantum circuit (PQC) with a single-layer classical neural network. The PQC encodes and processes state information through quantum mechanics for enhanced computational expressiveness, while the classical neural network transforms the quantum measurement expectations into actionable outputs for energy management. The trained hybrid quantum circuits are implemented on DSPs through quantum simulations. The method is validated through controller hardware-in-the-loop (CHIL) experiments, demonstrating superior performance in optimizing economic metrics compared to conventional methods.