以隨機森林方式建立臺北盆地之地盤放大係數模型

Development of Ground Motion Amplification Model for Taipei Basin Using Random Forest Technique

徐爲恩(Wei-En Hsu)

指導教授 ： 郭安妮

國立臺灣大學/工學院/土木工程學研究所/碩士(2021年)

https://doi.org/10.6342/NTU202103510

摘要

隨機森林(random forest)模型是由Breiman和Culter在2001年提出之機器學習演算法。通過在森林中建構一棵棵各自獨立之決策樹(分類樹或迴歸樹)，並以投票方式(眾數或平均數)決定最終結果，以此提高模型之預測精度。該建模方式之優點包含運轉速度快、處理大量數據時表現優異、可提供特徵參數之重要性以及便於計算特徵參數之非線性作用，並提供特徵參數間之交互作用。
綜觀歷年地盤放大係數之研究，多數研究者於不同地區提供之迴歸方程式中所使用之參數不乏用以表示線性區間之VS30以及非線性區間之PGAr。由於難以有效將其餘對地盤放大係數產生影響之因子進行量化，致使該作法對於特殊地形或不規則地下岩盤之特定區域無法順利奏效。此外，由於研究區域乃半地塹之構造盆地，在特定入射方向上之震波可能會引致盆地中彈性波之建設性干涉而產生表面波，使得盆地內部分地點之主頻與震幅發生變化。有鑑於此，本研究將基岩深度與地表地形納入考量，透過自由場測站之加速度資料求得譜加速度，並使用TAP071測站作為參考測站以計算臺北盆地之地盤放大係數。接續彙整各項場址參數及地盤放大係數，分別作為數據庫之特徵參數與預測值，並使用Python作為建模環境以建立隨機森林模型。利用傅立葉分析觀察不同地震下之各測站頻域，比較傳統迴歸分析與隨機森林演算法之地盤放大係數預估值。

關鍵字

場址效應 ； 臺北盆地 ； 反應譜 ； 地盤放大係數 ； 隨機森林

並列摘要

Random forest is a machine learning algorithm proposed by Breiman and Culter in 2001. To determine the model that can describe the relationship between the feature parameters and the (dependent) output variables, independent decision trees (also known as classification trees or regression trees) are constructed in the forest, and voting (mode or average of the results from the decision trees) is performed. The advantages of this modeling method include fast running speed, excellent performance when processing large amount of data, provision of the importance of feature parameters and the interaction between the feature parameters and the output values, and facilitation of the calculation of the nonlinear effect of the feature parameters. and provide.
For most ground motion amplification models which were developed using the regression methods, typical prediction parameters include Vs30 (for capturing the linear part of the site effect) and PGAr (for capturing the nonlinear part of the site effect). For complex sites with special surface topography or irregular basement geometry, additional parameters may be needed to effectively quantify the site amplification. However, these parameters may not be easily identified by using the classification regression technique. In this study, ground motion amplification model is developed for the Taiwan Basin using the random forest technique. In the model development process, in addition to Vs30 and PGAr, numerous potential parameters that describe the basement geometry, surface topography and source location, are considered. The parameters that can represent the source (e.g. magnitude, rupture depth) and path (e.g. site-source distance) effects are also utilized. The ground motion amplification model developed in this study is based on the reference site approach. The station TAP071, which is a rock site, is selected as the reference site. Python, which has built-in random forest algorithm, is used to develop the models. To evaluate the effectiveness of the ground motion amplification model developed by the random forest technique, the predictions (for the ground motion amplification of the test set) based on the random forest model and the classical regression model are compared. In general, the random forest model is found to have a great predictive power. Also, it has the advantage of not having to specify the functional form as required in the classical regression technique.

並列關鍵字

Basin effect ； Taipei Basin ； amplification factor ； random forest

參考文獻

Biot, M. A. (1941). A mechanical analyzer for the prediction of earthquake stresses. Bulletin of Seismological Society of America, Vol.60, No.1, 29-61.

Borcherdt, R. D. (1970). Effects of local geology on ground motion near San Francisco Bay. Bull. Seismol. Soc. Am., 60, 29–61.

Borcherdt, R. D. (1994). Estimates of site-dependent response spectra for design (methodology and justification). Earthquake Spectra, 10, 617–653.

Borcherdt, R. D. and Gibbs, J. F. (1976). Effects of local geological conditions in the San Francisco Bay region on ground motions and the intensities of the 1906 earthquake. Bulletin of the Seismological Society of America, 66 (2), 467-500.

Breiman L. (2001). Random forests. Mach. Learn., 45, 5-32.