A MACHINE LEARNING-BASED MODEL TO IMPROVE SHORTTERM FORECASTS OF FLOODING IN NOVA FRIBURGO-RJ

Abstract

Machine learning use in hydrological modeling has intensifi ed in recent decades given the potential of these techniques to produce in short time satisfactory solutions to support tasks such as early fl ooding warnings. In this context, this work reports the development and the results of a forecasting model built from a hydrometeorological database and using a regression tree. This regression tree-based model is intended to forecast, with hours in advance, the level of a river in Nova Friburgo-RJ, which was chosen as study area due to its recent history of major natural disasters. Rainfall and river level data used in the modeling were collected during the years of 2013 and 2014 in four stations located in the study area. The regression tree allowed more fl exibility in the model design. The fi rst regression tree results yielded Nash indexes above 0.75 indicating the feasibility of the approach. However, in order to be used as an operational decision-making support tool working in real time, the model should be improved with new studies and tests carried out with enlarged hydrometeorological databases.