Location: Hydrology and Remote Sensing LaboratoryTitle: Examination of multiple predictive approaches for estimating dam breach peak discharges Author
|Hood, K. - U.s. Military Academy|
|Hromadka, T. - U.s. Military Academy|
Submitted to: Journal Hydrologic Engineering
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 8/16/2018
Publication Date: N/A
Interpretive Summary: Predicting flood peaks originating from earthen dam failures is difficult because of the sparse nature of observations from this kind of rare event. This work employs an augmented database created from the aggregation of multiple small datasets relating to dam failures. Multiple traditional (regression-based) approaches are examined along with more innovative approaches based on machine learning algorithms that focus attention on most similar observations to an arbitrarily-chosen test location. Goodness-of-fit is quantified using relative standard error and relative bias. Calibration methods are examined and the importance of prediction being limited to within the convex hull of observations is discussed. Among the innovative methods examined, the Region of Influence method is observed to perform poorly while the k-Nearest Neighbor method using a simple arithmetic average is observed to perform well.
Technical Abstract: A database joining individual earthen dam breach failure studies is assembled and re-analyzed across all aggregate observations. Conventional regression methods are employed along with newer predictive approaches for making estimates of peak discharges resulting from an earthen dam failure. Goodness-of-fit is quantified through relative standard error and relative bias. These measures are computed and presented for previous predictive equations. Numerical optimization techniques are used to calibrate power law functions of 1, 2, and 3 predictors to estimate peak discharge from the aggregate database. Findings show that equations calibrated from the aggregate database have better goodness-of-fit metrics than those determined from their earlier, individual data sets. Two similar innovative techniques are applied to the aggregate database: Region of Influence (ROI) and k-Nearest Neighbor (kNN). Both of these methods identify a subset of most similar observations from the database, given a specific test location. The ROI approach performs poorly, indicating that better results are obtained as ROI size increases, contrary to the spirit of this method. In contrast, the kNN approach performs well, with best results obtained for a simple numerical average of the k nearest observations. Regression equation calibration via logarithmic transformation is briefly explored and the need to limit predictions to the test space within the convex hull of the observations is discussed.