The benefit of pre- and postprocessing streamflow forecasts for 119 Norwegian catchments, evaluated within the frame of an operational flood-forecasting system

In this work the performance of different pre- and postprocessing methods and schemes for ensemble forecasts were compared for a flood warning system.  The ECMWF ensemble forecasts of temperature (T) and precipitation (P) were used to force the operational hydrological HBV model, and we estimated 2 years (2014 and 2015) of daily retrospect streamflow forecasts for 119 Norwegian catchments. Two approaches were used to preprocess the temperature and precipitation forecasts: 1) the preprocessing provided by the operational weather forecasting service, that includes a quantile mapping method for temperature and a zero-adjusted gamma distribution for precipitation, applied to the gridded forecasts, 2)  Bayesian model averaging (BMA) applied to the catchment average values of temperature and precipitation. For the postprocessing of catchment streamflow forecasts, BMA was used. Streamflow forecasts were generated for fourteen schemes with different combinations of the raw, pre- and postprocessing approaches for the two-year period for lead-time 1-9 days.

The forecasts were evaluated for two datasets: i) all streamflow and ii) flood events. The median flood represents the lowest flood warning level in Norway, and all streamflow observations above median flood are included in the flood event evaluation dataset. We used the continuous ranked probability score (CRPS) to evaluate the pre- and postprocessing schemes. Evaluation based on all streamflow data showed that postprocessing improved the forecasts only up to a lead-time of 2 days, while preprocessing T and P using BMA improved the forecasts for 50% - 90% of the catchments beyond 2 days lead-time. However, with respect to flood events, no clear pattern was found, although the preprocessing of P and T gave better CRPS to marginally more catchments compared to the other schemes.

In an operational forecasting system, warnings are issued when forecasts exceed defined thresholds, and confidence in warnings depends on the hit and false alarm ratio. By analyzing the hit ratio adjusted for false alarms, we found that many of the forecasts seemed to perform equally well. Further, we found that there were large differences in the ability to issue correct warning levels between spring and autumn floods. There was almost no ability to predict autumn floods beyond 2 days, whereas the spring floods had predictability up to 9 days for many events and catchments.

The results underline differences in the predictability of floods depending on season and the flood generating processes, i.e. snowmelt affected spring floods versus rain induced autumn floods. The results moreover indicate that the ensemble forecasts are less good at predicting correct autumn precipitation, and more emphasis could be put on finding a better method to optimize autumn flood predictions. To summarize we find that the flood forecasts will benefit from pre-/postprocessing, the optimal processing approaches do, however, depend on region, catchment and season.