Forecasting Power and Wind Speed using Artificial and Wavelet Neural Networks for Prince Edward Island (PEI)

Wind power is a fast growing renewable energy technology world-wide. The power generated by wind turbines changes rapidly because of the continuous fluctuation of wind speed and direction. This results in the variability, unpredictability, and uncertainty of wind resources. Therefore, it is important for the electric power industry to predict power produced by the wind for power management and control.;The proposed models are used mainly for short-term forecasting of average values from 10-min up to 3-hrs ahead. The implementation of the forecasting models is based on different forecasting approaches. These include time series models, forecasting with multi-input variables and spatial correlation models. Time series models use past and current values in a time series to determine future values. They are based on single-step or multi-step ahead forecasting of averaged data. Moreover, ANN and WNN are trained with multiple input variables because wind turbine power production can be affected by many influence factors. Finally, forecasting models based on spatial correlation are investigated. These models are trained using data measured at neighboring sites up to 210 km away from the reference site. Choosing the appropriate number of inputs for a forecasting model is an important issue and it was done in this study by using trial-and-error and the linear correlation analysis methods;For verification purposes, actual historical data from two wind farms and four meteorological stations around the Maritimes region on the East coast of Canada are used to develop and test the forecasting models. The accuracy of the proposed models is evaluated by comparing their results with the persistent method. Computer simulations show that with appropriate input variables, both ANN and WNN forecasting models trained using the back-propagation algorithm can improve the forecasting accuracy compared to the persistent models. Moreover, the new proposed technique using the wavelet filtering method enhanced to improve the forecasting accuracy.;In this thesis, Artificial Neural Networks (ANNs) and Wavelet Neural Networks (WNNs) are proposed for wind speed and power output prediction. Focus is mainly on WNNs to examine their suitability and practicality to solve the problem of short-term wind speed and power prediction. Furthermore, a new method, based on wavelet soft-thresholding, is proposed. The method uses neural networks in combination with the wavelet filtering technique. The filtering process reduces the effect of noise and sharp edges in the measured wind data (time series), which are considered undesired parts or less effective in the developed forecasting system. The benefit of using filtered (smoothed) wind data series to improve the forecasting accuracy is investigated and compared with the results of a reference model, namely the persistent model. The comparison of different thresholding techniques show that the universal thresholding technique with Daubechies wavelet (db3) is effective and hence is recommended to be adopted for wind data series filtering (time series smoothing).