Reporter: Daniel Viviroli
Theme Block II consisted of presentations dealing with hydrological ensemble forecasts. Subjects dealt with were identification and decomposition of Ensemble Prediction System (EPS) uncertainties with Bayesian Theory [1], propagation of uncertainties in the chain of EPS constituents [2], operational EPS forecasts with updating in an Austrian meso-scale catchment [3], experiences from 30 months of operational EPS forecasts for 45 catchments in Sweden [4], EPS-based flood forecasting for the Swiss part of the Rhine River basin with examination of rare events [5], and a framework for demonstrating the ability of EPS for improved forecasts of heavy events in the Alps [6].
The subsequent discussion can be divided roughly in two main subject areas: On the one hand, uncertainties and errors of EPS were considered, showing that there are still significant issues to be investigated. On the other hand, the applicability of EPS results and their transfer to more practically oriented users was reviewed critically.
There was consensus that uncertainties in EPS have to be examined thoroughly, as was also suggested by Roman Krzysztofowicz in his presentation [1]. Especially the incorporation and further decomposition of hydrological uncertainties is a difficult task. Göran Lindström [4] proposed to start as simple as possible and then improve what needs to be improved. It also has to be borne in mind that EPS uncertainty is a function of basin scale and event characteristics and therefore is heavily space and time dependent. Furthermore, the lead time considered (cf. [3]) determines whether processes or data inputs are the major source of uncertainty. Such instationarities need to be stated clearly in any investigation of uncertainty. Furthermore, errors and uncertainties are propagated in the chain of observation, Numerical Weather Prediction (NWP) and hydrological model. Since hydrological variables are integrative in space and time, they are ideal for testing EPS results. There was agreement that the information content of hydrological data should be made more use of. Hydrological data should also be employed more frequently through updating procedures (e. g. Ensemble Kalman Filter). More generally speaking, data should be collected with view to the most crucial sources of uncertainty (see [2]).
A major concern is the identification of biases in EPS results; this needs to be investigated on basis of long-term series of observation data. With the same data, the probability of EPS results should be examined more thoroughly as well, as opposed to presenting only the number of EPS results that exceed a defined threshold. The estimated probability distributions should be conditioned on observations to increase consistency. Inconvenience for such investigations arises from the frequent change in EPS key features, e. g. in resolution, parameterisation and number of ensemble members. In order to increase the accuracy of forecasts, it was proposed to combine various EPS.
An important research question concerns the appropriate number of ensemble members to achieve reliable results. On the one hand, an ensemble should be large enough to allow a reasonable estimation of probabilities and to meet the requirements of the end-users as to reliability of the results. On the other hand, ensemble size is limited by the computational time available for a forecast.
Concerning the event size to be investigated more closely with EPS, it was proposed to increase the focus on big storm events, since their behaviour differs from the one of medium and small storms and therefore poses additional difficulties. The closer examination of interesting events and event types could also be used to improve the communication and co-operation between Hydrologists and Meteorologists involved with EPS techniques. However, an important question to be answered yet is how to deal with extreme events that are not represented in the atmospheric EPS.
A point that by general agreement seems important is how to combine the deterministic models with statistical methods. Since statistics fail to reproduce heavily non-linear processes (as they occur especially under extreme conditions) and cannot account for the dynamics in atmospheric and hydrological systems, a creative combination of statistical and deterministic approaches should be sought. For instance, statistical post-processing could serve to capture biases and reveal lack of representativity in EPS.
Although there was a separate theme block dealing with communication of uncertainties, the applicability of EPS techniques and the transfer to more practically oriented end-users was discussed to quite some extent. This transfer is difficult in that the communication between a forecaster (hydrologist or meteorologist) and an end-user is highly individual, i. e. that forecaster A and end-user B will end up with different conclusions and results than forecaster C and end-user D. This is also due to the fact that the actual cost/loss-ratio depends on the individual end-user. In view of above considerations, the reaction of end-users to (uncertain) results itself is subject to uncertainty. Therefore, the role of the "human factor" in the decision-making process (forecaster or end-user) should be further examined.
The question arose what the added value of EPS forecasts would be at all as compared to deterministic forecasts. In reply, it was stressed that the major benefit is the increase in information given about what to expect within the forecast period. The forecast values are complemented with a probability which helps the end-user in making his decisions. It was suspected that the demand for EPS-type results will increase when this added value is recognised more widely and the end-users get more accustomed to the interpretation of such forecasts. Although it is essential to instruct the end-users in utilisation of EPS results, the interpretation still will demand the involvement of experts. To facilitate the decision process, it seems desirable to establish joint data interfaces and platforms for visualisation and warning (cf. [6]).
To increase flexibility of EPS applications, there should be more possibilities for interaction, e. g. in changing initial conditions or adjusting the river stage manually. In case of extraordinary conditions (e. g. an ice jam in a region where this is not expected), model results may become invalid but may be "fixed" by an experienced forecaster or end-user. Human judgement could also be crucial for error identification.
From the end-user's perspective, the value of EPS results diminishes heavily if their spread increases. Therefore, the "spaghetti plots" seem not suitable for communication with end-users, and more appropriate solutions have to be sought, such as probability maps (cf. theme block III). It was stated, however, that NWP tend to cluster since they are based upon physical processes and that our impression of large spreads in EPS results may be caused by frequent examination of difficult cases. On the whole, only understandable results should be distributed, with respect to the end-user's crucial questions: when, where, how much, and with what probability? But what are consequences of a probabilistic forecast? What action is to be taken? While an EPS prediction shows a range of possibilities, the end-user's decision often is of a "yes/no"-type and therefore less flexible. Already the decision whether to issue a warning or not is a critical one since false alarms diminish credibility (which should be guaranteed with greatest priority), while omitted warnings may lead to fatalities and legal consequences. How to deal therefore with a probabilistic forecast that announces the possibility of a rare event, although with low probability (see [5])?
To conclude, it seems that research concerning hydrological ensemble forecasts makes progress. However, stable and reliable applications are difficult to establish at present, not least because of frequent changes in the underlying EPS. Co-ordination of research and exchange of experience is therefore of great importance. In order to reduce scepticism of end-users, the transfer of probabilistic results into a "yes/no"-world should be improved
[1] Roman Krzysztofowicz, Bayesian theory of ensemble forecasting (presentation not available as pdf)
[2] Florian Pappenberger, Cascading uncertainty in flood forecasting
[3] Christian Reszler, Operational ensemble forecasts of floods
[4] Göran Lindström, Evaluation of ensemble streamflow forecasting at SMHI
[5] Mark Verbunt, Ensemble flood forecasting in Switzerland: Selected case studies of extreme events
[6] Massimiliano Zappa, Towards (quasi-)operational demonstration of hydrometeorological ensemble prediction systems: The MAP D-PHASE and COST PROFIT projects