Estimating inelastic seismic response of reinforced concrete frame structures using a wavelet support vector machine and an artificial neural network

摘要

Modern building codes increasingly enforce evaluating the inelastic response of structures to ensure their safety in major seismic events. Although an inelastic dynamic analysis provides the most realistic and accurate measure for the seismic response, its application for large-scale structures is hampered by the excessive computational burden involved. This is particularly the case for the optimization of inelastic structures subjected to dynamic loads using metaheuristic algorithms where numerous analyses are required before the design converges to the optimum. In this regard, developing predictive models with sufficient accuracy will significantly help to reduce the computational demand, thus making the seismic analysis and optimization of large structures more feasible and common practice. Motivated by this need, this paper reports a study on the capabilities of a wavelet weighted least squares support vector machine (WWLSSVM) and a feedforward, backpropagation artificial neural network (ANN) to accurately predict the inelastic seismic responses of structures. The force- and displacement-based seismic responses of an 18-story reinforced concrete frame subjected to different earthquake ground motion records scaled to the design basis earthquake and maximum considered earthquake levels are used to train the models and examine their accuracies. The first three natural periods of the frame and combinations thereof are considered as the inputs for the model. The results indicate that both models exhibit satisfactory prediction performances, with the ANN model having a slight edge on accuracy in most of the cases studied, especially when a smaller number of samples are used for training. A parametric sensitivity analysis shows that the seismic responses predicted by the ANN model generally exhibit less sensitivity to the inputs than do those predicted by the WWLSSVM model. The results also indicate that force- and displacement-based responses exhibit the highest sensitivity to the first and second natural periods, respectively.