探討石門集水區降雨-逕流機制與類神經網路洪水預報模式

Investigating Shihmen Watershed Rainfall-Runoff Mechanisms and Modeling Flood Forecasting by Artificial Neural Networks

林桂宏(Guei-Hong Lin)

指導教授 ： 張斐章

國立臺灣大學/生物資源暨農學院/生物環境系統工程學研究所/碩士(2012年)

https://doi.org/10.6342/NTU.2012.00786

摘要

掌握集水區颱風暴雨時期降雨-逕流機制是一個重要的議題，因此探討集水區降雨-逕流機制及建置一精確的流量預報有其必要性。本研究探討颱洪時期石門水庫集水區上游至下游降雨-逕流特性及建置其未來1~5小時流量預報模式。研究分成二大部分，首先為集水區降雨-逕流機制之探討，結果顯示: (1)流量歷線上升點至洪峰點，下游延時(20~25小時)較中上游延時 (15~20小時)長;洪峰點至平穩點，下游延時 (20~25小時)較中上游延時(15~20小時)長; (2) 相關性分析結果，大流量場次之降雨-逕流時間稽延較小流量場次之降雨-逕流時間稽延短(短1~2小時); (3) 各雨量站前30天累積雨量與流量站之基流量呈現正相關趨勢。本研究並利用模糊推論系統推論開始降雨至上升點延時，其推論結果與真實值趨勢一致。本研究第二部分為使用倒傳遞類神經網路(BPNN)建立流量預報模式，以流量/雨量資訊作為輸入項，同時亦考慮雨量站對流量站及流量站對流量站之不同的時間稽延關係，預報未來1~5小時流量，研究顯示除了以流量訊息作為輸入外，加入考量稽延的雨量資訊確實能提昇模式預測的精確性(RMSE約有20%之改善)。而退水段部分，利用迴歸分析尋找最佳迴歸函數，並將其預報結果與BPNN之流量預報結果比較，結果顯示利用線性迴歸預報退水段結果較類神經網路為佳，可有效降低類神經網路於退水段之震盪情形。
關鍵詞:降雨-逕流機制、模糊推論系統、倒傳遞類神經網路、多時刻流量預報

關鍵字

降雨-逕流機制 ； 模糊推論系統 ； 倒傳遞類神經網路 ； 多時刻流量預報

並列摘要

Investigating the rainfall-runoff processes over watersheds during typhoon periods is an important subject for water resources management. Therefore it is necessary to investigate the rainfall-runoff processes of watersheds in detail for establishing a precise flow forecast model. This study aims to investigate rainfall-runoff processes from the upstream to downstream areas of the Shihmen Reservoir watershed and then construct a multi-step-ahead flow forecast model for predicting the future 1-5 hour flow for the Shihmen Reservoir during typhoon periods. First of all, the characteristics of the rainfall-runoff processes in this watershed are investigated, results indicate: (1) based on the flow hydrographs of the areas from upstream to downstream in this watershed, it shows time lags at each stages are different: the time lag (20-25 hours) between the starting point of the rising limb and the peak point in the downstream area is longer than that (15-20 hours) in the up- and mid-stream areas while the time lag (20-25 hours) between the peak point and the starting point of the recession limb in the downstream area is longer than that (15-20 hours) in the up- and mid-stream areas; (2) according to the correlation analysis, the time lags between rainfall and runoff for big flow events are shorter than those of small flow events by 1-2 hours; and (3) it shows a positive correlation between the accumulative rainfall within a month prior to a typhoon event and the base flow of the flow station. In the second part, the artificial neural network (ANN), an effective data manipulation and prediction tool, is introduced in this study. The Back Propagation Neural Network (BPNN) model is developed for multi-step-ahead (1-5 hours) flow forecasting. Flow and rainfall data are used as the inputs to the BPNN. The time lags between rain gauge stations and flow stations and the time lags between flow stations are taken into consideration simultaneously as well. The results indicate that the inclusion of rainfall data, besides flow data, into the BPNN indeed helps to increase the accuracy of flow forecasts (the improvement rate for the RMSE is about 20%). In the recession limb, the regression analysis is used to find the optimal regression function and its forecast result is compared with that of the BPNN. The results indicate the regression analysis performs better than the BPNN in the recession limb, and the regression analysis can effectively reduce the vibration occurred in the recession limb for the ANN model.
Keywords: Rainfall-Runoff mechanism, Fuzzy Inference System (FIS), Back Propagation Neural Network (BPNN), Multistep-ahead flow forecasting.

並列關鍵字

Rainfall-Runoff mechanism ； Fuzzy Inference System (FIS) ； Back Propagation Neural Network (BPNN) ； Multistep-ahead flow forecasting

參考文獻

54. 鄭力嘉，2005，「地理資訊系統於石門水庫集水區降雨逕流分析之應用」，中原大學土木工程研究所碩士論文。

44. 許清炫，2005，「流域集水區降雨量與逕流量之分析與探討」，淡江大學水資源及環境工程學系碩士論文。

29. 李南慶，2007，「颱風影響集水區流域的降雨量與逕流量關係之研究」，淡江大學水資源及環境工程學系碩士論文。

52. 蔡文柄，2007，「應用類神經網路推估溪流之生物多樣性」，臺灣大學生物環境系統工程研究所碩士論文。

46. 陳永祥，2009，「演化式類神經網路於水文系統預測之研究」，臺灣大學生物環境系統工程研究所博士論文。