Online Remaining Useful Life Prediction of Lithium-ion Batteries Based on Hybrid Model

A hybrid model based on black-winged kite algorithm and dual-attention mechanism optimized temporal convolutional network (TCN) with simple recurrent unit (SRU) is proposed to improve the accuracy of online remaining-useful-life (RUL) prediction for Li-ion batteries (LIBs). Health indicators (HIs) correlated with battery capacity are extracted from calculated variables and verified with Spearman correlation coefficient and the model is constructed, applying TCN with multi-head self-attention mechanism to capture correlation in spatial dimension and the decay pattern in temporal dimension of HIs, introducing temporal pattern attention to improve the ability of SRU to capture timing patterns in input sequences as well as applying BKA to further optimize hyper-parameters, enhancing model performance. Experimental data are used to validate the proposed hybrid model’s predictive performance for LIBs at different usage levels under complex conditions such as capacity regeneration, sharp fluctuations, and plunges. The prediction results of the hybrid model achieve the MAE of less than 3.66%, MAPE below 2.02%, RMSE not exceeding 5.03%, and R 2 greater than 0.96, with the absolute error in RUL prediction not exceeding 5. The experimental results demonstrate that the proposed hybrid model can achieve accurate RUL online prediction and perform good robustness.