Uncertainties in Results of Measure-Correlate-Predict Analyses

Summary This paper explores the uncertainties in measure-correlate-predict (MCP) predictions. Prediction uncertainties as a function of concurrent data length are determined using 12 pairs of reference and target sites for which long term concurrent data sets exist. The data are then used to test methods for estimating uncertainties of MCP results when long-term data are not available. Uncertainties resulting from a linear regression analysis significantly underestimate the uncertainties due to the serial correlation of wind data. The jackknife estimate of variance, which estimates variance using MCP results based on multiple estimates, each with a segment of the original data missing, are also considered. The jackknife estimate of variance is an improvement, as shown by estimates of the uncertainty of predicted mean wind speed and Weibull parameters at the 12 paired sites. The jackknife estimate of variance appears to correctly estimate the uncertainty based on the concurrent data but still often underestimates the overall uncertainty in the relationship between the two sites because of the unknown variability at time scales longer than the concurrent data length. An MCP model with monthly terms and the use of an empirical correction factor are considered to remedy this situation. The empirical correction factor shows promise, but may not be applicable to all sites. An alternate approach, the relationship between MCP prediction uncertainty and the correlation coefficient of the data sets, does not provide any better results than the jackknife estimate.