Posters

Abstract

Wake steering simulation and optimisation tools rely on steady state models of the farm in discrete wind conditions to allow rapid iteration of setpoints and gradient based methods. A frequency-based approach is often used for AEP uplift estimation: the power gain in each discrete wind condition is weighted by the expected annual hours in that condition, summed over all expected conditions, and divided by the gross energy to provide a total uplift estimate. This method has a number of implicit assumptions, for example: turbines will yaw instantaneously to new setpoints as wind conditions change and maintain these setpoints exactly, the time taken for wake propagation through the farm is ignored, wake meandering is ignored, and no variation in wind conditions and wake directions across the site is assumed, beyond what can be represented in a steady state model. Typically, all turbines are expected to experience the same global, un-waked, site wind condition in steady state models. Field tests and dynamic simulations of wind farm control have shown that energy gains are often significantly lower than predicted by steady state models, and the assumptions listed provide an obvious justification for these real-world losses. DNV have developed an innovative methodology to characterise dynamic losses using limited simulations of the farm in an in-house dynamic wind farm simulator, LongSim. Ideally the farm might be simulated over a time-series representing a typical year of production, but this is computationally unfeasible for applications that require results quickly. •The methodology sorts discrete wind conditions in order of their contribution to the total AEP gain predicted with steady state methods. •A subset of discrete conditions is selected to represent the majority of the gain (this aspect is complex and the subject of investigation in the presentation). •A time series of representative wind conditions (e.g. from a mast or large forecasting model) is searched for periods with mean conditions which match the subset. •Wind fields representing the whole site area are generated for each selected time history, and the farm is simulated with and without WFC. Turbine supervisory logic which represents expected yaw/induction tracking in both control scenarios is modelled including wind condition estimation methods, averaging times and dead-bands. •The results are post-processed to find the energy gain from wind farm control over each simulation, and compared to the expected gain in steady state conditions to calculate a dynamic loss factor associated with the discrete condition. •The loss factor is weighted by its contribution to the total gain and iteratively added to a total loss factor estimate, which converges as more dynamic simulations are added to the subset. •A logarithmic model is fitted to the loss factor estimate as a function of the proportion of the gain modelled, and solved to find an estimate of the total loss factor for 100% of the gain. The presentation will include the methodology, validation against site trials of wind farm control, and characterise the key aspects of uncertainty which need to be resolved to build confidence in this calculation approach.