Keywords

Supervised machine learning, hydrological prediction, ensemble, hybrid model, flash flood

Start Date

26-6-2018 3:40 PM

End Date

26-6-2018 5:00 PM

Abstract

Application of supervised classification to short-term forecasting of hydrological events demonstrated that combining outputs of several individual learners into a final judgement improves the accuracy of predictions and creates more robust models. Given that predictions are generated as categorical values corresponding to a class label of a future hydrological event, the ensembles perform better when they incorporate black box models which disagree on the same subsets of data. To further extend the ‘diversity of opinions’ of ensemble members, black box models of a different type can be added to an ensemble of classifiers. Given that regression methods are aimed at accurate calculation of future magnitudes of hydrological characteristics as opposed to determining a class label denoting future hydrological conditions, the extension of the ensemble approach to regression and hybrid models looks promising to further increase the lead time of reliable predictions. The study investigated regression models applied for short-term predictions of hydrological events, such as flash floods, at a highly urbanized small watershed and their inclusion into ensembles of classifiers. The predictions were generated solely on readily available data collected by stream and rain gauges. The heterogeneous measurements of water levels and precipitation were combined and transformed into phase spaces using time-delay embedding. The potential for developing a hybrid model incorporating both classification and regression approaches was analysed. The results of this study are presented in the paper.

Stream and Session

Stream B: (Big) Data Solutions for Planning, Management, and Operation and Environmental Systems

Session B2: Hybrid modelling and innovative data analysis for integrated environmental decision support

Share

COinS
 
Jun 26th, 3:40 PM Jun 26th, 5:00 PM

Data-Driven Hybrid Approach to Short-Term Predictions of Hydrological Events

Application of supervised classification to short-term forecasting of hydrological events demonstrated that combining outputs of several individual learners into a final judgement improves the accuracy of predictions and creates more robust models. Given that predictions are generated as categorical values corresponding to a class label of a future hydrological event, the ensembles perform better when they incorporate black box models which disagree on the same subsets of data. To further extend the ‘diversity of opinions’ of ensemble members, black box models of a different type can be added to an ensemble of classifiers. Given that regression methods are aimed at accurate calculation of future magnitudes of hydrological characteristics as opposed to determining a class label denoting future hydrological conditions, the extension of the ensemble approach to regression and hybrid models looks promising to further increase the lead time of reliable predictions. The study investigated regression models applied for short-term predictions of hydrological events, such as flash floods, at a highly urbanized small watershed and their inclusion into ensembles of classifiers. The predictions were generated solely on readily available data collected by stream and rain gauges. The heterogeneous measurements of water levels and precipitation were combined and transformed into phase spaces using time-delay embedding. The potential for developing a hybrid model incorporating both classification and regression approaches was analysed. The results of this study are presented in the paper.