以XGBOOST演算法探討台北地區地震動預測研究

Abstract

在地震工程的研究中得知影響地震動的因素相當複雜，主要分為震源效應、路徑效應和場址效應。尤其是地震波進入盆地後，會受到盆地幾何形狀及其鬆軟質沉積物的影響，會產生明顯的地震波振幅放大現象。台北地區為地質構造特殊的盆地，因此台北地區的地震動反應為本研究的目標，在這項研究中主要探討兩個地震動預測目標，第一是地震動振幅，第二是場址放大因子。 傳統的地震動放大預測模型利用迴歸方程式進行分析，常用的預測參數包括 Vs30和 PGAr。對於具有盆地地形或不規則幾何形狀的場址，可能需要額外的參數來有效量化場址效應，但是這些參數不容易透過迴歸方程式進行量化分析。 隨著電腦科技的進步，利用機器學習來開發地震動的預測模型也成為一種有效的方法。本研究在眾多機器學習演算法中主要使用 XGBOOST 演算法，利用地震資料特徵和測站資料特徵建立數據庫，來建立地震動的預測模型進行預測，並與其他演算法的預測結果進行比較。機器學習與回歸模型相比，機器學習在預測時往往具有更好的準確性，但它也失去了線性模型的可解釋性。因此機器學習的可解釋性需要通過 SHAP 值(一種廣泛適用的解釋模型方法)來改善，並針對輸入-輸出依賴關係進行說明。最後比較機器學習和傳統迴歸分析所建立地震動的預測模型的分析結果。

In ground motion modeling, the influences from earthquake source, path, and site are usually considered separately. For sites with special geometry, such as basin, the wave propagation process can be complex as the ray path would be affected by the basin geometry and the properties of the soft sediments within the basin. The City of Taipei is located in a basin with special geological structure, so the site effect on the ground motion is always the topic of interest. Conventionally, empirical ground motion prediction models were developed using regression method. Typical prediction parameters for capturing site effect include the average shear wave velocity in the upper 30 meters (Vs30) and peak ground acceleration expected on rock site (PGAr). For sites with irregular topography or subsurface geometry, additional parameters may be required to effectively quantify the site effect. However, these parameters may not be easily identified. With the advancement of computer technology, the use of machine learning to develop ground-motion prediction models becomes plausible. In this study, the XGBOOST technique is used to develop models for predicting the ground motion amplitudes (such as peak ground acceleration and spectral acceleration at a particular period) and the amplification factors. The definition of amplification factor adopted in this study is the ratio of an intensity measure for a particular site to the same intensity measure for the reference site. Compared to the models developed by the regression method, the XGBOOST model tends to have a better prediction accuracy. The interpretability of the models developed by the XGBOOST method is achieved by examining the SHAP value, which can be used to explain the physical and quantitative interpretation of the input-output dependencies.