Learn more about the use case:

Reasonance GmbH

Forecasting / Prediction


Optimization of a district heating network

What drove the decision to develop an AI-based use case?
AI methods were essential for various aspects of optimizing the parameters of the district heating network. For example, we had to predict the energy consumption per subnetwork. In other words, how much energy all households connected to a distribution network will need in the future (aggregated). Sensor data such as temperature or cloud cover were given and could also be used for predictions by external service providers. Purely statistical methods can only be of limited use for these concerns, as they can map the effects of different components, but in our case could not describe future values accurately enough for high-dimensional input data. Therefore, we decided to use AI methods. For this, we used a neural network that learns probability distributions based on the high-dimensional temporal sensor data.

What were your expectations or requirements for the AI application?
The most important requirement was to have a low level of uncertainty. This means that our model had to have a high overall accuracy, but also not allow strong deviations from the target values. In the case of consumption prediction, we can imagine that the network consists of 10 distribution networks. Assuming that the prediction yields a total error of 10 megajoules (MJ), various scenarios could be envisioned. On one hand, each distribution network could provide an error of 1 MJ. On the other hand, 9 of the 10 distribution grids could also achieve a perfect prediction, while the tenth distribution grid has an error of 10 MJ. While the prediction provides the same accuracy in both cases, it would be fatal for our prediction if much too little or too much demand is predicted for one distribution network. Thus, we decided to use a model that also predicts uncertainty as an additional value in addition to overall accuracy.

Which partner did you rely on with which technology for the development?
We were reliant on our customer as they had to provide us with the necessary data. Without a meaningful data set, which we were guaranteed before evaluating the technology, the project could not have been successfully completed in this capacity. Additionally, we used local weather data as well as weather forecasts. While historical and current weather data are facts, forecasts are prone to risk.

The use of AI solutions must also be accompanied by a change in competences. How have you dealt with this challenge in respect of your employees?
As a research-driven technology company, we know the added value of AI solutions. The implementation at our client's site went without a problem, as they were able to weigh up the pros and cons and were convinced of the added value of this AI solution based on our technology evaluation. A solution without AI was no longer an option for them. From our experience, a clear weighing of the pros and cons of different technologies is helpful here, so that companies whose expertise lies in other areas can shape the choice of technology themselves. Knowledge transfer is also very important, so that the technology provided also helps the people who are supposed to use it. For us, this means that we actively design workshops and training courses on this topic.

In what timeframe did you implement the AI solution in your company?
The project started in May 2019 and was rolled out in December 2019. This includes the design of the project, the digital transformation of the district heating network and its parameters, the estimation of the parameters not given, the optimization of the system, the deployment in a production environment and the provision of real-time monitoring.

A conclusion to your AI solution:
Our AI solution enabled us to make the district heating network more sustainable by reducing transport losses through the use of our software. While we were already convinced by the technology, our client was able to clearly see the added value of this application. It also confirmed that relatively simple models are often sufficient to successfully solve complex problems. In addition, it was important for us that the model could improve over time. This means that it can be re-trained with the real-time data at a later point in time. Thus, our model always represents values up to the current point in time and provides our client with a more accurate prediction for future measured values.

One final piece of advice/tip to other entrepreneurs looking to apply AI:
AI often offers advantages, but is also often misused to appear hip and modern. Many companies offer consulting on possible use cases with this topic, so that you can get an idea of whether the use of this technology is worthwhile for your company. Often, much simpler methods are sufficient to achieve the desired results. We are happy to offer our expertise in this area and look forward to you contacting us.