Assist in real-time risk evaluation induced by electrical cabinet fires in nuclear power plants: A dual AI framework employing BiTCN and TCNN

Abstract

Electrical cabinet fires in nuclear power plants pose significant threats to reactor safety. While numerous studies have investigated cabinet fires, risk real-time evolution induced by high-temperature smoke layer has not received sufficient attention. Consequently, this study proposes a dual AI framework, integrating the Bidirectional Temporal Convolutional Network (BiTCN) and Transposed Convolutional Neural Network (TCNN), to predict temperature field in advance. Database is constructed by Fire Dynamics Simulator (FDS), featuring various burner heights, heat release rates, and ventilation conditions. Temperature beneath ceiling and temperature field are recorded. Thermocouple data is used to train BiTCN for forecasting ceiling temperature with a lead time of 60 s. The trained BiTCN model achieved an exceptional R² value exceeding 0.999. Compared to other methods, BiTCN has advantages in accuracy and computational efficiency. The TCNN takes output from BiTCN as input and FDS temperature slice results as output labels to deduce real-time changes in two-dimensional temperature field. It achieves an R² value of 0.973. Although some discrepancies exist, results indicate strong predictive capability and reliability in capturing spatial and temporal dynamics of temperature field. This work demonstrates potential of using Artificial Intelligence (AI) to predict dynamic evolution of cabinet fires and represents a significant exploration of applying AI in nuclear safety.