颱風時期水庫系集流量預報之發展

Development of ensemble reservoir inflow forecasts during typhoon periods

張鈞凱(Chun-Kai Chang)

指導教授 ： 林國峰

國立臺灣大學/工學院/土木工程學研究所/碩士(2020年)

https://doi.org/10.6342/NTU202002066

摘要

受到氣候變遷影響，預報準確度在近年來更加重要。在水庫管理方面，提前的蓄洪、洩洪的預防性減災遠優於致災後的處理，水庫操作多仰賴準確的入流量預報，現行模式僅使用觀測降雨和入流量預報未來6小時水庫入流量，但在近年因極端天氣響，需要更長時間的入流量預報，提供給決策者更充裕的調控空間。此外，現行水庫操作大多使用定率式降雨預報，而定率式預報容易給決策者確定性的假象，故近年來提倡系集雨量預報以涵蓋雨量的不確定性，以供決策者評估預報的風險。然而機率式預報在使用上無法明確給使用者指南，通常會額外選擇以簡單平均法(Ensemble Mean , EM)來整合數量眾多的系集預報結果，但簡單平均法會將極端值平滑化。
本研究提出交替預報(Switch Prediction Method, SPM)整合系集降雨產品，並與EM及各系集產品進行比較，同時提供機率式預報和定率式降雨預報，當作降雨逕流模式輸入項，預報未來72小時水庫入流量。本研究採用在水文領域上表現良好的支援向量機(Support Vector Machine, SVM)預報水庫入流量，結合多步階預報(Multi-Step Forecasting, MSF)，透過反覆迭代的過程預報出未來72小時水庫入流量。
本研究以石門水庫集水區證明模式的準確性，搜集2004年至2019年共18場颱風事件的雨量資料與入流量資料。經比較各系集產品介接水庫入流量模式過後，SPM在整合系集資料方面表現良好。在SVM t+1模式方面，預報值貼近觀測值並沒有明顯峰值誤差，SVM結合MSF模式在預報未來72小時水庫入流量，其評鑑指標表現良好，可供水庫洩洪、蓄洪操作決策者參考使用。

關鍵字

水庫降雨逕流模式 ； 交替預報 ； 支援向量機 ； 系集降雨預報 ； 多步階預報

並列摘要

Affected by climate change, the forecast accuracy is more important in recent years. Reservoir operations rely on accurate inflow forecasts. Previous studies used observed rainfall and inflow to forecast reservoir inflows for 1– to 6–h lead times. However, in recent years, the long lead time forecasts are required to provide decision-makers with ample time for regulation and control due to the impact of extreme climates. Also, most of the current reservoir operations use the deterministic rainfall forecast which gives the illusion of certainty to decision-makers. Therefore, in recent years, it has promoted an ensemble rainfall forecast to cover the rainfall uncertainty for decision-makers to assess the risk of forecasts. However, the ensemble forecast cannot be used as a guide for users. The ensemble mean (EM) is conventionally used to integrate a large number of ensemble forecast results, but the EM smooths out extreme values.
This study adopts a switch prediction method (SPM) to integrate ensemble rainfall products, and compares with EM and each ensemble forecast. The rainfall forecasts are then used as input to the rainfall-runoff model. The support vector machine (SVM) combined with multi-step forecasting (MSF) is used to forecast reservoir inflow for 1– to 72–h lead times through the iterative process.
Rainfall and inflow data of the Shihmen Reservoir for 18 typhoons from 2004 to 2019 are employed to demonstrate the advantages of the proposed methodology. The SPM performs well in integrating the ensemble rainfall products. Regarding the SVM t+1 model, the forecasts are close to the observed and there is no obvious peak error. The SVM combined with MSF performs well in forecasting the long lead time (72 hours) inflow.

並列關鍵字

Reservoir rainfall-runoff model ； Switch prediction method ； Support vector machine ； Ensemble rainfall forecast ； Multi-step forecasting

參考文獻

1. Buizza, R., Palmer, T.N., 1998, Impact of ensemble size on ensemble prediction, Monthly Weather Review, Vol. 126, pp. 2503–2518.

2. Chen, Y.H., Chang, F.J., 2009, Evolutionary artificial neural networks for hydrological systems forecasting, Journal of Hydrology, Vol. 367, Issues 1–2, pp. 125–137.

3. Cristianini, N., and Shawe-Taylor, J., 2000, An introduction to support vector machines and other kernel-based learning methods, Cambridge university press.

4. Ebert, E. E., 2001, Ability of a poor man's ensemble to predict the probability and distribution of precipitation, Monthly Weather Review, Vol. 129, pp. 2461–2480.

5. Fahimi, A., Baktir, M. A., Moghadam, S., Mojabi, F. S., Sumanth, K., McNerney, M. W., ... and Salehi, A, 2017, Physical exercise induces structural alterations in the hippocampal astrocytes: exploring the role of BDNF-TrkB signaling, Brain Structure and Function. Vol. 222, pp. 1797–1808.