Google's DeepMind Can Predict Wind Patterns a Day In Advance (engadget.com) 57
technology_dude writes: Google's DeepMind can predict wind patterns one day in advance. "Beginning last year, [Google and DeepMind] fed weather forecasts and existing turbine data into DeepMind's machine learning platform, which churned out wind power predictions 36 hours ahead of actual power generation," Engadget reports. "Google could then make supply commitments to power grids a full day before delivery." According to the report, this makes the energy generated by its wind turbines more valuable (by roughly 20%). Is this a blow to Big Blue who purchased The Weather Channel's Weather.com to showcase Watson, or is it news just because it's Google?
But can it predict.. (Score:5, Funny)
Re: (Score:2)
So what? I can do that. (Score:4, Funny)
My predictions rarely turn out to be correct, but that's besides the point.
So can any other prediction company (Score:2)
Actually we don't call it prediction, we call it prognosis.
And there are plenty of companies like https://www.windfinder.com/ [windfinder.com] who do this since decades.
Re: (Score:2)
Then the energy market has to find and price in other types of energy at great cost.
eg at night if the wind stops and the sun is not out for solar.
Power has to stay on and has to be generated.
A correct prediction allows for better use of energy production that can take time to set up and get ready due to complex solar and wind energy production problems.
Re: (Score:2)
A correct prediction allows for better use of energy production that can take time to set up and get ready due to complex solar and wind energy production problems.
Yes, and we work with correct prognoses since decades. My old customer EnBW, buys prognosis data from about 10 providers. The average, sometimes a weighted one, from the 3 providers that were the most accurate over the last 3 or 4 hours are used for the next hour. This is corrected every 15 minutes.
So your feared "other power plants react to slow
Re: (Score:2)
What you are talking about? Most weather pattern prediction services just use the NWS's GFS, including Windfinder (it even says so right in their FAQ).
The DeepMind approach is completely different. A 36 hour forecast isn't exactly amazing, but it's just the beginning.
Re: (Score:2)
You also need to take more account of surface topography.
that really does not matter much, because for one wind turbines are erected in open wind area's, and two : they are gigantic (up to 200m as of now) and clearing all buildings and trees.
Re: (Score:2)
Absolutely. And if they had done that I would have been dutifully impressed, but they didn't. They just calculated average kWh output for a block of wind turbines, which is just using the weather forecast and monitoring turbine output. I didn't realize this until I read the article more carefully (see my other comment). Small steps, I suppose....
Weather Prediction is BIG MONEY (Score:1)
Re: (Score:1)
it's more than just that. More wind means cheaper energy on the market. If you are a huge industrial electricity consumer (say, a central Walmart food distribution center) and you need to keep your freezers at least at -10F, you could go to -20F when energy is cheap, and just let it warm up to -10F while energy is more expensive. So, being able to predict wind and sun means predicting supply and demand on the energy market. If you do it well, you can save millions.
I've got a question (Score:4, Insightful)
Weather forecasting is a hard mathematical problem with thousands of variables which needs to be calculated to be precise.
The question is: does DeepMind AI/algorithm calculate the weather or it is just guessing it? If it's doing the latter then this guesswork is going to be pretty random and equally worthless, and I see no way it's gonna reach the precision of the known mathematical weather models. It might guess well in the short term (relatively few initial parameters), but in the long term I don't see it working well.
Re: (Score:2)
Re: (Score:2)
It's not guessing, but it's not strictly a model-based calculation either. Wind speed and direction, in particular, is a very complex system of equations (computational fluid dynamics) that doesn't scale reasonably at all if you try to use a physical model. So the field has been trending toward statistical models and stochastic simulations for a number of years. DeepMind can effectively do both, but on steroids, and it works well because of the enormous amount of data available.
"trained on widely available weather forecasts" (Score:3)
Best go direct to the source [www.blog.google] for your answers:
Using a neural network trained on widely available weather forecasts and historical turbine data, we configured the DeepMind system to predict wind power output 36 hours ahead of actual generation.
So no, it doesn't do weather forecasts itself, but it is based on those.
Re: (Score:2)
Weather forecasting is a hard mathematical problem with thousands of variables which needs to be calculated to be precise.
