Applying Nonparametric Methods and Statistical Learning to Hydro-Climatic Analysis and Groundwater Level Prediction

• Main Authors: Meng-Kai Yu, 游孟楷
• Other Authors: 陳佳正
• Format: Others
• Language: zh-TW
• Published: 2017
• Online Access: http://ndltd.ncl.edu.tw/handle/19288944071062211655

碩士 === 國立中興大學 === 土木工程學系所 === 105 === Owing to the high applicability and robust performance of statistical methods, various hydrology-related applications have been regarded as a salient research track. This study aims to demonstrate the application of nonparametric methods and statistical learning for the analysis and forecasting of groundwater level in Pingtung alluvial plain in Taiwan. Data acquired in this region include groundwater level (151 monitoring wells), river flow (14 stations) and water level (5 stations), average daily temperature (2 stations) and rainfall (20 stations). This study consists of two major components: The first component is to adopt nonparametric methods to perform analysis of long-term yearly data, including using the Mann-Whitney-Pettitt test on detecting if there is a statistically significant change point in a yearly time series, and using the Mann-Kendall test in conjunction with the Sen Slope test on calculating the data trend. The second component is to apply support vector regression originated from Support Vector Machine (SVM) to forecasting monthly groundwater level in Pingtung alluvial plain. To examine optimal lead time, selected predictors are shifted from one to twelve months preceding groundwater level data. In the first component, trend analysis reveals that no significant trend is found in annual precipitation amount or daily intensity, whether in the dry or wet period, indicating the basic recharge to the alluvial plain stays very much invariant in the recent decades. By contrast, annual temperature and potential evapotranspiration exhibit significant increasing trends. Annual runoff amount shows a decreasing trend, which might be caused by the lack of low-flow data prior to 2000. Regarding groundwater level, overall, a decreasing trend is found in the study region. In the second component, SVM regression shows that sources of predictability of monthly groundwater level can be identified. However, two different selections of target monitoring wells, one based on the spatial geographical characteristics and the other based on the results of trend analysis, both lead to the same conclusion that lead times can be extended to at longest three months to issue skillful forecasts.