Truace
Subscribe
52/100

Moderate evidence

Cite
More
Export CSVExport JSONEmailLinkedInX

Simulation suggests AI cooling control cuts data-center cooling energy by 25–28%

Study title: Towards Greener Data Centers: Integrated Optimization of Cooling and Resource Usage via Machine Learning

Evidence type
Benchmark or lab evaluation
Sources
1 sourcePeer-reviewed
Published
2 min read
Truvace record
TRV-2026-2626cite or check it

Study title: Towards Greener Data Centers: Integrated Optimization of Cooling and Resource Usage via Machine Learning

Researchers at the Pune Institute of Computer Technology built a simulated data center that uses a Random Forest model to predict workload and a Deep Q-Network agent to set cooling temperatures and place virtual machines. In the main scenario it cut total energy 20.7%, from 12,450 kWh to 9,870 kWh, and cooling energy 28.4% against a static 18°C setpoint.

Read the study abstract

Data centers form the backbone of the global digital economy, yet their exponential growth has led to significant energy consumption, with cooling systems alone accounting for 30 to 50% of total energy usage.This paper proposes an integrated framework that combines machine learning based workload prediction with dynamic cooling control to achieve holistic energy optimization.The system employs a multi-layered architecture comprising real-time sensor telemetry, predictive analytics (Random For-est and Reinforceme…

Towards Greener Data Centers: Integrated Optimization of Cooling and Resource Usage via Machine Learning
"Front of server racks at NERSC data center" by D Coetzee is marked with CC0 1.0. To view the terms, visit https://creativecommons.org/publicdomain/zero/1.0/.

The gain

In a simulation of a data center, a machine learning controller cut cooling energy by 28.1% on average and lowered PUE from 1.58 to 1.42 compared with a fixed-setpoint baseline.

In brief

Cooling can take 30 to 50% of a data center's energy, and the paper cites projections that data centers could use 4-6% of global electricity by 2030. Savings held at 26.5% in a bursty workload and 25.3% in a scenario based on a 24-hour Google cluster trace, averaged over 10 runs each.

The results come only from simulation. The authors say real-world sensor noise, actuator delays and mixed hardware are not fully captured, and they plan to test on a physical testbed.

Main points

  1. The simulated facility had 100 racks of 20 servers; the main model was scaled from a 10-server setup representing 2,000 servers.
  2. In the diurnal-plus-spikes scenario, cooling energy fell from 4,980 kWh to 3,565 kWh and PUE from 1.58 to 1.42.
  3. A rule-based controller saved 17.2% of cooling energy but ran hotter, with average maximum temperatures of 34.8°C versus 33.3°C for the ML controller.
  4. Across three workload scenarios, cooling savings were 28.1%, 26.5% and 25.3%, each averaged over 10 runs with different random seeds.
  5. A paired t-test across the 10 diurnal runs gave p < 0.001, and varying the Random Forest settings changed savings by less than ±1.5%.
  6. The controller moves virtual machines onto fewer servers at low demand, which lets idle servers enter low-power states and reduces IT energy.
  7. In the bursty scenario PUE improved only from 1.61 to 1.47, and the authors note the controller can sometimes raise PUE when IT power drops faster than total power.

The full story

The system has four layers: a sensor network, a data fusion engine using Kalman filters and Bayesian inference, an ML predictive controller, and an actuation layer that adjusts cooling unit setpoints, fan speeds, chilled water valves and virtual machine placement. Safety constraints keep temperatures within ASHRAE recommended limits.

The Random Forest, with 100 trees and depth 10, was trained on 70% of 10,000 simulated time steps and predicts workload 15 minutes ahead. The controller adjusts setpoints between 18°C and 27°C. Simulations ran 24 hours in 5-minute steps.

In the main scenario, IT energy fell from 7,470 kWh to 6,305 kWh because of workload consolidation, and average maximum temperature dropped from 35.4°C to 33.3°C. The authors say the Google trace results show the approach generalizes beyond synthetic patterns.

The authors say future work will add renewable energy availability and electricity pricing, extend the agent to multi-zone hyperscale facilities, test on a physical testbed and add liquid cooling, which they say could cut cooling energy by up to 40%.

Sources

1 on file
  1. International Journal of Computer Applicationsdoi:10.5120/ijca2e7b62930f25
    Peer-reviewed2026-05-30

Talk it over

Four questions to start onWho gets this benefit, and who was left out of the study or story?

Start with a question

Simulation suggests AI cooling control cuts data-center cooling energy by 25–28%https://truvace.com/item/towards-greener-data-centers-integrated-optimization-of-cooling-and-resource-usa

  1. Who gets this benefit, and who was left out of the study or story?

    The gain In a simulation of a data center, a machine learning controller cut cooling energy by 28.1% on average and lowered PUE from 1.58 to 1.42 compared with a fixed-setpoint baseline.

    Answer in the thread

  2. How do we know? What can this kind of evidence show, and what can’t it?

    The evidence This rests on a benchmark or lab test, from International Journal of Computer Applications.

    Answer in the thread

  3. What does it leave open? What else would you want to know before acting on it?

    The record’s own caveat The evaluation is simulation-based; real-world sensor noise, actuator delays and hardware differences are not fully captured, and the reinforcement learning agent uses a simplified state space.

    Answer in the thread

  4. Have you seen AI used this way where you study, work or live? Did it go the same way?

    Answer in the thread

The questions are the same on every record. The details under them are taken from this record as written.

What readers say

More from Climate