Diagnosis and Prediction for Loss of Coolant Accidents in Nuclear Power Plants Using Deep Learning Methods

Jingke She, Tianzi Shi, Shi-Yu Xue, Yan Zhu, Shaofei Lu, Peiwei Sun, Huasong CaoFrontiers in Energy Research 2021 · DOI

Abstract

A combination of Convolutional Neural Network (CNN), Long-Short Term Memory (LSTM), and Convolutional LSTM (ConvLSTM) is constructed in this work for the fault diagnosis and post-accident prediction for Loss of Coolant Accidents (LOCAs) in Nuclear Power Plants (NPPs). The advantages of ConvLSTM, such as effective feature determination and extraction, are applied to the classification of LOCA cases. The prediction accuracy is enhanced via the collaborative work of CNN and LSTM. Such a hybrid model is proved to be functional, accurate, and adaptive, offering quick accident judgment and a reliable decision basis for the emergency response purpose. It then allows NPPs to have an Artificial Intelligence (AI)-based solution for fault diagnosis and post-accident prediction.

Introduction

In a nuclear power plant (NPP), a Loss of Coolant Accident (LOCA)—where a pipe break causes a rapid loss of reactor core cooling—is a high-stakes event. Responding correctly requires two critical capabilities: an accurate diagnosis of the break size and a reliable prediction of how system parameters, such as pressure and flow, will evolve over time.

Traditionally, these tasks are handled by complex physical models or statistical methods. However, the physical processes inside a reactor are highly non-linear and involve dozens of interacting variables, making traditional models computationally expensive or prone to inaccuracies due to oversimplifying assumptions. This research proposes a data-driven Artificial Intelligence (AI) solution designed for real-time emergency response.

Methodology

The researchers developed a two-stage hybrid architecture to manage LOCAs.

1. Diagnosis: Improved ConvLSTM

To identify the severity (break size) of the accident, an improved Convolutional LSTM (ConvLSTM) was employed. Unlike standard LSTMs that handle sequences of data using matrix multiplications, ConvLSTMs utilize convolution operations. This allows the model to capture spatial relationships between different sensors (e.g., pressure and temperature) while simultaneously tracking their temporal evolution.

Figure 1: Structure of the improved ConvLSTM model for fault diagnosis.

2. Prediction: CNN + LSTM Pipeline

Once the break size is identified, a second model predicts future trends. This model uses a CNN layer as a feature extractor to distill complex multi-dimensional data into its most salient components. These features are then fed into a deep LSTM network (comprising two layers) to forecast long-term system behavior.

Figure 2: Structure of the Post-LOCA prediction model combining CNN and LSTM layers.

3. The Fault Dictionary

A key implementation detail is the "Fault Dictionary" lookup table. The system stores five sets of pre-trained weights, each optimized for a specific break size. Once the diagnosis stage identifies the break size, the prediction model loads the corresponding optimized weights to ensure maximum forecasting accuracy.

Datasets and Simulation

Due to the rarity of real-world LOCA data, the study utilized an industry-grade simulation platform:

  • Source: An engineering simulator coupled with MATLAB/Simulink.
  • Scope: Five break sizes (0.2, 0.4, 0.6, 0.8, and 1.0 cm²).
  • Operational States: Simulations were conducted at five reactor power levels (60% to 100%) to ensure model robustness.
  • Preprocessing: Noise was added to simulated signals to mimic real-world sensor interference, and data was normalized to a 0–1 scale.

Key Results and Insights

The hybrid framework demonstrated significant improvements over traditional and standalone deep learning models:

  • High Diagnostic Accuracy: The ConvLSTM classifier achieved a 96% accuracy rate, correctly identifying the break size in 24 out of 25 test cases.
  • Precise Forecasting: The CNN+LSTM prediction model maintained a Mean Square Error (MSE) as low as $10^{-3}$.
  • Spatio-Temporal Synergy: By treating sensor data as a multi-dimensional structure, the ConvLSTM recognizes physical relationships between disparate sensors (e.g., pressure drops corresponding to flow changes).
  • Feature Extraction: The CNN layer proved essential for filtering noise and simplifying input for the LSTM, preventing the model from being overwhelmed by raw data complexity.

Limitations and Future Directions

Despite the promising results, several challenges remain for deployment in operational control rooms:

  • Non-linear Inflection Points: The model occasionally struggles with "underfitting" at sharp turning points where system parameters change violently and suddenly.
  • Data Diversity: Future work will expand the training set to include a wider variety of break sizes and locations.
  • Real-World Validation: Transitioning from simulated data to historical real-world NPP data is a necessary step to bridge the simulation-to-reality gap.

Summary updated

Abstract

A combination of Convolutional Neural Network (CNN), Long-Short Term Memory (LSTM), and Convolutional LSTM (ConvLSTM) is constructed in this work for the fault diagnosis and post-accident prediction for Loss of Coolant Accidents (LOCAs) in Nuclear Power Plants (NPPs). The advantages of ConvLSTM, such as effective feature determination and extraction, are applied to the classification of LOCA cases. The prediction accuracy is enhanced via the collaborative work of CNN and LSTM. Such a hybrid model is proved to be functional, accurate, and adaptive, offering quick accident judgment and a reliable decision basis for the emergency response purpose. It then allows NPPs to have an Artificial Intelligence (AI)-based solution for fault diagnosis and post-accident prediction.
BibTeX
@inproceedings{she2021diagnosis,
  author = {Jingke She and Tianzi Shi and Shi-Yu Xue and Yan Zhu and Shaofei Lu and Peiwei Sun and Huasong Cao},
  title = {{Diagnosis and Prediction for Loss of Coolant Accidents in Nuclear Power Plants Using Deep Learning Methods}},
  booktitle = {Frontiers in Energy Research},
  year = {2021},
  doi = {10.3389/fenrg.2021.665262},
}
Figure