Signal-to-Noise Ratio Analysis for a Reinforcement Learning Based State of Charge Estimation of a Hybrid Supercapacitor

An accurate study on the influence of frequency and Signal-to-Noise Ratio (SNR) on the State-of-Charge (SoC) estimation for a cylindrical hybrid supercapacitor is presented in this paper based on Electrochemical Impedance Spectroscopy (EIS) data. A data driven approach based on a reinforcement learning model is applied, incorporating a Twin Deep Delayed Deterministic (TD3) neural network-based policy gradient agent. This model is followed by a residual error correction phase which incorporates an ensemble Least Square Boost (LSB) algorithm, followed by a linear three-point interpolation phase. Five EIS frequency points and five different percentages of multiplicative gaussian noise have been taken into account. Magnitude and phase of the internal impedance are selected as the input features to train and test the model using only one frequency point at a time, in order to minimize the timings related to the measurement procedure. The best results were achieved with the 0.1 Hz frequency, which features a Root Mean Squared Error (RMSE) of 0.42%, a very promising result when compared with literature. Moreover, the proposed algorithm also showed a good tolerance to the degradation of the SNR when compared with the existing literature.