Real-time pipe break extent prediction for loss of coolant accidents in pressurized water reactors using a multi-step forecast-assisted long short-term memory model

Real-time operator support systems are essential for improving accident management and decision-making in complex systems such as nuclear reactors. However, conventional physics-based models often exhibit computational inefficiency, highlighting the need for AI-driven approaches to accelerate solution processes. This study proposes the multi-step forecast-assisted long short-term memory model as a surrogate for real-time prediction of pipe break severity in loss of coolant accidents (LOCA) from hot leg and cold leg piping breaks in pressurized water reactors (PWR). The proposed model demonstrated a 67–74% reduction in accumulated real-time prediction error and 76–82% decrease in time-weighted error when compared to other diagnostic schemes. Additionally, the model yields predictions with throughput of 25.6 Hz and median latency of 26.2 ms, highlighting its scalability for real-world applications. The proposed model also exhibited robustness when artificially perturbed with low amplitude global noise and random burst noise, rapidly adapting to restore its baseline real-time performance. The analysis also revealed that the model is most sensitive to hot leg temperature perturbations, particularly under higher noise levels (e.g., 50% and 70%), while showing minimal sensitivity to fluctuations in coolant flow rate. Overall, the study pioneers AI-based real-time accident diagnosis to enhance nuclear safety and support collaborative decision-making between operators and AI systems. • AI framework was developed for real-time LOCA pipe break diagnosis in PWRs. • Multi-step forecasts boost real-time accuracy of LOCA pipe break diagnosis models. • Model runs much faster than the data source, showing strong real-world scalability. • Model maintains accuracy despite multiple artificial noise scenarios.