TGT: Battery State of Charge Estimation with Robustness to Missing Data

In recent years, with the large-scale application of new energy sources, whole-life battery state-of-condition (SOC) estimation has become increasingly important in power systems and electric vehicle range. However, existing SOC estimation methods such as ARIMA and machine learning methods such as Transformer have insufficient accuracy to solve estimation problems in long-term sequences. Based on this, this paper proposes a novel TCN-GRU-Transformer (TGT) model that aims to optimize the performance of SOC estimation in scenarios with different degrees of missing data. The model captures the global dependencies in the battery data by extracting useful features from the time-series dataset to understand the trend of the battery state over time. Through experiments on publicly available datasets, it is demonstrated that the MAE of the TGT model is 0.0192 while classical machine learning models is 0.030 in battery soc estimation. The error rate decreases by 36.1% in MAE.