That was the old way of doing it. There are limitations to that method though, because there are limits on the accuracy of measurements and problems with the vast number of hidden variables, unknown unknowns that are constantly changing.
To get around that the old systems did a large number of predictions with random variations, and based on how like the variations were deemed to be and how many of the models resulted in high winds in area X, they would make a forecast.
AI appears able to do a better job at t
Re: (Score:2)
I would not do ANY math to make my predictions.
There are small weather stations set up all over the place reporting in. I'd program a neural net to track down the ones that most correlate with a 36 hour prediction, and then just track those.
Point is, I don't need a mathematically exhaustive model of the atmosphere to predict wind two days from now. I just need to look where the wind is coming from to see what's coming. When I was a kid, I learned:
Red sun at night, sailors delight.
Red sun in morning, sail
This works every time, until it doesn't (Score:2)
Re: (Score:2)
It'll work better over time. Each windmill will get more and more training to fine tune its specific location.
Weather forecasting? (Score:2)
Re: (Score:2)
When to turn on gas power, state back up when wind drops.
What new price for energy to set when wind drops, for how long and where.
Oooh, goodie! (Score:2)
This means that if a calm day is predicted for tomorrow, you would know to put off running the oven and the dryer for another day.
Re: (Score:2)
This means that if a calm day is predicted for tomorrow, you would know to put off running the oven and the dryer for another day.
What would in fact happen is that the wind farm, knowing it will have less power to deliver, will offer less for sale so the power grid will contract with baseload providers for the difference. What messes up electricity service is rapid and unexpected change.
Not wind prediction (Score:5, Informative)
Ah, ok, so I actually read the article (no I'm not new here). And it even says it right there in the summary, but somehow it didn't register. They are predicting wind power generation, not the wind itself!
Definitely not as impressive as my first reading. They are even using existing weather forecast data, so nothing particularly innovative here. I think there is a lot of potential to do something really cool in this area with DeepMind, though. There is a ton of historical and real-time data available.
Re: (Score:2)
Replying to myself, but just wanted to answer the question posed in the summary:
Is this a blow to Big Blue who purchased The Weather Channel's Weather.com to showcase Watson
No. If this is the extent of Google's interest in this area (hopefully not), then Watson has nothing to worry about.
, or is it news just because it's Google?
It's not news. It's a blog post on one of Google's websites.
Re: (Score:2)
You've clearly never built a deep network architecture and trained it. Success is highly dependent on both domain expertise for the problem you are trying to solve, as well as a strong math and statistics background. Sure, any Joe can use TensorFlow and make a small CNN to classify their family photos, but building something that can beat world champions at Go requires teams of experts.
Re: (Score:2)
Oh damn-it - I was hoping it could tell me when to avoid the second floor toilets :-(
the big question (Score:2)
This is Google we talking about. So there's an obvious big question, now that this service has proven useful: when will they discontinue it?
No one read the article (Score:3, Insightful)
We have been doing this for years and this is nothing new. This is news because it has Googles name in it and the AI buzz word.
Simply astonishing what people will write up as an article and it makes as "news"
Algorithms vs. models (Score:5, Interesting)
A little lesson in scientific literacy:
Algorithms, like Google's, that trawl through historical data & make predictions based on that work fine, as long as everything continues as before & nothing changes. It's a complex equivalent of continuing a straight line on a graph & calling it a prediction. It's completely useless & can even be destructive if you use it in combination with exerting influence on the system itself because biased or wrong predictions (which is always the case) create positive feedback loops that are self-reinforcing. A prime example of this abuse of data is so called predictive policing, which suffers from a negative form of the "Matthew Effect."
In contrast, a theoretical model makes predictions based on the properties, forces, & constraints on a system, be it complex (probabilities) or simple (rules). Even if conditions change, anomalies appear, etc., the model is more likely to remain reasonably predictive.
BTW, looking for patterns, regularities, etc. in data without testable hypotheses or research questions is a form of scientific malpractice called p-hacking.
Conclusions: AI is only predictive as long as nothing changes. Models or more reliable & flexible. We shouldn't stop thinking about how the world works & give it all over to mindless algorithms.
Technology Review has a nice summary... (Score:2)
There is a concise, succinct summary at Technology Review's “The Download” page (link below):
Re: Technology Review has a nice summary... (Score:1)