A Lightweight and Term-Arbitrary Memory Network for Remaining Useful Life Prediction of Li-Ion Battery

Accurately predicting the remaining useful life (RUL) of power battery packs is crucial for enhancing the lifespan and safety of electric vehicles (EVs). Traditional neural network (NN) methods often face challenges in modeling long-term and complex time sequence. While recent transformer-based methods tackle this issue by self-attention mechanism, it inevitably introduces high model complexity and requires large amounts of training data. To combine the high efficiency of traditional NNs with the temporal modeling advantages of transformers, we propose a novel RUL prediction method based on a lightweight and term-arbitrary memory network (LTM-Net). The core design is to incorporate a linear dot product to replace highly computational softmax approach in self-attention block. Moreover, a selective forgetting module is incorporated to filter input temporal information based on their importance in a learned manner. As a result, our proposed LTM-Net could efficiently model time sequence data with arbitrary length. Extensive experiments are conducted using real EVs’ operation datasets and battery charge/discharge datasets. The RUL prediction average absolute error (AE) for four batteries is less than 5.5, and the average mean absolute error (MAE) for four vehicles is less than 0.5%, outperforming existing NN-series and transformer-series methods. Note that regardless of the window size (e.g., term), our LTM-Net exhibits consistent performance and efficiency advantages.