Author(s): Isa Ebtehaj; Hossein Bonakdari

Linked Author(s): Isa Ebtehaj, Hossein Bonakdari

Keywords: Early-warning flood forecasting; Generalized Structure of Group Method of Data Handling; Machine learning; Quebec; Water resource management

Abstract: Accurate early-warning flood forecasting is mandatory in the province of Quebec, Canada, as a result of its susceptibility to flooding. In the current study, a new scheme of machine learning models is developed for early-warning flood forecasting known as Generalized Structure of Group Method of Data Handling (GSGMDH). The developed method not only overcame the limitation of the most existing machine learning approaches (i.e., failure to provide simple equations for practical application), but also enhanced the performance of the classical Group Method of Data Handling (GMDH). The enhanced performance was obtained by providing second and third order polynomials with two and three input neurons, for which the inputs of each neuron could be selected from both adjacent and non-adjacent layers. Data from the Huron hydrometric station located in the province of Quebec, Canada, were extracted at 15-minute intervals and converted to an hourly dataset. The data was collected from April 08, 2008, to September 30, 2021, with more than 78000 samples. The forecasting performance ranged between R [0.946, 0.998], NSE [0.889, 0.996], NRMSE [0.033, 0.193], RMSRE [0.023, 0.147], and MARE [0.009, 0.091] for models with one to six lead times, indicating acceptable performance of the GSGMDH multi-step ahead flood forecasting. Although the general performance of the models in all ranges of samples was acceptable, the peak flow results showed that only models with one to three hours ahead have reliable performance. Consequently, the developed machine learning model could be applied as an alternative for the existing model for flood forecasting in Quebec, Canada.

DOI: https://doi.org/10.3850/IAHR-39WC252171192022627

Year: 2022