Neural networks to improve the performance of high-power wind turbines

Wind energy has become an important source of electricity generation, capable of helping to achieve a cleaner and more sustainable energy model. However, it is necessary to improve the performance of wind turbines to be able to compete with conventional energy resources.

In order to achieve this improvement, flow control devices are implemented in the aerodynamic profiles, to improve the aerodynamic efficiency of the wind turbine rotors: “Thus, with the same wind turbine more megawatts can be produced, the cost per megawatt hour reduced, and that transferred, for example, to a wind turbine located in the sea (which are huge), makes the implementation cost negligible, but on the other hand the aerodynamic improvement can be around 8 or 10 percent”, explains Unai Fernández Gámiz, professor at the Department of Nuclear Engineering and Fluid Mechanics at the University of the Basque Country (UPV/EHU).

Computational fluid dynamics (CFD) simulations are the most popular method used to analyze this type of device: “This is software that simulates the movement of fluids, which requires a large computational capacity, that is to say, very powerful computers and a lot of computing time”, explains Fernández Gámiz. But in recent years, with the growth of artificial intelligence, prediction of flow characteristics using neural networks is becoming more and more popular; In this sense, the UPV/EHU student Koldo Portal Porras has implemented a convolutional neural network () that determines a series of parameters used to control the flow of wind turbines.

The results show that the proposed convolutional neural network for field prediction is capable of accurately predicting the main characteristics of the flow around the flow control device, showing very slight errors. Regarding the aerodynamic coefficients, the proposed convolutional neural network is also able to predict them reliably, being able to correctly predict both the trend and the values. “Compared to computational fluid dynamics simulations, the use of convolutional neural networks reduces computational time by four orders of magnitude,” says researcher Portal Porras. “Quick, almost immediate results have been achieved, with an error of between 5 or 6 percent, in some cases. A fairly acceptable error for an industry that fundamentally seeks quick results”, adds Fernández Gámiz.

Unai Fernández with two researchers from her team. (Photo: Nuria González, UPV/EHU)

“First we have launched the simulations in computational fluid dynamics with two different flow control devices (rotary microtabs and Gurney flaps), and from there we obtain the output data, which we take as real and which we use to train the convolutional neural network — Portal Porras explains. What we do is enter the geometry as input and the results obtained with computational fluid dynamics as output. In this way the network is trained, and then if we insert another geometry, with the results it had before, it is capable of predicting the new velocity and pressure fields”.

In the opinion of Fernández Gámiz, Portal Porras has achieved “a fast, flexible and cheap tool. The industry today requires quick solutions. To apply this type of networks, it is not really necessary to resort to large computers, or computer clusters, etc. And, in addition, we have achieved a flexible tool, because it is applicable to any aerodynamic profile, to all types of device systems and even to other types of geometries”. Portal Porras states that the network is suitable for all types of wind turbines, “but the training data that we have entered was for a specific aerodynamic profile. Therefore, if you enter another aerodynamic profile, you would have to do the entire training process, that is, enter the input and output data of the other wind turbine”. Both agree on the importance of artificial intelligence: “It is a fundamental step if we want our industrial environment to be competitive. If we do not enter into artificial intelligence issues, we are not going to advance in competitiveness in international markets”.

Koldo Portal Porras, Unai Fernández Gámiz, Ekaitz Zulueta, Alejandro Ballesteros Coll and Asier Zulueta present the technical details of their convolutional neural network for wind turbines in the academic journal Scientific Reports, under the title ” based fow control device modeling on aerodynamic airfoils”. (Source: UPV/EHU)