Systemic models for supporting flood warning procedures in river sections with no rating curves: application to a set of Emilia-Romagna watersheds

The present work analyses the possibilities offered by the implementation of artificial neural networks (ANNs) for the flood warning purpose. In particular, four river basins located in the Emilia-Romagna region have been selected, analysing the forecast performances of the models at each lead-time from +1 to +18 hours. Specifically, the models, run at hourly time-scale, have been calibrated considering as input variables past river levels and past and future precipitation values. The ANN calibration has been done considering continuous simulation periods (therefore including peak, average and low flow regimes). In a first phase the models selected for each case study have been evaluated according to the Nash-Sutcliffe efficiency in order to understand the general behavior of the ANNs. Secondly, a further assessment is added considering the skill score indices (POD, FAR, ETS) of the exceedance of the assigned flood thresholds; in this way, it is possible to understand the reliability of the ANNs for flood warning purposes. For this second evaluation, ANN models which make use only of the past river levels hence with no precipitation input values, are included in the analysis as an additional standard of reference. Considering the simulation results on all the basins, the Nash-Sutcliffe efficiency highlighted acceptable results for all the forecasting lead-times (1 to 18 hours). On the other hand, the skill score indexes on the exceedance of the flood thresholds are satisfactory only over shorter time-horizons.