Accurate estimation of the state of charge (SOC) of lithium batteries is crucial to improving the dynamic performance and energy utilization of batteries. Aiming at the problems of low accuracy and poor stability of existing neural network SOC estimation methods under complex working conditions, this paper proposes an improved GRU model algorithm to estimate SOC. Firstly, combine 1DCNN and Bi-GRU and add attention mechanism to build 1DCNN-Bi-GRU-ATT model. Secondly, in order to eliminate the phenomenon that the ReLU activation function is prone to dead neurons, it is improved to PReLU activation function. At the same time, in order to solve the problem that MSE-Loss is easily affected by abnormal battery data in complex working conditions and the convergence speed of MAE-Loss is slow, Huber-Loss is used instead as the network loss function. Finally, the Adam algorithm is improved to Nadam algorithm using Nesterov accelerated gradient. The experimental results of lithium battery SOC estimation show that the average values of root mean square error and mean absolute error of the model algorithm under 12 complex operating conditions are 1.181 7% and 0.924 1%, respectively. Compared with the model before improvement and other models, the comprehensive performance of this model in 12 cases is more stable and accurate, and it has higher generalizability.