Aiming at the trend prediction of the Loss of Coolant Accident (LOCA) in nuclear power plants, this paper proposes a simulation prediction method optimized by deep learning. This method adopts the Long Short-Term Memory (LSTM) model to optimize the traditional Model Predictive Control (MPC) method, thereby achieving the computing strategy combining time series feature extraction and optimal dynamic adjustment. The LSTM model running in the deep learning environment provides real-time reference of predicted values for MPC, helping MPC calculate the predicted values in the subsequent timesteps. In the verification experiment of LOCA coolant flowrate prediction, the overall average optimization ratio of the prediction accuracy reached 68.30%, and it demonstrates stability and robustness in multiple-timestep tests. This study shows the feasibility of using deep learning models to optimize the prediction accuracy of MPC, which may assist further studies of enhancing intelligent prediction performance in nuclear power plants.