State of charge estimation and error analysis of lithium-ion batteries for electric vehicles using Kalman filter and deep neural network

Abstract

The lithium-ion battery has a great significance in meeting the growing demand for Electric Vehicles (EVs) due to its higher energy density, longer life cycle, and notable nominal voltage and capacity. One crucial parameter for lithium-ion batteries is the State of Charge (SOC), which represents the available capacity and ensures that the system operates in a secure and reliable mode for EVs. SOC plays a significant role in the Battery Management System (BMS). This research aims to propose an Equivalent Circuit Model (ECM) based on Kalman filtering method and a data-driven technique, Deep Feed-Forward Neural Network (DFNN), for accurate SOC estimation of Electric Vehicle Battery (EVB). Initially, lithium-ion battery parameters are identified using a second-order RC (2-RC) Equivalent Circuit Model (ECM). Subsequently, battery modeling is performed, and various operating conditions such as terminal voltage, load current, and temperature are measured to obtain initial values for the filtering method used in SOC estimation. These operating conditions are crucial for ensuring safe and efficient charging and discharging of lithium-ion batteries. Based on the identified ECM parameters, SOC estimation error and bound error are then minimized using Kalman Filter (KF), Extended Kalman Filter (EKF), and Unscented Kalman Filter (UKF) techniques. These filtering methods are employed to accurately estimate the SOC of the battery. The results demonstrate that the proposed model based on KF and EKF algorithms estimates SOC bound error within 2.5 % - –2 % and estimation error <1.5 % - –0.7 %. On the other hand, the UKF estimates a SOC bound error of 1.5 % and an estimation error of 0.5 %, proving the algorithm's efficiency and reliability. Particularly, this estimation error rejects measurement noise and parametric uncertainties for lithium-ion batteries to drive EVs with efficacy using UKF. Hence, the UKF algorithm estimated SOC has low estimation error, ensuring more accurate results. Finally, data-driven, DFNN method is implemented for accuracy enhancement of SOC estimation with trained 20 iterations and epochs data. Using this method, the SOC estimation accuracy is satisfactory with only 0.04 % Root Mean Squared Error (RMSE). The validation results indicate that the model-based filtering method is an effective method for SOC estimation to be applicable. In contrast, in a novel data-driven approach, SOC estimation accuracy is improved by approximately 0.46 %.