Heating network: Grenoble district heating company (CCIAG), France
Enhancing economic and environmental performance through more accurate temperature forecasts.
Energy
07/11/2023
by
Damien Raynaud
10 min
Customer needs
null
As a public operator, the Intercommunal Heating Company of the Grenoble Urban Area (CCIAG) is responsible for managing the district heating network of Grenoble-Alpes Metropolis. This extensive district heating network, spanning over 180 kilometers, is the second largest in France. It provides heating and domestic hot water to both public and private buildings, including residential complexes and offices. In total, it provides heat to the equivalent of 100,000 houses, with an average annual energy sales of 800 GWh. CCIAG operates five different production sites and uses a variety of 10 fuel types.
The choice of fuels is primarily guided by the goal of controlling energy consumption and reducing greenhouse gas emissions and air pollution in its area. CCIAG is actively working to gradually reduce the use of coal, which accounted for 12% of the energy mix in 2022, while considering the specificity of each production site. The definitive elimination of coal usage is planned for 2026.
Figure 1: CCIAG's energy mix Distribution for the 2021-2022 heating season (Source: CCIAG, Entre Nous #38, Dossier "Heat Production: Understanding the Choice of Energy Sources").
null
null
The types and proportions of each fuel are chosen based on technical, economic, and environmental criteria. Decisions are regularly made, especially during significant fuel price variations. Thermal storage is used to provide flexibility, regulate production, and reduce the use of backup boilers.
CCIAG needs to anticipate heat requirements to adjust its energy production and define the optimal utilization of its production resources to meet these needs. As heat demand is directly influenced by outdoor temperature, CCIAG places significant importance on temperature forecasting. They seek forecasts with maximum accuracy, aiming for an average error level of less than 1.2°C during the winter season, where conventional forecasting systems offer a precision of around 2°C. Steadysun provides CCIAG with outdoor temperature forecasts four times a day, up to fourteen days in advance.
null
Description of the forecasting solution
null
The unique geographical location of Grenoble, with its atypical bowl-shaped topography, poses a real challenge for temperature forecasting. Meteorological models exhibit systematic biases and specific biases for certain atmospheric conditions. For example, temperature inversions, where cold air remains trapped in the valley bottom while warmer air lies above, are common in the region. These inversions are often associated with anticyclonic conditions and persistent low-level clouds that are poorly modeled in forecasts.
Providing forecasts directly from raw meteorological model outputs is unsatisfactory for CCIAG. For most individual models, mean absolute errors (MAE [1]) exceeding 2°C were observed in the first 24 hours of forecasts (winter 2019 to 2022). Steadysun developed a customized solution to reduce these errors. The focus is on reducing the average error over a winter season and decreasing the occurrence of very large errors (daily MAE exceeding 2°C).
null
null
Multimodel forecasting
Our Frogcast forecasting platform utilizes forecasts from the world's best numerical weather models, including those from Météo-France (AROME, ARPEGE), the German DWD (ICON-D2, ICON-EU), the U.S. National Oceanic and Atmospheric Administration (NOAA, GFS, GEFS), the European Centre for Medium-Range Weather Forecasts (ECMWF, IFS-HRES), and the Canadian Meteorological Service (GDPS). These information sources are optimally combined on each grid point of the globe. The primary advantage of this multi-model forecast is to reduce large errors by giving more weight to the most common meteorological scenario, favored by the majority of models.
null
null
Post-Processing and correction by analogy
CCIAG has a temperature sensor near its Poterne production site. These measurements are an essential source of information, as they allow the assessment of forecast quality over an extended period and the development of advanced post-processing tools. The chosen method is based on the principle of analogy, as illustrated in Figure 2. It requires access to meteorological forecast archives and atmospheric reanalyses at a continental scale for certain parameters (geopotential, temperature, relative humidity, etc.), as well as local forecasts for the variable being corrected. For each new forecast, the meteorological situation is analyzed and described using large-scale weather fields called predictors. Days with similar weather patterns are selected. The method then relies on the assumption that for similar meteorological conditions, models will have similar biases. This allows the anticipation of errors in the current forecast and the application of an appropriate correction. This post-processing method has two major advantages:
1) it adapts to the meteorological situation
2) it improves over the years as the depth of archives increases.
