Research On Explainable Prediction Model For Site Response Based On SHAP And Machine Learning

Site seismic response analysis is one of the important research contents in seismic engineering,and it is an important basis for evaluating the seismic safety of engineering sites.At present,one-dimensional soil seismic response analysis methods are mainly used in engineering,but research has shown that there are significant errors in the calculation results of existing soil seismic response analysis methods on soft soil and strong seismic.Based on the horizontal seismic records of more than 40 stations in the KiK-net network,this paper uses machine learning to establish two prediction models for peak ground acceleration(PGA)and surface ground acceleration response spectra.The prediction results of machine learning are evaluated based on measured records and the calculation results of four numerical simulation methods.At the same time,SHAP(SHapley Additive exPlanations)is used to analyze the impact of each input feature of the machine learning model on the prediction results,and the reliability of the model’s prediction results is judged based on physical laws and existing experience.The main work and research results are as follows:1.Based on the KiK-net network,more than 40 horizontal site stations are selected.And combined with the soil profile information provided by the strong motion stations,11 commonly used and easily obtainable parameters representing site and ground motion characteristic are selected.Then two datasets are established for predicting PGA and surface acceleration response spectra.2.Using 14 feature selection methods to rank the importance of 11 features and obtain a comprehensive score,10 feature combinations were given based on their importance from high to low.The feature combination that maximizes the model evaluation index is selected as the input feature for the model The XGBoost algorithm is used to establish a PGA prediction model based on the 4 input features.At the same time,decision tree and random forest algorithm are used to predict PGA,and the results of three machine learning prediction models and four numerical simulation methods are compared and analyzed.The results show that the prediction results of machine learning models are better than that of numerical simulation methods,and the prediction results of XGBoost are better than that of random forest and decision tree.3.Using random forest for feature importance ranking and feature selection,the optimal feature combination of the model is screened,and the multi output random forest prediction model of surface ground acceleration response spectra is established.Regression model evaluation indicators and DTW(Dynamic Time Warping)distance are used to evaluate the calculation results for test dataset,different site classes,and different seismic intensity.From the evaluation indicators of the regression model and the DTW distance,the overall prediction performance of the machine learning model is better than that of numerical simulation methods.4.Using SHAP to interpret and analyze the PGA prediction model and surface acceleration response spectra prediction model.The prediction results are visualized from three perspectives:feature global analysis,feature dependency analysis,and single sample analysis.The impact mechanism of each input feature on the model prediction results is analyzed.The results indicate that the influence of input features on output results is consistent with objective laws,which verifies the reliability of the models prediction results and the rationality of the machine learning methods in this paper.