- 學位論文
結合機器學習及主成分分析建置都市區域淹水多時刻預測模式-以臺北市為案例
Explore Machine Learning and Principal Component Analysis to Construct a Multi-Step-Ahead Urban Flood Forecasting Model - Taking Taipei City as a Case Study
摘要
受到氣候變遷與都市化的影響，近幾年洪水災害有頻率增加、災害程度更加劇烈的趨勢，為減輕災害的損失，各國政府對於洪災預警與災害應變更加重視，以提早針對災害發生地區進行應變，藉此降低災害所造成的影響。 本研究蒐集近年來臺北市實際淹水事件之雨型，透過SOBEK二維淹水模式，模擬實際暴雨事件與設計暴雨事件(共51場)，得到2047筆模擬淹水資料，每筆資料包含45101個10m×10m的網格淹水深資料。使用主成分分析、自組特徵映射網路(SOM)與非線性自回歸外因輸入模式(R-NARX)針對抽水站集水區建立都市區域淹水預測模式；以淹水模擬資料進行主成分分析，透過四個主成分值代表不同淹水空間分布的特性；SOM根據各時刻淹水特性進行聚類分析，將淹水模擬網格資料映射到二維拓樸圖上，各神經元可以表示不同淹水大小與淹水空間分布之狀況；R-NARX以回饋項加上雨量資料作為輸入資料，分別建立以10分鐘為時距之未來一小時預測模式(T+1~T+6)，並且以兩種模式進行預測結果比較，模式一為預測平均淹水深的單輸出模式，模式二則是預測平均淹水深與主成分值的多輸出模式；最後之整合模式先將R-NARX預測結果比對SOM拓樸圖各神經元，模式一選擇與預測平均淹水深最相近的神經元，模式二選擇各主成分值最相近的神經元，以及被選總次數最高神經元，並將所選神經元依照不同比例計算各網格之權重，模式一與模式二選出神經元後，以平均淹水深預測結果微調各網格權重值，藉此得到區域淹水預測結果。 本研究分析結果顯示9×9拓樸圖大小較能完整呈現不同淹水空間分布；R-NARX建立之預測模式，比較模式一與模式二於平均淹水深之預測結果，可發現模式二之R2數值較小，模式二預測較難以掌握平均淹水深之趨勢，顯示多輸出模式因訓練時根據多個輸出項誤差調整參數，相較於單輸出模式預測結果較差；由模式二各主成分預測結果可知，部分主成分值因趨勢與平均淹水深較不同，因此預測模式較難以掌握其趨勢，於各預測時距之R2數值較小。比較模式一與模式二之整合模式結果，模式二於各預測時距之結果相較於模式一RMSE數值較小；並以SOBEK模擬淹水深與T+1模式一、模式二之預測淹水深進行比較，結果顯示模式一使用平均淹水深篩選神經元，較適合用於全面降雨的颱風事件，模式二使用各主成分值篩選神經元，因具有淹水空間分布特性之指標，較適合用於降雨空間分布不均的豪雨事件，相較於模式一有較穩定準確的淹水空間分布預測結果。 本研究結果顯示以主成分值代表淹水空間分布特性，並結合SOM與R-NARX模式，能夠掌握不同降雨空間分布所造成之淹水情形，並且即時提供都市區域的淹水預測，可以幫助決策者提前針對預測淹水地區進行應變。
並列摘要
Due to climate change and urbanization, the frequency and severity of flood disasters have increased in recent years. In order to reduce damage losses, governments in the globe have paid high attention to flood warnings and disaster response so as to respond to floods earlier for reducing the impact of disasters. This study collected the rainfall patterns of flood events in Taipei in recent years. The SOBEK model was implemented for making 2-D inundation simulation of actual rainfall events and designed ones (51 events in total), where a total of 2047 datasets were generated and each datasets contained 45101 grids of inundation depths in the study area for illustrating the rainfall-flooding process. This study combines Principal Component Analysis (PCA), Self-Organizing Map (SOM), and Nonlinear Autoregressive with Exogenous Inputs (R-NARX) to establish urban flood forecast models. The PCA was performed on flood inundation simulation data to obtain four principal components representing the different spatial distributions of inundation. According to the characteristics of flood events, the SOM was used to cluster the gridded inundation simulation data into the neurons of a two-dimensional topological feature map, where each neuron in the topological map represented the average depth and spatial distribution of inundation grids. The R-NARX model used its feedback value and rainfall data as inputs to establish a flood forecast model for the next hour, with a forecast horizon of 10 minutes (i.e. T+1 – T+6). Two R-NARX models were established: Model 1 made the forecast of the average inundation depth; and Model 2 made the forecasts of the average inundation depth and principal components. Then, we compared R-NARX results with the average inundation depth of each neuron in the SOM topological map. Model 1 selected the neuron with average inundation depth the closet to the forecasted average inundation depth. Model 2 selected the neurons the most similar to the four principal components and the neuron with the highest frequency of being selected. In each model, the weight of each selected neuron was adjusted by its forecasted average inundation depth. Consequently, the regional inundation depth forecast can be obtained. Comparing the forecasted results of the proposed model, the RMSE value of Model 2 at each horizon is smaller than that of Model 1. The inundation depth simulated by SOBEK was compared with those forecasted by Model 1 and Model 2 at T+1. The comparative results show that Model 1 using the average inundation depth to select the neuron is more suitable for forecasting flood inundation caused by evenly-distributed rainfall such as typhoon-induced rainfall. Model 2 that uses the principal components to select neurons is more indicative of the spatial distribution characteristics of inundation, and therefore is more suitable for forecasting flood inundation caused by torrential rainfall events (uneven spatial distribution of rainfall). The results show that Model 2 produces a more stable and accurate forecast of the spatial distribution of inundation than Model 1. The results of this study demonstrates the proposed approach that integrates PCA (for characterizing the spatial distribution of flood inundation) with the SOM and R-NARX models (for forecasting flood inundation depths) can grasp the inundation status caused by the different spatial distribution of rainfall to provide real-time flood inundation forecasts at urban areas. The proposed approach can help decision makers respond to floods earlier and reduce the impact of disasters.
參考文獻
延伸閱讀
- 林筱翊（2017）。應用空間迴歸分析方法於都市洪汎風險評估與防災決策分析-以台北市為例〔碩士論文，中原大學〕。華藝線上圖書館。https://doi.org/10.6840/cycu201700335
- 陳樹群、吳俊毅、謝政道（2012）。崩塌危害分析模型之建立－以臺北水源特定區為例。中華水土保持學報，43(4)，332-345。https://doi.org/10.29417/JCSWC.201212_43(4).0006
- 紀柏全、沈哲緯、冀樹勇、黃立遠、林士淵、潘信元（2014）。都會區住宅邊坡崩塌潛感評估模式初探－以臺北市為例。中華水土保持學報，45(4)，217-224。https://doi.org/10.29417/JCSWC.201412_45(4).0001
- 顧嘉安、劉家彤（2020）。探討都市土地利用型態與淹水潛勢之空間關聯－以原臺中市為例。都市與計劃，47(1)，27-52。https://doi.org/10.6128/CP.202003_47(1).0002
- 陳姿伶、林子閎（2023）。Exploring the Relationship between Risk Indicators and Flooding Events-A case study in New Taipei City, Taiwan。都市與計劃，50(1)，57-84。https://doi.org/10.6128/CP.202303_50(1).0003