博碩士論文 108522086 完整後設資料紀錄

DC 欄位 語言
DC.contributor資訊工程學系zh_TW
DC.creator王振綱zh_TW
DC.creatorZhen-Gang Wangen_US
dc.date.accessioned2021-7-26T07:39:07Z
dc.date.available2021-7-26T07:39:07Z
dc.date.issued2021
dc.identifier.urihttp://ir.lib.ncu.edu.tw:88/thesis/view_etd.asp?URN=108522086
dc.contributor.department資訊工程學系zh_TW
DC.description國立中央大學zh_TW
DC.descriptionNational Central Universityen_US
dc.description.abstract地層下陷在世界各地發生,已然造成許多的災害,濁水溪沖積扇亦深受其害,為了減緩地層下陷帶來的影響,預測地層下陷便成為重要任務;台灣先前地層下陷預測的研究主要著重在數值模擬模型,文獻上較少發現應用機器學習方法於地層下陷預測的研究,相比於數值模擬模式建立時須提供水文地質參數、補注量與抽水量資料,機器學習方法可直接對給定資料集建模;本研究欲建立濁水溪沖積扇地區建立地層下陷預測模型,先後探討輸入模型資料的時間尺度、特徵重要性、不同模型效能差異、特徵影響半徑,最後再建立空間解析度100 x 100 (單位公尺)的地層下陷預測模型;本研究使用機器學習方法的隨機森林(Random Forest)和長短期記憶(Long Short-Term Memory)神經網路,觀測井地下水位、雨量、溫度與濕度作為特徵,深層樁與Global Navigation Satellite System (GNSS) 為地層下陷資料,為使資料趨勢平穩,地下水位與地層下陷資料以一階差分計算變化量。實驗一與實驗二的結果顯示,透過兩種評估模型預測值與真實值偏離程度的指標,Root Mean Square Error (RMSE)最低可達4.27(單位mm),Mean Absolute Error (MAE)則是2.97(單位mm),亦即當隨機森林模型以90天為訓練時間尺度,30天為訓預測時間尺度時,模型的預測表現會最好,此結果可用於決定模型輸入的特徵與輸出的水文地質資料的維度,同時,根據實驗二之中的特徵重要性實驗,地下水位特徵為地層下陷問題裡最重要的特徵,而在地下水位特徵中,第二與第三含水層的重要性最高。zh_TW
dc.description.abstractLand subsidence (LS) occurs all over the world and has brought many disasters to people. Choushui River alluvial fan is also suffered from it. Thus, in order to mitigate the impact of LS, prediction of LS becomes an important goal. Previously, researches of LS prediction in taiwan were mainly focused on numerical model, while machine learning method was seldom applied. In comparison with numerical model, machine learning method don′t necessarily need hydrogeological parameters when building model, it can be directly built with given dataset. This study aims to establish a LS prediction model of Choushui River alluvial fan area. We discussed the time scale of data fed into the model, importance of features, performance of different models, influence radius of feature, and establishment of a model with a spatial resolution of 100 x 100 (in meters). This study uses Random Forest (RF) and Long Short-Term Memory (LSTM) of machine learning method to build the model. Several hydrogeological data is regarded as feature. Deep leveling pile and Global Navigation Satellite System data as LS data. In order to stabilize the data trend, the groundwater level and LS data are processed by first-order difference. According to the results of experiment one and two, RF model has the best performance when training and prediction time scale are 90 and 30 days respectively, with two measurements which are used to calculate the deviation between predict and true value, the Root Mean Square Error (RMSE) can be as low as 4.27 (mm) and Mean Absolute Error (MAE) is 2.97 (mm). The time scale result can be used to determine the dimension of the input features and output LS result of the model. Besides, the feature importance experiment shows that the groundwater level is the most important feature. Among the groundwater level feature, the first confined and second confined aquifers are the most important.en_US
DC.subject地層下陷預測zh_TW
DC.subject濁水溪沖積扇zh_TW
DC.subject機器學習zh_TW
DC.subject隨機森林zh_TW
DC.subject長短期記憶zh_TW
DC.subjectLand subsidence predictionen_US
DC.subjectChoushui River Alluvial Fanen_US
DC.subjectMachine learningen_US
DC.subjectRandom Foresten_US
DC.subjectLong Short-Term Memoryen_US
DC.title利用機器學習方法基於多類型地層監測資料預測濁水溪沖積扇地區之地層下陷zh_TW
dc.language.isozh-TWzh-TW
DC.titlePrediction of Land Subsidence in Choushui River Alluvial Fan Area Using Machine Learning Methods Based on Multiple Types of Ground Level Monitoring Dataen_US
DC.type博碩士論文zh_TW
DC.typethesisen_US
DC.publisherNational Central Universityen_US

若有論文相關問題,請聯絡國立中央大學圖書館推廣服務組 TEL:(03)422-7151轉57407,或E-mail聯絡  - 隱私權政策聲明