Battery state-of-charge estimation using data-driven Gaussian process Kalman filters

Abstract

This paper proposes methods of battery state-of-charge (SoC) estimation using Gaussian processes (GPs): GP-Unscented Kalman filter (GP-UKF) and GP-Particle filter (GP-PF). Based on Bayesian filtering frameworks, both GP-UKF and GP-PF use data-driven battery GP prediction and observation models trained via machine learning. In contrast to the classical GP-based UKF, our implementation of GP-UKF incorporates a straightforward technique of calculating sigma points of predicted voltage. In addition, the noise covariance matrices are automatically adjusted in the proposed GP-UKF using the GP prediction and observation models. In comparison with the existing GP-based PF, our implementation of GP-PF involves calculating the initial samples and weights of the previous SoC and estimating the SoC mean and variance from the samples and weights of the SoC posterior distribution. Furthermore, the likelihood function of the present voltage is automatically adjusted in the proposed GP-PF using the GP prediction and observation models. The SoC estimation performances of the proposed methods were evaluated via electric vehicle battery charging and discharging simulations according to the urban dynamometer driving schedule speed (UDDS) and highway fuel economy test (HWFET), in which the proposed GP-UKF achieved 60.13% and 35.17% increases in the estimation accuracy, and 88.53% and 79.92% decreases in the estimation uncertainty compared with a classical UKF, respectively. In addition, the proposed GP-PF achieved 47.72% and 58.97% increases in the estimation accuracy, and 99.27% and 97.16% decreases in the estimation uncertainty compared with a classical PF, respectively in the same UDDS and HWFET EV simulations.