Development of a Neural Network Predictor for Wind Power

Increasing public concern about environmental impact of emissions from fossil fuel generating stations, disposal of hazardous materials and threat of nuclear mishap is causing great undertainty in approval and installation of traditional large generating stations to meet future energy needs. Additionally federal regulation regarding the level of CO2 and SO2 emission is urging utility to considernon-polluting sources of power in their overall generation mix to meet continued increasing demand for electric power also as they retire old generation stations. PURPS is requiring utilities to provide access to their transmission and distribution systems for independent power producers. All of this is forcing utility industry to consider alternate sources of power especially renewable andnon-polluting sources such as wind power. Because of its nature wind is very unpredictable and thus the power from wind electric power systems. As the wind power sources gain more reliability in performance and become cost competitive, the penetration level of wind power in the overall generation mix will increase. Reliable wind power prediction methods are needed. Short time predictions areneeded for economic generation scheduling and long term wind power prediction will be useful for planning purposes. Neural networks have been used and are very good for pattern recognition and when fully trained their response is extremely fast and they are very robust. Neural networks have been investigated for load prediction for utility. This paper investigates different multilayerfeedforward neural network architectures using back-propagation algorithm for the development of a neural network-based wind turbine power output predictor. This multilayer feedforward neural network was trained and tested on a sample wind turbine power output data. This paper will also discuss control and communication issues related to interface of wind electric power system with utility.