時間與空間最佳化資訊於降雨逕流模式分析之研究

Abstract

極端降雨事件與其對應之洪水預報在學術上一直為相當具有挑戰性的議題。由於台灣地區獨特的地形條件與氣候型態，對於雨量與洪水量預報的需求更為迫切。雨量估計已存在有多種降雨觀測產品，例如地面雨量站、雨量筒、雷達估計雨量、衛星影像推估雨量等，各自有其獨特的空間與時間特性。若能整合地面雨量站、雷達觀測雨量及衛星影像之推估雨量，提供高精確度之雨量之時空間變化趨勢，便能有效增加降雨預報模式之精確度。因此透過對於各種雨量資料的了解與評估，使用適當之雨量資料，以資料驅動模式(data-driven model)進行預報，並透過輸入資料的時空間整合流程，建置最佳預報模式為本研究之重點。 本研究以石門水庫集水區為例，首先蒐集中央氣象局QPESUMS系統提供之雷達雨量產品、美國加州大學水文氣象及遙測中心建置之PERSIANN-CCS衛星觀測系統提供之雨量產品，以及集水區之地面雨量觀測紀錄，以倒傳遞類神經網路方法校正不同觀測資訊之推估誤差，接著利用遺傳演算法融合地面雨量、雷達及衛星影像推估雨量三種資訊；再以ANFIS架構降雨預報模式，預測未來1及2小時之降雨。結果顯示QPESUMS經由倒傳遞類神經網路校正後，大幅改善校正前降雨推估誤差；而PERSIANN-CCS在雨量誤差校正方面，由於模式建置是以大陸型氣候為訓練背景，因此使用於台灣海島型氣候時產生較大之系統推估誤差。另經由GA進行最佳融合權重搜尋，其融合雨量即t時刻之推估雨量與實際降雨量相當接近，相關係數高達0.99，並且以融合了3種雨量資訊的降雨預報模式表現最佳，可證明融合雨量之有效性。 由於臺灣河川坡陡流急，逕流滯留時間短暫，上游洪水往往在數小時內便抵達下游及注滿水庫。在這種特殊的降雨-逕流特徵下，為期能減輕洪水災害，掌握「時間」該項關鍵因素，發展一精確多時刻流量預報模式、提供未來水庫入流量資訊及替決策者爭取更多時間實有其必要性。本研究使用不同雨量資訊及流量資訊分別探討集水區降雨─逕流關係，並利用DEM資料，在符合集水區物理特性條件下，依據高程、坡度資訊將集水區分別劃分成1、4、8及12個子集水區，進行空間時間整合；接著設計6組不同輸入變數方案(S1、S2、S3(n), n=1,4,8,12)，使用ANFIS架構流量預報模式，預報未來1~5小時石門水庫入流量。由降雨─逕流分析結果可知降雨與水庫入流量之延遲時間約為5~7小時，當水庫入流量預報時間為未來1小時，6個方案幾乎無差距，然而當流量預報延時大於3小時，加入雷達雨量資訊對於水庫入流量預報之精確度有相當顯著之改善；另在畫分集水區個數上，以劃分4個子集水區之模式表現最佳，其對於流量預報模式而言不會過於複雜或簡單，且在長延時有最好的預報能力，顯示石門集水區適合簡單的空間整合，而使用12個雨量站資訊雖有較多的雨量資訊輸入，但因為單點式資料，缺乏集水區整體而全面之資訊，無法有效提高雨量預報精確度。 由於颱風暴雨具有空間變異特性，其所造成之逕流量間之關係必非以簡單線性相關即可描述，故本研究以非線性方法應用於石門水庫入流量之預報，先利用SOM得到最佳的空間整合方式，接著應用2階段GT演算法，選擇最佳稽延雨量資訊。由結果可知2階段GT演算法可有效率的降低電腦計算量，大量減少計算時間，且可挑選出最佳輸入變數組合；使用(SOM+2階段GT)模式在流量預報時間為t+3至t+4時，相關係數仍可高於0.94及CE值高於0.88，且預報結果在t+1小時預報上可做到洪峰到達時間零延遲，t+3小時與t+4小時預報上，洪峰則皆僅有1~2小時的延遲，這在使用類神經網路架構入流量預報模式上是非常顯著的改善；最後本研究由使用者(或決策者)的角度出發，除了單一的模式預報值，亦提供預報值的可靠度，使用機率的概念呈現預報結果。

