Prediction and sensitivity analysis of hysteretic energy spectra for self-centering systems using artificial neural network considering Chinese seismic scenarios

This study proposes a deep learning-based artificial neural network (ANN) model for predicting the hysteretic energy spectra for self-centering systems. Ground motion records are selected based on the site classifications specified in the Chinese seismic code. A procedure for constructing hysteretic energy spectra for self-centering systems is developed to generate the dataset for the model training, resulting in a total of 81000 samples with varying input features and corresponding output labels. Bayesian Optimization is employed to determine the optimal ANN configuration. Additionally, model robustness is evaluated by introducing Gaussian noise into the input data. SHapley Additive exPlanations (SHAP) analysis is further performed to quantify the contributions of different input features. The results show that the ANN model using acceleration response spectra as input seismic features exhibits superior predictive accuracy and generalization capability. Meanwhile, the developed model remains reliable even in the presence of noise, particularly for stronger earthquakes. SHAP analysis demonstrates that while earthquake characteristics exert only a marginal impact on predictive accuracy, neglecting seismic features still impairs the model's performance. Additionally, the results highlight a negative impact of the damping ratio ξ on hysteretic energy spectra, whereas the energy ratio η has the opposite influence.