State of charge estimation of lithium-ion batteries based on a combination of Convolutional Neural Networks and Temporal Kolmogorov–Arnold Networks

Abstract

Traditional methods cannot effectively address the nonlinear challenges in State of Charge (SOC) estimation of lithium-ion batteries for electric vehicle(EV). The emerging Kolmogorov–Arnold Networks (KAN) have demonstrated strong performance in handling nonlinearity and time series problems. However, the learnable spline functions in the KAN and the increased parameters contribute to a slower training speed. This study proposes a new neural network configuration combining Convolutional Neural Networks (CNN) and Temporal Kolmogorov–Arnold Networks (TKAN) to solve these problems. This configuration not only enhances the estimation accuracy of the SOC for lithium-ion batteries but also accelerates the training speed. The proposed model outperforms CNN+(Gated Recurrent Units)GRU and CNN+(Long Short-Term Memory)LSTM with the same training strategy and hyperparameters across three metrics: RMSE, MAE and R${}^2$. Additionally, the model demonstrates good robustness when the initial SOC is not 100%, further demonstrating the potentiality of KAN for SOC estimation of lithium-ion batteries. The paper also discusses the impact of pooling layer configurations in the CNN on the performance of the model. Compared with a single TKAN, the proposed model not only enhances the training speed but also improves the accuracy. Furthermore, when comparing different pooling layer configurations, our model demonstrates greater suitability for practical SOC estimation of lithium-ion batteries.