Machine learning based design flood forecasting for the Ribb embankment dam in the Tana sub-basin, Abay River Basin, Ethiopia

Abstract

Design flood forecasting plays a significant role in assuring the safety of embankment dams. Worldwide, embankment dams have been prone to different collapses originating from hydrologic and hydraulic design errors, notably in design flood forecasts. Globally, the major problems related to embankment dam failures are overtopping, seepage, piping and instability due to horizontal water pressure issues. The aim of this study is to apply machine learning approaches to predict design floods for the Ribb embankment dam situated in the Abay basin, Ethiopia. Five machine learning methods, including random forest, gradient boosting, support vector machines, decision tree, and artificial neural network, were utilized on the study catchment. Historical annual maximum flows from two gauged catchments were utilized to obtain the target variable for the research. The Support Vector Machine model demonstrated superior predictive performance among the models evaluated based on three statistical assessment indicators. Specifically, the Support Vector Machine achieved a coefficient of determination of 0.9989 for training and 0.9914 for testing, a root mean square error of 1.867 m3/s for training and 2.143 m3/s for testing, and a mean average error of 1.403 m3/s for training and 1.849 m3/s for testing. Based on these results, design floods for multiple return periods were estimated and evaluated across the models. The findings demonstrate that the Support Vector Machine consistently outperformed the other machine learning approaches, confirming the capacity of such techniques to substantially enhance design flood prediction accuracy in data scarce regions. In particular, Support Vector Machine based forecasts provide reliable inputs for the safety assessment and hydraulic design of embankment dams in the Abay Basin.