A Novel SoC Estimation Approach Using Kolmogorov-Arnold Networks with Shapley-Based Interpretability

In lithium-ion battery management, accurately estimating the State of Charge (SoC) is crucial for ensuring the performance and reliability of energy storage systems. This study proposes an advanced method that combines the Kolmogorov-Arnold Network (KAN) with Shapley analysis to enhance SoC estimation accuracy and provide interpretability for the model's predictions. The KAN network is employed to capture complex nonlinear relationships between input parameters such as voltage, current, and temperature, while Shapley analysis, a cooperative game theory concept, quantifies the contribution of each parameter to the SoC estimation results. Experimental results demonstrate that the model performs optimally when using the Butterworth (BWF) filter, compared to methods without filtering or those using other filters such as the Moving Average (MA) filter or the Hampel (H) filter. Specifically, the model achieves $R^2=0.9996$, MAE = 0.40, and RMSE = 0.56. Through Shapley analysis, the study identifies the influence of input parameters on SoC estimation. The results indicate that the current processed with the Butterworth filter (Current_BWF [A]) has the most significant impact, with a Shapley value of 0.00332, followed by the raw current (Current [A]) at 0.00219 and the bottom temperature (TempBottom [°C]) at 0.00188. Other parameters contribute at lower levels but still play an important role in stabilizing the prediction model. This research offers several practical applications. First, achieving high-accuracy SoC estimation helps optimize battery usage, thereby extending the battery's lifespan. Second, using Shapley values to explain the model allows engineers and researchers to understand the system's operation better, providing a solid foundation for more effective design and operational decisions. Finally, this method can be easily integrated into modern Battery Management Systems (BMS), enhancing the overall efficiency and reliability of the system.