使用權重組合模型預測雲林縣地層下陷;Using Ensemble Model Predict The Land Subsidence in Yunlin County

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/92643

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/92643

Title:	使用權重組合模型預測雲林縣地層下陷;Using Ensemble Model Predict The Land Subsidence in Yunlin County
Authors:	唐子傑;Tang, Zi-Jie
Contributors:	資訊工程學系
Keywords:	地層下陷預測;雲林縣濁水溪沖積扇;克利金插值法;機器學習;深度學習;長短期記憶模型;組合權重模型;Land subsidence prediction;Alluvial fan of Zhuoshuixi, Yunlin County;Kriging Interpolation;Machine learning;Deep learnging;Long Short -Term Memory;Ensemble model
Date:	2023-07-27
Issue Date:	2023-10-04 16:07:24 (UTC+8)
Publisher:	國立中央大學
Abstract:	地層下陷是一種不可逆的地理現象，世界各地發生的地層下陷已經嚴重影響了人身安全，而台灣也不例外，所以預測地層下陷並做到提前防治下陷成為目前重要的議題。在早期台灣地層下陷預測較著重在數值研究模擬，相較於數值模擬，人工智慧方法不需提供推估的水文參數和抽水量資料，可以直接透過給定的資料集進行建模，而其他應用人工智慧方式預測的研究也礙於紀錄資料不夠充足無法做到有效的預測，隨著紀錄資料越來越豐富和人工智慧技術越發進步，在此議題上更能夠取得不錯的效果。本研究使用克利金插值法(Kriging Interpolation)對 Global Navigation Satellite System(GNSS)自動固定站資料建立雲林縣全域的資料集和以測站為單位的資料集，分別測試雲林縣全域地層下陷敏感性模型對地層下陷的描述能力和預測數座需要特別關注的 GNSS 測站未來數周地層下陷量以每彌補測站測量資料時所需的時間延遲。在使用了七種特徵建立模型 – 地下水水位、濕度、氣溫、降雨量、日照時數、土地利用、地質組成，在雲林縣全域地層下陷敏感性測試中已經可建立有效描述真實下陷量的模型，指標中 R2 (決定系數 R 平方, Coefficient of Determination) – 0.954、Cor (決定系數 R 平方,Pearson Correlation Coefficient) – 0.979、MSE( 平均平方誤差, Mean Square Error) – 2.20E-05(單位:平方公尺)；特定 GNSS 測站預測中，使用 8 周訓練資料建立預測未來8周的地層下陷模型有最穩定的預測結果，組合模型中的 R2– 0.221 、Cor – 0.519、RMSE(平均平方根誤差, Root Mean Square Error) – 0.00207(單位:公尺)。;Land subsidence is an irreversible geological phenomenon that has had a significant impact on human safety around the world, and Taiwan is no exception. Therefore, predicting land subsidence and implementing early prevention measures have become important issues. In the early stages, land subsidence prediction in Taiwan focused mainly on numerical research and simulations. Compared to numerical simulations, artificial intelligence methods do not require estimated hydrological parameters and pumping data. They can directly model the given dataset. However, other studies that apply artificial intelligence methods for prediction have been hindered by insufficient recorded data, making it challenging to achieve effective predictions. With the increasing availability of recorded data and advancements in artificial intelligence technology, more promising results can be achieved in this field. This study utilized the Kriging interpolation to establish a dataset for the entire Yunlin County using Global Navigation Satellite System (GNSS) automatic reference station data, as well as station-based datasets. This study utilized the Kriging method to establish a dataset for the entire Yunlin County using Global Navigation Satellite System (GNSS) automatic reference station data, as well as station-based datasets. By using seven features, including groundwater level, humidity, temperature, rainfall, sunshine hours, land use, and geological composition, the model was able to effectively describe the real subsidence levels in the Yunlin County-wide land subsidence sensitivity test. The evaluation indicators were as follows: R 2 (Coefficient of Determination) – 0.954, Cor (Pearson Correlation Coefficient) – 0.979, MSE (Mean Square Error) – 2.20E-05 (unit: square meter). Regarding the prediction of specific GNSS stations, the model showed the most stable results when using an 8-week training dataset to predict subsidence for the next 8 weeks. The combined model yielded the following results: R2 – 0.221, Cor– 0.519, and RMSE (Root Mean Square Error) – 0.00207(unit: meter).
Appears in Collections:	[Graduate Institute of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	47	View/Open

社群 sharing

Loading...