An Improved CNN-LSTM Model Based on an Attention Mechanism for Estimating the State of Health of Lithium-ion Batteries

The nonlinear nature of battery degradation, and the complex operational environment, present significant challenges for traditional SOH estimation methods. In this study, we integrate a one-dimensional convolutional neural network, a multihead self-attention mechanism, and a long short-term memory network into a model designed for estimating the SOH of LIBs, hereinafter referred to as CNN-ALSTM. The model inputs were chosen with practical application in mind; discharge voltages and charging currents were utilized. Incorporating the multihead self-attention mechanism significantly enhances the depth of analysis concerning time series data, enabling it to focus more closely on critical time points that influence outcome prediction and capture long-range dependencies. It can process different sequence segments in parallel, improving feature representational power and prediction precision. The model demonstrates a marked precision in tracking the variations throughout battery charging and discharging cycles. The calculated results confirm the accuracy and robustness of the proposed approach. Specifically, under the scenario involving Cell3 from the Oxford battery degradation dataset, the CNN-ALSTM model achieved the best estimation results, with a root mean square error and mean absolute error of 0.121% and 0.095%, respectively.