The human hydrological model
by Marie-Amélie Boucher, a HEPEX 2015 Guest Columnist
Very recently, my fellow HEPEX columnist Jan Danhelka discussed the importance of the human forecasters in the whole process of weather and streamflow forecasting. His blog reminded me of a visit I made last year which somewhat challenged my views on flood forecasting. It was a technical tour about water management on Spencer Creek watershed that took place at the end of a conference. And there I met the first (and so far only) “human hydrological model” I’ve ever met.
Setting the scene: Spencer Creek
Before telling you what I mean by “human hydrological model”, let me introduce a bit of context. I’ll be going from a larger scale (the city of Hamilton and its area) to a smaller one (Christie dam on Spencer Creek).
Hamilton is a fairly large city (approx. 520 000 hab.) located on the western end of the Niagara Peninsula in the province of Ontario, Canada. The area is known for its many waterfalls all around the Niagara Escarpment, as can be seen on this map where every blue dot is a waterfall.
Figure 1: A warning sign on the downstream side of Christie Lake dam (picture: M-A Boucher)
Spencer Creek watershed covers a small area of 291 km2 and spreads all the way from the relatively flat portion above the Escarpment all the way down to the city of Hamilton. Consequently, Lower Spencer Creek watershed has a staircase-like topography which is one factor that makes it vulnerable to flash floods: if the infiltration capacity is exceeded in the area above the escarpment, the excess runoff water goes downstream very fast (see warning sign on Figure 1).
Besides, the downstream portion of the small watershed is largely impervious and the last 7 km of the main tributary have been channelized to accommodate urbanization (the upstream area is mostly agricultural). In addition, according to the city of Hamilton “Over the past five years Hamilton and other southern Ontario municipalities have experienced significant rainfall events with intensities, durations and frequency of re-occurrence that is unprecedented for [the] area.” This makes water management on Spencer Creek very challenging. It also raises questions relative to the human and nature relationship but that’s a bit off topic.
Water management for Spencer Creek is under the responsibility of the Hamilton Conservation Authority (HCA). This includes the operation of many small dams and reservoirs devoted to flood control, among which Christie Lake and Christie dam which were part of my visit.
Streamflow forecasting without a model
So last year, during the annual Canadian Water Resources Association’s conference in Hamilton, I enrolled on a tour of some of HCA’s facilities, including a dam removal site which was my primary interest. Although the dam removal site was interesting, it turned out that the Christie Lake Dam (see Figure 2) visit was amazing. Not really for the dam itself, but because this dam is managed by one person who:
- Collects data from meteorological stations all over the watershed
- Analyzes the meteorological forecasts for the next few days
- Gathers data from streamflow and level gauges upstream and downstream from the dam
- Relies on his experience and knowledge about the behaviour of the catchment to merge all of this information in his head and issue his own personal streamflow forecast without any model.
- Decides on whether or not he should have the gates opened, and if so, how much.
Figure 2: A picture of the downstream side of Christie Lake dam (picture: M-A Boucher)
So this guy’s mind is both the hydrological model and the decision model. He is supported by a small team (2 technicians) but in the end he is the one who issues the forecasts and takes decisions. Since he must answer for his actions when problems arise or during audits, he explained that he carefully keeps a logbook of all his decisions, supported by a summary of the reasoning behind.
Of course we discussed about uncertainty, especially on how one’s confidence in his or her forecast can vary from day to day, depending on the available information and specific conditions. He said that in his opinion, the best way for him to deal with this uncertainty is to be ready to revise his decisions and to adapt to real-time observation. It is a 24/7 job. People do actually call him when the level of the river is too high downtown…
I asked him if he would like to use a hydrological model sometimes and his answer was mixed: a computer based model might provide additional information, but one has to learn how to use the model efficiently, which could be time consuming. In addition, he said that after many years of working as a water manager for HCA, he felt pretty confident in his own forecasts and would not trust a model just that easily.
Figure 3: Forecasting with a perceptual model only (source: Pixabay, noncommercial reuse authorized)
Over the years, it seems like he has developed a good perceptual model of Spencer Creek catchment. According to Beven (2012), every hydrologist has such a model built in his or her mind. The building of this model evolves over time, according to experience and training.
