多時刻河川流量系集預報模式

Abstract

台灣因地理位置特殊，每年平均約有3.5個颱風會侵襲台灣，加上台灣的特殊地形，使得河川坡陡流急，河川的集流時間相當短，集水區降雨往往於數小時內就會抵達水庫。藉此，如果能於颱風期間，提供數小時後水庫入流量的資訊，有助於水庫操作人員有更充分的時間做應變。 本研究採用地面雨量站之雨量資訊及雷達雨量資訊(QPESUMS)分別探討曾文集水區之降雨逕流特性，結果顯示集水區上游至下游的逕流延時一致為兩小時具有最高相關性，並根據此結果作為模式輸入之依據，以雷達雨量資訊(QPESUMS)和當時刻流量作為模式輸入建構類神經網路之多時刻水庫入庫流量預報模式，結果以BPNN、ANFIS及RNN三種類神經網路所建置之流量預報模式具有優異的表現，其中BPNN在1~3小時內預報有最好的結果；ANFIS則對於流量峰值的預測具有最好之結果；而RNN在長時距4~6小時的預報效果最為優異，此部分結果顯示不同理論之類神經網路模式具有不同之優點。 系集預報模式起源於大氣領域，目前於大氣領域當中已有相當多的應用，而水文模式對於系集預報之應用目前尚在發展當中，本研究以系集預報之概念，整合前述三多種流量預報模式之結果，以倒傳遞類神經網路建置系集流量預報模式，並結合敏感度分析找出各流量模式於各時刻所代表之權重，透過系集預報整合之後提供更穩定及更精確之水庫入流量預報模式。

Due to unique geographical location of Taiwan, an average of 3.5 typhoons attack Taiwan each year. In addition, the particular topographical terrains of Taiwan make rivers short and steep such that rivers rapidly flow from catchments to reservoirs within a few hours during typhoon events. It will be very helpful and useful for reservoir operation management if reservoir inflow information can be provided in the next few hours after typhoons initially arrive. This study investigates the rainfall-runoff process of reservoir catchment by using two different rainfall information: rain gauge precipitation data and radar rainfall data (QPESUMS: Quantitative Precipitation Estimation and Segregation Using Multiple Sensors). The Zhengwen Reservoir catchment is the study area. The correlation analysis results show that it takes only two hours for rainfall to travel from catchment to reservoir in the study area. Therefore, this study aims to build up multi-step-ahead reservoir inflow forecast models through artificial neural networks based on QPESUMS and inflow information. The results indicate that all the BPNN, ANFIS and RNN models have excellent estimation performance. The BPNN model performs the best for one- to three-hour-ahead forecasts, while the RNN model has the best performance for four- to six-hour-ahead forecasts. The ANFIS model is superior to the other models for peak flow forecasts. The results demonstrate that each neural network has its own distinct advantages from others. Ensemble forecasting was originated from Atmospheric sciences and has been developed for years. In this study, we build up an ensemble forecast model by incorporating the outputs of three constructed forecast models into the BPNN to produce multi-step-ahead reservoir inflow forecasts, and further conduct the sensitivity analysis to summarize the weights of individual models incorporated in the ensemble forecast model for each time step. The results demonstrate that the ensemble forecast model can provide more reliable and accurate multi-step-ahead reservoir inflow forecasts than individual models incorporated.