Author(s): Amirhossein Salimi; Tadros Ghobrial; Hossein Bonakdari
Linked Author(s): Tadros Ghobrial, Hossein Bonakdari
Keywords: Ice Jam Flood; Artificial Intelligence; River Ice; Occurrence
Abstract: Many rivers in the northern hemisphere experience threat of Ice-Jam Floods (IJFs). The inherent complexity in managing IJFs arises from their chaotic nature. Various flood forecasting systems have been developed. Recent Machine learning (ML) approaches have been developed to forecast IJFs occurrence, severity, timing, and location across different scales. These approaches face limitations due to the scarcity of historical flooding records and observational data at gauging stations. The extensive dataset on IJFs on the Chaudiere River, Quebec, Canada, presents an opportunity to assess the abilities of ML models to forecast IJFs with a reliable dataset. The current study uses an extract of this dataset, which contains a series of IJFs dates and peak water levels in Saint-Joseph-de-Beauce from 2015 to 2021. The primary objective of this research is to evaluate the effect of normalization and data splitting methodologies, such as k-fold cross-validation and the hold-out method, on IJF occurrence prediction by Extreme Learning Machine (ELM). The results show that ELM performance in prediction of IJF occurrence was notably improved by employing these ensemble pre-processing methods and, with up to 25%. In addition, using 1-day lag of air temperature inputs can have more correlation. The best model achieved superior results by useing Z-score and k-fold cross-validation.
Year: 2024