Data-driven streamflow simulation: the influence of exogenous variables and temporal resolution

Data-driven modelling approaches, like artificial neural networks, are particularly sensitive to the choice of input and output variables. This study focuses on the size of the temporal observation interval of input and output data in a river flow prediction application, analysing the simulation performances when considering increasing time aggregations of different input variables. The analyses are carried out on the data registered in a medium-sized (1,050 km2) watershed located on the Apennine mountains, where hourly meteorological data and streamflow measurements are available over an 8-year period. Four modelling approaches are considered for the prediction of river flow in the closure section: (1) without exogenous inputs, (2) with the additional input of past precipitation data, (3) with the additional input of past streamflow data measured in the upstream section, (4) with the additional input of both past precipitation and past upstream flow. For each modelling approach, using both (a) input data and output data at the same time scale and (b) input data at a temporal resolution finer than that of the output data, optimal modelling networks are identified and forecast performances are compared. The results highlight how the simulation improves with the addition of exogenous inputs, in particular upstream flow data, and with the use of input data at a temporal resolution finer than that of the output data. The results also show how both such benefits increase for larger temporal aggregation of the forecasts