| Report | |
| Title | A Deep Learning Approach to HVAC Control Systems |
| Author(s) | Hovaldt Højrup, Andreas (Aalborg U.) ; Fink, Xaver Eugen (dir.) (CERN) ; Fernandez Adiego, Borja (dir.) (CERN) |
| Publication | 2026 - 124. |
| Note | Project Oriented Internship Report |
| Subject category | Computing and Computers |
| Abstract | Model Predictive Control offers good performance for HVAC systems through its ability to optimize control actions over a prediction horizon, while satisfying operational constraints. However, the computational burden of solving online optimization problems limits deployment on resource-constrained industrial controllers. This project investigates the approximation of an economic MPC solution for HVAC control at CERN’s SR1 facility using neural networks, with the goal of reducing computational requirements while preserving control performance. A complete pipeline was developed, comprising data extraction and dataset creation, a modular training framework supporting multiple neural network architectures, and integration with an EcosimPro digital twin for closed-loop evaluation. MLP-, TCN- , and transformer-based models were implemented and compared. The transformer architecture achieved the best performance, with an R 2 of 0.995 for actuator regression and 94.03% class-averaged accuracy for mode classification. In closed-loop simulation, the neural network controller maintained zone temperatures within specified bounds over a week-long evaluation period, while reducing computation time by a factor of approximately 460 compared to the MPC. |
Zapis kreiran 2026-03-09, zadnja izmjena 2026-03-09
