HYDROLOGICAL PROCESSES SCIENTIFIC BRIEFING Hydrol. Process. 17, 3791 – 3801 (2003) Published online in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/hyp.5168 Impact of meteorological predictions on real-time spring flow forecasting Paulin Coulibaly* Department of Civil Engineering and School of Geography and Geology, McMaster University, Hamilton, Ont., Canada *Correspondence to: Paulin Coulibaly, Department of Civil Engineering/School of Geography and Geology, McMaster University, 1280 Main Street West, Hamilton, Ontario L8S 4L7, Canada. E-mail: couliba@mcmaster.ca Abstract Meteorological predictions, such as precipitation and temperature, are com- monly used to improve real-time hydrologic forecasting, despite their inherent uncertainty and their absence in the model calibration stage. In this study, we quantify the effect of meteorological prediction errors on the accuracy of daily spring reservoir inflow forecasts using weather predictions in both the model calibration and testing phases. Different modelling experiments are compared using an operational conceptual model and nonlinear empirical models to assess the effects of using daily numerical weather predictions as opposed to the use of historical observations. It is found that, even with large prediction errors, meteorological forecasts can provide significant improve- ment of spring flow forecast for up to 7 days lead time, particularly for low flows. Spring flow prediction errors associated with the type of hydrological model used are significantly larger than those related to the meteorological predictions, particularly for 1 to 4 days ahead forecasts. The experimental results also indicate that multiple model-based forecasting using an iterative prediction approach appears to be the most effective method for an adequate use of weather predictions. Copyright  2003 John Wiley & Sons, Ltd. Introduction Cost-effective and reliable operation and scheduling of hydro-systems require accurate real-time forecasts of river flows. In many cases, ade- quate short-term site-specific predictions remain a difficult task because of various sources of uncertainty: (1) errors in meteorological input data (historical records and forecasted values); (2) errors in observed hydro- logical variables; (3) errors and simplification inherent to the prediction model structure; and (4) errors due to the use of non-optimal parameters. In the model calibration process, only error source (4) is minimized. Observations errors due to the recording context and/or material may be unavoidable, but error source (3) is apparent in comparative model studies. Here, we focus on analysing error source (1), which is par- ticularly important for real-time spring runoff forecasting in cold and snowy regions. In general, historical meteorological predictions are not available for hydrologic model calibration owing to the fact that all numerical weather predictions are not usually archived. Thus, recorded observations are commonly used for model calibration and then a correction procedure is applied to accommodate the use of meteorological predictions and to improve the real-time forecast accuracy. This has been considered a good alternative for improving conceptual model forecast skill (Kitanidis Received 2 July 2002 Copyright  2003 John Wiley & Sons, Ltd. 3791 Accepted 8 August 2003