Adaptive Generalized Logit-Normal Distributions for Wind Power Short-Term Forecasting

TL;DR

This paper introduces adaptive generalized logit-normal distributions for improved very short-term wind power forecasting, addressing nonstationarity and non-linearity to enhance point and probabilistic predictions.

Contribution

It proposes a novel flexible distribution model with maximum likelihood estimation, including online adaptation for non-stationary wind power data.

Findings

Effective in modeling non-linear, double-bounded wind power data

Improves 10-minute-ahead forecast accuracy

Handles non-stationarity via online parameter updates

Abstract

There is increasing interest in very short-term and higher-resolution wind power forecasting (from minutes to hours ahead), especially offshore. Statistical methods are of utmost relevance, since weather forecasts cannot be informative for those lead times. Those approaches ought to account for the fact that wind power generation as a stochastic process is nonstationary, double-bounded (by zero and the nominal power of the turbine) and non-linear. Accommodating those aspects may lead to improving both point and probabilistic forecasts. We propose here to focus on generalized logit-normal distributions, which are naturally suitable and flexible for double-bounded and non-linear processes. Relevant parameters are estimated via maximum likelihood inference. Both batch and online versions of the estimation approach are described -- the online version permitting to additionally handle non-stationarity through the variation of distribution parameters. The approach is applied and analysed on the test case of the Anholt offshore wind farm in Denmark, with emphasis placed on 10-min-ahead point and probabilistic forecasts.