Seismic Response Time Series Prediction of a 500kV Transformer Bushing

Transformer bushings are among the most seismically vulnerable components in high-voltage electrical systems, yet post-earthquake response evaluation is often constrained by the lack of field measurement data. This study constructs a large-scale ground motion-response dataset for a 500 kV bushing using finite element simulations and applies a Long Short-Term Memory (LSTM) network to perform response timeseries prediction. The effects of key hyperparameters are systematically evaluated. Furthermore, various regression models are employed to generalize prediction performance across different parameter combinations and to identify the optimal configuration. Results show that window length and the number of hidden units have the most significant impact on model performance. Under representative strong ground motion scenarios, the proposed model achieves high-accuracy predictions with peak response error controlled within $1.1 \%$. The findings validate the applicability of LSTM-based modeling for postearthquake response prediction of electrical equipment and provide data-driven guidance for parameter tuning and dataset design.