Comparison of Multilayer Perceptron and Long Short-Term Memory for Plant Parameter Trend Prediction

Abstract

Human operators always have the possibility to commit human errors, and in safety-critical infrastructures such as a nuclear power plant, human error could cause serious consequences. Since nuclear plant operations involve highly complex and mentally taxing activities, especially in emergency situations, it is important to detect human errors to maintain plant safety. This work proposes a method to predict the future trends of important plant parameters to determine whether a performed action is an error or not. To achieve this prediction, a recursive strategy is adopted that employs an artificial neural network as its prediction model. Two artificial neural networks were selected and compared: multilayer perceptron and long short-term memory (LSTM). Model training was accomplished using emergency operation data from a nuclear power plant simulator. From the comparison results, it was observed that the future trends of plant parameters were quite accurately predicted through the LSTM model. It is expected that the plant parameter prediction function proposed in this work can give useful information for detecting and recovering human errors.