The benefits of pre- and postprocessing streamflow forecasts for an operational flood-forecasting system of 119 Norwegian catchments
Abstract. The novelty of this study is to evaluate the univariate and the combined effects of including both precipitation and temperature forecasts in the preprocessing together with the postprocessing of streamflow for forecasting of floods as well as all streamflow values for a large sample of catchments. A hydrometeorological forecasting chain in an operational flood forecasting setting with 119 Norwegian catchments was used. This study evaluates the added value of pre- and postprocessing methods for ensemble forecasts in a hydrometeorological forecasting chain in an operational flood forecasting setting with 119 Norwegian catchments. Two years of ECMWF ensemble forecasts of temperature (T) and precipitation (P) with a lead-time up to 9 days were used to force the operational hydrological HBV model to establish streamflow forecasts. Two approaches to preprocess the temperature and precipitation forecasts were tested. 1) An existing approach applied to the gridded forecasts using quantile mapping for temperature and a Bernoulli-gamma distribution for precipitation. 2) Bayesian model averaging (BMA) applied to catchment average values of temperature and precipitation. BMA was also used for postprocessing catchment streamflow forecasts. Ensemble forecasts of streamflow were generated for a total of fourteen schemes based on combinations of raw, preprocessed, and postprocessed forecasts in the hydrometeorological forecasting chain. The aim of this study is to assess which pre- and postprocessing approaches should be used to improve streamflow and flood forecasts and look for regional or seasonal patterns in preferred approaches.
The forecasts were evaluated for two datasets: i) all streamflows and ii) flood events with streamflow above mean annual flood. Evaluations were based on reliability, continuous ranked probability score (CRPS) and -skill score (CRPSS). For the flood dataset, the critical success index (CSI) was used. Evaluations based on all streamflow data showed that postprocessing improved the forecasts only up to a lead-time of two to three days, whereas preprocessing T and P using BMA improved the forecasts for 50 %–90 % of the catchments beyond three days lead-time. However, for flood events, the added value of pre- and postprocessing is smaller. Preprocessing of P and T gave better CRPS for marginally more catchments compared to the other schemes.
Based on CSI, we found that many of the forecast schemes perform equally well. Further, we found large differences in the ability to issue warnings between spring and autumn floods. There was almost no ability to predict autumn floods beyond 3 days, whereas the spring floods had predictability up to 9 days for many events and catchments. The results indicate that the ensemble forecasts have problems in predicting correct autumn precipitation, and the uncertainty is larger for heavy autumn precipitation compared to spring events when temperature driven snow melt is important. To summarize we find that the flood forecasts benefit from most pre-and postprocessing schemes, although the best processing approaches depend on region, catchment, and season, and that the processing scheme should be tailored to each catchment, lead time, season and the purpose of the forecasting.
Trine J. Hegdahl et al.
Trine J. Hegdahl et al.
Trine J. Hegdahl et al.
Viewed (geographical distribution)
The manuscript by Hegdahl et al. attempted to apply a set of pre and post-processors with a HBV model on 119 catchments in Norway to find out general features of improving streamflow forecasts. Introduction is highly motivating and it covers a set of very important questions to be answered in this area of research. In fact that actually made me read this very long manuscript. However, from Section 2.3 onwards, I started noticing that there is one line of thinking that Authors are trying to pursue in this manuscript, which may be the serious limitation of the scope of their work. For instance, why should one use only HBV model, why to use ECMWF ENS data only, how can one interpolates 25 km spatial resolution forecasts to 1 km observation data without any downscaling technique, how Authors estimate the aggregated average values for each catchment, why not one use log-sinh instead of Box-Cox transformation, how justified it is to use Ensemble Coupla-Coupling, why not use Schaake Shuffle, aren't there other pre- and post- processors than CAL and BMA, etc. etc. There are many such questions which are not addressed here. In other words, I couldn't find what is novel here, knowing very well that there are several papers on this topic already published. Practically, every month we find new publication on pre- or post-processing in different journals.
When I started reading the results section, very long writeup in sections 5 and 6, I wondered why Authors needed to look at 119 catchments. Why not pick few catchments and present definite answers to the two questions which Authors have summarized in the conclusion. Since the study tried to capture so many different aspects, physiography, seasonal, snow-melt vs. rainfall based flood, etc. etc. that lead to Authors having fairly standard conclusions. Instead, I would look into few aspects but with rigorous analysis and try to derive some conclusions which can actually benefit the hydro-meteorological forecast community, not only in Norway but other places also.
It is very important to have plots which can be interpreted easily. In this manuscript, almost all the results are shown through box-plots. A set of time series plots showing how good the pre- and post-processors are improving the forecast would be highly beneficial. Similarly to show the improved flood forecasting, a time series plot would make things very clear. But I can only imagine the difficulty one would face in summarizing the results of 119 catchments, 51 ensembles, 9 lead time etc.. Therefore, a small number of catchments from different parts of the country may be the way forward. In one sense, Authors are already doing this by summarizing results of only 6 catchments. Then please limit the work to only pre-processors, that could be one option. In summary, Authors have to find a way to focus on novelty of this study rather than trying to cover all possible aspects on this topic.