Estimation of Floods Related to Extreme Precipitations through a Machine Learning Approach

The study is geared towards the implementation of a workflow based on a Support Vector Regression Machine Learning (SVR-ML) approach which is conducive to estimates of flowrates across a given cross-section of a target stream in the presence of extreme precipitation events. The work is motivated by the observation that damages ensuing flash floods are a matter of global concern. A broad set of evidences suggests the ecosystem is experiencing changes of precipitation extremes, a causality relationship between increasing extreme floods and global climate dynamics being evidenced. In this context, practical tools associated with analyses of floods caused by extreme precipitation events can assist the design of early alert strategies across vulnerable regions. Physically and conceptually-based models have been extensively employed to link stream flowrates to precipitation events. These kinds of models are formulated and validated upon relying on continuous monitoring of flowrates as well as hydrometeorological variables associated with the area of the watershed related to a target stream. The typically high uncertainties underlying (a) the description of the physical processes governing the rainfall-runoff relationship as well as (b) monitoring and quantification of quantities and attributes characterizing the system behavior tend to propagate to outputs of interest of a given model. When considering well instrumented watersheds, data-driven modeling approaches grounded on machine learning (ML) algorithms can be an attractive alternative/complement to physically-based modeling approaches to analyze extreme flood events. Here, we rely on a Support Vector Regression ML (SVR-ML) algorithm that makes use of a linear kernel to provide estimates of hourly flowrate at a stream upon relying on observations of hydrometeorological variables across the watershed associated with the stream. The analysis encompasses three watersheds differing in size (ranging from about 25 to 250 km2) and located in the North of Italy and is structured across three steps: (i) identification of variables that are most informative to the target quantity (i.e., the flowrate in the stream), a step relying on cross-correlation and partial auto-correlation analyses; (ii) training of the SVR-ML algorithm, comprising the estimation of the optimal hyperparameters and parameters of trained models and the ensuing validation; and (iii) analysis of the anticipation time at which an early alert is effective, model performance being then quantified through the typical Mean Average Percentual Error (MAPE) metric. Our results suggest that, as expected, precipitation is the main driving force in a rainfall-runoff process, quantities such as temperature and relative humidity being least informative to the construction of the ML model considered. The predictive capability of the model (quantified through MAPE) is influenced by the desired anticipation time (i.e., the distance in time between the inputs and the output of the ML model). In general, one can note that (i) predictions of enhanced quality (MAPE smaller than 10%) are obtained for shorter anticipation times and (ii) models associated with low values of MAPE are obtained if the anticipation time is equal to or smaller than the time of concentration of the watershed.