Power prediction, Long short-term memory, Self-attention, Fault diagnosis, Artificial intelligence, Nuclear power plants, Electrical engineering. Electronics. Nuclear engineering, TK1-9971, Computer software, QA76.75-76.765

Monitoring key parameters in nuclear power plant control rooms is critical, as human errors can result in severe safety and operational consequences. This study proposes a hybrid framework for power prediction and fault detection that integrates multi-head self-attention mechanisms with long short-term memory networks, combined with a dual-monitoring system. The framework is evaluated using real-time data from two pressurized water reactor units (Units 5 and 6) under four realistic operational scenarios. In the most informative case, the model achieves a 56.6% reduction in root mean square error and a 36.8% reduction in mean absolute error, with a coefficient of determination (R2) of 0.9924—significantly outperforming the next-best benchmark. For fault diagnosis, the dual-monitoring system reduces the false negative rate to 18.73% and improves recall to 81.27%, demonstrating strong anomaly detection under complex conditions. By combining short-term fluctuation sensitivity with long-term trend stability, the proposed approach offers a robust and generalizable solution for intelligent monitoring. These findings advance the development of artificial intelligence–enhanced systems for secure and efficient operation of critical energy infrastructure.