| Summary: | One of the largest challenges with the utilization of wind as a renewable resource, is its natural variability and intermittent nature. To achieve a sustainable integration of wind power into the power grid, a precise and reliable prediction method is therefore required. In this study, several short-term wind power prediction models based on statistical time series analysis were developed and tested, focusing on five wind power parks in northern Norway. All prediction models were applied to each of the five complex terrain sites, Havøygavlen, Kjøllefjord, Nygårdsfjellet, Fakken and Raggovidda wind park. The models apply meteorological forecast data, provided by the Norwegian Meteorological Institute, and the measured hourly total power output, for the time period 1. January 2017 – 31. December 2017, for each wind park. Five Markov chain models have been trained and tested using different sets of input parameters, such as wind speed, wind direction, temperature, surface air pressure and power output. Additionally, a meteorological data-customized power curve function by polynomial regression was developed and tested, using the on-site power output and forecasted wind speed and direction. The performances of all models were measured in terms of the NRMSE, and compared with that of a persistent model, by an improvement parameter. All Markov chain models were found to have lower NRMSE than the persistent model, for all five wind parks. The best performing Markov chain model at each wind park, in terms of improvement with reference to the persistent model, was found to be 6.17%, 4.86%, 9.31%, 9.48% and 12.01%, for Havøygavlen, Kjøllefjord, Nygårdsfjellet, Fakken and Raggovidda, respectively. A linear combination of the meteorological data-customized power curve function model and the persistent model, was found to outperform all Markov chain models at all five sites. A turbine-wise prediction for 15 turbines at Havøygavlen wind park, by the use of Markov chains, was found to attain an improvement parameter value of 8.07%. This suggested a substantial improvement gain by the turbine-wise approach, compared to the 1.98% improvement of using the same Markov chain model for the whole park. Furthermore, the wind regimes and seasonal variations at all sites are investigated by an analysis of the statistical properties of the applied wind data.
|
|---|