Accurately predicting the long-term Remaining Useful Life (RUL) of lithium batteries is critical for ensuring the reliability of lithium battery-powerd systems.This study constructs an LSTM-KAN prediction model by integrating the sequential data processing capability of Long Short-Term Memory (LSTM) network with the superior nonlinear fitting ability of Kolmogorov-Arnold Network (KAN).To enhance model performance,the Harris Hawk Optimization (HHO) algorithm is employed to determine its optimal hyperparameters.Using datasets from the University of Maryland (batteries CS2_35 and CS2_36) and NASA (battery B0007),features correlated with battery capacity are first extracted from operational data such as voltage,current,and temperature.Features with a Spearman correlation coefficient greater than 0.9 are then selected as model inputs to reduce data complexity.Subsequently,the HHO algorithm optimizes the hyperparameters of the LSTM-KAN model,and the processed data are fed into the optimized model for RUL prediction.Experimental results demonstrate that the proposed HHO-LSTM-KAN model effectively predicts the long-term degradation trend of lithium batteries.Its performance,in terms of mean squared error and the required number of prediction samples,is superior to that of other benchmark battery life prediction models.