Machine learning-based energy management and bidirectional power control for battery-fed electric vehicle traction drive

Abstract

Abstract Efficient energy management in a battery-fed electric vehicle (EV) traction system requires coordinated control of propulsion power, battery operating conditions, Direct Current (DC)-link voltage, and regenerative braking under rapidly changing driving conditions. This study develops a machine learning (ML)-based energy-management and bidirectional power-control framework integrating a lithium-ion battery, bidirectional direct current–direct current (DC–DC) converter, DC link, inverter-fed Permanent Magnet Synchronous Motor (PMSM), battery-state estimation, and regenerative-braking control. The ML controller receives motor torque demand, motor speed, State of Charge (SoC), State of Health (SoH), DC-link voltage, battery current, battery temperature, and road-gradient information and generates duty-ratio correction, motor-torque reference, and regenerative-power reference. The distinguishing feature of the framework is that a single supervisory network produces converter-level, traction-level and battery-level references simultaneously from one multivariable input vector, whereas reported ML, deep-reinforcement-learning and neural-network energy managers generate either a power-split command or a state estimate, and delegate converter regulation and torque limiting to separate loops. A multilayer feedforward neural network with three hidden layers is trained using 10,000 operating samples generated by Latin-hypercube sampling of the traction, cruising, braking and battery operating envelope; the target labels are the constrained-optimal duty-ratio correction, torque reference and regenerative-power reference obtained offline from a single-step constrained optimisation of the same multi-objective cost used for training. A second, separately trained network performs SoC and SoH estimation and is described independently of the energy-management controller. Simulation studies are conducted for a 1200-kg EV equipped with a 50-kW PMSM and a 352-V, 50-Ah battery pack. The battery–converter–motor power balance is verified explicitly: at the peak traction instant of the 180-s cycle the pack delivers 37.37 kW at 344.5 V, corresponding to 108.5 A (2.17 C), and at the strongest braking instant it absorbs 15.30 kW at 354.4 V, corresponding to − 43.2 A (0.86 C). Comparative assessment against conventional Proportional–Integral (PI), rule-based, fuzzy-logic, and Model Predictive Control (MPC) methods, tuned and tested under identical conditions, demonstrates reductions in speed Root Mean Square Error (RMSE), torque RMSE, DC-link RMSE, and SoC RMSE to 0.54 km/h, 0.96 N m, 0.86 V, and 0.082% SoC, respectively, from 1.84 km/h, 2.76 N m, 3.84 V and 0.438% SoC for PI control, corresponding to reductions of 70.7%, 65.2%, 77.6% and 81.3%. The regenerative energy-conversion efficiency of the electrically braked energy, defined explicitly in (41), reaches 87.91% against a hardware ceiling of 87.98% set by the product of the motor, inverter and converter efficiencies, and the overall braking-energy recovery, which additionally accounts for the friction-braked share, reaches 82.90% compared with 51.80% for PI control. The trained controller requires 1531 parameters and 1456 multiply–accumulate operations per control step, corresponding to approximately 25 µs on a 200-MHz fixed-point digital signal controller and 0.25% processor utilisation at the 10-ms control interval. The ageing relation used in this work is a mathematical simulation model and has not been calibrated against cell test data; its role is limited to providing a slowly varying SoH input to the supervisory controller.