The perceptual model describes the response of a watershed to a particular rainfall event, depending on how the hydrologist understands and interprets the initial state of the watershed and the interconnected processes which generate runoff. When choosing a computer-based hydrological model, it would certainly be coherent to choose one that agrees with our perceptual model. However, in practice, sometimes we just choose the model which is the most convenient, because it is (for instance) easily available or just because there is a tradition of using this particular model at a given workplace. Maybe the water manager at Spencer Creek felt that his perceptual model was enough or didn’t think an existing computer based model could fit with it.
I personally think that hydrological models can be really helpful in water management and ease the process. Are they always essential for efficient water management though? Can they replace the knowledge and experience of human forecasters? I definitely don’t think so.
References:
- Beven, K. (2012) Rainfall-Runoff Modelling: The Primer 2nd edition, Wiley-Blackwell, 488 pages.
- City of Hamilton (2009) Flood Aware Preparedness Program, available here.
April 24, 2015 at 14:47
I have met a number of “human hydrological models” before, including both “lumped” (one individual) and “distributed” (a committee). So, even though, over the years, these forecasters may have developed a sense for how a watershed may respond to a particular precipitation event, one major problem remains: what will happen when these people retire. And retirement for them may not be too far away. One surprising thing is that human hydrological models in developing countries are much more open to start using numerical models instead of their own forecast, not as a tool to just consult. That applies not only to the hydrological model itself, but also to the reservoir operations models.
It’s really astonishing the number of public utilities that manage multi-million, even multi-bilion, dollar operations by the seat of their pants. In the power generation sector, this may be due to the fact that, in the US, power generation was a highly regulated activity, which meant that the profit from the utilities was essentially fixed. That lead the utilities to operate in a fashion that made it irrelevant to operate in the most efficient way, since regardless of what they did, the benefit to the shareholders was fixed, either by increasing or decreasing the electricity rates. So, even though power generation has been mostly de-regulated, the legacy operation remains in the minds of the personnel.
April 25, 2015 at 00:15
I agree that managing multi-million dollar operations by gut feeling doesn’t sound right, but relying completely on a numerical model does not seem right either. Middle ground is probably the best option here.
Actually, Spencer Creek watershed was modelled in a few hydrological studies (see for instance Grillakis et al., 2011). So the model exists, but isn’t used operationally.
Most forecasters I’ve met so far use a numerical hydrological model as a complement of their own perceptual model. In fact, they interact quite a lot with the model, especially for manual data assimilation. So it is really a sort of collaboration between a person and a model. Sometimes the person trusts the model, sometimes not so much.
As for knowledge transfert and retirement issues, unless you rely entirely on the output of your numerical model, you inevitably get to learn a lot about the behaviour of a watershed you work on for many years. So I think in many organizations, knowledge transfert from the retiring forecaster to the new forecaster remains a very important issue even if a numerical model is involved.
Grillakis M.G., Koutroulis A.G. and Tsanis I.K. (2011). Climate change impact on the hydrology of Spencer Creek watershed in Southern Ontario, Canada, Journal of Hydrology, 409(1-2), 1-19.
May 18, 2015 at 15:58
I am afraid I am not as impressed by “human” forecasters. It might of course be different in hydrology compared to meteorology, but I doubt.
This is not intended to be a criticism of the persons involved. One problem is that models unavoidably change, so what was true in 2013 might not be true, or not as much true in 2015.
Secondly, a lot of “experience” goes unverified. Nobody checks if the “thumb rule” is true. From my time at ECMWF I found that about half of the comments we got about systematic errors or systematic behaviours from forecasters in the Member States were just not true. They were believed because “everybody” was convinced they were true.
Thirdly, there are, as I have shown in my many presentations, a lot of “psychological” traps, things that look absolutely true, but nevertheless are wrong or meaningless.
But as I have also shown in my many presentations, I believe in the “Human Forecaster” but he or she should be carefully trained, not only in model characteristics, but also, or rather much more, in how to handle an incoming flow of data during time constrains.