Energy storage container is the core equipment of a power plant for lithium battery energy storage . Each container is composed of thousands of cells connected in series and parallel. Therefore, the accurate estimation of the state of charge (SOC) of lithium batteries in container cores becomes the core and basic parameter to characterize the operation of a power plant for energy storage. Moreover, in order to assist the new energy to be connected to the grid efficiently, the operating state of the energy storage system is randomness, fluctuation and uncertainty, which requires higher accuracy of the cell state estimation. In this paper, the Thevenin model of battery is firstly established on the basis of the Kirchhoff's circuit laws. The state and observation equations of the system are listed according to the ampere-time integration method, and then as the study object for the extended Kalman filter (EKF) algorithm. The EKF algorithm is used to update and iterate the estimated SOC of battery. The updated error values of the Kalman matrix and state variables derived from the EKF algorithm, and the battery data under UDDS conditions are as a training data set for long-term and short-term memory (LSTM) neural network algorithm. The joint algorithm of LSTM-EKF is thus completed to achieve an optimized estimation of the SOC of batteries in container cores. The SOC error can be reduced to less than 1% by the proposed LSTM-EKF algorithm. The optimization algorithm applied in the safe operation and monitoring platform of energy storage power station is finally introduced.