Learning to Cool Off With DeepMind

Last week, I attended the WIRED Smarter event in London. My main reason for going was the cluster of sessions from key companies in the sustainable energy business. With this week’s news about climate change, clean energy is clearly an area of increasing importance in the technology space, as well as a topic of abiding interest to me. But before I offer a summary of one of the day’s presentations, I would like to congratulate the WIRED events team on curating a group of speakers on energy technologies that interleaved so well; there was clearly much consideration given to the challenge of presenting on a broad range of topics in a coherent way. As a side note, it was also impressive to see a significant number of women speakers chosen to represent their technologies.

I was particularly looking forward to a presentation by Sims Witherspoon, of Google’s DeepMind team (I’m hoping that the recording will soon be available online). Witherspoon covered the use of Deep Learning to reduce data center energy usage. As Nature News reported last month, data center energy usage this year will exceed the national energy consumption of some countries. Currently, data centers account for approximately 1% of global electricity demand, but usage is predicted to rise rapidly within the coming years — particularly if computationally intensive cryptocurrency mining continues to grow. This usage will make a significant contribution to global carbon emissions, since only approximately 20% of the electricity used by data centers comes from renewable sources, according to Greenpeace.

The main cause of energy consumption in a data center is cooling, which is typically performed by pumps, chillers, and cooling towers. Traditionally, it has been difficult to optimize the cooling process manually because of the complexity of the interactions between the combinations of necessary equipment. The rules and heuristics needed for every scenario have been difficult to define, particularly when interactions with the surrounding environment (such as the weather) are also considered. The result was that human operators were unable to calculate changes to settings that could respond sufficiently quickly to variations within the data centre environment in order to optimise electricity efficiency.

To investigate whether AI could do better, Google turned to DeepMind, and in 2016, the team blogged about a Deep Learning model, trained with sensor data from Google data centres, that was able predict the impact of environmental factors on performance and energy consumption. At the time, the model made recommendations to human operators to suggest optimisation settings for improving cooling efficiency and thus reducing power consumption. In one particular Google data centre, the model effected a 40% drop in energy usage for cooling.

In her presentation, Sims echoed a recent DeepMind blog post, which explained that the Data Center teams wondered if the AI could shortcut the human operators and work directly to control the cooling systems. With significant levels of safety measures to overcome, new AI was developed and rolled out some months ago. Since deployment, Google reports that the AI is delivering consistent energy savings of around 30%, which are expected to improve as more data is gathered (the reason the energy reduction is lower when automated, compared to the savings made for manual intervention, is that Google has constrained the system somewhat to prioritize safety and reliability over power savings).

Here’s a graphic from the DeepMind blog post:

The graph above plots AI performance over time relative to the historical baseline before AI control. Performance is measured by a common industry metric for cooling energy efficiency, kW/ton (or energy input per ton of cooling achieved). Over nine months, our AI control system performance increases from a 12 percent improvement (the initial launch of autonomous control) to around a 30 percent improvement.

Every little helps, and it is exciting to think that this technology could be shared so it can be used by other companies, and beyond in other industries, to reduce CO2 emissions and tackle climate change. At present, it is described in a Google patent, but we can hope that they will waive their rights to compensation as the fight against climate change intensifies.