Research on multi-model prediction of skeleton curves of prefabricated concrete columns based on Residual fusion Long Short-Term Memory -Transformer

This study aims to investigate a deep learning-based model that permits quick evaluation of the seismic performance of prefabricated concrete columns under bidirectional seismic influences. To capture the time-series properties of the skeleton curves and simultaneously solve the difficulty of predicting the randomness of the initial points of the curves, two models, the fusion model ResMLP-LSTM-Transformer and the single ResMLP, are used to predict the skeleton curves for different concrete columns. For this objective, a skeleton curve dataset of 300 concrete columns with 14 physical properties of each column as input parameters and coordinates of each point of the skeleton curve as output was created. In addition, after utilizing 10-fold cross-validation and hyper-parameter tuning to the models, the predictions of each model were compared with curve results obtained from experiments and finite elements, and the accuracy of the predictions from each model was quantified using RMSE, MSE, MAE, MAPE, and R2. Finally, a thorough investigation was conducted into the impact of the ResMLP sequence length on the fusion model's accuracy. The results demonstrate that despite the diverse and variable physical characteristics of concrete columns, both the fusion model and ResMLP are capable of producing accurate predictions, with the fusion model having a better prediction performance. In addition, the evaluation indexes demonstrate that the fusion model's prediction error increases as ResMLP's sequence length increases, while the RMLT_3 model's prediction performance at a sequence length of three is optimal.