Abstract: To meet the requirements for real-time, accurate, and reliable prediction of safety-critical parameters under nuclear power plant accident conditions, this study proposed a physics-informed and data-driven intelligent prediction method for safety-critical parameters in nuclear power plants. The thermal-hydraulic system codes can describe complex accident progression with strong physical consistency. However, their relatively high computational cost limits their application in rapid prediction and online early-warning scenarios. In contrast, purely data-driven models can improve computational efficiency, but they often suffer from insufficient physical constraints, uncertain prediction reliability, and limited interpretability. To address these issues, this study developed a time-series surrogate modeling framework by combining thermal-hydraulic mechanism knowledge with data-driven learning. First, the RELAP5 code was used to model and simulate a small-break loss-of-coolant accident (SBLOCA) in the main coolant pipe of a pressurized water reactor, generating high-fidelity accident data under multiple operating conditions. Then, mutual information analysis and mechanism-based understanding were combined to select key input features from system operating parameters, coolant properties parameters, and dimensionless physical parameters. Dimensionless parameters, including the Reynolds number (Re), Prandtl number (Pr), and Nusselt number (Nu), were introduced to enhance the model representation of flow and heat-transfer processes. Based on these features, a Transformer-based time-series surrogate model was constructed to perform multi-step prediction of peak cladding temperature (PCT). Furthermore, Monte Carlo dropout was used to obtain predictive distributions, and the interval score (IS) was adopted to quantitatively evaluate the quality of prediction intervals and the level of uncertainty. SHAP analysis was also employed to identify the contribution of different input features to the model outputs, thereby improving the interpretability of the prediction results. The SBLOCA case study shows that the model incorporating dimensionless physical parameters improves both prediction accuracy and uncertainty performance compared with the model without dimensionless parameters. The results of mean squared error, root mean squared error, mean absolute error, and mean absolute percentage error measures are 21.995, 4.690, 1.957, and 3.171%, respectively. Compared with the approximately 30 min computational time required by a single RELAP5 simulation, the proposed surrogate model reduces the inference time for a complete accident sequence to (355.778±12.521) ms, significantly improving computational efficiency. The uncertainty evaluation results show that the global average IS of the physics-enhanced model is 9.16, which is 27.8% lower than that of the model without dimensionless parameters. The SHAP analysis indicates that core water level and break area are important factors affecting PCT prediction, while dimensionless parameters such as Re, Pr, and Nu at the hot spot also contribute to the prediction. These results suggest that incorporating appropriate physical mechanism parameters can improve the accuracy, efficiency, and interpretability of data-driven surrogate models under complex accident conditions, providing a useful reference for rapid prediction of safety-critical parameters and auxiliary safety-state assessment in nuclear power plants.