Google Deepmind AI Project to Improve Renewable Energy Reliability Quietly Disappeared

Guest essay by Eric Worrall

A Google project to use artificial intelligence to improve the predictability of wind energy appears to have failed. Engineers who worked on the project have reportedly removed mention of their work from their linked in profiles.

Climate change falls down the agenda at A.I. lab DeepMind after energy heads leave



  • The DeepMind Energy unit, which has attracted its fair share of attention over the years, has vanished and none of the company’s staff mention it on their LinkedIn profiles. 
  • At one point, DeepMind wanted to use its AI technology to optimize National Grid, which owns and operates the infrastructure that provides electricity to homes and businesses around Britain.
  • The organizations spent a considerable amount of time working together, but there were many hurdles to overcome if anything was ever going to be implemented.

LONDON — Artificial intelligence (AI) lab DeepMind has shifted its focus from climate change to other areas of science and is pursuing its original mission of creating artificial general intelligence (AGI), which is widely seen as the holy grail of the nascent technology, according to several people familiar with the matter.  

One of the DeepMind’s early and perhaps most successful projects involved slashing Google’s enormous electricity bill, instantly reducing the company’s carbon footprint. The search giant, technically a sister company to DeepMind as they’re both run by Alphabet, announced in July 2016 that it had been able to reduce the energy consumption of its data center cooling units — used to stop Google’s servers from overheating — by as much as 40% with the help of a DeepMind AI system.

While the company’s achievements are significant, it has yet to publicly confirm where and how its energy-saving AI has been applied beyond Google’s data centers and wind farms.

National Grid nightmare?

At one point, DeepMind wanted to use its AI technology to optimize National Grid, which owns and operates the infrastructure that provides electricity to homes and businesses around Britain.

“We’re early stages talking to National Grid and other big providers about how we could look at the sorts of problems they have,” DeepMind Chief Executive Demis Hassabis said in an interview with The Financial Times in March 2017. “It would be amazing if you could save 10% of the country’s energy usage without any new infrastructure, just from optimisation. That’s pretty exciting.”

Last March, it emerged that talks had broken down between DeepMind and National Grid. The organizations spent a considerable amount of time working together, sometimes at a National Grid facility near Reading, in the English county of Berkshire, but there were many hurdles to overcome if anything was ever going to be implemented.

Gary Marcus, CEO of robotics company Robust AI, and co-author of “Rebooting AI,” which critically analyzes the industry and suggests how to advance it further, told CNBC that the technology may not have worked well enough for National Grid to justify the cost.

Their primary technique, deep reinforcement learning, works best in well-controlled environments, such as board games, and may struggle with the complexity and unpredictability of the real world,” Marcus told CNBC.

Sheikh added: “The technology might not have worked because it’s not really that mature.”

Read more:

This is not the first time a high profile Google renewable energy project failed.

Back in 2016 Google Engineers announced the failure of an engineering project to find a feasible path to 100% renewable energy. The 2016 project Google engineers concluded renewable energy “simply won’t work”.

5 13 votes
Article Rating
Newest Most Voted
Inline Feedbacks
View all comments
December 17, 2020 2:04 pm

You need artificial intelligence to tie the output of your coolers to how hard the CPUs are being driven???????

Reply to  MarkW
December 17, 2020 2:15 pm

I remember back in the early days of my career. Back when I didn’t know that certain things just weren’t possible. We had to load sort patterns up to a document sorter. The sort patterns were written in ascii using a proprietary language.
Since our link to the sorter was using 1200 baud, and the sort patterns were often 10’s of thousands of bytes, my manager asked us to find a way to strip out unnecessary white space from the patterns.
Since I didn’t know that some things were just too hard to do in ‘C’, I proceeded to write a filter in ‘C’. In the meantime our very expensive contractor set out to write a filter using awk and lex. I’ve had to maintain awk and lex scripts, so if you don’t know what they are, count yourself lucky.
Any way. I had my program up and running in about 3 hours. It took our contractor over a week to get his scripts working. When we ran them side by side, they both reduced the size of the sort patterns by the same amount, and mine ran about 10 times faster. Which was important when your desktop is a 12MHz 80286. My program was also a lot smaller, which is also important when your disk drive is 120Mbyte.

