State of Charge Estimation of Lithium-Ion Batteries: Comparison of GRU, LSTM, and Temporal Convolutional Deep Neural Networks

Accurate state of charge (SOC) estimation of lithium-ion batteries (LIBs) is essential to ensure safe and reliable battery operation. A simple way to obtain the SOC is by integrating the charge withdrawn or inserted into the battery. However, this coulomb counting method is often inaccurate due to accumulated error associated with current sensor offset or bias error. Hence, estimators are often used to report the SOC values by monitoring the battery's measured parameters. In this study, three deep neural network (DNN) -based SOC estimators are benchmarked, including a gated recurrent unit (GRU), long short-term memory layer (LSTM), and temporal convolutional neural network (TCN). The networks are trained to estimate the SOC of a prismatic battery at five ambient temperatures ranging from -20 to 40°C. The results show that the optimum configuration of the three DNN types estimates SOC with less than 2% root mean square and 60% maximum error. The GRU shows slightly higher error and lower computational resources than the LSTM and TCN, which was most evident for challenging cases such as drive cycles at -20°C. The TCN is shown to require around 250,000 learnable parameters and thus seven times higher execution time to achieve similar accuracy as an LSTM or GRU RNN with just 3,000 learnable parameters.