Data-driven estimation method for the state of charge(SOC) of lithium batteries still relies on a large amount of calibration data and shows poor performance in dealing with dynamic changes and complex operating conditions. Therefore, an improved locust algorithm optimization combined with time-domain convolutional networks and multi-head attention mechanisms was proposed to estimate the SOC of lithium battery. A time-domain convolutional network was first used to model the long-term dependency relationships in the time series data of lithium battery charging. Meanwhile, multi-head attention was used to learn the long-term dependency relationships of data features, and each attention head was used to calculate the dependency relationships of different tensors in the sequence to assist the time-domain convolutional neural network in enhancing the capture of dependency relationships and reducing its dependence on a large amount of calibration data. In addition, the chaotic locust algorithm was improved to optimize the hyperparameters of model to maximize the performance of model. The experimental results show that, compared with other methods, the optimized model can exhibit better accuracy and stability in the task of estimating the SOC of lithium battery under different temperature conditions.
