Exciting German study bodes well for heat networks

Date: 15 November 2017

Professor Matthias Finkenrath is leading a team in southern Germany dedicated to producing more effective district heating networks. He spoke to Decentralized Energy about the KWK Flex project.

Finkenrath and his colleagues are using a software they have named DeepDHC (acronym denotes district heating and cooling) to facilitate thermal load prediction on the Ulm-Kempten heat network.

The project had been in their minds but took on some accidental impetus, when members of the University of Kempten’s mechanical engineering faculty happened to strike up a conversation on the subject with their colleagues in the informatics department.

“We met colleagues from the field of deep learning in informatics and fortunately the subject resonated. We worked on the subject for over a year and this is the culmination of those tools from informatics and from our side the application of knowledge of thermal load – this combination has proven very successful and similar things could happen in the coming years in many fields”.
What has resulted is the cutting edge of predictive technology being applied to the heat network between Ulm, a large city of 200,000 and Kempten, a smaller city of 70,000. The focus is on identifying more accurate prediction of future heating loads.

“It’s about dispatch optimisation and we are using novel tools. If they are operating a waste incineration plant or a biomass plant, there is large thermal energy storage involved and the idea was to come up with a tool that gives the better options to operate their plants, to dispatch them to get additional flexibility so they are not just providing thermal power but electrical power if possible.”

The variables involved in making such predictions on load are well-known. Cold weather needs more thermal load for example and some companies use simpler tools, Excel-based tools to predict that maybe a sunny day will require less thermal load. However, Kempten’s technology is comfortable with much more data.

“The differentiator with this novel approach is that these tools are so powerful in extracting correlations that usually standard algorithms or humans are unable to match. For example, we feed them with information on weather forecast, temperature, wind direction, humidity and percentage cloud cover. It can also filter out sensitivities, so for example if you think of temperature and humidity, you know that, in a given day, you might have a certain temperature but on the second day you may have the same temperature but a different humidity. It feels much different and the thermal load need is much different. This technology is very accurate in these cases.”

DeepDHC is based in speech and image recognition technology. That technology has become increasingly effective over the last year in particular, according to Professor Finkenrath. “In terms of speech recognition or translation, tools can now translate the words that are spoken immediately yet also understand the context of the whole sentence or sentence before. That’s why they are so powerful. We essentially use the same tools to extract certain patterns from thermal heating networks. The tools look at the actual thermal load and the gradient and gets an idea as to the direction it is developing but it also looks at weather conditions and a little back to previous days, whether it’s a weekend day, day or night, and so on. These advanced neural networks are very powerful so you can provide many data points, over a million in our case, whether data from the grid itself, the load gradient and trace it with data from the last 15 years. This neural network, or artificial intelligence is trained with much historical data and establishes which patterns are most sensitive and is able to then predict thermal load much more accurately. To give you an idea, if you use the neural network today and ask it what will be the thermal load in three days, it will provide you with an answer within 3 per cent of accuracy.”

The platform being used by the team is based on Google software called TensorFlow, an open source deep learning software similar to that used by Google’s speech algorithms. Using heat network data from the last 15 years, that software is entered into the network’s master control station.

“The main benefit is in this context. If they run a biomass plant to provide heat, it’s CO2 neutral and they would run it that way in general. But if there is a sudden thermal load requirement for a peak, what they would do is typically run a peaking boiler using oil or natural gas that being supplied from conventional fuels. This is very expensive, needs to start up quickly and has a lot of emissions. If they can predict the load accurately, they can operate much more precisely in advance with the heat supplied by the biomass plant, filling it with carbon neutral heat and then once they reach the peak requirement or load they can supply the heat from the thermal energy storage and no longer from the fossil boiler. That is one advantage. Another is that if they have to order fossil fuels a few days before, backup fuel like natural gas, it is better to know how much they would need if they need it at all.”

The Kempten faculty is heartened by results so far but there is a way to go before any talk of a national or international roll-out can be justified.

The German District Heating Association (AGFW) is participating in the project and if it turns out to be successful, it would be aiming for delivery to other municipalities, and indeed, there has already been requests from interested utilities. “We are surprised and happy with testing results so far”, Finkenrath says, “but we’re still at research level. Next Spring, we will have really implemented this tool and it will be interesting as the fluctuations are much higher in the grid in Winter. The information will have been validated by Spring and we will have learned a lot about how well it works as well as any shortcomings.”

The fact that the University has filed a European patent application for this novel method for load prediction is an indication of their confidence. “In terms of thermal load predictions focusing on district heating networks, no one has ever used these deep learning algorithms before. They have been used since maybe 2012 in the communications industry, through speech and image recognition, but we are the first to have the idea to use these most advanced technologies for this purpose.”

Source: Decentralised Energy