An approach of state-of-charge prediction for lithium-ion batteries using hybrid Pyraformer-LSTM learning network

Abstract

Accurate prediction of the State of Charge (SOC) of lithium-ion batteries remains difficult under complex operating conditions. This difficulty arises from strong nonlinearity and long-term temporal dependency in battery data. In this study, we developed a hybrid prediction model that combines a pyramidal attention network with a Long Short-Term Memory (LSTM) network. The key innovation of this work lies in the integration of a pyramidal attention mechanism with a multi-scale sliding window design, enabling hierarchical extraction of temporal features ranging from local dynamic variations to long-term evolutionary trends. Specifically, the pyramidal attention module captures multi-granularity temporal dependencies through dynamic weighted fusion of features from different window scales (8, 16, and 32), while the LSTM network captures nonlinear dynamics and preserves long-term dependencies through gated temporal modeling. We evaluated the proposed method using battery datasets collected under multiple temperatures. We compared its performance with six representative SOC estimation methods. The experimental results show that the proposed model achieves RMSE and MAE of 1.8554% and 1.3021% on the test set, significantly outperforming baselines, demonstrating high prediction accuracy and validating the effectiveness of multi-scale attention and recurrent structures in SOC prediction.