by Jan Danhelka, a HEPEX 2015 Guest Columnist
I have to start my first HEPEX blog post by a short introduction of myself. I have more than 10 years of experience in real time hydrological forecasting and modelling. Those years (un)fortunately included quite a few large (in some cases truly catastrophic) floods of different nature including the 2002 large scale summer flood, the 2006 winter flood and many flash floods events, in particular in 2006 and 2009. What I have learned was that every flood is specific and surprising in some of its aspects.
When it comes to me, I consider myself to be a practician, not a hard scientist. In addition, I like to ask provocative questions and provide controversial answers and I don’t mind to play the devil’s advocate. Going back to the issue of flood forecasting, let me start in far history.
The Hydrological Service in Bohemia was established in 1875 (we are celebrating 140 years this summer) making it one of the oldest such services in the world.
The head of the service, Prof. Harlacher, published his novel method of forecasting based on discharges and travel times in 1886 and 1887. To my knowledge, it was the first conceptual forecasting method in the world, as previous methods, e.g. for the Seine River in Paris, used just statistical methods to estimate changes in water stages based on changes in water stages and in precipitation upstream. By the way, Harlacher’s method started to be used operationally in 1892, after the telegraph reports of water observation were exempt from charges (and is used until today, in parallel to hydrological models).
I gave this large introduction to illustrate the long history of forecasting under different conditions and circumstances. I have joined the Czech Hydrometeorological Institute in the time of the implementation of models into every day forecasting practice, so I have experienced the initial distrust of forecasters as well as their discovery of the possibilities of this new tool. I have realized then that new model implementations cannot succeed without the full support from the forecasters.
We have moved in time and a lot of research has been done; obviously, the future of hydrological forecasting is in probabilistic modelling and multi-models. From my point of view, these approaches are fantastic, except one little thing – the role of the forecaster.
Lets have a look at the analogy in weather forecasting. It seems to me that the old generation of weather forecasters, who based their forecast on the synoptic analysis, knowledge and experience without meteorological model outputs, was replaced with the generation that grew up in the age of NWP models.
Forecasters are fully trained in the use of the model but, more and more, they tend to accept the model as a source of information similar to the measured data, and become fully dependent on it. Unfortunately, some forecasters, not all of them nor the majority of them, have turned to be the interpreters (skilful and educated) of model outputs, and perceive the model as a given black box think, created somewhere in a previous step of the forecasting chain (again similar to e.g. radar estimates or satellite pictures – someone else’s job and responsibility).
The forecaster experience then tends to be more about the model behaviour than the behaviour of the atmosphere over the given area (they know that COSMO has overestimated temperature for the last three days by 2°C, but they do not think about the physical process in reality). It happens to me several times that when I ask “What will be the precipitation in the next two days?”, the response I get starts with “Well, models say…”, instead of “Well, the synoptic situation is…” or “The wind flows from…”.
I remember a HEPEX workshop in Italy a couple of years ago. Someone (probably Eric Wood) presented a result from another workshop where a question was raised (not the precise wording): “What would you prefer to have in order to achieve skill in forecasting?” a) a better model, b) better data, c) a “better” forecaster. Participants there voted for better data (they were mostly modellers and model developers, so they trusted their models very much and they tried to make automatic modelling systems without a forecaster, therefore the vote for better data was not surprising). It was John Schaake who said “I would prefer a better forecaster”. I absolutely agree with John.
One of the specifics of the real time forecasting is immediate verification of your forecast (in opposite to e.g. climate change modelling). Everyone who prepares a forecast knows the unpleasant feeling you have when facing the reality completely different from what you have forecasted few hours ago. Typically, one tries to find what was wrong and what she could/should done better. However, I think the situation is a bit different if you ran the model, tried to tune the parameters, to optimize the estimation of initial conditions, or to interact with the model somehow on one hand, or if you simply received the model output to interpret it. In the first case, you think about model and real processes and your own knowledge of these. The reflection in the second case is obviously a bit different.
The progress in modelling technology cannot be stopped (nor there is any reason to do that), however we should not lose the forecaster and her/his role from our scope when developing a new forecasting system, to prevent the turn from ‘forecaster’ to ‘forecast interpreter’.
I do not know a solution nor have an idea on how to do that, but I am fully convinced that hydrologists/forecasters should be kept in the heart of the process; they should remain the ones who make the forecasts. If you have an idea let me know.
Let me finish this post with a quotation of Vít Klemeš from his marvellous book “Common sense and other heresies”:
“The modelling technology has far outstripped the level of our understanding of the physical processes being modelled. Making use of this technology then requires that the gaps in the factual knowledge be filled with assumptions which, although often appearing logical, have not been verified and may sometimes be wrong.”
That is why we need a forecaster to be skillful, experienced, and why a forecaster should understand well the model structure and its limitation.