We are really proud to have taken part in and contributed to this wonderful project.
Marinette Loisy
After graduating with a Master's Degree in Business Decision Techniques (with a focus on statistics and IT) at the Université Paris 1 Panthéon-Sorbonne, I joined the Group in 2008, specifically the Lab Crigen Research Centre. I worked on extremely interesting projects covering a wide variety of problems, including one that involved usage forecasting for gas. The aim was to optimise the size of the gas pipes in line with forecast usage. I also worked with the H2 lab analysing the behaviour of hydrogen-powered vehicles, with the Environment and Society lab to study the correlation between the weather and the presence of bats near wind turbines, and with the Liquefaction lab to identify the LNG carrier routes that generate the least LNG loss.
I also helped to detect the behaviour that heralds breakdowns in biomass boilers, with a view to predictive maintenance, by analysing data from sensors and breakdowns recorded in previous years.
In 2020, I obtained a Certificate of Specialised Studies (CES) in Data Science from Télécom Paris.
Marinette Loisy
My degree says Data Scientist, but for me it's just a sort of 'buzzword' for anything to do with statistics. In my head, I'm still a statistician.
It's true that we now have access to increasingly large volumes of data and new techniques for mining it, which wasn't the case before. And data is becoming increasingly important.
Slawomir Pietrasz
I have an engineering degree from Ecole Centrale Paris (now CentraleSupélec), majoring in applied mathematics. My work first focused on biology, and then on energy when I joined Gaz de France (now ENGIE) in 2005.
I've worked on a number of gas-related projects and on decision-support tools.
Like Marinette, I also got a Certificate of Specialised Studies (CES) in 2020 , but it was more focused on artificial intelligence than data analysis. I wanted to update my knowledge of statistical learning techniques, artificial intelligence in the broad sense, possible methods and architectures and their applications, such as autonomous cars, robotics based on reinforcement learning, convolutional neural networks, ontologies, recommender systems, etc. The course allowed me to understand what lies behind all these keywords.
Predity for Asset is a tool of the Energy Solutions GBU, which manages the energy mix of eight cooling networks and over 120 heating networks, including some ten international networks. Heating networks face volatile energy prices depending on the weather, and choosing one energy source over another can have serious financial and environmental consequences. Predity for Asset collects over 600 million pieces of data a day, from sensors on the networks, but also from external sources such as the weather or market prices.
For the past 6 months, Predity for Asset has incorporated two tools developed by the Crigen Lab - Predi Demande and Predi Scenario-, which serve to simplify and automate this decision-making process. Our main contribution involved these two building blocks.
Predi Demand is an AI tool that uses 12 months of production data for each network to correlate production with the external variables of temperature and calendar. It can thus be used to predict demand over the next 14 days. These predictions are generated automatically every night for all networks.
Predi Scenario takes this result and allows engineers to add technical and financial characteristics for each asset: amount and types of equipment, installed capacity, fuel costs, CO2 emissions and operational constraints such as maintenance or contractual obligations. This tool uses a powerful algorithm and a solver to calculate an optimal scenario within a few minutes.
This is one of the first practical applications of AI to improve performance, developed through internal collaboration between the R&I and GBU teams. It has direct impact: more reliable forecasts lead to better energy management, improved business performance and a reduced carbon footprint, by maximising the use of green energy and reducing costs.
Slawomir Pietrasz
Last year, with the Cylergie Lab and the GBU we had already submitted this project for consideration, but from a technical point of view. The project was not selected because it was not mature enough.
This time, the project was resubmitted at the GBU’s behest. The GBU chose to give the project a slightly broader scope, stressing the new forecasting modules that Marinette had developed and Predi Scenario, the tool to help manage heat production assets, which have since been deployed operationally.
Slawomir Pietrasz
The initial project -the one submitted in 2023- concerned a specific use of biomass power plants. The GBU asked for the scope of the forecasting function to be extended and for new equipment to be taken into account for the production function. After several tests, we decided that it would be more interesting to rewrite the code using a well-honed method for the prediction function.
Marinette studied a wider range of networks and produced a different forecasting model, adapted to the broader functional scope being sought.
Other Crigen colleagues took part in implementing the model, which Marinette tested on the data and then transcribed so that it could be integrated into the customer environment.
Our aim to stay as close as possible to the customer's needs led us to sometimes rewrite and sometimes compile the implemented code to integrate it into the Predity platform. The bulk of this work concerned the forecasting tool that Marinette was working on.
A fine example of collaboration between Labs
Originally, it was about responding to a very old problem: will it be cold? How much wood will I need?
Applied to the professions of Cylergie contributors, the need was a little more precise: how many trucks of wood to order, day by day, taking into account the characteristics of the biomass boiler room and the demands of the heat network? To anticipate variations in the network's heat demand, Cylergie had developed a forecasting model, based on neural networks, which proved to be very effective. The tool (BM Conso) was successfully tested on 5 sites during the winter of 2021. But, to ensure a massive deployment, Cylergie suggested to Lab Crigen that the product be industrialized.
In order to do that, Cylergie resumed the development of the application while Lab Crigen resumed its integration and hosting, strengthened the forecast and grafted a much broader and more complex functionality: the optimization of network asset commitments: Predi Scenario, a functionality prioritized by the client.
After many refinements, the forecasting tool (Predi Demand) and the commitment optimization tool (Predi Scenario) were integrated into Predity for Assets from ENGIE Solutions. Since December, the range has been completed by Predi Biomass, thus returning to the original question.
A fine example of collaboration between Labs for sometimes complicated research projects, with unexpected… but concrete results.
Slawomir Pietrasz
The metaphor that comes to mind is 'backpropagation', an algorithm used in neural network training.
The first project was developed based on the need to manage the supply of biomass boilers. To do so, we needed to develop a tool that could predict the number of trucks and the quantity of biomass required. But to do that, we needed to predict what the network's energy demand would be, and we needed a function to predict the quantity of heating energy, with a very local scope, just a few networks.
I suggested to proceed by optimisation rather than by simulation, to help them make the best possible decisions in the shortest possible time. Instead of calculating different scenarios by hand, the idea was to enter constraints that the tool would use to calculate the best scenarios.
It turned out that to run, this tool needed a forecast, which gave rise to Predi Demand, that could be used in part for Predi Biomass.
In short, we started from a requirement and realised that there were intermediate building blocks to be created before we could meet that final requirement. And these intermediate building blocks were integrated into Predity, enabling it to win an OEA in the process.
Marinette Loisy
When you're working on a lot of projects looking for new ideas, but in the end these projects are not deployed, you sometimes feel frustrated. A project that ends up being set aside is frustrating for those who have done the research, and it's also a problem for managers, who invest in projects for which they get no return.
The new R&I creation strategy requires us to work with the aim of industrialising, deploying and being profitable. And indeed this project matches that goal, because we worked on it, it was deployed and it is enabling the Group to make savings while also meeting sustainable development objectives.
Slawomir Pietrasz
I’d like to add that when you work on a project or an idea for a bit, it continues to mature in your mind, and when you find yourself in a situation where you see similarities, you think: “I could try to apply this solution, I can easily see how to apply that technique to this problem.”
This was the case for my optimisation model, and was also the case for Marinette, who had already worked on a forecasting model and was sure she could adapt it.
Obviously, we had to research the data and analyse it to identify the relevant factors, but thanks to our previous research, we knew and mastered the type of model we wanted to use to tackle this problem.
The model I've rewritten is inspired by what I did for a multi-energy network in Singapore for the REIDS-SPORE project. The way I designed the optimisation code is more generic, simpler as regards the model, and a little more elaborate as regards the data, so it requires a little more work on the customer’s side, but a little less maintenance for the developer. I tried to design it so that it could be deployed on the customer's site as simply as possible, so that it didn't involve any major IT engineering that the customer wouldn't want.
We re-use the experience of what didn't work and why it didn't work. There are bound to be failures and unsuccessful projects, some of which end up being set aside, but every time we get back to work, we say to ourselves, “OK, what can I do to make this a success, to get it to the customer? What did we miss?”
Marinette Loisy
Yes indeed, every time you work on a project, even if it doesn't come to fruition, you acquire new techniques, learn new algorithms, new ways of coding that are going to be more efficient. And of course, if we need them later on and they make our work easier, we reuse them. When I worked on this forecasting model, I based it on what I'd learned, on what I'd been able to do when I did gas forecasting, because it's more or less the same principle.
Of course, there were discussions with the trade to discover the specific features of their networks, but there is a methodology and a way of moving forward that is somewhat similar.
Marinette Loisy
It was a great moment, especially as we weren't expecting it. We'd been led to believe that we weren't going to win, so it was a big surprise.
It's also nice to be recognised for the work we do, and to be able to showcase our Lab and our skills.
And it shows what we can do and how useful it is. The fact that the project is being rolled out across the world and not just in France also gives it weight.
So we were really proud to have taken part in and contributed to this wonderful project.
Slawomir Pietrasz
What I'm pleased about is, as Marinette says, the international roll-out for a large number of users.
Deploying it on a French scale with some twenty users in different regions of France was already a success. It's very satisfying to tell yourself: it’s deployed, it's going to be useful. And particularly because Marinette's model represents the basis, the building block of demand forecasting that feeds other tools that may be used less frequently.
What's more, there will undoubtedly be adaptations, and some countries will probably have even more specific needs that we may be asked to work on.
It's satisfying to know that the model I developed and then adapted into a more compact version is integrated into an operational platform. I'm also pleased to have brought on board people whose skills I value and who are keen to contribute. And for them to be recognised for this success, for the project to shine a light on them.
Marinette Loisy
We're talking to the Energy Solutions GBU, and we're going to be meeting the teams shortly to see if there are things that could be improved after a few months of using the tool, or things that they think would be more ergonomic or relevant to upgrade.
We are also working on developing demand forecasting models for cooling networks, since the present tool is designed for heating networks.
And as the current model was developed with a degree of urgency, we had quite a few other ideas that we didn't have time to explore to make it even better, and this we'll be able to do in the future.
