An improved Transformer incorporating fuzzy information entropy and average input strategy for SOC estimation of lithium-ion battery

Accurate state of charge (SOC) estimation is crucial for the prolonging lifetime and ensuring safety of lithium-ion batteries. The nonlinear and time-varying characteristics of lithium battery datasets make accurate SOC estimation challenging for traditional neural networks. To improve the performance of SOC estimation, a novel Transformer integrating bidirectional long short-term memory (Transformer-BiLSTM) is proposed. The global relationships between data are learnt through the self-attention mechanism of Transformer, and then local dependencies based on battery characteristics are captured from the global pattern using BiLSTM. The proposed Transformer-BiLSTM can efficiently process long sequence data at the same time accurately understand the nonlinear relationship between features and SOC. During preprocessing, fuzzy information entropy (FIE) is used to comprehensively evaluate the impact of each feature subset on SOC. Slow time-varying information is introduced through a sliding window-based averaging input strategy, which suppresses SOC output fluctuations. Moreover, transfer learning is applied to enhance the generalization ability of model. The experimental results show that the proposed algorithm is able to accurately estimate SOC under various operating conditions and a wide temperature range, with accuracy consistently maintained within 1.2 % and the root mean square errors (RMSE) all below 0.35 %, significantly improving SOC estimation performance.