Wind farm power forecasting: New algorithms with simplified mathematical structure

Abstract

Nowadays wind power forecasting research is becoming more and more important in economic operation and safety management of power systems. Moreover, with a huge increase of data availability, data-driven models, especially for short-term forecast, represent a good compromise between precision and computational loads. This paper deals with a data-driven algorithm consisting of a combination of three main steps: gis-based data collection and pre-treatment, classification wind turbines step (Twostep cluster analysis) and short-term wind power forecast with artificial neural networks (ANN). The wind turbines were geo-referenced, the initial variables (single turbines power output, wind direction and speed) were pre-processed, a recursive optimized cluster analysis (CA) was performed to obtain a simplified mathematical structure for an artificial neural network. A windfarm (48 MW) placed in the eastern part of Sicily (Italy), was tested as case study. It consists of two different sites each one with 28 wind turbines at different rotor heights and different orography. The time series dataset consists of almost four years data (sampling time of 10 minutes). The consequent simplified mathematical structure leaded to perform good results in in a short-term wind power output forecast for both sites of the case study.