State of charge estimation method for lithium-ion battery based on informer model combining time domain and frequency domain attention

Accurate state-of-charge (SOC) estimation under various working conditions and ambient temperatures is crucial for ensuring the safety of the battery management system. To overcome this challenge, a multi-model framework based on combining time and frequency domain attention with the sparrow search algorithm optimization Informer (TAFDA-SSA-Informer) is proposed, and better SOC estimation results are obtained from the following four aspects. Firstly, based on the relevant characteristics of lithium-ion batteries (voltage, current, temperature, discharge energy, charge energy, discharge capacity, and charge energy), the optimal input features of the battery (discharge energy and discharge capacity) are selected by using the random forest (RF) method. Secondly, a novel SOC estimation model based on TAFDA and the Informer model (TAFDA-Informer) is proposed in this paper. The proposed model captures the correspondence between input features and lithium-ion battery SOC over a long period in the time domain and enhances the continuity of high-frequency and low-frequency signals in the frequency domain, which enhances the multi-level attention of the Informer model to input features. Thirdly, to overcome the disadvantage of the Informer model being prone to local parameter optima, the sparrow search algorithm (SSA) is applied to optimize the hyperparameters of the proposed model, which improves the estimation performance of the proposed model. Finally, the effectiveness of the proposed model is validated by training with small samples under different temperatures and working conditions. The experimental results show that under the set optimal parameters, high accuracy of SOC estimation for lithium-ion batteries is achieved. Specifically, the mean average error (MAE) and root mean square error (RMSE) are controlled within 0.2 % and 0.18 %, respectively, and the calculation burden of the proposed model takes 21.7 s. Compared with other models, the average accuracy of the proposed SOC estimation method is enhanced by 24 %.