Estimating long-term annual energy production from shorter-time-series data

Abstract

Models used in wind resource assessment (WRA) range from engineering wake models and computational fluid dynamics models to mesoscale weather models with wind farm parameterizations and, more recently, large-eddy simulation (LES). The latter two produce time series of wind farm power of a certain period. This simulation period is, in the case of LES, mostly limited to ≤ 1 year due to the computational costs. However, estimates of long-term (O(10 years)) power production are of high value to many parties involved in WRA. To address the need to calculate long-term annual energy production from ≤ 1-year model runs, therefore, this paper presents methods to estimate the long-term (O(10 years)) power production of a wind farm using a ≤ 1-year simulation. To validate the methods, a 10-year LES of a hypothetical large offshore wind farm is performed. The methods work by estimating the conditional probability densities between wind farm power from the LES and wind speed from reanalysis data (ERA5) from a short (≤ 1 year) LES run. The conditional probability densities are then integrated over 10 years of ERA5 wind speed, yielding an estimate of the long-term mean power production. This "long-term correction"method is validated on varying simulation periods, selected with four different day-selection techniques. When applied to a simulation period of 365 consecutive days, the methods can estimate the 10-year mean power production with a mean absolute error of around 0.35 % of the long-term mean. When choosing the simulation period with day-selection techniques that represent the long-term climate, only roughly 200 simulation days are needed to achieve the same accuracy. Finally, a method to also include wind observations in the long-term correction is presented and tested. This requires an additional "free stream"LES run without active turbines and gives estimates of long-term power and wind that are corrected for a potential LES bias. Although validation of this final approach is difficult in the employed modeling strategy, it gives valuable insights and fits within the common WRA practice of combining models and observations. The presented techniques are based on physical arguments, computationally cheap, and simple to implement. Furthermore, they are not limited to LES but can be applied to other time-series-based models. As such, they could be a useful extension for the diverse set of modeling, observational, and statistical techniques used in WRA.