Mahmood, Y. 2026 Data-driven digital twin for structural seismic response prediction using temporal convolutional networks. Lower Hutt, NZ: Earth Sciences New Zealand. GNS Science report 2026/10. 25 p.; doi: 10.21420/XAHA-K088
Abstract
In this research, the performance of a Temporal Convolutional Network (TCN) as a data-driven component of a structural digital twin in predicting seismic response has been investigated. First, the capability of the proposed algorithm has been examined in a single-degree-of-freedom (SDOF) structure and then applied to an instrumented Earth Sciences New Zealand Avalon building. In the SDOF example, the model has achieved high accuracy (95.41% correlation) in capturing time and frequency domain responses and learning key system properties such as natural period and damping. For the Avalon building, the accuracy in predicting responses under recorded earthquakes is relatively high; the coefficient of determination (R²) is equal to 0.926 and 0.802 for Units 1 and 2, respectively. There was good agreement between the predicted and measured responses, while differences in structural complexity and sensor locations were also reflected. Overall, the results demonstrate the potential of TCNs as fast and accurate modelling tools for structural digital twins in earthquake engineering, supporting real-time monitoring and performance-based assessment (auth)
