When there exists a cause-effect relationship de-scribed by an input-output structural equation, for instance the power curve of a wind turbine, the effect (generated power) is obviously predicted by feeding the cause (wind speed) into the structural model (the power curve). But what is going to happen if the wind speed is not directly observed and is surrogated by a guess, as happens with meteorological forecasts? Such a kind of errors-in-variables framework is well understood in the linear Gaussian case: the straightforward application of the structural equation does not give optimal predictions, a phenomenon known as regression dilution. In the present work, the practical significance of regression dilution effects in the nonlinear regression of wind power from forecasts of wind speed is assessed through data taken from the Global Energy Forecasting Competition 2012 (GEFCOM2012). It is found that the effect is relevant and some lessons are learned, in particular about the benefit of using regression models tailored to the specific prediction horizon.
