WakeOpColl has endowed flow devices with artificial intelligence in order to make decisions about their own operating point and collaborate with others when in a group.
Two approaches can be followed. A first, physics-blind one is an uninformed strategy, learned through training. Through so-called reinforcement learning or supervised learning, our flow devices learn through trial and error how to either fly in the beneficial zone of wakes, or better handle the incoming turbulence. Such schemes are flexible and quite adaptive but can be less effective.
A second approach, a physics-aware one, endows our aircraft or turbine with some knowledge about the flow physics. The device has a mental image of the flow and will refine it through its flow measurements in a so-called data assimilation process. This mental image requires to have a simplified model of the flow and can then support a control algorithm, which will exploit this knowledge for additional performance.
First, for formation flight, we have used both approaches for an aircraft to exploit a leader’s wake. A physics-blind aircraft was first successfully trained to position itself in the upwash region. However, a control ambiguity arose: diametrically different wake encounters can indeed lead to sensed information. Furthermore, such an ambiguity constitutes an unacceptable risk when considering an operational deployment. WakeOpColl addressed this challenge with the physics-aware approach. In a first breakthrough, we demonstrated the feasibility of data assimilation-based wake sensing using the existing sensors aboard an aircraft to locate the wake of the leading aircraft to exploit it safely. This has led to a patent application, a collaboration with Airbus and contributions to the ongoing project SESAR JU GEESE (Airbus coord.) to make extended (i.e. with a large distance between leader and follower) formation flight an operational reality for trans-Atlantic flights by 2027. The fuel savings amount to about 10%; this gain is simply enormous when one is reminded that no new hardware development is necessary to reach it, just software and operational procedures!
A second aspect was larger formations where we proved their feasibility through large scale simulations. These simulations showed that (1) the vortices of a formation of several aircraft remain “usable” for extended formation flight if the participating aircraft and (2) control instabilities can occur, just as in a platoon of cars, but can be mitigated through an adequate choice of controllers.
On the topic of wind energy, WakeOpColl also used both AI approaches in a complementary fashion. At the level of a single turbine, we have used reinforcement learning to train the turbine at reducing its fatigue loads in a turbulent flow. The result is a novel control scheme that is completely physics-agnostic and therefore quite flexible; it is another stand-out result of the project. Still for a single turbine, we have used data assimilation to extract wind properties from the loads on the rotor, such as the wind profile, the turbulence intensity, etc. These quantities can then be leveraged toward a global picture of the wind flow through the wind farm.
WakeOpColl precisely did that by expanding this data assimilation technique to the scale of the wind farm. This entails exploiting the wind turbine-level data assimilation technique to produce the information necessary to feed a flow estimation in the space between the wind turbines. This flow estimation relies on seeding the deduced flow information from the wind turbines into the farm through two separate fields: the sensed free-flow information and the estimated wake signature. Together, these two pieces of information allow to reconstruct a flow field from information obtained from the turbine, therefore again without the need to add meteorological masts or LiDAR systems, etc. The resulting model and flow estimation modules were made public at the leading conference in wind energy in 2022, shared as the open source software, Onwards (github), and is now part of an internationally-developed and widely used Floridyn library. Most importantly, this software runs at speeds several orders of magnitude faster than real-time and can therefore be used to estimate the flow also accounting for uncertainties, or estimating the state of the wind farm in an alternate control scenario where it would for instance be curtailed to offer a reserve of power for the grid operator. Such capabilities will be crucial to increase the penetration of wind energy in our energy mix and reduce the cost of renewable energies further through grid services and their added value.
