Bachelor and Master Theses

To apply for conducting this thesis, please contact the thesis supervisor(s).
Title: Learning-Based Control of Hybrid Cooling Systems for Data Centers
Subject: Computer science, Embedded systems, Software engineering, Applied Artificial Intelligence
Level: Basic, Advanced
Description:

Master's Thesis Proposal 

Title: Learning-Based Control of Hybrid Cooling Systems for Data Centers 

Background 

Cooling accounts for a large share of a data center's energy use. Many cooling plants combine dry coolers, which use outside air, with chillers, and often have a water storage tank. These plants are usually run with fixed rules. The rules are simple, but they waste energy and ignore weather and load forecasts. Machine learning specially  reinforcement learning can perform better, but common methods require extensive trial and error, are hard to trust, and often fail on a new plant without retraining. 

Problem statement 

There is no simple and reliable way to run a hybrid cooling plant hour by hour so that it: 

  • costs as little as possible to run, 

  • keeps the water temperature within safe limits, and 

  • works on new plant layouts and in new climates without retraining. 

Aim 

To develop and evaluate a control method that learns a physics-constrained model of the cooling plant and uses it to plan  reliable and low-cost operation. 

Objectives 

  • Develop a physics-constrained predictive model that estimates the thermal and operational behavior of a hybrid cooling system. 

  • Develop a planner that uses forecasts and keeps a safety margin based on the model's measured error. 

  • Assess generalization and robustness across different climatic conditions, cooling demand profiles, and plant configurations 

Research questions 

  • How accurately can a physics-constrained model predict plant behavior on new layouts and climates? 

  • How much running cost can model-based planning save compared with rule-based and reinforcement learning control? 

  • Can the model-measured error be used to maintain the temperature limits? 

Methodology 

  • Generate operating data from the simulator. 

  • Train the physics-constrained model on this data. 

  • Build the planner, for example with dynamic programming or model predictive control, and add a safety margin from the measured model error, for example with conformal prediction. 

  • Test all controllers across many climates and layouts and compare the results with simple statistical tests. 

Expected results 

  • A working control method with open-source code 

  • A test bench that others can reuse 

Scope and limitations 

  • Simulation only, with no tests on a real plant 

  • One type of cooling plant (dry coolers, chillers and a water tank) 

  • Decisions are made once per hour 

 

 
Start date: 2027-01-25
End date: 2027-06-29
Prerequisites:

Required background 

  • Good Python programming 

  • Basic machine learning 

  • Interest in optimization or control 

  • Basic heat transfer or thermodynamics is helpful but not required 

Tools and technologies 

  • Python, NumPy, pandas 

  • PyTorch 

  • Building energy simulation (for example, TRNSYS) is helpful but not required 

 
IDT supervisors: Shahina Begum Shaibal Barua
Examiner: Mobyen Uddin Ahmed
Comments:

PhD student Md Mohsin Kabir will provide the domain and data.
Company contact: