Truace
Subscribe
52/100

Moderate evidence

Cite
More
Export CSVExport JSONEmailLinkedInX

Data-center cooling and power AI cut energy costs 25.8% in tests on Alibaba traces

Study title: PoCo: Joint Optimization of Power and Cooling for Low-Carbon CCHP-Integrated Data Centers

Evidence type
Benchmark or lab evaluation
Sources
1 sourcePeer-reviewed
Published
1 min read
Truvace record
TRV-2026-2627cite or check it

Study title: PoCo: Joint Optimization of Power and Cooling for Low-Carbon CCHP-Integrated Data Centers

Researchers propose PoCo, a deep reinforcement learning framework that controls power generation and cooling together in data centers fed by combined cooling, heating and power (CCHP) systems. Tested on real-world Alibaba cluster traces, it cut energy costs by 25.8% and carbon emissions by 15.3% against baseline algorithms, and improved power usage effectiveness by 4.9%.

Read the study abstract

The recent explosive growth of artificial intelligence (AI) applications has sharply increased the energy consumption and carbon emissions of data centers. Conventional energy-saving strategies often regulate power and cooling subsystems separately, resulting in inefficiencies and overlooking lifecycle carbon impacts. To address this issue, this article proposes PoCo, a deep reinforcement learning-based framework for the joint optimization of power generation and cooling control in combined cooling, heating, and…

The gain

In experiments replaying real Alibaba cluster traces, a reinforcement-learning controller for combined cooling, heating and power data centers cut energy costs by 25.8% and carbon emissions by 15.3% versus baseline algorithms.

In brief

The work responds to rising data-center energy use and emissions driven by the rapid growth of AI applications. Conventional strategies tend to regulate power and cooling subsystems separately, which the authors say causes inefficiencies and ignores lifecycle carbon impacts.

Main points

  1. PoCo reduced energy costs by 25.8% compared with baseline algorithms in experiments on Alibaba cluster traces.
  2. Carbon emissions fell by 15.3% and power usage effectiveness improved by 4.9% against the same baselines.
  3. The framework jointly controls power generation and cooling in CCHP-integrated data centers instead of treating them separately.
  4. A parameterized action space handles hybrid discrete-continuous decisions that arise in joint optimization.
  5. A long- and short-term time-series network with attention enables predictive cooling control and mitigates thermal inertia effects.
  6. A parameterized deep Q-network regulates power generation and cooling together, resolving thermoelectric coupling conflicts.

The full story

PoCo is built on three components. A parameterized action space handles the mix of discrete and continuous choices in joint power and cooling decisions, and a long- and short-term time-series network with attention predicts cooling needs to counter thermal inertia.

A parameterized deep Q-network then regulates power generation and cooling provision together. According to the authors, this addresses the high-dimensional, dynamic nature of CCHP systems and the thermoelectric coupling conflicts between heat and electricity production.

The evaluation used real-world Alibaba cluster traces and compared PoCo with baseline algorithms. Results were a 25.8% cut in energy costs, a 15.3% cut in carbon emissions and a 4.9% gain in power usage effectiveness.

Sources

1 on file
  1. IEEE Transactions on Industrial Informaticsdoi:10.1109/tii.2026.3692684
    Peer-reviewed2026-05-27

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

Data-center cooling and power AI cut energy costs 25.8% in tests on Alibaba traceshttps://truvace.com/item/poco-joint-optimization-of-power-and-cooling-for-low-carbon-cchp-integrated-data

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

    The gain In experiments replaying real Alibaba cluster traces, a reinforcement-learning controller for combined cooling, heating and power data centers cut energy costs by 25.8% and carbon emissions by 15.3% versus baseline algorithms.

    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 IEEE Transactions on Industrial Informatics.

    Answer in the thread

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

    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