Session: 15-06

Paper Number: 134780

134780 - Investigation of Lstm and Automl-Based Models for the Real-Time Diagnosis of Pwr Loca Accident Progression 

Abstract: 

The loss of coolant accident (LOCA) in a nuclear reactor is an abnormal state where the coolant is leaking out of the reactor vessel which deprives the core of its primary cooling function. In the worst-case scenario, LOCA can progress to a severe accident which happened in Three Mile Island and Fukushima Daiichi highlighting the need for prompt and reliable LOCA accident management. One of the ways to mitigate LOCA is to help the operators assess the situation more accurately through the real-time diagnosis of the accident progression. Specifically, the real-time diagnosis of LOCA progression involves the forecasting of key plant variables and the identification of the accident scenario such as the location of pipe break and the extent of pipe damage. Although the real-time diagnosis of LOCA improves the decision-making process of the operators, it is complex due to the non-linear behavior of the plant variables, the difficulty of predictive diagnosis, and the required processing speed. With the availability of big data and processing power, machine learning (ML) models are used to perform predictive tasks faster than first principles-based models due to their capability to learn non-linear patterns from the actual data. In this study, the potential of ML models, specifically LSTM and AutoML, in the real-time diagnosis of LOCA scenarios were investigated. The real-time diagnosis of LOCA is modelled as a combination of a forecasting model for the water level, cold leg temperature, and hot leg temperature, and a predictive model that identifies the location of the pipe break and extent of pipe damage. The LSTM and AutoML models were trained using 428 different LOCA scenarios that were generated from RELAP5/SCDAPSIM. The model hyperparameters were tuned using Bayesian optimization and the models were chosen based on their validation performance. The best performing LSTM and AutoML models were compared based on their accuracy in forecasting the chosen plant variables and the location and extent of pipe break, as well as the prediction speed. Moreover, the progression of the prediction through time was analyzed for both the LSTM and AutoML models. The study was able to show that both the LSTM and AutoML models were able to provide fast and reliable real-time forecasting of the chosen plant variables. Furthermore, the models were able to correctly identify the location and extent of pipe damage at the early stages of the LOCA progression which implies the potential of the ML-based real-time LOCA diagnostic tool in aiding the decision-making process of the operators to swiftly mitigate the LOCA scenario. Lastly, when compared to a non-real-time LSTM-based predictive model, the real-time counterpart has more stable predictions over time which proves the advantage of a real-time LOCA diagnostic tool.

Presenting Author: Johndel Obra The University of Tokyo

Presenting Author Biography: Mr. Johndel B. Obra is a first-year master's student in Nuclear Engineering and Management at The University of Tokyo under the supervision of Dr. Shuichiro Miwa. Mr. Obra is a member of the Visualization Laboratory in UTokyo headed by Dr. Koji Okamoto. He is conducting research on the application of machine learning and artificial intelligence in nuclear engineering problems, especially real-time plant diagnostics during loss of coolant accidents. Prior to his stay in UTokyo, Mr. Obra finished his bachelor's and master's degrees in Chemical Engineering at the University of the Philippines Diliman where he did research on the applications of machine learning in energy complementarity and emission prediction.

Authors:

Johndel Obra The University of Tokyo
Shuichiro Miwa The University of Tokyo
Koji Okamoto The University of Tokyo