Comparative study of ANN/KF for on-board SOC estimation for vehicular applications

Abstract

Abstract Nowadays, bidirectional power transfer is a necessity for Vehicle-to-Grid technology. Making accurate and fast SOC estimation more crucial than ever by battery management systems. Striving to establish the optimal strategy for real-time SOC estimation on-board of an electric vehicle for the newly established V2G technology, this paper presents a detailed and critical comparative study of the unscented Kalman filter and artificial neural network. Unlike most conventional comparative studies in literature, this paper presents a critical view of the design and implementation processes of each strategy. All the design aspects are addressed and critically compared including predesign and design requirements. Their accuracy and sensitivity to the erroneous initial SOC values, alongside their tolerance to unpredicted operational conditions, are investigated using different load scenarios typically encountered by the V2G environment. Furthermore, a hybrid nonlinear least square algorithm is presented for battery internal parameter identification, needed for UKF. The experimental results indicate that UKF presents the best overall performance and compromise between accuracy, robustness, computational power and required resources; 91% fewer data were required than the ANN, robustness against erroneous initial SOC for up to 100% deviation, and more accurate under fast changing current and unpredictable conditions; 1.5% compared to 3%.