State-of-charge estimation of lithium-ion battery: A fusion model combining unidirectional and bidirectional long short-term memory networks and multi-head attention mechanism

Conventional battery state of charge (SOC) estimation methodologies frequently depend on isolated point-in-time measurements, thereby struggling to accurately capture the highly non-linear dynamic behaviors and long-term temporal dependencies inherent in lithium-ion batteries. To overcome these limitations, this paper proposes an advanced hybrid estimation framework, termed LBM, which architecturally integrates Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), and a Multi-Head Attention (MHA) mechanism. The core innovation of the proposed LBM architecture lies in its hierarchical spatiotemporal feature extraction mechanism. Furthermore, a targeted, physics-informed feature engineering strategy is introduced, incorporating multi-time-step sliding windows and average voltage parameters to significantly enhance the model's capacity to map underlying electrochemical and thermodynamic states. To rigorously validate the methodological advantages, extensive comparative evaluations were conducted against six benchmark models (CNN, RNN, GRU, LSTM, LSTM-BiLSTM, and LSTM-MHA). The results demonstrate that the LBM framework consistently achieves superior estimation accuracy and systemic robustness. Quantitatively, the proposed framework yields an average Mean Absolute Error (MAE) of 1.02% (ranging strictly from 0.64% to 1.34%) despite diverse operating temperatures, variable time-step configurations, and induced noise conditions. This sustained high precision validates the LBM model as a highly reliable SOC estimation solution for complex real-world engineering applications, including electric vehicles.