The state of charge (SOC) of lithium batteries is one of the core state parameters of Battery Management Systems (BMS), and accurate estimation of the battery SOC is of great significance for the development of electric vehicles. Traditional methods rely heavily on the accuracy of battery models and struggle to adapt to the highly nonlinear and time-dependent characteristics of batteries. With the development of deep learning theory, estimation methods based on neural networks have been widely applied. This paper proposes an improved sparrow search algorithm (ISSA) model, which combines chaotic mapping, sine-cosine algorithm, and firefly disturbance methods to optimize the backpropagation (BP) neural network (ISSA-BP) for high-precision SOC estimation. The model was validated using the public experimental dataset from the University of Maryland, which includes various complex operating conditions and different temperatures. The prediction results were evaluated from the perspectives of mean absolute error, mean square error, and root mean square error. The results show that the ISSA-BP model can control the SOC estimation error within 2% under various conditions and temperatures, demonstrating better accuracy compared to single neural network models, as well as excellent robustness and generalization ability.