Accident Diagnosis and Autonomous Control of Safety Functions During the Startup Operation of Nuclear Power Plants Using LSTM

Abstract

Accident diagnosis is regarded as one of the complex tasks for nuclear power plant (NPP) operators. In addition, if the accident occurs during the startup operation, it is hard to cope with the situation appropriately because the initial conditions are different from the normal operation mode. Although operating procedures are provided to operators, accident diagnosis and control for recovery are difficult tasks under extremely stressful conditions. In order to achieve safe operation during the startup operation, this study proposes algorithms not only for accident diagnosis but also for protection control using long short-term memory (LSTM), which is an advanced version of recurrent neural networks, and functional requirement analysis (FRA). Using the LSTM, the network structures of algorithms are built. In addition, FRA is performed to define the goal, functions, processes, systems, and components for protection control. This approach was trained and validated with a compact nuclear simulator for several accidents to demonstrate the feasibility of diagnosis and correct response under startup operation.