| Summary: | 博士 === 國立臺灣大學 === 生物環境系統工程學研究所 === 95 === The major purpose of this dissertation is to effectively construct artificial neural networks-based multi-step-ahead flood forecasting using radar and numerical weather prediction information. To achieve this goal, three investigations by using neural networks for rainfall estimation and/or rainfall-runoff process simulation have been performed to explore their accuracy and applicability. The first topic investigates the model forecasts through static and dynamic neural networks by using four sets of training data which consist of different sample sizes and contents. Performance of these two types of networks suggest that the dynamic neural network generally could produce better and more stable forecasts than the static neural network, and the static model could produce satisfactory results only when sufficient and adequate training data are provided.
The second topic focuses on the evaluation of effectiveness and stability of three neural networks-based multi-step-ahead forecasts in terms of model structures. The results indicate that a neural network with a serial-propagated structure can help in improving the accuracy of forecasts. This concept not only provides a possibility of finding better solution for multi-step-ahead forecasts but enhances the predictive reliability. Results from above two studies are further utilized in the third topic which is to construct a precise and feasible multi-step-ahead flood forecasting. For better multi-step-ahead flood forecasting, there is a necessity to conduct the predicted meteorological information. Therefore, an improved quantitative precipitation forecasting is obtained from a merging procedure that combines radar-derived predictions and precipitation forecasts extracted from a numerical weather prediction model.
The comparison of multi-step-ahead flood forecasting derived from the serial- propagated structure and the merged precipitation prediction is made by estimating the timing and the percent error of a predicted peak flow relate to observed peak flow and the corresponding improvement. Based on the comprehensive comparison, the merging procedure successfully demonstrates the capability of efficiently combining the information from both rainfall sources and improves the accuracy of 1-6 h precipitation predictions. For multi-step-ahead flood forecasting, an important finding is the hydrologic responses seem not sensitive to the precipitation predictions in short lead times (in our case 1 to 3 hours) but dominate by previous runoff information, whereas the model forecasts are highly dependent on predicted precipitation information for lead time greater than 3 hours. Overall, the results strongly demonstrate that accurate and stable multi-step-ahead flood forecasting can be obtained from a serial-propagated structure and enhanced by the proposed precipitation predictions.
|
|---|
