Design of 1D CNN Kernels for Predicting Seismic Response of Nuclear Power Plants

Abstract

Nuclear power plants (NPPs) play a critical role in energy generation; however, their structural integrity is vulnerable to seismic events, necessitating precise seismic response predictions to ensure operational safety. Recent advancements in artificial intelligence have enabled the development of highly effective models for this purpose, with one-dimensional convolutional neural networks (1D CNNs) demonstrating notable success in modeling the complex relationships between seismic excitations and structural behavior. As seismic responses are fundamentally governed by the frequency characteristics of ground motions, it is essential for prediction models to effectively extract relevant frequency-domain features. 1D CNNs, which apply localized convolution operations analogous to finite impulse response (FIR) filters, are influenced by the kernel size when determining which frequency components are captured. To improve the ability of the model to extract meaningful features across a broad range of frequencies inherent in seismic signals, this study proposes a hierarchical 1D CNN architecture that systematically decreases the kernel sizes across the network layers. By combining multiple kernel scales within a single framework, the hierarchical design improves the ability of the model to extract diverse frequency components and enhances the overall prediction accuracy. The proposed model is validated using a case study of NPP structures, where the experimental results demonstrate its ability to achieve significantly lower prediction errors compared with fixed-kernel CNNs, accurately capturing both long-term structural behavior and short-duration seismic events.