Intelligent 5th generation heating and cooling networks (5GDHC) are a key component of the energy transition in urban areas. However, controlling them is one of the biggest scientific hurdles on the way to a sustainable future. Alexander Schlüter addressed this topic in his presentation at the international SDEWES 2025 conference. Instead of presenting a ready-made solution, he offered an honest assessment of the complex and fundamental challenges that need to be overcome on the way to AI-controlled regulation of these future grids.
The vision: a network of thermal prosumers
The key innovation of 5GDHC grids is the shift from passive consumers to active prosumers. Instead of a centralised producer supplying heat unidirectionally, the network participants themselves become actors: they can draw (consume) thermal energy from the network, but also feed (produce) their own waste heat - for example from industrial processes or building cooling - into the network. This paradigm creates a highly dynamic, bidirectional energy system whose complexity makes conventional control impossible.
A whole spectrum of challenges
This new dynamic, driven by a multitude of interacting prosumers, brings with it a whole spectrum of profound scientific problems. In his presentation, Prof Schlüter focused on the fact that the development of an AI-controlled system is far more than just a programming task. The hurdles presented as examples include:
- The inertial physics of grids: In contrast to power grids, thermal grids react extremely slowly. AI decisions often only have an effect after a long time delay. It is extremely difficult for an algorithm to correctly learn these complex cause-and-effect chains and act with foresight.
- The dilemma of AI agents (local vs. global goals): The approach pursued at the chair uses decentralised AI agents. This creates a classic dilemma: how do you ensure that each individual agent not only selfishly optimises its own task, but also acts in the interests of the overall network? Designing the right incentive systems ("reward shaping") is one of the most subtle and at the same time most important tasks here.
- The sheer scale of the problem: The number of possible states in a network full of prosumers is astronomically high. An AI must efficiently learn to distinguish good from bad decisions in this huge solution space. This can lead to extremely long training times ("sample inefficiency") and requires highly efficient algorithms.
However, these points are just a few of the dozens of core problems identified, which also include coupled hydraulic dynamics, the constant change in environmental conditions ("non-stationary environment") and the fundamental issue of error-free coordination between the AI agents.
Conclusion: Understanding the problem is the key to success
The decisive contribution of the research presented at SDEWES 2025 therefore lies not in a ready-made "plug-and-play" solution, but in analysing this complex problem area. This fundamental work is central to the mission of the NIWI Chair to actively shape the industrial and urban transformation through excellent research.