http://scholars.ntou.edu.tw/handle/123456789/20001
Title: | 應用機器學習技術於淹水預測之研究 | Authors: | 張雅惠 | Issue Date: | Oct-2019 | Abstract: | 本申請者數年前與水理專家合作研發洪氾預警系統,以預測即將淹水的地區,在計畫執行中得知相關的水理模組需要大量的資料與複雜的運算,導致費時甚久且無法避免誤差。另一方面,資料探勘或機器學習的技術,近年來已經有很多令人耳目一新的應用。所以,在本計畫中,我們希望將成熟的人工智慧與資料處理等相關技術,應用於淹水預測的研究中。具體來說,我們將蒐集近年來完整的淹水資料,大規模的分析哪些區域易於淹水,並將其分類。我們同時也將蒐集與淹水相關的氣象觀測資料,然後針對各個不同淹水類型的區域,探索其間的相關性,以建立對應的分類預測器,未來協助進行淹水預測與洪氾預警。在此研究中,針對資料的時間特性,由於淹水現象通常與一段時間的降雨有關,所以我們以時間軸將觀測資料和淹水資料切割,但保留其前置時間的觀測資料以表示其連續性。至於資料的空間特性,我們先行將空間切割成網格,以網格為單位分析其淹水情況,然後進一步將相鄰且同類的網格加以合併,並找出每類型的代表性網格,以降低後續進行機器學習的資料量。我們將根據真實的氣象觀測資料和淹水資料建置淹水預測模式,並利用不同的歷史資料和水理模式的推算結果加以驗證比對,以確認模式的可行性。我們相信以真實歷史資料為主的觀點配合機器學習的技術,將會對淹水預測的相關研究帶來新的貢獻。 Previously, we have cooperated with several hydraulic experts to develop a flood forecasting system for predicting the area to be flooded. We have observed that it takes a lot of time for the hydraulic modules to get the results since it requires complicated computations on a lot of data. On the other hand, the techniques of data mining and machine learning have succeeded in many applications. In this project, we propose to apply the techniques of artificial intelligence and data processing on the task of flood prediction. Specifically, we will collect the real flood data and meteorological data to identify their relationship. For the temporal properties of these data, we will represent them based on the timestamp, but consider the meteorological data observed within the lead-time 1-6 hours. For the spatial properties, we will first divide the space into a set of fix-sized cells, determine the flood situation of each cell and classify all cells into several groups accordingly. We then further merge adjacent cells and identify the representative cells for each group, to reduce the amount of data which need to be processed. We will build a classifier for each group of cells and use different history data to verify the effectiveness of the proposed model. |
URI: | http://scholars.ntou.edu.tw/handle/123456789/20001 |
Appears in Collections: | 資訊工程學系 |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.