Overview and details of the sessions of this conference. Please select a date or location to show only sessions at that day or location. Please select a single session for detailed view (with abstracts and downloads if available).
Session Chair: Daniele Giuffrida, Federal Authority for Nuclear Regulation - FANR, United Arab Emirates Session Chair: Hyun Gook Kang, RPI, United States of America
Key learnings of Full Scope Simulator’s Nuclear Steam Supply System (NSSS) model re-hosting project at EDF
David Pialla, Alexandre Bonne, Nicolas Delanghe, Nicolas Bardiaux, Karine Vareille
EDF is the largest worldwide nuclear power plant operator. With 19 reactor sites and 58 nuclear power plants in operation and one power plant under commissioning, training and swift licensing is a key factor to maintain and improve the operating factor of the nuclear fleet. In order to comply with regulatory requirements and commitments, Full Scope Simulators (FSS) are critical tools.
In order to maintain and improve FSS physical representativeness and availability rate, EDF has launched a continuous process of maintenance of its FSS since the beginning of the industrial use. This continuous maintenance addresses various fields of skills: models corrections and updates, taking into account plant evolutions according the periodic safety reviews, operating software and hardware upgrades etc.
The present paper deals with the ambitious project about the re-hosting of Nuclear Steam Supply System (NSSS) model. Firstly, the motivations for the project are explained. The choice has been made to retain the ‘State-Of-The-Art’ CATHARE_2 thermal-hydraulics system code for the next NSSS model. This challenging choice is possible thanks to a robust and industrial methodology developed at EDF, in order to integrate safety code in real-time simulators. In another part, the global management of the project is described, including the different actors and contributors. Finally, the key learnings and the prospects are highlighted. In a few words, the integration of ‘State-Of-The-Art’ codes instead of simplified models, leads to challenging tasks and work for all the contributors of the project, but contributes actively to strengthen the FSS simulator predictions and opens a new field of applications for the users.
Relevant Plant Parameter Tracking for Abnormal State Diagnosis
Ji Hyeon Shin, Seung Jun Lee
Ulsan National Institute of Science and Technology
In the abnormal situation in a nuclear power plant (NPP), operators have to check alarms and plant parameters and to enter appropriate abnormal operating procedure (AOP). However, it is very difficult for operator to see hundreds of parameters and to select the appropriate of 82 AOPs in a short time. According to a prior re-search, a convolutional neural networks (CNN) model among many deep learning algorithms is developed to diagnose abnormalities. In this paper, the contribution degree of feature values trained in this CNN model was calculated. Based on this contribution degree, relevant plant parameters to the classification of abnormal conditions were identified. The developed research will support the operator to diagnose abnormal conditions more easily.
2. Relevant Parameters with Explainable Artificial Intelligence (XAI)
In this study, high relevant parameters are extracted using XAI for more rapid diagnosis. Three kinds of abnormal states are predicted with high accuracy using the CNN model. Each feature or pixel in the input image represent each plant parameter in each second. The contribution of each feature is calculated using the gradient information corresponding to each classification in the trained CNN model. This calculation is used by methods, Gradient-weighted class activation mapping (Grad-CAM) or saliency map. The heatmap is shown for the contribution mean value of each abnormal state data. Based on feature values of each heatmap, the top 5 relevant parameters corresponding to each abnormal state are tracked.
This study proposed method extracting plant parameters related to a certain abnormal condition. The XAI can identified parameters that contribute significantly to each abnormal state classification. In the future work for the thousands of variables, it is expected to reduce the burden of abnormal states diagnosis by specifying parameters that need to be noted by the plant operator.
Analysis of RPSs with Multiple Platforms from Safety and Security Point of View: Case Study for AGN-201K
Jiye Jeong, Gyunyoung Heo, Ibrahim Ahmed
Safety-related instrumentation and control (I&C) systems of nuclear facilities have been steadily implemented using software based several digital platforms.
In most of cases, the common cause failures (CCFs) of the digital platforms have been pointed out as a major factor affecting the safety of NPPs. In addition to this, as the possibility of cyber attack on digital I&C systems should be checked while maintaining compliance with safety aspect. One of key technical assets to address these problems underlies the selection of a platform.
To propose a potential solution in a conceptual level, the digital hybrid model that consists of two platforms with Programmable Logic Controller (PLC) and Field Programmable Gate Array (FPGA) has been investigated. The case example was demonstrated on the basis of the reactor protection system (RPS) structure of an educational research reactor, AGN-201K.
Multiple independent platforms can have pros and cons from safety and security viewpoint, so this paper focused on how much the digital hybrid model is different than other single platform including cost-benefit analysis.
It presents quantitative information for failure rates, CCFs and cyber security evaluation as unavailability, and the unavailability of the digital hybrid model was compared with the current model (Analog) and single platform models (PLC or FPGA, respectively).
Simultaneous monitoring of a large number of process values and early detection of signs of abnormalities with a two-stage model composed of a time window autoencoder and a deviation autoencoder
1Toshiba Corporation; 2Toshiba Energy Systems & Solution Corporation
In a large-scale plant such as a nuclear power plant, many process values are measured for the purpose of monitoring its plant performance and the health of various systems. It is difficult for plant operators to constantly monitor all of the process values. We aim to enable monitoring of a large number of process values and early detection of signs of abnormalities including unprecedented events with high detection performance, without limiting the monitoring targets to specific instruments, systems, and operating conditions. To achieve this objective, a detection algorithm is required to (1) handle a large number of the process values simultaneously along with evaluating their complex correlations and (2) handle the plant transient states with high accuracy, and (3) tune the models with low man-hours. Autoencoder, one of the deep learning methods, was selected as a detection algorithm that can meet these requirements. The autoencoder model is trained with the plant normal state values, and hence the deviation between the actual process value and the predictive one given as output of the model indicates the signs of abnormalities when it exceeds its threshold. In order to handle the transient state and reduce the false detection, a time window structure examining the time correlation between the process values and a deviation structure predicting the normal fluctuation signal superimposing on individual process value which cannot be eliminated by input signal filter, such as the electrical noise signal, respectively, were introduced. The effect of this two-stage model consisting of them was confirmed with simulated process values of a virtual nuclear power plant (1.1 GW, analog input points of 3100). The two-stage models constructed for all systems of the plant showed good detection performance of signs of abnormalities during both steady and transient states.