Multi-dimensional features based data-driven state of charge estimation method for LiFePO4 batteries

The flat open-circuit voltage (OCV) curve of LiFePO4 (LFP) batteries poses a significant challenge to state of charge (SOC) estimation. To solve this problem, this paper proposes a data-driven SOC estimation method based on multi-dimensional features, especially incorporating force signals. The significant force variation at the middle SOC region section compensates for the flat OCV problem. A long short-term memory (LSTM) neural network model is established to estimate SOC. Battery voltage, current, temperature, and force data sampled only in 5 s are taken as input. The proposed method is validated under different dynamic testing profiles and different temperatures. Experimental results indicate that this method can highly improve SOC estimation accuracy in the middle SOC region, with less than 0.5% root mean square errors and less than 2.5% maximum errors. The validation results at different temperatures also maintain high accuracy with the same model, showing strong robustness and excellent generalization performance. Additionally, the model training process of this method only takes 1.5 h, and the online estimation time is less than 1 s, considerably reducing time cost.

Automated summary

This paper proposes a data-driven SOC estimation method based on multi-dimensional features, especially incorporating force signals, to solve the challenge of the flat OCV curve in LiFePO4 (LFP) batteries. A LSTM neural network model is established to estimate SOC, with battery voltage, current, temperature, and force data as input. The proposed method shows high accuracy in the middle SOC region, with less than 0.5% root mean square errors and less than 2.5% maximum errors, and maintains robustness and generalization performance at different temperatures.