DescriptionAccurate State of Charge (SOC) estimation is very important for safe and reliable use of lithium-ion battery, which is widely installed as a new energy storage device in electrical vehicles. It is one of parameters in Battery Management System (BMS) to ensure good working condition of battery in case over-charging or over-discharging problems. There are a lot of methods used in SOC estimation such as open circuit voltage, coulomb counting and Kalman filter. Kalman filter is very popular method used recent years, but the accuracy needs to be improved a lot.
Researchers start to work on predicting SOC based on neural networks. In this thesis, three different types of deep neural networks are used to estimate SOC of lithium-ion battery. they are Artificial Neural Network (ANN), Nonlinear Autoregressive Neural Network (NARX) and Recurrent Neural Network (RNN). Performances of accuracy are tested among three networks to identify ability of predicting SOC.
RNN performs best by measuring Mean Square Error (MAE) 1.21 of dataset UDDS, 3.43 of dataset NYCC and 1.51 of dataset UN. NARX has the lowest performance in estimation, but it is visual friendly in plotting trend of battery discharging process. Performance of ANN is fundamental compared with accuracy of RNN and display ability of NARX.