Abstract: This study aims to develop a data assimilation framework based on dynamic time warping (DTW) and ensemble Kalman filter (EnKF) to solve the problem of insufficient prediction accuracy of steady-state and transient conditions in thermal-hydraulic analysis of nuclear reactors. Through the collaborative mechanism of parameter optimization and real-time data correction, the global prediction ability of the sub-channel model is improved, while meeting the high efficiency requirements of real-time monitoring of nuclear power plants. Two types of data assimilation frameworks were constructed. One is parameter-driven static assimilation. By constructing a multi-dimensional parameter space of heat transfer coefficient α and correction coefficient β, grid scanning is adopted, and the optimal parameter combination under steady-state conditions was screened with the minimum DTW distance as the optimization goal. The other is data-driven dynamic assimilation. Based on the EnKF algorithm, real-time temperature observation data are fused every 0.01 s, and the model state covariance matrix is dynamically updated to suppress noise and correct transient deviations to obtain assimilation results. Based on the COBRA-EN subchannel program, a thermal-hydraulic analysis example of the initial core life of a pressurized water reactor was constructed, using a 1/8 symmetric geometric model. The tent function was used to increase the transient power multiple, and a fast transient analysis of 1.5 s was completed. Under steady-state conditions, static optimization reduces the mean absolute error (MAE) by 15%. The assimilation curve has a response lag of about 200 ms compared with the measured data, indicating that the global optimal parameters are difficult to adapt to the local dynamic mutations of transient conditions. Dynamic assimilation further compresses the peak deviation of the transient stage from 4.2 K to 1.8 K. In terms of noise suppression, dynamic assimilation reduces the standard deviation of the prediction noise by 37%, effectively suppressing the divergence caused by the accumulation of model errors. In terms of real-time verification, the DTW-EnKF framework completes the calculation within a time step of 0.01 s, meeting the monitoring requirements of nuclear reactors. However, the DTW-EnKF data assimilation method still has certain limitations, and the lumped parameter assumption of the subchannel program leads to the loss of local heat exchange details. The model validation results show that the data assimilation technology effectively suppresses the model prediction error, especially under transient conditions such as power step, the fluctuation amplitude of the assimilated path is reduced by 30% compared with the independent prediction. The DTW-EnKF data assimilation framework proposed in this study significantly improves the prediction accuracy and real-time performance of the nuclear reactor thermal hydraulic model. Static parameter optimization provides a basic guarantee for steady-state conditions, while dynamic data assimilation effectively copes with the uncertainty of transient processes. However, the model’s simplified assumptions and high-frequency signal processing capabilities still need to be improved. Future work can introduce particle filtering and adaptive Kalman gain strategies to enhance robustness under complex conditions.