State of Charge Estimation in Lithium-Sulfur Cells Using LSTM Recurrent Neural Networks

This paper presents a framework for all-state estimation of Lithium-Sulfur (Li-S) battery cells based on a Long Short-Term Memory Recurrent Neural Network (LSTM RNN) model. Under the proposed framework, the LSTM RNN model is calibrated into the single task of State of Charge (SoC) estimation for fresh Li-S prototype cells. The Adaptive Moment Estimation (Adam) solver is used. Data sets for training and testing are derived from experiments using the WLTP duty cycles. The calibrated LSTM RNN structure is described for the purposes of training and testing with experimental datasets, so as to generate a network that can be deployed in real-time system. The demonstration of the training and testing results has shown robustness of the proposed approach against nonlinearities of the experimental datasets and uncertainty in initial SoC. The approach gave satisfactory estimation performance with an acceptable tradeoff between estimation accuracy and convergence speed.