Analysis of the measure-correlate-predict methodology for wind resource assessment

Twenty-one data sets with widely varying wind speed characteristics are analyzed to investigate the predictive capability of three measure-correlate-predict (MCP) methodologies: the Variance Ratio, Mortimer's, and artificial neural networks (ANNs). The MCP method involves correlation between observations at two or more monitoring stations and consequent prediction of wind speed at one station based on the historical period of record at the other station(s). Results show that an increase in the averaging interval of wind speed tends to increase the correlation coefficient (r) between neighboring stations, while average wind power density (WPD) decreases with increasing averaging interval. A method for extrapolating estimations of WPD from one averaging interval to another is proposed and found to produce reliable estimations. Artificial neural networks have the potential to generate robust predictions of wind speed at a target site, however, the optimization of the network weights must be closely monitored to avoid early exiting of the algorithm due to local minima. Of the evaluated measure-correlate-predict methodologies, only the Variance Ratio is recommended for general usage and only for data sets with a correlation coefficient greater than 0.8. Though more analysis is necessary to disaggregate the influence of topography on the success of these techniques, the correlation coefficient is highest between sites with uniform terrain such as ocean buoys and lower between sites with complex topography. All three MCP methods produce highly uncertain results for sites with lower correlation coefficients. But even for sites with higher correlation, special consideration should be paid to the uncertainty of estimated metrics introduced by the Variance Ratio method. Linear trends of root mean squared error (RMSE) vs. r estimated in this analysis can be used to indicate the error associated with the application of the Variance Ratio method to a new pair of stations. The jackknife methodology for estimation of variance is not recommended due to unacceptably high RMSE in the results.