Development of forecasting and control algorithms using machine learning.
In contrast to standard methods for controlling heat generators, the project will use a variety of internal and external input variables for control. These include weather data, electricity prices, electricity generation and correlation variables for occupancy. Not only current state variables are used in the decision-making process for control, but also future values. This means a predictive and forward-looking control instead of a reactive standard control. For such an optimization task with a large number of input variables and a longer observation horizon, machine learning methods such as reinforcement learning are suitable. The advantage of this method is a demand-oriented and economical heat supply. With the help of simulation models, a large number of scenarios can be simulated and used as training data.
The goal of the research project is to automate the optimization of cross-sector energy systems and to use machine learning and metadata to continuously adjust system parameters. Through transfer learning, measurement data from other buildings can be used as training data. Heating energy demand is predicted using neural networks. The developed prediction and control models are applied to several buildings in combination with the communication interface. Thereby generalization possibilities shall be found, by which the training can be done faster for new buildings. The unique selling point of the project idea is the symbiosis of algorithm development and its immediate validation in a practical context.
The Hermann-Rietschel-Institut focuses on the development of the heating load prediction model and the control algorithm for the heat generators.
Funding

Project term | May 2023 to April 2026 |
Funding code | 03EN1076B |
Project management | Alexander Neubauer |