A hybrid approach to predict battery health combined with attention-based transformer and online correction

Abstract

The lithium-ion battery is an important part of green energy systems, and battery aging will lead to the degraded performance of energy storage systems (ESSs). Therefore, accurate battery health prediction is crucial to guarantee the safe and efficient operation of ESSs. This paper proposes a hybrid battery health prediction method that fuses Transformer and online correction. First, the attention-based Transformer is taken as a global model to establish the nonlinear relationship between measured data and battery capacity decline. Second, a local model based on unscented particle filter is developed for the online correction of Transformer outputs. To characterize battery degradation behavior as much as possible, multi-scale health features are considered, including time-series and statistical features extracted from partial charging curves and operating data distributions, respectively. Then, feature dimension reduction is performed based on the maximum information coefficient method. To handle both types of features, a special filter layer is carefully designed in Transformer. Compared with the state-of-the-art algorithms, the proposed method achieves optimal health prediction performance with minimal computational resources for batteries with different aging conditions. When only 20% of the cell data is used for training, the predicted root mean square error can still be guaranteed to be within 0.72%.