Energy management of electric vehicles based on improved long short term memory network and data-enabled predictive control

Abstract

As a popular energy management strategy (EMS) in electric vehicles with hybrid energy storage systems (HESS), model predictive control (MPC) is vulnerable to model accuracy and parameter sensitivity effects with existing parametric modeling methods. This paper proposes a novel EMS based on hierarchical data-driven predictive control. The upper layer utilizes an optimized long short-term memory (LSTM) network for trajectory prediction, enabling the acquisition of cost-effective load power demands for the lower layer. In the lower layer, a data-enabled predictive control (DeePC) is proposed for the HESS to achieve optimal power distribution between the battery and supercapacitor while minimizing battery capacity loss. Unlike conventional MPC, DeePC is based on a non-parametric model built solely from input–output data of the HESS, enabling agile handling of diverse nonlinearities and uncertainties across different tasks and environments. Comparison with nonlinear model predictive control shows that DeePC reduces the total operating cost by 22.68%, with optimization results closer to offline dynamic programming results. Furthermore, the effectiveness of the proposed DeePC method is validated through hardware-in-the-loop (HIL).