Hybrid Kolmogorov-Arnold and xLSTM network for enhanced RUL prediction of lithium batteries

Accurate prediction of the remaining useful life (RUL) of lithium batteries is essential for ensuring efficient equipment maintenance and energy management, particularly as these batteries serve as a core driver in the new energy technology revolution. While deep learning models such as Convolutional Neural Networks (CNNs), Long Short-Term Memory Networks (LSTMs), and their variants have demonstrated significant success in RUL prediction, they often face challenges related to inadequate modelling of long-term dependencies in complex degradation data. To overcome these limitations, this paper proposes a novel hybrid architecture that integrates the Kolmogorov-Arnold Network (KAN) with an Extended Long Short-Term Memory Network (xLSTM). The KAN component enhances high-dimensional function approximation and improves parameter efficiency by substituting traditional linear weights with B-spline-parameterized univariate functions. Meanwhile, the xLSTM introduces exponential gating mechanisms and covariance update rules to more effectively capture high-order long-term dependencies. Experimental results on the NASA lithium battery aging dataset demonstrate that the proposed KAN-xLSTM model significantly outperforms CNN, LSTM, and xLSTM models in prediction accuracy, particularly for battery groups with large capacity fluctuations.