Performance Comparison of NARX & RNN-LSTM Neural Networks for LiFePO4 Battery State of Charge Estimation

The purpose of this research is to evaluate the performance of NARX network and an improved RNN with LSTM artificial neural networks to estimate the SOC of LiFePO4 batteries used in EVs, HEVs, UAVs, UUVs and energy storage systems for power electronics and solar power systems. Due to unequal self-discharge rates, difference in current leakage and temperature under repeated charge discharge cycles of the battery pack, made by series and parallel combinations of different cells to increase the power rating for different applications, the SOC of each battery cell mismatched with respect to others. Therefore, an accurate estimation of SOC is required for optimal power use of Li-Ion battery pack. In this research two neural network models for SOC estimation of LiFePO4 batteries are designed and their performance is evaluated on different parameters. These models are trained by optimal selection of hyper parameters using real time data of voltage, current and battery pack temperature as input and SOC as output of the models. The input dataset has been preprocessed as multivariate time series forecasting problem for RNN-LSTM neural network. The tested results showed that accurate results with an RMSE lower than 0.001 is achieved in NARX and lower than 0.002 is achieved in RNN-LSTM at different test datasets. However, RNN-LSTM has better performance than NARX in multivariate time series in seq2seq future forecasting configuration with an RMSE lower than 0.003 at different test datasets with a temperature range of 5 to 25°C.