Accurate state of charge estimation of lithium-ion batteries using Butterworth filter integrated with deep learning model optimized by the Bayesian optimization algorithm

Accurately estimating the state of charge (SOC) of lithium-ion batteries is crucial for ensuring their safety. However, there are still problems such as susceptibility to noise interference during battery data collection and inability of existing models to capture features of both short-term abrupt changes and long-term temporal dependencies during the dynamic variations of battery data. To address these problems, this paper proposes a novel SOC estimation method that integrates the Butterworth filter with the temporal convolutional network-long short-term memory-multi-head self-attention (TLM) model optimized by the Bayesian optimization (BO) algorithm. This method uses the Butterworth filter to filter the raw signal, aiming to reduce noise from the data. The TLM model first uses the temporal convolutional network to extract the local temporal features of input data, then uses the long short-term memory to capture the long-term dependencies of data sequences, and finally uses the multi-head self-attention to model the global temporal dependencies. Furthermore, the BO algorithm is used to optimize the model’s hyperparameters, ensuring the model achieves optimal performance. Experimental testing of the proposed method is conducted using lithium-ion battery drive cycle datasets from the University of Wisconsin-Madison and the Technical University of Berlin under fixed and varying ambient temperatures. Experimental results show that the maximum average root mean square error (RMSE) and maximum average mean absolute error (MAE) of the proposed method are 0.812% and 0.652%, respectively. In contrast, the maximum average RMSEs of the comparative models, including TCN, LSTM, TCN-MHSA, LSTM-MHSA, and TCN-LSTM, are 0.903%, 1.207%, 0.836%, 1.148%, and 0.840%, respectively, while their maximum average MAEs are 0.738%, 0.994%, 0.690%, 0.948%, and 0.687%, respectively. Furthermore, under the LA92 drive cycle at 25 \(^∘\) C, when using the filtered data, the RMSE and MAE of this method under fixed ambient temperature are 0.409% and 0.306%, respectively, and under varying ambient temperature, they are 0.475% and 0.371%, respectively. In contrast, when using the raw data, the RMSE and MAE under fixed ambient temperature are 1.984% and 1.475%, respectively, and under varying ambient temperature, they are 2.285% and 1.734%, respectively.