GRADS – Generating artificial district heating supply systems
Due to the limited number of district heating supply systems in Germany and their individuality, their data are difficult to use for research purposes and unpublishable. The sensitive data reveals the affiliation to specific providers, posing a risk of abuse. Artificial intelligence, specifically deep reinforcement learning (DRL), is used for research purposes to create artificial models of district heating supply systems that are both representative and abstract enough to be able to investigate current issues in the district heating industry and politics through modeling and simulation. For example, this may include analyses for decarbonizing district heating regarding cost-optimized transformation pathways, the influence of regulatory framework conditions or the significance of hydrogen in district heating.
The project is interesting to:
District heating suppliers, utility companies
Using deep reinforcement learning to generate characteristic parameters of synthetic district heating supply systems
Developing methods and proof of concept based on the use of artificial systems to identify cost-optimized pathways for decarbonization with investSCOPE
Reinforcement learning is the third group of machine learning methods (with the other two being supervised and unsupervised) where the target solutions are generated based on self-learned strategies and as feedback to rewards received in a trial-and-error process. This requires no training data. Instead, a large number of simulation repetitions are performed to improve the strategy and maximize the reward in order to find the optimum solution.
investSCOPE optimizes transformation pathways for the decarbonization of generation portfolios while taking into account financial ratios, determining which technology to invest in at what time and at what costs to achieve an economically optimal pathway to portfolio development.
Robust and viable investment strategies are identified by taking into account anticipated cost and price developments of technologies, fuels, electricity and CO2, a change in heat or power requirements, and incentive mechanisms.
The Python- and Pyomo-based tool uses the method of mixed-integer linear programming (MILP) as well as the net present value method.
Städtische Werke Kassel
TU Berlin, Institute of Energy Technology
Project participants: Martin Wiemer, Pedro Gíron, Nazgul Asanalieva, Britta Zimmermann
A publication is planned as part of the ECML PKDD 2022.