Solutions

Analog ensemble (AnEn)

As part of NASA and NREL funded projects, a new method has been proposed and demonstrated for the long-term estimate of the wind speeds at a target site, a key step in wind resource assessments (Vavyve et al. 2013, Zhang et al. 2015). Analog ensemble (AnEn) techniques have been used with success for short-term weather predictions (e.g., Delle Monache et al. 2013). In the context of the wind resource assessment, the analog-ensemble method draws on the information contained in the historical data of multiple physical quantities over the period these data overlap with the observations (known as training period; typically 365 days) of the quantity of interest (known as predictand; the wind speed in this study). The relationships derived within the training period are then applied to reconstruct the on-site wind speed over the period for which there are no observations (hereafter referred to as reconstructed period, e.g., the past 20 years before the measurement campaign started).

Figure 1 Sketch of the functioning of the analog ensemble method for one analog predictor, the analog trend reduced to one time step, and when retaining the best three analogs.
Figure 1 Sketch of the functioning of the analog ensemble method for one analog predictor, the analog trend reduced to one time step, and when retaining the best three analogs.

More precisely, this is a three-stage process that is executed independently at every target site for every hour of the reconstructed period, as sketched in Fig.1. First, the historical value of multiple physical quantities (known as analog predictors; e.g. wind speed itself, wind direction, pressure, etc.) is retrieved for a time window (known as an analog trend) centered around time t (black dot in Fig. 1). The analog predictors are selected beforehand based on their known or anticipated correlations to the predictand. Second, other historical cases with conditions similar to those in the target window are identified (known as analogs) by looking at a time window (known as analog search window) centered around the same hour of the day for every day in the training period, and ranked by closeness of match. Analogs may therefore come from any day the training period. Using multiple predictors helps distinguish the analogs by identifying specific weather regimes relevant to the predictand. Third, the K best analogs (the number of analogs; black circles) are selected, and the corresponding observed values of the predictand are retrieved (black squares). The latter constitute the ensemble members for hour t.

The final result is the analog ensemble, i.e., a set of K wind speed values for every hour t of the reconstructed period. The assumption is that if analogs are found, their errors will likely be similar to the error of the historical time step to reconstruct, error that can then be inferred from them. As shown by Vavyve et al. (2013) and Zhang et al. (2015):

  • The AnEn can be used effectively for wind resource applications;
  • The AnEn provides an accurate long-term wind resource estimate at target sites;
  • The AnEn reliably quantifies the uncertainty allowing for cost-effective decision making;
  • The AnEn is a computationally inexpensive method.