Identification of low flow events by machine learning algorithms

Abstract

An improved forecast of low flow events in catchment basins could be a valuable tool for the operation and decision making of dependent infrastructure (e.g. wastewater discharge, water abstraction) along corresponding rivers. Therefore, the classification of 6642 independent low-flow-events (being the Q347 as the discharge less than the 95%- exceedance quantile of the FDC) from 55 catchment basins within the Kanton Solothurn (Switzerland) was performed by five different machine learning algorithms (i.e. knn, decision tree, random forest, support vector machine, logistic regression). Herein, each low flow event was characterized by 47 static and dynamic parameters (i.e. description of catchment and event history), being supplemented by differently defined (near) non-low-flow events, leading up to a total population of approx. 18000 discharge events.

The validation and verification showed different qualities of the classification accuracy for the forecast of low-flow events, being dependent on the selection of the defined event populations, the selected machine learning algorithm and the definition of classes. In general, the support vector machine and random forest may lead, with the presumption of carefully selected classes, to forecast accuracies of >90%.