Accurate forecasting of extreme rainfall event and its corresponding flow is still a challenging issue for most of hydrologists. Due to unique topographic feature and weather pattern in Taiwan, this issue is even more critical. Thus, there is an urgent need to develop an accurate forecasting of rainfall and discharge. The major aims of this study are two-folds. First, the study is to compare various rainfall products, such as rain-gauge measurement, radar rainfall, and rainfall estimation from satellite imagery, and evaluate the ability of merging different combiuations of rainfall products to improve rainfall forecasting using an artificial neural network model. Secondly, different approaches for spatio-temporal lumping of radar rainfall are proposed here to evaluate the rainfall-runoff relationship using a data driven model for inflow forecasting in Shihmen reservoir. In this study, ground measurements and a radar rainfall dataset (QPESUMS) provided by Center Weather Beural (CWB) and a satellite-based rainfall dataset (PERSIANN-CCS) are collected. A BP model was developed to calibrate the estimation errors of the QPESUMS and PERSIANN-CCS, respectively. After calibration, a Genetic Algorithm (GA) is applied to merge these three rainfall datasets. The merged rainfall is further used as the input of an ANFIS model for rainfall forecasting at 1 and 2 hours horizons. The results showed that the BP effectively reduces the estimation error of QPESUMS dataset while only limit improvement can be made for PERSIANN-CCS. The reason for this may due to the PERSIANN-CSS was developed for Continental-scale climate modeling and may not be able applied directly to an island-scale climte pattern in Taiwan. After merged by GA, the merged rainfall has very high correlation with actural rainfall and has best performance for rainfall forecasting. With a better understanding of these rainfall products, the next focus of this study is to evealute inflow forecasting using proper rainfall dataset. Flood forecasting is an extremely crucial non-structural approach for real-time reservoir operation in Taiwan due to its unique topographical features and heterogeneous typhoon patterns. As a result of steep slope and short rivers in Taiwan, a flash flood occurs typically within few hours and reservoirs could easily and quickly be filled up with mass inflow in a typhoon event. Such conditions make real-time reservoir operation very challenging and reveal an urgent need for efficient and accurate multi-step-ahead inflow forecasting models. This study utilizes different rainfall datasets, such as rain-gauges and QPESUMS, and inflow data to evaluate the rainfall-runoff relationship. The spatial-lumping of QPESUMS is based on terrain analysis using DEM data and aggregrates the catchment into 1, 4, 8, and 12 sub-catchments. Six input strategies (S1, S2, S3(n), n= 1,4,8,12) were designed for a ANFIS model to forecast inflow of Shihmen reservoir at 1 to 5 hours horizons. From correlation analysis, it reveals that the time of concentration is about 5-7 hours in the catchment. For one hour forecasting, there is no significant difference between 6 strategies; while for 3 hours horizon, the improvement of using radar dataset is quite clear than using gauge-based rainfall. The spatial lumping to 4 sub-catchments has optimal performance in long-term, longer than 3 hours, inflow forecasting. These results suggest that using point-based ground measurements fails to catch spatial information in the catchment and leads to poor results of inflow forecasting; while simple spatial aggregration, 4 sub-catchments, of radar rainfall is more suitable for ANFIS model than complex spatial aggregration, 12 sub-catchments. The above inflow forecasting results may be further improved by using a non-linear spatio-temporal lumping approach. Here, the spatial lumping method based on terrain analysis using only DEM is replaced by a non-linear clustering method, Self-Orgnizied Map (SOM) using DEM and radar rainfall of typhoon events. The linear correlation analysis is replaced by a 2-staged Gamm test approach which is proposed in this study and is an efficient method to select non-trival input combination for ANFIS forecasting model. This novel spatio-temporal lumping method is termed as SOM+2-staged GT. There are several advantages in flow forecasting. First, it has best forecasting results for 3 and 4 hours horizons with correlation coefficient (CC) as high as 0.94 and coefficient of efficience (CE) as 0.88. Secondly, for one hour ahead forecasting, there is no time-lag between estimated and observed inflow peak; while for 3 and 4 hours horizons the time-lags are typically less than 2 hours. This is a remarked improvement in an inflow forecasting model based on ANFIS. From the perspective of end-users (or decision makers), this study suggested a confidence level of inflow forecasting using a pre-determined threshold of forecasting error. The confidence level of forecasts is presented by the percentage of forecast errors that fall within the designed error threshold.