Using support vector machines to improve reservoir inflow forecasting during typhoon-warning periods

本論文提出一種以支援向量機（Support Vector Machines, SVMs）為基礎的水庫入流量預報（Reservoir Inflow Forecasting）模式，支援向量機是一種新型的類神經網路（Neural Networks, NNs）。根據統計學習理論，支援向量機相較於傳統最常被使用的倒傳遞類神經網路（Back-Propagation Networks, BPN）有三個主要的優勢。第一，支援向量機具有較佳的能力。第二，支援向量機的最佳架構及權重保證會有唯一解，並為全域最佳解。第三，支援向量機具有更快速的學習效率。本研究利用18場颱風事件對倒傳遞類神經網路模式及支援向量機模式進行測試，測試結果很清晰的顯示出上述三項優勢。比起倒傳遞類神經網路模式，支援向量機有更好的預報準確度、更強健(Robust)且更為迅速。除了以支援向量機取代倒傳遞類神經網路之外，為了更進一步提升長時間的預報表現，颱風因子（Typhoon Characteristics）也被加入模式的輸入項。在以往的文獻中，颱風因子很少被當作水庫入流量的關鍵輸入項，本研究針對加入颱風因子與不加入颱風因子的模式表現進行比較，結果更加肯定颱風因子顯著地提升了長時間的預報表現。總結來說，颱風因子應該被當作颱風期間水庫入流量預報的輸入項。基於支援向量機的準確度、強健性及效率，本研究所提出的支援向量機模式可做為現有水庫入流量預報的替代模式，而本研究所提出的模擬技術對提升水庫入流量預報很有幫助。In this paper, effective reservoir inflow forecasting models based on the support vector machine (SVM), which is a novel kind of neural networks (NNs), are proposed. Based on statistical learning theory, the SVMs have three advantages over back-propagation netwoks (BPNs), which are the most frequently used convectional NNs. Firstly, SVMs have better generalization ability. Secondly, the architectures and the weights of the SVMs are guaranteed to be unique and globally optimal. Finally, SVM is trained much more rapidly. An application is conducted to clearly demonstrate these three advantages. The results indicate that the proposed SVM-based models are more well-performed, robust and efficient than the existing BPN-based models. In addition to using SVMs instead of BPNs, typhoon characteristics, which are seldom regarded as key input for inflow forecasting, are added to the proposed models to further improve the long lead-time forecasting during typhoon-warning periods. A comparison between models with and without typhoon characteristics is also presented to confirm that the addition of typhoon characteristics significantly improves the forecasting performance for long lead-time forecasting. In conclusion, the typhoon characteristics should be used as input to the reservoir inflow forecasting. The proposed SVM-based models are recommended as an alternative to the existing models because of their accuracy, robustness and efficiency. The proposed modeling technique is expected to be useful to improve the reservoir inflow forecasting.口試委員會審定書 i謝 ii文摘要 iiibstract iv一章 導論 1二章 理論方法 3-1倒傳遞類神經網路(BPN) 3-2支援向量機(SVM) 4三章 研究區域與模式架構 8-1研究區域與資料 8-2模式架構 9-3模式輸入項及參數設定 11-3-1輸入項確立 11-3-2參數設定 11-4交替驗證與評鑑指標 12-4-1 交替驗證 12-4-2 評鑑指標 12四章 結果與討論 14-1 BPN與SVM模式效能之比較 14-1-1模式的準確度 14-1-2模式的強健性 15-1-3模式的效率 16-2颱風因子對模式預報之影響 16-2-1 SVM-M1與SVM-M2的比較 16-2-2 雨量資料損毀 17五章 結論與建議 19考文獻 2