Search

In our intelligent society, water resources are being managed using vast amounts of hydrological data collected through telemetric devices. Recently, advanced data quality control technologies for data refinement based on hydrological observation history, such as big data and artificial intelligence, have been studied. However, these are impractical due to insufficient verification and implementation periods. In this study, a process to accurately identify missing and false-reading data was developed to efficiently validate hydrological data by combining various conventional validation methods. Here, false-reading data were reclassified into suspected and confirmed groups by combining the results of individual validation methods. Furthermore, an integrated quality control process that links data validation and reconstruction was developed. In particular, an iterative quality control feedback process was proposed to achieve highly reliable data quality, which was applied to precipitation and water level stations in the Daecheong Dam Basin, South Korea. The case study revealed that the proposed approach can improve the quality control procedure of hydrological database and possibly be implemented in practice. [ABSTRACT FROM AUTHOR]

With recent increases of heavy rainfall during the summer season, South Korea is hit by substantial flood damage every year. To reduce such flood damage and cope with flood disasters, it is necessary to reliably estimate design floods. Despite the ongoing efforts to develop practical design practice, it has been difficult to develop a standardized guideline due to the lack of hydrologic data, especially flood data. In fact, flood frequency analysis (FFA) is impractical for ungauged watersheds, and design rainfall–runoff analysis (DRRA) overestimates design floods. This study estimated the appropriate design floods at ungauged watersheds by combining the DRRA and watershed characteristics using machine learning methods, including decision tree, random forest, support vector machine, deep neural network, the Elman recurrent neural network, and the Jordan recurrent neural network. The proposed models were validated using K-fold cross-validation to reduce overfitting and were evaluated based on various error measures. Even though the DRRA overestimated the design floods by 160%, on average, for our study areas the proposed model using random forest reduced the errors and estimated design floods at 99% of the FFA, on average. [ABSTRACT FROM AUTHOR]