Multistep ahead streamflow forecasting: Role of calibration data in conceptual and neural network modeling

When choosing the rainfall-runoff modeling approach to be integrated in a river flow forecasting system, two crucial issues are the minimum data requirement for calibration purposes and the reliability of the predictions over different time horizons (lead-times). The paper presents an investigation of the real-time forecasting ability of a conceptual and a neural network model, comparing the performances obtainable for increasing lead-times and analyzing the influence of the amount of the calibration data over two real-data case studies. Neural networks proved to be an excellent tool for the real-time rainfall-runoff simulation of continuous periods (including low, average and peak flows), provided that an extensive set of hydro-meteorological data is available for calibration purposes. On the other hand, the comparison highlights that a conceptual formulation may allow a significant forecasting improvement in comparison with the data-driven approach when focusing on the prediction of flood events and especially in case of a limited availability of calibration data.