Vestas supercharges its wind farm control models for sustainable energy with Azure HPC

Passing by a wind farm on the road, you may not realize the intricate concert of physics playing out moment to moment beside you. The spinning turbines collect energy from rushing gusts of air and leave turbulent streams behind them, which build honeycombs of wakes that collide with the other turbines and reduce their overall power yield.

IEEE Spectrum estimates that 10 percent of potential wind power is lost to wake effects—the result of turbulence from one turbine negatively affecting others behind it. Even a 1 percent improvement could translate to millions of dollars in revenue and a more sustainable energy future. As a global leader in renewable energy and the largest wind turbine manufacturer in the world, Vestas has long sought to optimize all aspects of wind power to boost the business case for its customers as well as sustainability.

Dance of the wind turbines

By introducing changes in the alignment of individual turbines, a technique called wake steering can reduce turbulence on other turbines and increase the overall output of a wind farm. Imagine a reactive, collaborative dance where turbines adjust, in real time, to changing wind conditions. As one yaws to redirect its wake and prevent it from fully streaming into a turbine behind it, others also adjust to reduce their impacts on the turbines downstream of them. 

Wake steering requires complex control strategies, and Vestas wanted to reduce the long development time by using reinforcement learning on Microsoft Azure high-performance computing (HPC) as an alternative to its traditional modeling. Developing an approach to more fully optimize wake steering required advanced AI technologies, intensive simulations, and high computation capacity to overcome the complexities of fluid dynamics and the scale of wind farm mechanics. Vestas, which owns one of the largest supercomputers in Europe, has been working on the problem it calls the Grand Challenge, but it took a change in tactics to unlock the results it really wanted. 

“We do have a supercomputer on-premises, a rather big one,” says Chief Specialist and Modeling & Analytics Module Design Owner Sven Jesper Knudsen. But Vestas recognized that cloud-based HPC might bring a better solution within reach. 

In fact, the journey from idea to proof of concept was a remarkably short one. “It’s been quite an eye opener for us how easy it is to start developing with the minds.ai platform on Azure,” says Knudsen. “I was completely blown away that one week into the project we almost had a minimum viable product.”

The company worked with Microsoft Partner Network member minds.ai and its reinforcement learning platform DeepSim to generate an intelligent wind farm controller by taking advantage of the power of Azure HPC resources. Vestas uses this platform to run complex simulations at scale to rapidly train models (controllers) capable of automatically reacting to wind conditions and orchestrate the turbine orientations to minimize power loss from wake-effect turbulence.

Deep impact

Using the reinforcement learning capabilities of the DeepSim platform on Azure, the system continually reinforces decision making using data based on the results of the actions. Vestas uses this method of training controllers to take appropriate actions based on inputs from the wind farm environment, like adjusting turbine yaw in response to wind direction, speed, and wake effect to mitigate wake effect and increase wind farm efficiency and yield.

The training involves millions of simulations, a data lake containing petabytes of data, and a lot of trial and error. The vast resources available from DeepSim running on Azure HPC deliver a faster convergence—which reduces computing cost. 

While the approach is still early in development, Vestas is now confident that it works and has the potential to make meaningful differences for the company’s customers. “The primary driver is more value, more revenue—that is, of course, an indirect consequence of more kilowatt hours,” says Vestas Senior Modeling Specialist Ewan Machefaux. 

Vestas expects that this optimization will create significant improvements for existing wind farms, and the company also wants to help customers build future wind farms with these formulas in mind so that they can deliver more energy from more closely positioned turbines, designed to anticipate and minimize wake effects. “Basically, the ability to design your product differently at lower cost but still realize the same yield,” Machefaux notes.

Measuring the impact of wakes on downstream turbines and how adjustments to their orientation might improve their effectiveness requires complicated models running thousands of simultaneous simulations. The HPC solution draws on Azure HBv3 virtual machines equipped with third-generation AMD EPYC™ processors and scaled up to 15,000 cores, with the capacity to go higher as needed.

“The latest AMD CPUs offered high compute power and memory bandwidth, which was crucial for simulations that involved large amounts of weather data. On top of that, the cost/performance ratio was amazing—it allowed us to both scale up and keep the cost reasonable,” says Jeroen Bédorf, PhD, Senior Systems Architect at minds.ai. “This project would have taken much longer without the AMD based Azure HBv3 nodes.”

DeepSim uses Azure HPC infrastructure along with Kubernetes to scale up and run the complex simulations. It then applies reinforcement learning to extract findings from these tests. The reinforcement learning engine employs a neural network to extract optimizations while testing millions of parameters. Partner minds.ai has employed DeepSim to improve device and system performance in other industries as well, including cruise control in hybrid electric vehicles.

Promising prototypes, beautiful simulations

Vestas views this early technology as both extremely promising and often beautiful as it illustrates banks of orchestrated turbines moving in concert to direct the tails of turbulence away from each turbine as they all adjust to changes in wind direction and velocity.

“We taught the controller to let the turbines dance together,” says Thomas Soule, Business Development Manager at minds.ai. “When you see it in simulation, it’s really gorgeous.”

And with Azure—unlike with an on-premises supercomputer—Vestas only pays for the cycles it uses during the modeling process. It can use the Azure pay-as-you-go plan to scale up computing and storage resources when it needs them and scale back down when it doesn’t. When the company takes the reinforcement learning solution to market, that flexibility will help it control costs and deliver breakthrough technology to new customers without having to make big investments in infrastructure.

“Reinforcement learning has been around for a while,” says Soule. “But it requires a lot of compute cycles and intensive simulations. And until that reached a certain point of affordability and power, it just wasn’t practical.”  

Once a model is completed, DeepSim results will instruct the wind farm’s turbine controller to adapt to shifting wind conditions using simplified mathematical computations to execute specific turbine adjustments without the need for ongoing intensive HPC cycles.

A sustainable, prosperous future

Even small efficiencies in wake optimization have the potential to unlock significantly more power and thus higher profits for wind farm operators. 

“The value delivered by the technologies built on Azure HPC unlocks wake mitigation but also other use cases that make wind power even more competitive,“ reports Knudsen. “With it, we can optimize design and profitability of sustainable energy solutions for our customers.”

While still in the early stages of development as an actual product for working wind farms, this proof of concept is a vital step toward realizing a vision for a more sustainable future. Improvements based on this initial prototype may take years to roll out to turbines worldwide, but Vestas now has a clear route to these efficiencies.

“We have a big role to play in the global energy transition. We are now on par with oil and gas and are the most profitable source of energy already,” says Line Storelvmo Holmberg, Senior Vice President, Applications, Controls, and Electrical, at Vestas. “Our ambition is to be the global leader in sustainable energy solutions.”

Find out more about Vestas on Twitter, Facebook, and LinkedIn.

Find out more about minds.ai on Twitter and LinkedIn.