Machine Learning based 100-year flood flow prediction model using basin characteristics and multiple meteorological datasets in the northeast United States
Abstract
Estimating the value of extreme hydrological events, such as the 100-year flood at ungaged locations, remains a significant hydrological challenge. Conventional flood flow estimation techniques have significant shortcomings, including random and systematic errors that limit their effectiveness supporting decision making in water resources planning. This paper presents the use of random forests (RF), a machine learning (ML) technique that integrates multiple dynamic and static datasets to estimate the 100-year-flood flow and characterize the uncertainty of the model predictions. Inputs for the proposed model include precipitation, temperature, slope, watershed area, land cover, and elevation datasets. 98-gage locations over the northeast United States, with an availability of minimum 40 years of historic streamflow, are selected to evaluate the ML-based approach. This evaluation is based on 100-year peak flows obtained from U.S. Geological Survey measurements. A k-fold cross-validation technique is used to test the flood flow prediction model. The ML technique is shown to provide significant improvements in 100-year-flood estimates by substantially reducing systematic and random error. This estimation approach can support the accurate characterization of error in predicting 100-year-flood, which is essential in the development of flood flow prediction algorithms.
