A Hybrid GRU-MHA model for accurate battery RUL forecasting with feature selection

Abstract

The State of Health (SoH) and Remaining Useful Lifetime (RUL) forecasting of batteries is an important part of the health prognostics of electronic devices. This research presents a combination of Deep Learning (DL) and Machine Learning (ML) models to forecast the RUL battery’s linear and nonlinear behavior. The proposed method is separated into two phases. The initial phase separates the features into correlated and uncorrelated features, and the second phase determines the architecture of the proposed methods based on the two sets of correlated and uncorrelated features. The proposed method combines Gated Recurrent Units (GRU) and Multi Head Attention (MHA) with dense layers as the memory units. Finally, ridge regression is used for final forecasting. NMC-LCO 18650 and NASA lithium batteries datasets are used to evaluate the proposed method. Experimental results indicate 0.002 MAE, 0.044 MSE, and 99.99 R2 score for RUL forecasting of NMC-LCO 18650. The evaluation results on NASAB0005 indicated 0.005 MAE, 0.0706 MSE, and 99.64 R2 score for capacity estimation. The proposed method is compared with similar research and ML models such as Support Vector Regression (SVR) and DL models like Transformers and outperformed them. This research indicates superior performance in predicting declining trends of battery capacity using GRU and MHA with feature selection.