利用系集決策法將氣象預報不確定性導入防災決策

Introducing Meteorological forecast uncertainty into disaster prevention decision Using Ensemble Decision Method

詹益淋(Yi-Lin Chan)

指導教授 ： 康仕仲

共同指導教授 ： 郭鴻基(Hung-Chi Kuo)

國立臺灣大學/工學院/土木工程學研究所/碩士(2019年)

https://doi.org/10.6342/NTU201903895

摘要

為了讓防災情資保有氣象預報的不確定性，本研究提出了系集決策法以輔助防災決策作業。系集決策法包含三個步驟，依序是建構防災情資數據模型，開發系集預警圖資，以及建立決策流程。在第一個步驟中，本研究直接使用氣象預報中各個系集成員的預報資料、降雨淹水門檻值以及防汛熱點進行災害模擬運算。在第二個步驟內，本研究開發基於資料視覺化的系集預警圖資，利用三種預警圖資來呈現災害模擬結果，包含縣市淹水預警威脅圖、鄉鎮淹水預警時序圖與多維淹水預警陣列圖。第三個步驟是建立決策流程，讓決策者可依循這個流程，依序進行威脅判斷、啟動評估以及確認應變強度三項作業。
為了驗證方法的有效性，本研究以2019年0517豪雨、0518豪雨與0520豪雨3場事件進行實驗，在臺灣設立淹水預警門檻值的363個鄉鎮市區裡面，3場事件的實際淹水數量分別為15個、1個與55個。驗證時以現有決定性預報及系集決策法達各門檻數量的一級淹水預警、及系集決策法達各門檻數量的二級淹水預警等三類方法進行比較。在淹水災情比較部份，利用淹水災情發生前24小時內的三次預報結果與實際災情做比較。決定性預報在0517豪雨與0518豪雨2場事件中無預警能力，在0520豪雨事件中的預兆得分分別是0.03、0.07與0.05。系集決策法在0517豪雨事件下的最佳預兆得分分別是0.13、0.44與0.71，其中051620初始場內的達4個二級淹水預警方法，其預兆得分是0.71，表示預警能力為71%，偏離指數是0.93，表示其接近無偏的預警，是本次研究統計資料內的最佳預警；在0518豪雨事件中各方法皆無預警能力；在0520豪雨事件中的最佳預兆得分分別是0.31、0.25與0.36。
在0517豪雨事件中，因降雨條件穩定，系集決策法隨預警時間逐漸接近實際淹水災情發生時間，對淹水災情的掌握度越來越好；在0518豪雨事件中，發生淹水災情的鄉鎮市區為1個，決定性預報與系集決策法對淹水災情的掌握度皆不足；在0520豪雨事件中，因降雨條件不穩定，加上淹水警戒無法充分表現實際發生淹水災情的鄉鎮市區，因此決定性預報與系集決策法下的各方法對災情掌握度皆不足。經過2019年的3場豪雨事件檢驗後，足可證明系集決策法產生的淹水預警較決定性預報所產生的淹水預警更能掌握實際淹水災情。

關鍵字

系集決策法 ； 氣象預報不確定性 ； 防災決策 ； 資料視覺化

並列摘要

This research proposes the Ensemble Decision Method to assist disaster prevention decision-making and to make the disaster prevention information keep the uncertainty of meteorological forecast. The Ensemble Decision Method included three steps: (1) build a disaster prevention data model to perform the disaster simulation operation; (2) develop an ensemble pre-alert map to visualize the ensemble alert map; and (3) establish a decision process to allow decision makers to conduct threat assessments, initiate assessments and confirm strain strength.
In order to verify the effectiveness of the method, this research carried out experiments with 0517 heavy rains, 0518 heavy rains and 0520 heavy rains in 2019. In the 363 townships where the rainfall threshold was set up in Taiwan. The actual flood disaster amount of the three events was respectively 15, 1 and 55. In the experiment, we used existing decisive forecast, various ensemble decision method for reaching the first-level flood pre-alert, and various ensemble decision method for reaching the second-level flood pre-alert. The results of the three predictions within 24 hours before the flooding occurred were compared with the actual disaster situation. The decisive forecast had no pre-alert ability in the 0517 heavy rain and 0518 heavy rain events, and in the 0520 heavy rain event, The threat score of decisive forecast was respectively 0.03, 0.07, and 0.05. In the 0517 heavy rain event, the best threat score of the ensemble decision method was respectively 0.13, 0.44 and 0.71. In the initial field of 051620, The threat score of the method which four modes reached the second-level flood pre-alert of the ensemble decision method was 0.71, indicating the pre-alert capability was 71%. The bias score was 0.93, indicating that it was close to unbiased pre-alert. This was the best pre-alert data in this research. The ensemble decision method had no pre-alert ability in the 0518 heavy rain. in the 0520 heavy rain, the best threat score of the ensemble decision method was respectively 0.31, 0.25 and 0.36.
In the 0517 heavy rain event, due to the rainfall conditions was stable, as the time was close to actual flood disaster occurred, the ensemble decision method had better mastery of the actual flood disaster amount. In the 0518 heavy rain event, there was one township with flood disaster. the decisive forecast and the ensemble decision method had insufficient mastery of flood disaster. In the 0520 heavy rain event, due to the unstable rainfall conditions and the flood warnings could not fully represent the actual flood disaster, the decisive forecast and the ensemble decision method had insufficient mastery of flood disaster. After the three heavy rain events in 2019, it is sufficient to prove that the flood pre-alert issued by the ensemble decision method can better grasp the actual flood disaster than the flood pre-alert issued by the decisive forecast.

並列關鍵字

ensemble decision method ； uncertainty of meteorological forecast ； disaster prevention decision ； data visualization

參考文獻

Bauer, M., & Johnson-Laird, P. 1993. How diagrams can improve reasoning. Psychological Science, 4: 372-378.

Bertin, J. 1981. Graphics and graphic information-processing. Berlin, Germany: Walter de Gruyter.

Burkhard, R. A. 2004. Learning from architects: The difference between knowledge visualization and information visualization. Paper presented at the 8th International Conference on Information Visualisation, London, UK.

Crapo, A. W., Waisel, L. B., Wallace, W. A., & Willemain, T. R. 2000. Visualization and the process of modeling: A cognitive-theoretic view. Paper presented at the 6th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Boston, MA.

Ebert, Elizabeth E., 2001: Ability of a Poor Man's Ensemble to Predict the Probability and Distribution. Mon. Wea. Rev., 129, 2461–2480.