Optimal Energy Management of Parallel Electric-Hydraulic Hybrid Vehicles based on Dynamic Programming and Recurrent Neural Networks

This paper presents an approach for optimal energy management in parallel electric-hydraulic hybrid vehicles using recurrent neural networks trained through dynamic programming results. While electric vehicles are increasingly recognized for their sustainability and zero tailpipe emissions, their implementation faces critical challenges, especially in heavy-duty applications. The low power density of lithium-ion batteries limits their ability to effectively capture regenerative braking energy, particularly during short braking periods. In addition, high peak current loads due to heavy torque demands can accelerate battery aging. To address these challenges, a hybrid vehicle that combines electric and hydraulic systems has been proposed in the literature. This vehicle integrates a hydraulic pump/motor and a hydro-pneumatic accumulator with an electric powertrain, aiming to achieve both high power density and energy density while mitigating peak loads on the battery. Like other hybrid vehicles, this vehicle requires optimal management of its two energy sources, i.e. electric energy and hydraulic energy, for given driving cycles. Dynamic programming is effective for optimizing control strategies in hybrid vehicles by evaluating all possible control actions to find a globally optimal solution. However, its high computational complexity makes it impractical for online implementation. To bridge the gap, this work explores the use of recurrent neural networks, trained on offline dynamic programming solutions, for online applications. Recurrent neutral networks, with their ability to use past outputs as inputs, effectively leverage past data to determine the current control action, making them well-suited for controls of dynamic systems. To demonstrate the effectiveness of the proposed online energy controller, it was implemented in a high-fidelity simulation model and validated using a practical heavy-duty driving cycle. Simulation results show that the recurrent neutral network-based controller achieves comparable performance to offline dynamic programming in terms of energy savings and battery stress mitigation.