Learning and Model Predictive Control Applied to Energy Optimization of Chiller Plants for Data Centers
(2025)Department of Automatic Control
- Abstract
- Buildings are responsible for approximately 30% of global energy use, with HVAC systems accounting for 40% of that demand. Among these, data centers have become significant energy consumers, requiring highly efficient and stable cooling solutions. This project aims to explore data-driven control methods to improve the energy efficiency of data center cooling systems as well as reduce the effort and expertise required to implement advanced control strategies, such as model predictive control (MPC). If this can be done while maintaining system stability and safety, the method would offer a scalable pathway to smarter and more sustainable cooling control.
The approach is centered on setpoint control of a simulated model of a data center... (More) - Buildings are responsible for approximately 30% of global energy use, with HVAC systems accounting for 40% of that demand. Among these, data centers have become significant energy consumers, requiring highly efficient and stable cooling solutions. This project aims to explore data-driven control methods to improve the energy efficiency of data center cooling systems as well as reduce the effort and expertise required to implement advanced control strategies, such as model predictive control (MPC). If this can be done while maintaining system stability and safety, the method would offer a scalable pathway to smarter and more sustainable cooling control.
The approach is centered on setpoint control of a simulated model of a data center cooling system with IT load variation and weather conditions as disturbances. The project involves system identification and an MPC variant, differentiable predictive control (DPC). The tool for the DPC implementation is NeuroMANCER, a differentiable programming library built on PyTorch, which enables the integration of neural networks with physics knowledge.
The results indicate that DPC has the potential to reduce energy consumption on the order of 1% compared to a baseline control strategy. In comparison, an estimated optimal setpoint controller achieves a 2% reduction. These findings suggest that DPC offers a promising, scalable approach to smarter and more sustainable cooling control in data centers. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9212170
- author
- Hübsch, Oskar and Horovitz, Simon
- supervisor
- organization
- year
- 2025
- type
- H3 - Professional qualifications (4 Years - )
- subject
- report number
- TFRT-6291
- other publication id
- 0280-5316
- language
- English
- id
- 9212170
- date added to LUP
- 2025-09-18 14:08:09
- date last changed
- 2025-09-18 14:08:09
@misc{9212170,
abstract = {{Buildings are responsible for approximately 30% of global energy use, with HVAC systems accounting for 40% of that demand. Among these, data centers have become significant energy consumers, requiring highly efficient and stable cooling solutions. This project aims to explore data-driven control methods to improve the energy efficiency of data center cooling systems as well as reduce the effort and expertise required to implement advanced control strategies, such as model predictive control (MPC). If this can be done while maintaining system stability and safety, the method would offer a scalable pathway to smarter and more sustainable cooling control.
The approach is centered on setpoint control of a simulated model of a data center cooling system with IT load variation and weather conditions as disturbances. The project involves system identification and an MPC variant, differentiable predictive control (DPC). The tool for the DPC implementation is NeuroMANCER, a differentiable programming library built on PyTorch, which enables the integration of neural networks with physics knowledge.
The results indicate that DPC has the potential to reduce energy consumption on the order of 1% compared to a baseline control strategy. In comparison, an estimated optimal setpoint controller achieves a 2% reduction. These findings suggest that DPC offers a promising, scalable approach to smarter and more sustainable cooling control in data centers.}},
author = {{Hübsch, Oskar and Horovitz, Simon}},
language = {{eng}},
note = {{Student Paper}},
title = {{Learning and Model Predictive Control Applied to Energy Optimization of Chiller Plants for Data Centers}},
year = {{2025}},
}