John in Oz
Reply to  MarkW
December 17, 2020 2:39 pm

I love it when you talk dirty – “awk’, “80286”, “disk drive is 120Mbyte”.

Next you’ll have us salivating with ‘acoustic modem’ and ‘core store memory’.

I go back to maintaining mechanical fire-control predictors in the Navy.

Julian Flood
Reply to  John in Oz
December 18, 2020 3:49 am

The Vulcan used a phonic wheel velodyne and a square rooting pinwheel. Forward throw was done by measuring the resistance of a steel tape.


It doesn't add up...
Reply to  John in Oz
December 20, 2020 10:16 am

Somewhere I think we still have a core memory circuit board acquired on a visit to an IBM facility. It’s 64 bits – an 8×8 array of threaded ring magnets, about 1 inch square, with wiring across and down, and looping through on the diagonals.

Reply to  MarkW
December 17, 2020 6:12 pm

I was amazed at how fast a very early version of MS Word runs on a version of Windows 95 I was running as a VM on Windows 10.

Reply to  MarkW
December 18, 2020 5:51 am

AWK, the consummate 3-line computer language

David Thompson
Reply to  MarkW
December 18, 2020 9:39 am

Sed would have done the job with one line buried in a shell script. Pulled this trick many times.

Reply to  David Thompson
December 18, 2020 10:14 am

Problem is that we had messages for the screens and ink jets built into sort patterns, whatever you used had to be smart enough to detect and skip over these strings.

Reply to  MarkW
December 17, 2020 4:01 pm

Its not “AI”.

Its missing the “I” part !!

Reply to  MarkW
December 17, 2020 8:35 pm

They believe the limitations “renewable” energy is in the process. So-called “AI” technology is selected as an optimizing or fitting mechanism to build control systems for what they presume is a nonlinear, perhaps chaotic response that is poorly matched or mismatched by conventional designs. They were wrong. The limitation is in the nature of their subject.

Richard Page
Reply to  n.n
December 18, 2020 9:15 am

None of the modern Ali’s are even close to being a true AI – they are still at a basic SI level (simulated intelligence – just mimicking basic responses) at the moment. SI won’t get you to an AI – it’ll take a real revolution in computing to do that and we’re nowhere even close to that sort of breakthrough yet.

Richard Page
Reply to  Richard Page
December 18, 2020 9:17 am

D’oh read AI for Ali – I do apologise for the blasted autocorrect feature!

Thomas Gasloli
Reply to  MarkW
December 18, 2020 5:05 am

But using AI is soooo much cooler (no pun intended) than just hiring a decent HVAC contractor. Although it cost more, it is more impressive on your resume, too.

Let’s face it Google isn’t what it used to be. Their search engine has declined in quality. I’ve been having a problem where, even though I have my location in my profile, they think I live in another state when I search businesses.

Microsoft has long been the GM of software; Google has become Chrysler without the Jeeps.

Reply to  MarkW
December 18, 2020 8:57 am

What’s really crazy is that the average employee can tell you this for free. But leadership doesn’t listen to one sensible voice. Instead, as Google has shown, they’d rather dump millions into what is essentially a excel spreadsheet to learn what the average employee would say for free. Bass ackward, you ask me.

Reply to  leowaj
December 18, 2020 10:20 am

I hear you. I had another job where I was director of software engineering for a product line. I told management that the CPU that was being used could not handle all the new features they wanted. They told me to do the best I could with the hardware I had.
A few months later they hired an “expert” at twice my salary and gave him my job, I was his sub-ordinate from then on. He spent a couple of weeks reviewing the system, then he told management the same thing I had been telling them for months, the frustrating thing is that they immediately ordered a transition from the existing CPU to the one that I had been recommending.
I left that job as soon as I finished training the new manager in his responsibilities.

Last edited 6 months ago by MarkW
December 17, 2020 2:09 pm

Like a lot of other Google ‘projects’ including AV research that drift away into nothing.

December 17, 2020 2:29 pm

Computer says,”NO”: “renewable energy “simply won’t work”.

Gregory Woods
Reply to  nicholas tesdorf
December 17, 2020 2:36 pm

I think that it depends upon what your definition of ‘renewable’ is.

