A Study of Building Regional Flood Inundation Forecast Models by Integrating Clustering Analysis and Artificial Neural Networks

• Main Authors: Hung-Yu Shen, 沈宏榆
• Other Authors: Li-Chiu Chang
• Format: Others
• Language: zh-TW
• Published: 2013
• Online Access: http://ndltd.ncl.edu.tw/handle/24470786829075197462

博士 === 淡江大學 === 水資源及環境工程學系博士班 === 101 === In recent years, the increasing frequency and severity of floods caused by climate change and/or land overuse has been reported both nationally and globally. Therefore, estimation of flood depths and extents may provide disaster information for alleviating risk and loss of life and property. The conventional inundation models need a huge amount of computational time to carry out the high-resolution spatial inundation maps. Moreover, for implementing appropriate mitigation strategies of various flood conditions, different flood scenarios and the corresponding mitigation alternatives are required. Consequently, it is very difficult to reach real-time simulation and/or forecast of the inundation extent by using conventional inundation models. In order to build real-time flood forecast systems, this study proposed two models, CHIM and SOM-R-NARX, for forecasting regional flood inundation depths and extents. The CHIM (clustering-based hybrid inundation model) model has two-stage procedure, including data preprocessing and model building stages. In the data preprocessing stage, K-means clustering is used to categorize the data points of the different flooding characteristics in the study area and to identify the control point(s) from individual flooding cluster(s). In the model building stage, R-NARX (recurrent configuration of nonlinear autoregressive with exogenous inputs) flood depth forecasting models are built for control point(s) and regional grids in each cluster(s). The SOM－R-NARX model is composed of SOM (Self-Organizing Map) and R-NARX. The SOM network categorizes different flood inundation maps of the study area to produce a regional flood topological map. The R-NARX model is built to forecast total inundated volume of the study area. To find the neuron with the closest total inundated volume to the forecasted total inundated volumes, the forecasted value is used to adjust the weights (inundated depths) of the closest neuron and obtain a regional flood inundation map. The major difference between these two models is that CHIM classify flooding characteristics, and SOM－R-NARX investigate the relationship between rainfall pattern and flooding spatial distribution. The practicability and accuracy of the proposed methodology is evaluated in Yilan county. The results show that the two proposed models can provide 3-h-ahead flood inundation maps efficiently and adequately. For comparison, SOM－R-NARX consistently outperform CHIM and can be adequately applied to online multistep-ahead forecasts of inundation depths in the study area. However, SOM－R-NARX needs more storage for model parameter than CHIM, and increases the loading of database as well.