2009-08-012024-05-18https://scholars.lib.ntu.edu.tw/handle/123456789/704416摘要：類神經網路(artificial neural network)模擬水文過程(hydrological process)的潛力已經被大量的應用實例所肯定，然而，由於大部份的類神經網路模式缺乏物理機制(physical mechanism)，因此在某些水文問題的應用上遭致失敗。此外，傳統上使用試誤法(trial and error procedure)來建構類神經網路，不但相當耗時，使用上也不方便。所以類神經網路在水文問題的應用亟需一套能夠提升效能的方法，而本研究之目的就在於建立一套有效的方法，使得類神經網路效能提升。 本論文提出兩個概念，概念一是根據已知的物理機制來設計類神經網路，而概念二則是只用高度相關的輸入項來建構類神經網路。在前兩年的研究中，吾人以概念一分別來設計倒傳遞類神經網路(back-propagation neural networks)與幅狀基底函數網路(radial basis function networks)，並將其應用在地下水含水層參數檢定之問題。改良式類神經網路係根據已知的物理機制來設計，因此與現有的類神經網路最大的不同就在於輸入項與輸出項的設計。本研究將根據1000組隨機資料的測試結果，比較改良式類神經網路和現有類神經網路之效能，以展現改良式類神經網路之卓越。而在第三年的研究中，吾人則以概念二來建立降雨-逕流類神經網路模式。為了只保留高度相關的輸入項，因此本研究提出一套系統化的方法來消除不相關的輸入項。本方法所建構的降雨-逕流類神經網路模式將應用於翡翠水庫集水區，以顯示本方法比傳統上所使用的試誤法更具優點，本方法對於建立降雨-逕流類神經網路模式有很大的助益。 <br> Abstract: Artificial neural networks (ANNs) have found increasing applications in various aspects of hydrology and previous studies have shown the potential of ANNs for modeling hydrological processes. However, ANN models failed to be applied to some hydrological problems, because the ANN architectures are usually lack of physical mechanisms. In addition, the ANN models were constructed by a trial and error procedure, which is inconvenient and time-consuming. Hence applications of ANNs in hydrology need an approach that is capable of improving the performance of ANN models. The objective of this project is to establish an effective approach to the construction of ANN models in hydrosystems. In this project, two concepts for constructing ANN models in hydrosystems are presented. The first concept is to construct ANN models based on known physical mechanisms and the second concept is to construct adequate ANN models only included highly relevant inputs. In the first two years, two ANN models, back-propagation neural networks (BPNs) and radial basis function networks (RBFNs), based on the first concept will be established to determine aquifer parameters from pumping test data. The major difference between the existing and the proposed ANN models is the design of ANN input and output components. The proposed ANNs are designed according to the analytical solutions, which express known physical mechanisms. Testing the existing and the proposed ANN models by 1000 sets of synthetic data will be performed to demonstrate that our design of ANNs is better than the existing ANN model. In the third year, a systematic approach based on the second concept will be used to construct ANN rainfall-runoff models. In order to construct adequate ANN models that only include highly relevant inputs, the irrelevant inputs will be trimmed by the systematic approach. An application to the Fei-Tsui Reservoir Watershed in northern Taiwan will be performed to show that the proposed ANN rainfall-runoff model has advantages over those obtained by the trial and error procedure. The proposed approach will be helpful to hydrologist to construct adequate ANN-based hydrological models.類神經網路地下水參數檢定降雨—逕流模式neural networksdetermination of aquifer parametersrainfall-runoff model