Self-Attention Transformer for Remaining Useful Life Prediction in Lithium-Ion Batteries

Precisely forecasting the efficiency of lithium batteries is crucial for addressing consumer worries regarding the safety and dependability of electric vehicles. However, existing research primarily focuses on a single degradation characteristic of the battery, neglecting the multiple degradation features present in real-world operation. To more effectively utilize the various degradation information of batteries, we propose a time series prediction approach based on a Transformer to forecast the remaining useful life of the battery. The proposed Transformer architecture comprises a sampling encoding layer and a temporal variable block. Initially, the battery data undergoes time series processing; the sampling encoding layer extracts degradation features from the sensor recordings and incorporates positional information. Subsequently, in the temporal variable block, the time step encoding layer and the variable encoding layer work in parallel to extract dynamic temporal information and degradation features. hey focus on the information of different features within the feature vector, capturing the correlations between these features through a multi head self-attention mechanism, and determining the relative significance of each feature in forecasting the current time step. These elements are then integrated to produce a feature vector. Finally, the feature vector is passed to a decoder based on a multi-layer perceptron to compute the prediction outcomes. The experimental outcomes demonstrate that, on the Center for Advanced Life Cycle Engineering battery dataset, the proposed method surpasses existing approaches in forecasting the remaining useful life of batteries.