Reply to  Gregory Woods
December 17, 2020 4:38 pm

‘renewable’ is… Replacing
Batteries every 8yrs
offshore Wind every 12yrs
on land Wind every 20yrs
Solar every 20yrs
burning trees quicker than you can grow them.
that’s what ‘renewable’ is !

Doug Huffman
Reply to  Gregory Woods
December 18, 2020 6:05 am

Yep. ALL energy renews on some time scale, the arrival of a photon, rotting of peat, or another Big Bang.

LOL. AWK – ward!

December 17, 2020 2:33 pm

There is a lot more money in ad deals with monopoly power or oligopoly partnering with Facebook.

Pat from Kerbob
December 17, 2020 2:36 pm

I would like details of how they supposedly decreased power requirements for cooling. Server farms and their cooling needs are clearly the biggest new sinks for electricity, cannot be intermittent, id like to see proof of savings

glen ferrier
Reply to  Eric Worrall
December 17, 2020 6:43 pm

Eric: They accomplished their savings by idling their servers during periods of high utilization. You may have noticed this if you tried to google up anything about Hunter and Joe.



Reply to  Pat from Kerbob
December 17, 2020 4:07 pm

Unlike household AC units, big chillers used for computers can be ramped up and down. I managed a computer room back in the day, if you set the displays for percent power, you could watch the AC units increase their output as more computers and big sorters were powered up in the morning, and watch them ramp back down as everything was shut off at night. In our lab, the logic was pretty simple. If the temperature was above the set point, the power was increased, if it was below the set point it was decreased, and if it was at the set point, do nothing. If the temperature kept dropping and the AC units were already at zero, turn on the heater,
We did have one problem that I was never able to solve and it came about because we had two units and each unit had separate logic and even though both units read 72 degrees, one probably was closer to 71.5 and the other to 72.5.
At night, when there wasn’t much heat being produced, we would get into a mode where one unit started heating, while the other unit was still trying to cool.
In this case, an intelligent controller that was able to coordinate between multiple units would have saved energy. But there was no need for AI to write the control software for something that simple.

Reply to  Eric Worrall
December 18, 2020 10:25 am

That feature was only accessible to the guys who maintained the units, and they wouldn’t change anything unless their management instructed them to.

Last edited 6 months ago by MarkW
D. Mix-O'Lydian
Reply to  Pat from Kerbob
December 18, 2020 1:04 am

With spinning disk RAID storage, optimizing data layout on the platters minimizes head seeks, lowers disk temperature and vibration, reducing cooling required and extending disk life.

Reply to  D. Mix-O'Lydian
December 18, 2020 10:27 am

I don’t know about other OS’s, but most flavors of UNIX do that in the background.

John Bell
December 17, 2020 2:38 pm

Reminds me of the hydraulic hybrid UPS truck i worked on for the EPA in Michigan. Tens of millions spent to develop something that was of no practical value, they were never fielded because the price of oil stayed low.

old engineer
Reply to  John Bell
December 18, 2020 4:07 pm


I remember that project! I worked for an organization that did contract work for the EPA at Ann Arbor. I was very interested because It looked like hydraulics had certain advantages over batteries for storing energy in a vehicle.

I think I retired before the EPA stopped work on hydraulics.

December 17, 2020 2:40 pm

It seems pretty stupid to think that AI could “solve” an obvious deficiency of an unreliable power source.

Joel O'Bryan
December 17, 2020 2:51 pm

“it had been able to reduce the energy consumption of its data center cooling units — used to stop Google’s servers from overheating — by as much as 40% with the help of a DeepMind AI system.”

DeepMind probably told them to change out the fluorescent lights for LED lighting and set thermostat wee bit higher.

Last edited 6 months ago by joelobryan
Reply to  Joel O'Bryan
December 17, 2020 5:23 pm

How much of the savings was actually from upgraded servers and memory? Without knowing everything they did, like changing out lighting, or just putting auto shut off on lighting when unoccupied, 40% means nothing.

Robert of Texas
December 17, 2020 3:02 pm

People seem to think you can “mine” wind like you can gold… You can’t use the wind if it’s too slow, too fast, or too variable. At least solar has the advantage of you KNOW it is useless 12 hours a day or more.Without a breakthrough in batteries or energy storage, Wind is pointless as an energy source.