Figure 2: Methodology of post-Processing by analogy.
Results
null
The Frogcast multi-model forecast, coupled with the analogy-based post-processing method, has been operational at the District Heating Company since the winter of 2021-2022. Moreover, meteorological archives have extended the method's evaluation over two additional winters.
Figure 3 presents an example of a forecast for 15 days provided to CCIAG. The top curve corresponds to the raw forecast from the Frogcast platform, while the bottom one is the forecast corrected using the analogy-based post-processing. The P20-P80 confidence interval is constructed using the various corrections provided by individual analogous days. It is noteworthy that the multi-model raw forecast relatively simulates the temperature in Grenoble well, with a MAE of 1.46°C for the first three days of the forecast. However, there are significant errors (> 3°C) on particular days, such as November 27. The application of the analogy-based correction significantly improves the forecast by reducing these errors. We achieve a reduction of approximately 0.74°C in MAE between Day+0 and Day+3 and a reduction of 0.6°C for a Day+15 forecast, resulting in improvements of 51% and 34%, respectively.
Figure 3: Example of Raw (Top) and Analogy-Corrected (Bottom) Frogcast Forecasts, 0 to 15 days ahead.
null
null
Performance improvement has been quantitatively assessed over four winters since 2019. Errors (MAE) in the first 24 hours of forecasts are presented in Figure 4 in the form of a probability density function. The application of the analogy-based post-processing not only reduces the mean error (1.2°C versus 2.1°C) but also reduces the occurrence of very large errors exceeding 3% (less than 3% versus 10%), which can be highly detrimental to the management of the urban heating network. These very significant errors, particularly problematic for urban heating network management, have thus been substantially reduced, meeting CCIAG's expectations regarding maximum mean errors.
Figure 4: Probability Density Function (PDF) of Forecast Errors from H+0 to H+24 (Winters 2019 to 2023)
Perspectives
null
As previously mentioned, one advantage of the analogy-based method is its automatic enrichment over time with the increase in available meteorological archive depth. Therefore, we can expect a constant improvement in forecast quality as the years go by.
Other sources of meteorological data can also be integrated into the current forecasting system. One such source is the immediate forecast product from the AROME model, with data updated every three hours, providing refined forecasts for the next six hours. The assimilation of a large number of observational data in this system (ground-based stations, radar, etc.) significantly reduces biases and could bring significant improvements to the performance in the very short-term forecasts.
The deployment of a sensor network around the Grenoble basin, both in the valley and at higher altitudes, could also refine forecasting and develop additional tools. Local phenomena, such as passing showers or the onset of the south wind (foehn), involve sudden temperature variations of up to ten degrees in a few minutes. An alert system based on measurements taken a few kilometers from Grenoble could provide a precise forecast of the timing of these rapid fluctuations minutes in advance.
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
Start for free!
Notes:
1. The normalized Mean Absolute Error (nMAE) is calculated based on half-hour ranges. This is consistent with the usual functioning of the electricity network. To quantify the forecasts or the daily production, we consequently have a set of 48 values corresponding to 48 time slots of 30’ each (1rst value = average between 00.00 am and 00.30 am, 2nd value = averages between 00.30 am and 01.00 am and so on).
Powers are normalized by the peak power in order to allow comparisons between a power plant and another one (see equation 1).
A value for the daily nMAE is then calculated for each day and for each year (see equations 2 and 3).
null
Where Pforecasted and Pproduced are the average power planned and realized on the considered 30’ time slot, Ppeak is the peak power and "48" is the number of half an hour time slots of the day.
The accuracy of the forecast system on a given power plant will then be estimated by averaging the daily MAE’s on all the available days (see equation 3).
null
null
null
Follow the Frog
Through a simple and efficient API, Frogcast promises to make it easy for you to integrate reliable weather forecasts! Join Frogcast now by connecting your application directly to the API!
