以決策樹與卷積神經網路驗證颱風路徑模擬

Verifying Typhoon Track Simulations using Decision Trees and Convolutional Neural Networks

蕭聖禾(Sheng-Ho Hsiao)

指導教授 ： 王人牧

淡江大學/工學院/土木工程學系碩士班/碩士(2021年)

https://doi.org/10.6846/TKU.2021.00823

摘要

淡江大學風工程研究中心致力於研究風對於建築物的受力關係，而在台灣颱風則是較大的風力來源，有鑑於現今颱風資料不足，為進行風力規範之修訂需進行大量颱風的模擬，將生成之颱風分為中央氣象局所定義的九類路徑為目的，AI人工智慧是目前各行業正熱門的部分，本研究利用機器學習與深度學習建立模型以便分類處理模擬生成的大量數據進而驗證颱風模擬之成效。
研究範圍主要以資料前處理，訓練資料調整和分類模型精進為重點，測試資料為中央氣象局的歷史數據，以模型準確率、混淆矩陣等分類指標做判斷，預測資料為淡江大學風工程研究中心的模擬數據，以卷積神經網路與決策樹的兩種系統測試效果做比較，驗證模擬數據的可信度，並比較兩種模型的差異。
本研究的主要概念是將中央氣象局從1975年到2020年期間影響台灣的颱風資料進行收集，範圍為北太平洋經度118° E ~ 126° E以及緯度 19° N ~ 28° N，之後再分成以決策樹(Decision tree)與卷積神經網路(Convolutional Neural Networks)以監督式學習的訓練模式來分類及辨識颱風路徑，進而建構一套可以自動分類颱風路徑的程式從而取代人工分類。
本研究決策樹模型經增加訓練資料的數量與多元性後，在預測中央氣象局的資料時準確率已到80%，卷積神經網路參考108年梁啟納所建構的模型，預測中央氣象局的資料準確率可到79%，但兩模型在特定幾類的準確率仍須提升，在預測淡江大學風工程研究中心模擬生成的數據也由一定成效，但在某些路徑容易出現混淆的現象，這部分仍須加入更多研究改善。
除了改進目前模型與比較兩模型之外，本論文之後也能融入其他風工程相關分析的應用中，將所有前處理、訓練到預測的部分以程式自動化，使得能讓之後生成的資料能得到快速且準確的分類，甚至未來能建立一套能自我分類的系統。

關鍵字

卷積神經網路 ； 深度學習 ； 類神經網路 ； 預測模型 ； 影像分類 ； 決策樹

並列摘要

Wind Engineering Research Center (WERC-TKU) studies on the interaction between wind and buildings. Typhoon is the most significant source of wind force in Taiwan. Due to the lack of typhoon data, a large number of typhoon simulations is needed to revise the wind specification. The generated typhoon is divided into nine types of paths defined by the Central Weather Bureau (CWB) for the purpose of the study. Artificial intelligence (AI) is a hot topic in the industry. Machine learning and deep learning are used to establish models to classify and process large amounts of simulated data to verify the effectiveness of typhoon simulation.
The scope of the study focuses on data pre-processing, training data adjustment and classification model refinement. Test data are historical data from the Central Weather Bureau. Model evaluations are judged by classification indexes such as model accuracy and confusion matrix. The prediction data are the simulation data of the Wind Engineering Research Center of Tamkang University. The test results of convolutional neural network and decision tree are compared to verify the credibility of the simulation data and compare the differences between the two models.
The main concept of this study is to collect CWB typhoon data that affect Taiwan from 1975 to 2020. Typhoons passing through 118 ° E ~ 126 ° E longitude and 19 ° N ~ 28 ° N latitude of the North Pacific Ocean are classified and identified by decision tree and convolutional neural networks in supervised learning training mode. Then a program that can automatically classify typhoon tracks is constructed to replace manual classification.
After increasing the number and diversity of training data, the decision tree model in this study has reached 80 % accuracy in predicting the data of CWB. The convolutional neural network predicts 79 % accuracy of data from CWB, based on the model developed by Leung Kai Nap in 2019. However, the accuracy of the two models in specific categories still needs to be improved. The classification of data generated by the typhoon path simulation model of WERC-TKU is also effective, but it is prone to confusion in some paths, and more research needs to be added to improve this part.
In addition to improving the current models and comparing the two models, this thesis can also be integrated into the application of other wind engineering related analysis. Programmatically automate all pre-processing, training to predicting tasks so that later generated data can be quickly and accurately classified, and even a self-grouping system can be established in the future.

並列關鍵字

Deep Learning ； neural networks ； Convolution Neural Network ； Decision Tree ； predictive model ； image classification

參考文獻

[1] 梁啟納(2019)。以卷積神經網路進行颱風路徑分類，淡江大學土木工程學系碩士論文。

[2] Quanshi Zhang , Yu Yang , Haotian Ma , and Ying Nian Wu , “Interpreting CNNs via Decision Trees”

[3] Hyeong-Seog Kim, Joo-Hong Kim, Chang-Hoi Ho, and Pao-Shin Chu, “Pattern Classification of Typhoon Tracks Using the Fuzzy c-Means Clustering Method”

[4] Dmitry Laptev, Joachim M. Buhmann ETH Zurich, “Convolutional Decision Trees for Feature Learning and Segmentation”

[5] Conner Chyung , Michael Tsang , Yan Liu , “Extracting Interpretable Concept-Based Decision Trees from CNNs”.