洪流之多變量時序分析

Multivariate Time Series Analysis of Flood Discharge

10.29417/JCSWC.199809_29(3).0003

陳昶憲(Chang-Shian Chen)；李姍燁(Shan-Yeh Li)

多變量模式 ； 向量模式 ； 時序分析 ； multivariate ARIMA models ； vector ARIMA models ； time series analysis

中華水土保持學報

29卷3期（1998 / 09 / 01）

199 - 209

繁體中文

於河川防洪分析，控制點一般皆定於人口聚集之中下游，故其集水區幾乎皆存在河系分佈，即除主流外，尚有其他支流匯入。因此，本文嘗試以多變量之向量時序分析，建立具河系分佈之集水區洪流推估時序模式，並由參數大小，分析上游各主支流對洪流推估之影響權重。文中以烏溪流域為分析實例，此流域除本流外，尚有大里溪及貓羅溪水系，本文探討下游控制點大肚橋及各水系上游測站間之流量時序關係，結果顯示在各測站之互饋影響下，空間上以溪南橋，時間上以大肚橋前一期流量影響最顯著。此外在流量歷線推估之趨勢掌握上，模式前進一階預測結果樣本判定係數（R^2）皆可達0.96以上，而代表準確度之效率係數（CE）則可達0.97，至於前進三階預測R^2仍可維持0.77，CE亦可維持0.77-故由模擬，預報結果，本文建立之多變量時序模式不但能有效的掌握流量變化趨勢，且其具備實用上之三階預測能力。

In flood-prevention analysis, control points are usually fixed on region with dense population. The mainstream in a watershed encomposed tributaries will be affected by these tributaries. In this paper, multivariate time series are ustilized to analyze the mainstream and tributary of the watershed. A infered time series model is bulit which has stream distribution, and parameter size analysis is employed to infer mainstream and tributary whose flood effect. Except mainstream, the Wu-Chi basin has Ta-Li Chi and Mao-Lo Chi as tributaries The downstream station of Da-Du bridge is used as a control point. Stochastis relationship of discharges between the control point and two tributaries were analyzed. The multivariate time series models are applied to examine weights of different parameters. The results of mutuality effect showed that station of the an bridge on space and station of the Da-Du bridge on time have more compicuous than other station in one step ahead discharge analysis. Additionally, in the tendency analysis, R-square value of the one-step ahead forecasting can attain 0.96, and coefficient of efficiency for accuracy can reach 0.965. For three-step ahead forecasting, R-square still could keep near 0.77, and coefficient of efficiency, 0.77. Considering the usefulness of forecast, multivariate time series model is obviously more superior to single-variable model on the tendency control. Being in capable of three-steps ahead forecasting multivariate time series model is more practical.