建構集水區多時刻降雨-逕流機制與類神經網路洪水預報模式

Abstract

颱風暴雨時期之水庫操作為水資源管理中相當重要的一環，因颱風往往造成大量洪水在短時間內自集水區匯流進入水庫中，若能提供精確之洪水預報資訊，則可有效提升水庫操作策略之可靠性，故水庫入流量之預報訊息，對於水庫操作的決策工作，具有極高的參考價值，因此本研究將以颱洪時期石門水庫入流量之多時刻預報為研究目標；研究內容主要分為兩部份，第一部份探討集水區降雨-逕流特性，針對雨量及入流量資料進行趨勢檢定，採用Pearson及Kendall相關係數以瞭解降雨及逕流歷線之間的線性及非線性關聯，並探討降雨的延遲特性。第二部份則根據第一部份之分析結果，加入相關性較高的雨量資訊，作為類神經網路輸入項，並以調適性網路模糊推論系統建立水庫入流量於颱風時期之多時刻預報模式；相關性分析結果顯示石門水庫集水區之平均雨量與入流量約有5∼7小時之稽延時間，而流量預報模式以預報時刻前5∼7小時之雨量及最近3小時之入流量作為輸入項，有最佳的預報結果，並可有效架構至未來1∼5小時之流量預報。對於多時刻的流量預報，除了以流量訊息作為輸入，加入適當的雨量資訊確實能提昇模式預測的精確性。

Reservoir operation is a prerequisite component for water resource management during typhoon periods. Due to the floods caused by typhoons often reach the reservoir within few hours, the need for a higher reliability of reservoir operation strategies is required. Therefore, an accurate hydrological model plays a key role and provides valuable flood information for reservoir operational decisions. The major purpose of this study is to construct a stable and reliable multistep ahead inflow forecasting of Shihmen reservoir. First of all, the characteristics of rainfall-runoff processes in this watershed are investigated by using trend detection methods for finding the relation between rainfall and runoff. In the second part, the artificial neural network (ANN), an effective data manipulation and prediction tool, is introduced in this study. The Adaptive Network-based Fuzzy Inference System (ANFIS) with model inputs consist of previous precipitation and streamflow information is developed for multistep ahead flood forecasting. The results indicate that both precipitation and flood patterns have the same trend with a shift of 5-7 hours. Accordingly, the ANFIS model provides accurate and effective flood forecasts to 5 hours-ahead. As far as the multistep ahead flood forecasting is concerned, appropriate combination of precipitation information indeed help to increase the accuracy of flood predictions.