Short-term electric energy production forecasting at wind power plants in pareto-optimality context

The paper discusses the possibilities of multi-criteria optimisation of a multi-layer perceptron (MLP) model applied to the short-term (intra- and next-day) wind power forecasting problem. The paper comprises two main parts: a theoretical background and study case using data (wind power production and historical weather forecast) obtained from two wind farms (at different power capacity levels). The problem stated in this paper is to formulate a method allowing for the estimation of a set of prediction models meeting the selected three model learning criteria: nBIAS, nMAE and nRMSE. The two-step NISE method has been used in order to estimate the non-dominated forecast evaluation set. The available data have been divided into three subsets for model learning, testing and validation. Than, a set of prediction model variants has been investigated considering different types of data subsets used for stopping the MLP learning process as well as calculating the forecast error. Additionally, different structures of MLP and learning algorithms have been analysed. Finally the paper is ended with a summary and conclusions.