Physics-Informed Deep Learning for Transmission Tower Response under Wind Loading

William Hughes,Wei Zhang

DOI: https://doi.org/10.1109/naps56150.2022.10012226

2022-01-01

Abstract:Under strong winds, structural failures of the power system can lead to cascading power outages. For damage prognosis, accurate and computationally efficient models of the transmission system structural response are necessary. Traditional methods, including pushover analysis, Monte Carlo simulations, or finite element analysis, can be computationally inefficient for such predictive purposes. Meanwhile, deep learning algorithms, such as long short-term memory (LSTM) networks, have become popular for efficiently forecasting time series data. However, such black-box methods fail to leverage the structural system’s underlying physics and are limited by the availability of sufficient training data. To expedite training and enhance accuracy, physics-informed deep learning techniques are proposed to integrate the physical governing equations as constraints on the deep learning algorithms. Presently, a deep recurrent residual neural network (DR-RNN) architecture is utilized for the prediction of transmission tower response under wind loadings. The DR-RNN implements the physical constraints through the state space representation and is found to outperform the traditional LSTM accuracy and the traditional numerical solution computational efficiency. Furthermore, the DR-RNN is shown to have superior generalization capacity to data outside the training range. Finally, the DR-RNN application for the development of fragility curves under various sustained wind durations is demonstrated.