Probabilistic forecasting of wind power ramp events using autoregressive logit models

A challenge for the efficient operation of power systems and wind farms is the occurrence of wind power ramps, which are sudden large changes in the power output from a wind farm. This paper considers the probabilistic forecasting of a ramp event, defined as exceedance beyond a specified threshold. We directly model the exceedance probability using autoregressive logit models fitted to the change in wind power. These models can be estimated by maximising a Bernoulli likelihood. We introduce a model that simultaneously estimates the ramp event probabilities for different thresholds using a multinomial logit structure and categorical distribution. To model jointly the probability of ramp events at more than one wind farm, we develop a multinomial logit formulation, with parameters estimated using a bivariate Bernoulli distribution. We use a similar approach in a model for jointly predicting one and two steps-ahead. We evaluate post-sample probability forecast accuracy using hourly wind power data from four wind farms.