Statistical Analysis of Environment Canada’s Wind Speed Data Someshwar Singh Department of Electrical and Computer Engineering University of New Brunswick-Fredericton New Brunswick, Canada Email: someshwar.singh@unb.ca James H. Taylor Department of Electrical and Computer Engineering University of New Brunswick-Fredericton New Brunswick, Canada Email: jtaylor@unb.ca Abstract—Wind energy utilities use wind speed modeling and prediction to forecast their power production in order to participate in electricity markets. Time-series models which are indirectly based on a Weibull Distribution (WD) are used extensively to predict wind speed. The WD is converted into an approximately Gaussian distribution, as there are no rigorously developed time-series models for random variables possessing a WD. This conversion is performed using the parameters of the WD, a procedure that may negatively impact the accuracy of the forecast – research has demonstrated that WDs under- or over-fit the lower and upper ranges of wind speed histograms. This paper reports on a study of the histories of wind speed forecasts and actual wind speed data available from Environment Canada and the resulting estimates of forecast error distributions and statistics. It is shown through statistical analysis that the hourly prediction error distributions are nearly Gaussian in nature. It also appears to show that the statistics of the wind-speed prediction error do not increase significantly as time increases, which is in contrast to other researchers’ arguments that the error increases over time. This result may warrant further investigation. I. INTRODUCTION The intermittent nature of wind power generation poses operational difficulties to electricity markets. An electricity market operated by an Independent System Operator (ISO) must always maintain a balance between supply and demand of electricity at each instant of time. If there is any variation in load, there must be reserves at the ISO’s disposal. To maintain stable operation of the grid, the ISO accepts hourly bids starting at 9:00 am and ending at 11:00 am Atlantic Standard Time (AST), for the following day (Delivery Day, 00:00 to 23:59:59), from buyers and suppliers [1]. The system operator then runs an optimization algorithm to calculate the price at which maximum demand has been fulfilled at minimum cost. The participants have to fulfill their obligation at the time of delivery. After the delivery day, deviations from the hourly accepted bid quantities are calculated for each market partici- pant and financial penalties will be charged to the defaulters. The wind energy (WE) utility thus faces the challenge of producing accurate power generation forecasts before entering into the electricity market, as power forecast errors could have a significant impact on the WE utility’s revenue. Wind power prediction requires wind speed forecasting because the kinetic energy in the wind is converted into electric power by the wind power generator. Stationary time-series models are used extensively for modeling and forecasting wind speed [2], [3], [4]. The wind speeds are recorded at the site, then their distribution is plotted; the statistical distribution of the series does not change over time. It has been assumed that the wind speeds follow a Weibull Distribution (WD) [5]. Since there is no rigorously developed time-series models for random variables possessing a WD, the data is transformed into an approximately Gaussian Distribution (GD) [2], [3], [4]. Garc ´ ia-Bustamante et al. [5] and Jamil et al. [6] have shown that the WD assumption for recorded wind-speed data is not appropriate – the WD under- or over-fits wind speed histograms, especially in lower- and upper-range wind speed intervals. The transformation from a WD to an approximate GD is carried out by raising each hourly wind speed to the power of m; the value of m is calculated using shape and scale parameters of a WD. Since a WD fit is not quite appropriate for recorded wind-speed data, the transformation from a WD to an approximate GD may not be a realistic characterization of the recorded wind-speed data. Holttinen [7] has shown that prediction error increases as time increases; we have not observed this in our data set, however. In such cases, the WE utility revenue could increase by 7% if wind power is traded 2 hours before actual delivery [7] rather than 16 hours, as is presently the case in New Brunswick. II. PREDICTION ERROR CALCULATION Environment Canada’s Fredericton station provides weather forecasts every day at 08:00 for the next 48 hours in 3 hour blocks in Gridded Binary (GRIB) format. The forecasts are