Operational trend prediction technology utilizes historical state data and current equipment operating conditions to forecast the future development trends of nuclear power plant parameters. In the event of an accident, it can promptly reflect the operational trend changes of certain parameters in the nuclear power plant, providing pre-alarm information to operators. This paper proposes a multi-step prediction method that combines a surrogate model based on phase space reconstruction with a real-time prediction model. The method is used to quickly predict the future 50-second development trends of key parameters following a fault in the nuclear power system. This paper first employs phase space reconstruction to transform univariate time series data into a high-dimensional space that maintains the same topological properties as the original dynamical system. Under this high-dimensional space, training and testing datasets are constructed using a sliding time window approach. The LSTM method is then applied to build both a surrogate model and a real-time prediction model, with the final prediction results obtained by combining the predictions from these two models. To validate the accuracy of the proposed method, various system fault data including pressurizer safety valve leaks and SGTR were collected from the Fuqing pressurized water reactor simulator. The method was applied to predict the trends of critical parameters such as coolant pressure and steam generator steam production. Comparative analysis with ARIMA, LSTM methods demonstrates that the proposed method achieves superior prediction accuracy.