透過您的圖書館登入
IP:18.116.36.192
  • 學位論文

整合數值模擬與社群資訊於洪水預報最佳化之研究

Integrating Numerical Simulations and Social Media to Optimize Flood Forecasts

指導教授 : 石棟鑫

摘要


臺灣河川坡陡流急,當颱風來臨時,如何準確推估河川洪水為本研究探討的重點,包含洪峰時間點與洪峰水位值。根據不同的模式與大氣物理參數,在颱風發布警報時,中央氣象局、國家災害防救中心等多個單位合作,每六小時會進行一次定量降雨系集雨量預報,預報未來三天臺灣降雨的時空間分布,供災害預警。本研究以臺灣秀姑巒溪作為研究區域,蒐集降雨預報結果,整合降雨逕流模式HEC-HMS與水文數值模式WASH123D,根據流域水文與地文資料進行數值模擬,推估颱風時的洪峰時間點與洪峰水位值,同時蒐集社群資訊如社交媒體公開貼文、網路論壇討論、網路新聞報導、部落格文章等,利用長短期記憶時間遞迴網絡LSTM,推估颱風時的洪峰時間點,最後提出數值社群整合方法,找出各系集數值模擬與社群資訊權重,發展機率式洪水預報,對於侵臺颱風進行有效推估。 研究顯示無固定最佳的系集成員,且數值模擬經TensorFlow迴歸校正後,可大幅降低洪峰水位值誤差,卻仍無法有效推估洪峰時間點,不過颱風海上警報發布後,透過其後24小時的社群資訊聲量變化,利用長短期記憶時間遞迴網絡LSTM,社群資訊可有效推估洪峰時間點,最後可整合數值與社群進行機率式洪水預報。案例測試結果顯示,對於侵台颱風,採用「颱風」關鍵字的社群資訊,數值社群整合方法可有效推估河川的洪峰時間點與洪峰水位值,但對於颱風外圍環流影響之降雨型態如艾利颱風,則建議採用其他關鍵字,未來建議利用TensorFlow建立社群資訊篩選自動化機制,優化洪水預報。

並列摘要


Rivers in Taiwan with steep slopes have the characteristic of high flow velocity. As the typhoons hit, how to correctly predict peak time and peak stage of the river is the most important aim in this research. Every six hours during typhoon warning period, according to different models and setting of physical parameters, the government including CWB, NCDR produces rainfall forecasts of the next 3 days for disaster warning in Taiwan. Then, collecting hydrologic and topographic data of Xiuguluan River, the numerical simulations can get peak time and peak stage by integrating the physically based models HEC-HMS and WASH123D. At the same time, collecting the social media information such as facebook and twitter, discussions in PTT, network news and so on, the social media can get peak time by using LSTM constructed by TensorFlow. In the end, find the weights between numerical simulations and social media. The research can develop probabilistic flood forecasts for typhoons. The research shows there is no best ensemble member. Moreover, the numerical simulations can reduce the errors of peak stage by regression. As typhoon warning released, the social media of the first 24 hours help produce a more precise peak time by LSTM. Overall, the test results shows the method can produce precise peak time and peak stage of the river as the typhoon hit in Taiwan by using the suitable keyword of social media. In the future, it's recommended to construct an automatically screening system for social media by TensorFlow to optimize flood forecasts.

參考文獻


1. 行政院,2014。《103年災害防救白皮書》,臺北:行政院。頁20~29、頁32~46。
2. 行政院,2015。《104年災害防救白皮書》,臺北:行政院。頁24~31、頁34~35、頁45~46、頁235。
3. 行政院,2016。《105年災害防救白皮書》,臺北:行政院。頁28~31、頁46~49、頁83~86。
4. 行政院,2017。《106年災害防救白皮書》,臺北:行政院。頁16~19、頁32~35、頁112~125。
5. 行政院,2017。《行政院公報》。行政院,第023卷第093期,財政經濟篇。

延伸閱讀