Optimizing deep neural networks for nuclear power plant temperature estimation: A study on feature importance and outlier detection

Abstract

Extending the operational life of the Angra 1 Nuclear Power Plant (NPP) requires an accurate estimation of historical temperature exposure for equipment within the containment area to assess the aging and degradation of critical components. This assessment is essential for extending the plant's license by 20 years. Since Mobile Temperature Sensors (MTSs) were installed only in 2015, this study employs Deep Neural Networks (DNNs), including Deep Rectifier Neural Networks (DRNNs), Convolutional Neural Networks (CNNs), and Long Short-Term Memory Networks (LSTMs), to infer historical temperature data before MTS deployment. The DNNs utilize time series data from Plant Fixed Sensors (PFS), monitored by the Angra 1 Integrated Computer System (SICA), as inputs. Feature importance and outlier detection methods are investigated to enhance DNN performance. Feature importance techniques, such as XGBoost, Random Forest, Principal Component Analysis (PCA), and outlier detection methods, including autoencoders, DBSCAN, and isolation forest, are evaluated. Results indicate that preprocessing significantly improves model accuracy. For instance, PCA without outlier detection combined with a CNN achieved a Mean Absolute Error (MAE) of 3.194, whereas the integration of Random Forest and XGBoost for feature importance with DBSCAN for outlier detection and a CNN reduced the MAE to 0.497.