Data-driven modeling for river flood forecasting based on a piecewise linear ARX system identification

•The objective of flood forecasting is to prevent from human and ecological disasters with a minimum lead-time based on rainfall-runoff modelling of rivers.•It is difficult to handle all the nonlinearities of a river model because of multiple factors such as the evapotranspiration phenomenon and the soil saturation.•Existing models are either hydrological, conceptual or data-driven.•Black-box modelling based on Piecewise Affine systems allows to handle nonlinearities by identifying a set of linear subsystems using only the rainfall and the streamflow measurements.•An unsupervised clustering-based approach based on Dempster–Shafer theory on the masses of belief is used to identify the Piecewise Affine model. Most of the studies related to the rainfall-runoff modeling of rivers consist of data-driven models, given that the corresponding physical modeling approaches are based on a thorough geological knowledge of the river in addition to a time consuming simulation. Indeed, flood forecasting services have the difficult task of avoiding natural and human disasters and choose for that to use input-output or grey box models for their simplicity and easy calibration updates. However, these models are not evolving according to the variations of environmental conditions or need at least the evapotranspiration and the soil humidity measurements in addition to the rainfall quantity. This paper gives an alternative approach to the existing rainfall/runoff linear and nonlinear models by the utilization of a hybrid system consisting in a Piecewise Auto-Regressive eXogeneous (PWARX) structure identified using an approach that alternates between data assignment and parameter estimation. The usage of this special kind of nonlinear systems bears a potential to handle the nonlinearities and varying-time delays mainly induced by the soil water storage and evapotranspiration.