Very short-term probabilistic forecasting of wind power with generalised logit-Normal distributions

Very short-term probabilistic forecasts, which are essential for an optimal management of wind generation, ought to account for the nonlinear and double-bounded nature of that stochastic process. They take here the form of discrete-continuous mixtures of generalised logit-Normal distributions and probability masses at the bounds. Both autoregressive and conditional parametric autoregressive models are considered for the dynamics of their location and scale parameters. Estimation is performed in a recursive least-squares framework with exponential forgetting. The superiority of this proposal over classical assumptions about the shape of predictive densities e.g. Normal and Beta, is demonstrated based on 10-minute ahead point and probabilistic forecasting at the Horns Rev wind farm in Denmark.