Outcomes from an Ensemble Forecasting Workshop in Brazil
Contributed by Walter Collischonn and Fernando Fan
IPH-UFRGS location, where the workshop was held
The first local Brazilian Ensemble Forecasting Workshop was held on 21 and 22 May at the Hydraulic Research Institute (IPH – Instituto de Pesquisas Hidráulicas), Brazil. IPH is part of the Federal University of Rio Grande do Sul (UFRGS – Universidade Federal do Rio Grande do Sul), located in Porto Alegre, Brazil.
The Workshop was organized by the Large Scale Hydrology Research Group of IPH. More than 80 participants from different parts of Brazil attended the workshop, including undergraduate and graduate students, professors, consultants, and employees from hydroelectric companies, research institutes, and forecasting services.
The workshop was organized around two final presentations of Master Thesis and PhD Dissertation, from IPH students (Vinicius Alencar Siqueira and Fernando Mainardi Fan). Both thesis were related to the use of ensemble forecasts in the Brazilian scenario.
The audience of the Brazilian Ensemble Forecasting Workshop
Besides the final thesis presentations (during mornings), the workshop accounted with four more guest speeches (during afternoons):
- Ensemble weather forecasts using regional meteorological models were discussed by Dr. Sin Chan Chou, from the Brazilian Center for Weather Forecasting and Climatic Studies (CPTEC – Centro de Previsão de Tempo e Estudos Climáticos) of the Brazilian National Spatial Agency (INPE – Instituto Nacional de Pesquisas Espaciais). She described CPTEC-INPE operational products and models, especially the Eta regional model, which is being used operationally for ensemble numerical weather predictions over South America (40×40 km resolution) and, with higher resolution (5 x 5 km), over South-east Brazil. These products can be followed here.
Presentation during the workshop
- Dr. Dirceu Reis from Brasilia University (UnB – Universidade de Brasília) described the efforts in developing operational forecasting systems at the Ceará State Meteorological and Hydrological Service (Funceme – Fundação Cearense de Meteorologia e Hidrologia). Dr. Reis presented the Funceme ensemble seasonal hydrological forecasting system that is used to reduce damages from droughts or excessive water incomes to reservoirs in the North-east state of Ceará, which is characterized by a semi-arid climate.
- Dr. Guilherme Marques, from IPH-UFRGS, talked about recent research outcomes of multiple reservoirs systems optimization using ensemble forecasts. The main focus was to discuss how Brazil can benefit from optimized reservoirs operation, since the energy generation of the country is almost 80% based on hydropower.
- Dr. Luis Gustavo Gonçalves, from CPTEC/INPE, talked about the use of remote sensing information to improve results from meteorological and hydrological models. Especial attention was given to soil moisture products, which may be used in data assimilation schemes and to improve models performance.
The main outcome of the workshop was a fruitful discussion about the next steps for the operational use of ensemble forecasts in the Brazilian context. The workshop provided a unique opportunity for sharing research expertise and networking.
Workshop program (in Portuguese)
For some local energy generation companies, ensemble forecasting systems are already a reality. This is the case of CEMIG (Minas Gerais state), where operational techniques based on works of Fan et al. (2014), Fan et al. (2015) and Schwanenberg et al. (2015) are being used.
For some other centres, this kind of operational system is still not a common practice, and for some participants, the workshop was the first opportunity to learn about the topic.
The interest shown by the audience suggests that the technical community in Brazil is opened to increase the use of ensemble forecasting techniques in operational practice during the next years.
However, there is still a long road to be covered, and more research based on specific conditions and challenges of Brazilian basins should be conducted to help pave it. For now, we can say that initiatives such as this workshop clearly seem like important steps to propagate ensemble forecasting knowledge.
July 4, 2015 at 09:35
I might have missed something when I scrolled through the presentation, but I didn’t see very much of the USE of ensemble information. I would in particular have been interested in how you approach some of the questions which are debated elsewhere:
1. The role of the ensemble mean or median. They are on one hand the optimum forecast in some deterministic sense, on the other hand they do not represent a “realistic” image of the atmosphere.
2. The use of the ensemble to estimate the reliability of some higher-resolution deterministic forecast. This “common sense” approach may, however, cause some severe inconsistencies (the ensemble says 80% probability of > 10 mm and the high resolution has only 5 mm).
3. The communication of uncertainty to customers and the public: should it be in the form of probabilities, intervals or variances etc. What experiences were conveyed at the workshop?
4. Some forecasters see the ensemble as a challenge to find out which one of the members is the “forecast of the day”. Was this discussed at the workshop?
5. Statistical post-processing in different form may still be needed to remove some systematic errors and enhance (or dampen) the variability. Was this aspect also discussed?
July 6, 2015 at 15:03
Hi Anders,
Indeed some of the questions you stated were discussed. But, as you can imagine, for a two day workshop we could not discuss everything.
Especially about the use of hydrological forecasts, we discussed a lot about optimization models. This topic was the the thematic of Guilherme Marques talk. Use of ensembles for optimization is a very popular topic high now in Brazillian applications, focusing on more hydropower generation coupled with flood control.
About points 1, 2 and 4, those were unfortunatelly not very discussed. Mainly because the reported experiences on ensembles uses does not have the objective of obtaining a best deterministic forecast. But have the objective of working with probabilities to issue alarms or operate hydraulic structures. And those points would be more related to “obtaining the best deterministic from ensembles”.
About point 3, this was definelly discussed. Dirceu Reis showed some samples about how this is done in Ceará State. Usually simple models such as box plots are efficient to show good information. Maps associated with probabilities are also used.
About pont 5, the statistical post-processing of forecasts was also debated. Overall, the common sense was that bias removal can improve results, this was even said by Sin Chan Chou (from the Brazilian Center for Weather Forecasting and Climatic Studies), who actually works developing meteorological models. But, among hydrological forecasting system operational users and developers, a main difficulty to apply those techniques is the short time series of observations and the low density of observed information over large river basins, what limits the possibilities of statistical post-processing use.
Thanks for the comment!
July 6, 2015 at 20:20
Hi Fernando,
Sorry, I didn’t notice it was just a two day workshop.
However, I still think it is important for the ensemble community to spend more time on HOW to use the information. To make an adventurous comparison: there is nothing wrong in being taught how cars are manufactured, but most of us are interested in learning how to drive them.
There is, in my view, no conflict between deterministic and probabilistic information. One supports the other. To be told that there is a 30% chance of dry conditions and 20% chance of > 20 mm is all right, but if it is added that the “most likely” rainfall is 10 mm, wouldn’t that make it easier to comprehend or relate to the probabilities? The “10 mm” doesn’t necessarily serve any operational objective, except as something to “hang up” the probabilities on.
Finally, statistical interpretation does not necessarily have to be of the MOS-type, where you need at least three years of data to evaluate the statistics. Recurring upgrades of the model(s) might make the data less representative and thus yield misleading statistics.
There are different types of Bayesian techniques where you hardly need any data at all! No, I didn’t mean that;-) What I meant was that the maths works in a way like the human mind, by daily “trial and error”. You adjust some correction equation from “daily experiences”.
See some of my Bologna lectures at https://hepex.inrae.fr/resources/a-course-on-probability/ on “Day 3” (Wednesday) and then in particular https://hepex.inrae.fr/wp-content/uploads/2013/12/Lecture_12.pdf