Fraunhofer Institute for Solar Energy Systems ISEMulti-family buildings offer largely untapped potential for energy savings and grid services in the US and in Europe. Against the backdrop of rapid electrification, the rise of heat pumps, and increasing regulatory pressure, the project "ABLM: Automated Building Load Modeling" project is developing physics-informed neural networks that can accurately predict the load flexibility of central consumers (heating, domestic hot water, cooling, e-mobility) with just a few input data. The hybrid approach overcomes the limitations of classic simulation and pure AI models and enables scalable, robust demand-side management.
Automated Building Load Modeling: Physics-Informed Neural Networks to Automate Load Flexibility Modeling for Multifamily Buildings
ABLM
Economic and ecological advantages in the operation of apartment buildings, such as those in the Gutleutmatten district of Freiburg, can be achieved through dynamic load management.
Initial Situation
Rapid electrification – in particular, the strong growth in heat pumps—and growing regulatory pressure to move away from fossil fuel heating systems underscore the need for scalable demand-side management solutions. Accurate modeling of load flexibility in multifamily buildings is crucial to decarbonizing space and domestic hot water heating while minimizing grid impacts. Existing physics-based simulation tools provide high-resolution insights but are expensive to calibrate and difficult to scale, while purely data-driven approaches often lack the robustness for reliable load forecasting under diverse conditions.
Objective
"ABLM: Automated Building Load Modeling" aims to develop validated neural networks that are capable of accurately simulating and predicting the load flexibility of different types of MFB in the US and Germany.
Approach
The project is developing a hybrid, physics-informed neural network (PINN) framework that combines real energy data from decentralized (US) and centralized (Germany) building systems with simulated data from validated white-box models. This allows realistic variations in heating, domestic hot water, and electric vehicle loads to be captured. The embedding of physical principles in the model architecture combines the scalability of AI with the precision of physics-based approaches and enables more accurate, cost-efficient predictions of load flexibility in different types of multi-family buildings.