Abstract: The structural integrity and in-service lifetime of critical components within advanced nuclear power systems are fundamentally governed by their performance under prolonged exposure to extreme conditions of high temperature and mechanical stress. 321H stainless steel is a principal material employed for such demanding applications, making the precise long-term prediction of its creep deformation behavior a paramount objective for both reactor safety authorities and design engineers. Conventional forecasting methods, which predominantly rely on empirical models fitted to relatively short-term data, are often inadequate for reliable multi-thousand-hour extrapolations due to underlying mechanistic transitions and complex microstructural evolution. In this paper, a new data-driven model based on a backpropagation neural network (BPNN) was developed to solve this problem, the main goal was to use short-term creep data to accurately predict the full creep strain curve up to 10 000 hours. The model training used experimental creep data at 550 ℃ under 190 MPa stress. Tests were specifically designed to study how the amount of training data affects prediction accuracy. For this purpose, the model was trained with different datasets: 3 000 hours, 4 000 hours, and 5 000 hours of creep data. After training, the model’s stability was tested by applying it to two other stress conditions, 170 MPa and 180 MPa, at the same temperature. This validation process checks whether the model can generalize well to similar high-stress scenarios. The results are highly positive. The model achieves a very low prediction relative error less than 2%, for the 10 000 hours creep strain when trained on 5 000 hours of data. It accurately captures the entire creep curve, including the initial primary stage and the critical steady-state stage. It also finds a clear relationship between training data length and prediction accuracy. Increasing the training duration significantly improves the model’s long-term forecast reliability. Furthermore, the model demonstrates excellent interpolation capabilities within the 170-190 MPa stress range, successfully predicting the creep behavior of stresses that have not been directly trained. In conclusion, this research proves that BPNN model is a powerful and reliable tool for the long-term creep assessment of 321H stainless steel. It provides a practical and efficient method for predicting material behavior, which can greatly accelerate the design and safety evaluation of nuclear reactor components. The study also offers clear guidance on the necessary amount of short-term test data needed for achieving dependable long-term predictions.