類神經網路在水庫操作的應用

Abstract

水資源為人類基本生存及社會發展的基本資源，在今天水資源不足的狀況 下，除開發新水庫以調蓄水源外，研究如何增加現有水資源的利用率亦為 重要的目標。 本研究主要分為兩部份，第一部份為利用模糊數學建立流量預測模式，第 二部份為利用倒傳遞類神經網路建立水庫操作模式。在流量預測模式中為 利用模糊數學中的模糊聚類分析法，對歷年之流量曲線做分類，以建立具 有不同流量特徵的流量曲線，再對已知的部份流量曲線，利用模糊識別法 和各已完成聚類的流量特徵曲線做識別，來預測未來的流量。第二部份為 利用類神經網路之學習的能力，以過去石門水庫歷年操作紀錄為訓練資料 ，建立一個類神經網路水庫操作模式，用以預測石門水庫石門大圳及下游 （包括桃園大圳）之放水量，並推算石門電廠的發電策略，並和實際之操 紀錄作一比較。在比較民國81、82年之實際操作紀錄和由類神經網路模擬 的操作方法中可知，在流量預測準確度在0.5以上時，由類神經網路模擬 的操作方法與實際的操作方法相當接近，證明類神經網路的確有學習到其 中的操作規則。 Water resources are essential to sustain human life and society development. Nowadays, the water resources in Taiwan area are not enough to meet demands especially when drought occurred. Hence it is important to improve the operational efficiency of existing reservoirs, other than to build new reservoirs. This research first presents the development of a streamflow forecast model based on fuzzy mathematics. The streamflow forecast model utilizes fuzzy clustering to classify historical streamflow curves into a number of patterns with different characteristics. This model also takes advantage of fuzzy recognition process to forecast future streamflow trends. The partially known streamflow curve is matched with the classified streamflow patterns. This model selects the most similar curve as the forecast forhe future streamflow. The second part of the research is to develop a reservoir operation simulation model that is based on artificial neural network. A back-propagation network (BPN) is utilized to learn the rules of reservoir operation from historical operation records. After the training, the reservoir operation can be performed with the rules learned by the BPN. The operation of Shihmen reservoir is used as the case study in this research. Comparisons between the past records and the neural network simulation results have been made. Sensitivity analysis shows that the model's operation performed with BPN is close to the recorded operations when the percentage accuracy of streamflow forecast is above 0.5.