Data assimilation for the prediction of wake trajectories within wind farms

Abstract

In this paper, we formulate a physics-based surrogate wake model in the framework of online wind farm control. A flow sensing module is coupled with a wake model in order to predict the behavior of the wake downstream of a wind turbine based on its loads, wind probe data and operating settings. Information about the incoming flow is recovered using flow sensing techniques and then fed to the wake model, which reconstructs the wake based on this limited set of information. Special focus is laid on limiting the number of input parameters while keeping the computational cost low in order to facilitate the tuning procedure. Once calibrated, comparison with high-fidelity numerical results retrieved from Large Eddy Simulation (LES) of a wind farm confirms the good potential of the approach for online wake prediction within farms. The two approaches are further compared in terms of their wake center and time-averaged speed deficit predictions demonstrating good agreement in the process.

A reinforcement‐learning approach for individual pitch control

“…Even though it would be interesting to train the NN in realistic turbulent flow fields, it is quite clear that the cost of the LES used in this work is prohibitive. However, multiple compromises can be envisioned for the training in a higher fidelity environment such as (1) generating synthetic turbulence data bases, typically using the Mann algorithm 50 or atmospheric boundary layer precursor simulations, 61 to train the NN in turbulent wind at affordable cost; (2) using dynamic wake models like the one proposed in Lejeune 62 to produce much more realistic meandering waked flow conditions, (3) using LES at coarser spatio‐temporal resolutions, which implies using disk approaches to model the rotor and thus handling IPC within an actuator disk framework, as proposed in Moens, 63 or (4) use the low‐fidelity environment to first train the NN and then refine the training in a LES framework. The best option should be chosen based on some cost‐benefit analysis, which leaves a variety of perspectives open for the present methodology.…”

“…With an improved load sensing and an enlarged action space, the missing element is a realistic training environment. Multiple possibilities towards that purpose were proposed before, such as using wake models, 62 synthetic, 50 or precursor‐generated 61 turbulence or disk‐based LES 63 …”

Extension and validation of an operational dynamic wake model to yawed configurations

“…Other popular DWM implementations include FAST-Farm [8] which was shown to capture accurately the loads of the turbine and the associated wake dynamics [9] providing full knowledge of the inflow. The work presented here extends the capabilities of an operational dynamic wake model [10,11] to yawed cases. The proposed framework bridges together flow sensing and wake modeling: the flow features are inferred at the wind turbine location directly from the wind turbine measurements and then advected across the wind farm thereby reconstructing the flow field.…”

“…A yaw module is deployed within an online wake dynamic estimation tool developed as part of previous works [10,11]. Information gathered and processed at the wind turbine location is propagated across the wind farm thereby reconstructing approximate snapshots of the flow field consistent with the observed data.…”

Application of an open-loop dynamic wake model with high-frequency SCADA data

“…A conventional wake model is used to calculate the velocity deficit at the discretized positions. The idea was extended with flow sensing turbine observers [19]. However, apart from that, the concept has not been widely tested.…”