A Hybrid Multi-Head Attention and Harris Hawks Optimized Long Short-Term Memory Model for High-Precision Lithium-Ion Battery Remaining Useful Life Prediction

Lithium-ion battery Remaining Useful Life (RUL) prediction with high accuracy is essential to guarantee the reliability, safety, and economic efficiency of energy storage systems (EESs). However, lithium-ion battery (LIB) degradation presents strong nonlinearity and dependency on time, which greatly affects the application of traditional data-driven prediction models, such as conventional Machine Learning and single-feature Deep Learning (DL) approaches. To meet this challenge, the present study introduces a novel hybrid DL approach named Multi-Head Self-Attention Long Short-Term Memory with Harris Hawks Optimization (MHA_LSTM_HHO) for high-precision lithium-ion battery RUL prediction. The proposed framework combines the MHA mechanism and an LSTM network to learn from long-range temporal dependencies and degradation-sensitive features in a battery charging and discharging pattern. Moreover, the HHO algorithm is used to automatically optimize the critical hyperparameters of the DL model, which enhances the stability of convergence and robustness of prediction. The proposed model is tested for the benchmark HNEI NMC-LCO 18650 lithium-ion battery degradation dataset, and tested with various battery aging conditions. The experimental results show that the proposed MHA_LSTM_HHO model achieves considerable performance improvement compared to the traditional LSTM model, the attention-based LSTM model, and other optimization-assisted models. The proposed model gives the best results with a minimum MAE of 0.8265 and an RMSE of 1.3924, while the conventional LSTM model yields an MAE of 1.4621 and an RMSE of 4.1854, and an R² score of 0.9997. Furthermore, the proposed model provides a reduction of 62.1% in RMSE and 25.5% in RMSE when compared with the standard LSTM model and 34.2% when compared with the SHA_LSTM_HHO model, respectively. The main innovation of this research work lies in the fact that multi-head attention-based temporal feature learning is effectively integrated with HHO-based hyperparameter optimization for predicting battery prognostics with higher accuracy, robustness, and generalizability than existing RUL prediction techniques.