動態系統與類神經網路之研究及其應用於降雨–逕流預報模式之整合

Abstract

本研究之目的係探討狀態空間降雨–逕流模式之系統識別，並結合線性動態理論與智慧型控制理論發展狀態空間類神經網路降雨–逕流預報模式。最後研析合適之狀態空間類神經網路生成法，並應用於流域之洪水預報。 模擬水文循環系統中之降雨–逕流歷程為一相當困難之工作。為考量精簡演算過程及增進模式之適用性，本研究應用動態系統理論以研析水文模式之轉換系統，並採用間接系統檢定方法，對水文歷程作深入之探討。文中進一步結合類神經網路發展出具狀態空間特性之狀態空間類神經網路模式，採用整合多種遞迴式類神經網路演算法後所得之統合演算法進行模式參數訓練學習之工作，以期即時更新、校正模式，並對模式參數之變化作深入之探討。一般水文模擬結果之好壞端賴模式之架構及參數之正確性，因此狀態空間類神經網路之生成法有其研究之重要性。本研究研析間接系統檢定法與子空間檢定法之優劣，並深入探討架構模式之過程，期冀能提高模擬降雨–逕流歷程之精確度。 研究中選取基隆河中上游五堵集水區民國55年至86年間颱洪事件之記錄降雨與逕流資料，分析定率性降雨–逕流模式之機制。間接系統檢定法乃依據最佳化理論求得系統之單位歷線，進一步估算狀態空間方程式與觀測方程式之參數矩陣，以確知系統之轉換過程。最後，狀態空間類神經網路生成法之研析過程中，考慮間接系統檢定法與直接子空間檢定法進行系統識別。兩種系統檢定法皆採用奇異值分解之數值運算。而藉由有系統之測試瞭解間接系統檢定法與直接子空間檢定法之優缺點後，本研究提出結合兩種檢定法優點之狀態空間類神經網路生成法。研究中選取部分歷年來颱洪事件之記錄降雨與逕流資料，訓練檢定生成之狀態空間類神經網路並進行模式之驗證。本研究所採用之狀態空間類神經網路生成法及獲致之成果，期冀可提供臺灣集水區防洪規劃及水土保持研析之參考應用。

The purposes of this study are to discuss the system identification of a state space rainfall-runoff model, and to integrate linear dynamic theory with intelligent control theory to develop a state space neural network rainfall-runoff forecasting model. Finally, an appropriate generation of state space neural network is established, and the model is applied to the flood forecasting of a river basin. Modeling the rainfall-runoff process is always a difficult task in whole hydrological cycle. For the purpose of simplifying calculation and enhancing the adaptability, dynamic systems theory is adopted to explore the transformation system of the hydrological model. The hydrological process is analyzed by using the indirect system identification. Furthermore, dynamic systems theory is integrated with neural netoworks to develop a state space neural network model, and a unification of the algorithms for recursive neural networks was applied to update the parameters of the model by training and learning process. However, the performance of simulation depends on the model’s correct structure and its accurate parameters. Therefore, it is important to study the approach to generate state space neural network. Finally, a comparison between indirect system identification and subspace identification is achieved, and the process of constructing the model is also discussed in this study. The deterministic model is calibrated and validated by the rainfall-runoff records of selected typhoon events occurred at the upstream watershed of Wutu, Keelung River Basin between 1966 and 1997 via indirect system identification that estimates the parameters of state space model based on the unit hydrograph derived from the constrained deconvolution step. Finally, a new approach that combines indirect system identification with direct subspace identification to generate state space neural networks is proposed in this study and results achieved herein can play a referential role on the planning of flood mitigation in Taiwan.