Estimating Design Floods using Machine Learning and Assessing Future Flood Risk considering Climate Change Scenarios

Abstract

With recent climate change, rainfall patterns became irregular with increased frequency of heavy rainfalls and guerilla rainstorms (a short, torrential rain), which leads to increased flood damage. In case of Korea, with the increase in rainfall during the summer rainy season and rainfall intensity, the country is hit by the substantial flood damage every year. In order to reduce such flood damage and secure the stability of hydraulic structures, development of reliable design flood estimation method is necessary. Despite the ongoing efforts to develop a design flood estimation method in Korea, it has been difficult to develop a standardized method that can be applied in practice due to the quantitative insufficiency of hydrologic data, in particular the flood data, and insufficient reliability Internationally, flood frequency analysis (FFA) and regional frequency analysis (RFA) based on flood data are often used in combination, and in Europe, flow data is shared and guidelines on RFA is prepared and utilized. In this regard, the trend of research is changing from point frequency analysis to RFA and from design rainfall-runoff analysis (DRRA) to FFA. However, research on the domestic application of RFA is still under progress, and there are limitations the applicability of RFA in that there are no standard for regional division in place and a methodology to taken into account the current status of rainfall data in Korea has not been established. In addition, DRRA is used in the comprehensive flood management plan of river basin, the largest-scale plan in the water resources sector in Korea, and FFA results are used for data verification. However, due to the tendency of overestimation in the probabilistic flood calculated by DRRA compared to the probabilistic flood calculated by FFA, there is a limit in its use as an adequate method of data verification. The hydrologic unit map provided in Korea shows the regional division of watershed, mid-sized basin, and basins, and has been developed for joint utilization in relation to the collection and analysis of data used in the water resources projects for the efficient operation of water resources development, planning and management. However, in the comprehensive flood management plan of river basin, the plans are established based on the confluences of streams rather than to the hydrologic unit map, and thus the applicability of the hydrologic unit map is limited. Therefore, it is necessary to develop flood estimation method considering the results of FFA method and DRRA method as well as the characteristics of watersheds taking into account its applicability in practice, and the risk analysis according to the climate change scenario is required. In this study, a probabilistic flood prediction model was developed with application of machine learning techniques, and risk analysis was performed according to the future climate change scenario. For the probabilistic flood prediction model, depending on whether to apply principal component analysis (PCA), general linear regression, decision tree, random forest, support vector machine, deep neural network, Elman recurrent neural network, and Jordan recurrent neural network were used. With the comparative review with the FFA analysis results of observed flood data, we have developed probabilistic flood prediction model. With the developed model, thirteen climate change scenarios were applied to compare and evaluate the change in the risks by future time period. The future flood risk of Korea was analyzed to increase by 0.53%(representative concentration pathway (RCP) 4.5) and 8.68%(RCP 8.5) on average than the current level of risk when 100-year frequency was used. Therefore, it is thought to be the time when the estimation of design return period of hydraulic structures for future climate change and accurate probabilistic flood estimation are required.