AI4Energy

Data-driven modeling enables the development of predictive models that can uncover patterns, trends, and insights in data. These models are crucial for making informed decisions, optimizing processes, and gaining a deeper understanding of complex systems based on empirical insights derived from the data. Methods used in data-driven modeling include machine learning algorithms and statistical analyses that employ various approaches such as regression, clustering, and neural networks.

Energy systems are complex infrastructures that enable and optimise the generation, storage, distribution, and use of energy in various forms. Our research group focuses primarily on the modelling and control of building energy systems – ranging from single-family homes and office buildings through to entire districts.

Two complementary approaches are employed for the optimal control of building energy systems. Model Predictive Control (MPC) is a modern control strategy that uses a dynamic model of the system – for example, a data-driven neural network – to predict future behaviour and, on that basis, optimise the control signals over a finite time horizon.

Complementing this, Reinforcement Learning (RL) is applied, whereby an agent independently learns optimal control strategies through continuous interaction with its environment and feedback in the form of rewards or penalties. Both methods enable the development of intelligent control systems that continuously adapt to environmental conditions and user behaviour, thereby maximising both the energy efficiency of buildings and the comfort of their occupants.

BuilDa is a tool for generating building data for data-driven methods in building energy and control research. The focus lies on developing simulation and data frameworks that can realistically replicate thermal building dynamics and produce them in large quantities. This data forms the foundation for machine learning methods, in particular for transfer learning and predictive models in intelligent building control.

The laboratory develops flexible simulation environments in which various building parameters, weather conditions, usage profiles, and control strategies can be systematically varied. This makes it possible to investigate a wide range of building types and operating conditions, and to develop robust, data-driven models for energy-efficient buildings. The aim is to research and validate new methods for intelligent energy management, predictive control, and data-driven optimisation of buildings.

More information at: https://ieeexplore.ieee.org/abstract/document/11118811
and arxiv.org/abs/2512.00483

The Time Series Lab is dedicated to the development of data-driven models for time-dependent data. The focus lies particularly on neural networks such as Multi-Layer Perceptrons (MLP), LSTM networks, and Transformer architectures, which are used for the modelling, prediction, and analysis of time series.

The laboratory is flexibly structured and can be operated with any dataset – for example, with synthetic building data from BuilDa as well as real measurement data from a variety of applications. This enables models to be systematically trained, compared, and validated in order to develop robust, data-driven solutions for diverse research questions.

The RL Lab is a laboratory for the development, training, and evaluation of reinforcement learning agents for energy-related applications in buildings. Within a flexible simulation environment, agents learn various control tasks – ranging from temperature control to energy-efficient operational optimisation of building systems.

The agents can be trained on both data-driven models and simulation models, and subsequently compared against conventional control strategies in order to objectively assess performance, robustness, and energy-saving potential. In addition, the RL Lab supports the investigation of transfer learning approaches, in which trained agents are transferred to new buildings and further developed there efficiently.

The MPC Laboratory enables the control of building energy systems using Model Predictive Control. Models of all complexity levels are employed – from physically grounded white-box models and grey-box approaches through to data-driven black-box models. A key advantage: the models developed in the AIMD Lab can be evaluated directly as predictors within the MPC control loop. Control is currently implemented in an FMU-based building simulation and is designed with a view to future deployment in real buildings. In addition to the in-house BuilDa simulation, the environment also supports the integration of external simulation tools.