Machine Learning Applications in District Heating

in our District Heating Research Trends we briefly included the topic of machine learning (ML) and artificial intelligence applied to the district heating domain. In this issue of our newsletter, we want to have a deeper look into this topic to better understand the potential as well as the barriers for applying machine learning techniques to district heating. We think this is an interesting topic because on the one hand, machine learning has already had a major impact in many other fields and there are significant research efforts to apply it to district heating, but on the other hand the actual impact in the field has argueably been limited so far.

In this context, we think that the paper Opportunities for Machine Learning in District Heating by Gideon Mbiydzenyuy et al. offers many useful insights into the current state of machine learning research with relation to the district heating domain and its relevance for the industry. Furthermore, the paper sets out to define future goals for ML applications in district heating, analyzes barriers on the path towards these goals and makes suggestions on how to potentially overcome those barriers.

Background: A Changing District Heating Landscape

As the background for applying novel technologies like machine learning, the authors describe several fundamental changes in the district heating landscape which introduce uncertainty and challenges for the industry, but may also result in opportunities and motivation to develop and apply innovations to the field. These changes include:

• Increasing market competition instead of old monopolies
• More distributed heat supply instead of centralized heating plants
• Collaboration beyond the district heating sector with other parts of a city's infrastructure and city planning
• A more service-centered approach
• The trend towards lower network temperatures

In general, the paper argues that along the entire district heating value chain, including supply, distribution, heat transfer, and heat consumption, these changes drive a growing pressure to avoid inefficiencies. The paper names several challenges along this value chain to improve the efficiency of the system operation. This ranges from the network supply, including the unit commitment of different heat generators and the control of network supply temperatures and pressure heads over time, through the monitoring of the distribution and supply on to the control of the heat transfer in the substations and the consumption in the building heating systems. This leaves many possible applications for ML to improve the system operation.

Review of Current Machine Learning Research for District Heating

Against this background, the authors did a thorough review of the literature on ML applications for district heating. They considered 179 papers published during the time between 2010 and 2021. During this period, the authors show a clear increase of research papers related to machine learning in district heating. Interestingly, they also show that 72 % of all papers dealing with a concrete ML application to district heating focused on a single such application: forecasting the heat demand in the network. And because this area has already been covered by other reviews and the aim of the selected paper is to better show the range of potential ML applications, the authors put their focus more on an overview of other ML applications.

These other applications include as the second-most researched topic anomaly and fault detection, with individual papers covering different aspects along the district heating value chain. This can range from leakage detection for the pipe network to identifying faulty substation behavior. Beyond that, another application for machine learning is to cluster different customers according to consumption behavior patterns. This can help to better understand their heat use as well as their consumption patterns over time. Other interesting application include the scheduling of domestic hot water heating cycles, or evaluating the conditions of valves in the network by analyzing the emitted sounds.

Regarding the most researched topic of demand forecasting, the paper emphasizes the value of such forecasts e.g. for scheduling the heat production. Yet, they also stress the distinction between *analytics* of the data and actual *control* of the system. And for that, the paper argues that for a major impact on the actual system realizing the full potential of this ML application, closing the loop towards actual control remains a significant challenge. Furthermore, from a practical perspective, they emphasize that for more practical relevance, forecasting technology should focus not only on accurate results, but also consider the asymmetry between overpredicting the demand (leading to a mere inefficiency) and underpredicting it (leading to unacceptable failures to supply all customers).

A Roadmap for Machine Learning Applications in District Heating

Based on this review of the current state of ML research for district heating, the paper identifies a set of 10 goals for future applications. These range from scalable solutions for very large networks, over standardized data protocols, data benchmark sets, and on to automated monitoring and predictive maintenance. Together they form a useful roadmap for future ML applications, all within the overarching idea to work towards a detailed optimization of the district heating operation along the entire value chain.

As for the barriers on the way to achieving these goals, the paper lists among others technical barriers like the limited availability of data and its quality, business barriers like missing incentives for customers to embrace ML-based applications, missing readiness of the industry to embrace innovations, and organizational barriers like the owner structure which currently often leaves the operators of a network cut off from the operation of the substations.

To overcome these barriers, the authors suggest several solutions. These include a better communication between researchers and the district heating industry to better understand actual real-world challenges, evaluate the potential, and find novel solutions. In addition, digitalization of district heating offers a path towards overcoming some of the barriers, but this still requires a good understanding of its challenges and the potentials. And last but not least, the paper suggests an interdisciplinary approach to address these complex challenges.

Further Information

The paper is available as open access at https://doi.org/10.3390/app11136112 and we recommend to read it fully as it includes many interesting aspects beyond what we could cover in this newsletter issue. Furthermore, the paper mentions the (Digital road-map for district heating and cooling) which makes for an interesting complementary reading regarding the digitalization of the DHC sector.

The next issue of our newsletter will be published July 6.