A Kolmogorov-Arnold neural networks approach to state of charge estimation and confidence assessment for Li-ion batteries

This study aims to thoroughly examine the potential of the Kolmogorov-Arnold Network (KAN) and its application to improving energy management efficiency, particularly in lithium-ion batteries. The study employs a novel method that utilizes one-dimensional adaptive activation functions parameterized by splines, in contrast to traditional neural networks, where activation functions are fixed. Traditional methods for activation function selection are based on empirical approaches and do not guarantee accurate approximation, potentially leading to suboptimal results. This approach enables the KAN to flexibly adapt to complex data structures, ensuring precise state-of-charge estimation. To objectively evaluate the algorithm's effectiveness, experiments were conducted on real datasets, focusing on analyzing the accuracy of state-of-charge estimation at confidence intervals of 95%, 90%, and 85%. The test results for various charge-discharge cycles demonstrated that the proposed method achieves high accuracy and maintains stability throughout the operation. The proposed method reduces the maximum error by at least 4.26% and significantly improves key performance metrics such as Mean Absolute Error, Root Mean Square Error. Thus, the obtained results confirm the efficiency and innovative nature of the KAN in energy management. This method holds great potential for energy management and can be effectively implemented in areas requiring precise time-series forecasting, including smart home systems, electric vehicles, and industrial devices. Future research will optimize the network architecture and expand its practical applications. Signifi-cantly, this method can be flexibly adapted to different types of batteries and energy systems, broadening its applicability in real-world conditions.