You also can’t make any good use of wind if there is too much power already available. Oh wait, they are going to make hydrogen when that happens… So then you have a fuel that has too low an energy density to be of much use costing too much to compete.

If they want a gaseous fuel, why not just capture the methane ALREADY being flared and wasted and use that? Oh yeah, costs too much to compete.

December 17, 2020 3:09 pm

Was it like some of the BOM forecasts (Australia)? For every prediction point there’s a 50% chance of the real value being higher and a 50% chance of it being lower than the previous ie. not useful.

Last edited 6 months ago by tygrus
December 17, 2020 5:41 pm

Computer says …
‘Britain is getting cooler. Get more reliable power generation ready to cover outages!’

Kevin A
December 17, 2020 7:52 pm

I used human “I” to read the manual of a programmable furnace thermostat trying to figure out how to set the ‘ON’ time to find it’s called ‘Swing’ or ‘Comfort’, the number of degrees a thermostat allows between on/off cycles. The default was set at 0.5F which gave a 1° temperature swing, my indoor reading after I went to sleep looked like a sawtooth and over 6 hours had the furnace running 3 hours. Changing the ‘Swing’ to the maxim for the 3M-22 thermostat, 2°, gave a 4° cycle (62° to 58°) with only 30 minutes of furnace run time. The latest model programmable thermostat allow a 8° ‘Swing’ setting, I ordered one.
A 4X saving in energy without using AI

Mark Pawelek
December 17, 2020 9:18 pm

<blockquote><i>“It would be amazing if you could save 10% of the country’s energy usage without any new infrastructure, just from optimisation. That’s pretty exciting.”</i></blockquote>

Google will only be applying savings to electricity which is about 30% of Britain’s energy use. To save 10% of our energy use they’s needs to save a third of electricity. But the smart, renewable economy is supposed to use more electricity!

I think they need to acquire basic intelligence before they move on to AI.

Writing Observer
Reply to  Mark Pawelek
December 17, 2020 10:05 pm

Looks like you can’t blockquote without using the widgets. Disappointing, as I tend to just type the HTML myself.

Writing Observer
Reply to  Writing Observer
December 17, 2020 10:06 pm

Okay, that just looks stupid – an inverted quote over a bold text for the “blockquote” widget. Ah well, Anthony needs something to keep him busy over the holidays…

December 17, 2020 11:02 pm

Anthony the new format is poor. Mid grey text is difficult. It’s almost as bad as reverse text. Suggest you read Claude Hopkins ‘Scientific Advertising’. Old advertising guys have a list of things that instantly switch your wife’s brain off. If you want to test, test, test just do a focus group of disinterested folks off the street and you’ll soon find the appearance that’s good. When there’s evidence of a scientific approach to your new look, I’ll start reading again. In the meantime I’m very worried for you and your team . . .

Reply to  Warren
December 18, 2020 8:21 am

Mid grey text is difficult.

Agree — very hard on my eyes. Hopefully these things will be improved.

Reply to  Warren
December 18, 2020 10:28 am

Looks like the default text color has been darkened a bit.

Rainer Bensch
December 18, 2020 3:37 am

“It would be amazing if you could save 10% of the country’s energy usage without any new infrastructure, just from optimisation. That’s pretty exciting.”

Last March, it emerged that talks had broken down between DeepMind and National Grid.

Someone at National Grid must have realized that DeepMind was about to reduce their sales by 10%.

Russ Wood
December 18, 2020 6:58 am

Well judging by the problems that Google Home users had when the Google network went down – are we adding yet ANOTHER point of failure to an already unstable grid?

December 18, 2020 9:38 am

Google also funded Makani Power, who were promoting windmills suspended from giant kites. They actually bought Makani. It took them many years to figure out that this collossaly stupid idea wouldn’t work, something that I suspect any reader of WUWT could have told them in minutes. They spun off Makani into a wholly owned subsidiary, Makani’s assets were sold off in June 2020.

December 18, 2020 11:13 am

aka Deep Fake

December 18, 2020 3:34 pm

Isn’t this the second or even third failure of Google to acheive their Green Utopia?

December 24, 2020 1:26 pm

Heres an idea for optimizing the National Grid – turn all the friggin windmills off!!

%d bloggers like this: