Machine learning to cut cooling energy in Google's data centers
In 2016, DeepMind and Google's data center team trained neural networks on historical sensor readings from a Google data center, then used them to recommend how to run pumps, chillers and cooling towers. In live tests, cooling energy fell by up to 40%.
What they learned
AI recommendations cut cooling energy by up to 40% in live tests
From public sources: DeepMind reported cooling energy savings of up to 40%, which works out to about a 15% cut in total energy overhead once other losses are counted. The authors noted that every data center is built differently, so a model tuned for one site may not work at another; they built a general framework for that reason.
The problem
Data center cooling systems are complex, and each building behaves differently, so fixed rules waste energy.
What the AI does
Neural networks learn from thousands of sensors (temperatures, power, pump speeds) to predict energy efficiency an hour ahead, then suggest settings that use less energy while keeping temperatures and pressures safe.