2001-08-012024-05-17https://scholars.lib.ntu.edu.tw/handle/123456789/690487摘要：水資源係天然寶貴的資源，不但能提供都會區與工商業區的大量供水與水力發電外，而且能夠供給密集農業的灌溉水、涵養各類生物以維持生態系統的平衡，因此，水量的多寡影響了水資源的營運與規劃；是故，流量推估與預測在水資源的開發與管理上成為關鍵性的工作之一。然而在流量推估與即時預測方面，由於流域內降雨－逕流間的因果關係，係受氣候、地質、地形、土地利用情況、植物覆蓋等複雜因素所影響，致使其水文特性兼具空間與時間的非線性變異。因此，如何建立合適的水文模式來模擬這種特性，一直是水文研究者的目標。1995年由Sirosh所發展的互動連結自我組織類神經網路(LISSOM)，包括輸入層的縱向傳入連結與神經元間的橫向連結，其理論建立於大腦的視覺皮質層神經發展過程的基礎上，視神經的發育與成長是透過視覺經驗的學習使得皮質層的傳入神經形成平滑的地形圖、儲存視覺世界中主要特徵向量，同時，橫向連結神經儲存這些特徵向量間的相關性；此一模式成功地運用在向量最小距離分類、自動與異質關聯記憶以及聚類合成。本研究在初步探討中，以台灣高程圖驗證LISSOM模式具有描述輸入向量分佈的特性及其關聯性。惟此依網路僅利用輸入資訊無法直接應用於<br> Abstract: Water is one of our most precious natural resources. It drives all human systems (ex. public water supply, hydroelectric energy, and irrigation systems) and other ecosystems as well. The availability of water significantly influence water resource planning and management. Consequently, streamflow estimation and/or forecast are vitally importance works for water resource development and management. In order to accurately forecast the streamflow, one must have a good command of the causality between the rainfall and runoff. Nevertheless, this relation is very complicated. It is a lumped effect of climate, geology, geography, land use conditions, and plant types, as a result, the hydrological characteristics are nonlinear and great variability both on time and space domains. How to construct a suitable hydrological model to simulate this characteristic is constantly a major goal of hydrologist.The laterally interconnected synergectically self-organizing map (LISSOM) was first introduced by Sirosh (1995). The mapping algorithm convert patterns of m dimensionality into two dimensional arrays of neurons. Each neuron contains two types of connection, i.e. the afferent and lateral connections, respectively. The theorem is based on visual cortex learning process. The networks have been used successfully in many areas, such as auto- and heteroassociative memory, and cluster synthesis. Our preliminary study has also demonstrated that the LISSOM could represent the topography of Taiwan by (through) properly describing the main characteristics of input vector distribution and their association. In this stage, the major shortcoming of LISSOM is not sufficient for continuous function reproduction and simulation. This is primary due to the model does not have output layer.The main purpose of this study is, hence to develop a hybrid learning scheme which will incorporate the unsupervised learning scheme of LISSOM with supervised learning method to build a new type of artificial neural network. This method not only can maintain the topological characteristics of input vector through LISSOM, but also can derive the project relation of output vector. As a result, the relationship of input-output pairs would be constructed. The scheme is intended to be used for modelling the watersheds rainfall-runoff relation. The data of streamflow records in Dai-Chi will be used as case study to evaluate the models availability and effectiveness.流量推估即時預測互動連結自我組織圖類神經網路非監督式學習監督式學習幾何形態關係laterally interconnected synergetically self-organizing mapsupervised learningunsupervised learning