SOC Estimation of Lithium-ion Battery Based on Attention Mechanism of EMD-Bi-LSTM Improving by Bayesian Optimization

Data-driven methods are widely applied to predict state-of-charge (SOC) of lithium-ion batteries, as they do not require the detailed knowledge in terms of complex processes occurring inside the battery, but merely learn the underlying nonlinear link between SOC and outside measurements. However, the method has two limitations: (1) Feature information is not fully utilized; (2) Model parameters are numerous and difficult to determine; To address these challenges, this paper proposes a new deep learning approach using feature enhancement and adaptive optimization. Firstly, the empirical mode decomposition (EMD) is used to predict the features of the present situation, eliminating noise points of the data and improving the dimension of the input data to extract the effective decomposition characteristics. Secondly, the feature enhancement technology for box-cox is used to enhance the correlation between features and SOC. With these methods, the potential of features can be made full use of. Then, a two-way long-short-term memory (Bi-LSTM) network combined with dual-stage attention mechanism is used to establish the SOC prediction model. The input attention method independently chooses valuable characteristics as input at each time step in the first stage. In the second stage, the temporal attention mechanism fully considers the correlation of time series and selects the hidden layer state to accurately estimate SOC. Finally, the model's parameters are improved via Bayesian optimization (BO). The experimental findings demonstrate that the model's accuracy is superior than that of other prediction techniques. The average absolute error can be less than 0.50%, and the prediction accuracy can reach 99.91%.