The estimation of a windfarm’s powercurve, which links the wind speed to the power that is produced by the whole windfarm, is a challenging task because this relationship is nonlinear and bounded, in addition to being non-stationary due for example to changes in the site environment and seasonality. Even for a single wind turbine the measured power at different wind speeds is generally different than the rated power, since the operating conditions on site are generally different than the conditions under which the turbine was calibrated (the wind speed on site is not uniform horizontally across the face of the turbine; the vertical wind profile and the air density are different than during the calibration; the wind data available on site are not always measured at the height of the turbine’s hub).
The paper presents a data-driven approach for building an equivalent steady state model of a windfarm under normal operating conditions and shows its utilization for the creation of quality control charts at the aim of detecting anomalous functioning conditions of the windfarm. We use and compare three different machinelearning models – viz. a self-supervised neural network called GMR (Generalized Mapping Regressor), a feed-forward Multi Layer Perceptron (MLP) and a General Regression Neural Network (GRNN) – to estimate the relationship between the wind speed and the generated power in a windfarm. GMR is a novel incremental self-supervised neural network which can approximate every multidimensional function or relation presenting any kind of discontinuity; MLPs are the most widely used state-of-the-art neural network models and GRNNs belong to the family of kernel neural networks.
The methodology allows the creation of a non-parametric model of the powercurve that can be used as a reference profile for on-line monitoring of the power generation process, as well as for power forecasts. The results obtained show that the non-parametric approach provides fair performances, provided that a suitable pre-processing of the input data is accomplished.