Speaker
Description
Abstract – A flexible reward-based reinforcement learning (RL) approach for the state of charge (SoC) estimation of energy storage systems (ESSs) i.e. a Lithium-ion battery (LiB) and a hybrid supercapacitor (HSC), is presented in this research. The data acquisition for the Panasonic 18650 LiB is carried out through constant current – constant voltage (CC-CV) procedure whereas, for the cylindrical HSC electrochemical impedance spectroscopy (EIS) is used. ESS. A three-stage approach is presented i.e. reinforcement learning (RL) training, least square boosting (LSB) stage and three-point linear interpolation. RL incorporates a twin-delayed deep deterministic (TD3) neural network-based policy gradient agent along with a reward function specifically designed to take into account a flexible use in possible different contexts. The RMSE, along with other performance metric, is reported and analysed for each discharge cycle in the test datasets. The overall minimum average RMSE per cycle for the HSC and LiB are reported to be 0.33% and 0.87% whereas the overall average RMSEs are 0.71% and 1.10%, respectively. The ESSs results are progressive in comparison with the existing literature and validate the efficacy of the multi-term flexible reward function approach.