No need to be a statistician to see that averages are often useful to compare different things. And yet, they often fall short as a true measure of performance.
For instance, we use averages to compare the current weather against the climate normals or to get a sense of how our favorite sport team or player is doing. Team (player) statistics are a concise way of combining the scores from many games into a few numbers and assess their performance. Averages are convenient to compare two teams (players) or to compare the performance of a team (player) over time.
But averages can also hide important information. For example, Roger Federer had a better average record than Rafael Nadal at the peak of their professional tennis careers. And yet, Nadal almost always beat Federer on clay, and won the majority of their matches overall. So, which player was “better”? Averages tend to be rather insensitive to the frequency distribution of the underlying data. With a highly skewed distribution, the most likely outcome may be very different from the expected or average outcome.
Averages are often used in wind energy when estimating the energy production of a proposed wind farm. Typically, all plant losses are disclosed as annual averages in an energy production report. The report states that the predicted availability will be X % (typically X is between 90 to 97 %, Meaning an availability loss of 3 to 10 %) and that the electrical losses will be Y % (typically Y is around 2 %) on an annual basis. The chart below provides a somewhat standard way of categorizing the wind plant losses. A more natural way of taking into account plant losses would be to look at how each loss varies in a chronological order, minute by minute (or hour by hour), rather than looking at annual averages. Time-series energy modeling represents the next frontier of wind plant design and energy estimation.
Pie chart of wind plant losses with typical proportions for cold climates i.e. where icing losses non-negligible. The environmental losses typically include icing, temperature shutdown or derating, lightning, high wind hysteresis and blade degradation.
A time series approach (which is called time series energy modeling) has many potential advantages over the existing, standard method based on annual averages (which is called frequency domain energy modeling). Fundamentally, time series energy modeling allows us to capture a wider range of dynamic atmospheric conditions. Of course wind speed is not the only atmospheric variable worth considering in energy production estimates. For instance, (i) the wind shear and turbulence greatly impact wake and turbine performance losses, (ii) temperature and thermal stability play an important role in predicting wake, high wind hysteresis and icing losses, while (iii) relative humidity and precipitation also have major influences on icing losses. It should come as no surprise that energy production estimates would gain in accuracy in predicting plant losses by knowing how those meteorological conditions vary over time. Furthermore, there are benefits in knowing the prior conditions as well given that they impact the current plant losses. A time series approach to wind energy modeling would also allow us to properly model the gross energy production, that is the energy before accounting for any plant losses. Basically, the wind turbine power curves change with wind speed as well as several other environmental conditions, e.g. air density, wind shear, turbulence, inflow angle. Accounting for the variation in each turbine’s gross power generation on a minute by minute (or hour by hour) time interval leads to less simplifications, and potential errors, than using a distribution for each environmental variables (often times, only wind speed and air density are accounted for). Last but not least, time series of plant losses can provide the level of details necessary to understand under which circumstances a wind farm loses energy. Being able to pin point the source of some of the errors in plant losses, and consequently in energy production estimates, is the first critical step towards improving current methods.
In an effort to bring the science to the next level regarding wind energy production estimates, modifications were made to AWS Truepower’s Openwind software to implement a full time-series energy estimation model. Advances were made possible through the analysis of SCADA data from 18 operational wind farms in Québec, Canada, as part of a project to support grid operations, integration and reliability. Efforts were devoted to understanding the time-varying plant losses related to wakes, availability, environment, and electrical systems and developing ways to model them. Particular attention was paid to icing losses, which can be severe in the Canada and the northern US states mainly. Historical time series of wind power production and associated plant losses were generated for the 1979-2015 period (37 years). The long-term, hourly meteorological time series were created with the Weather Research and Forecasting (WRF) model initialized by the ERA-Interim reanalysis data set. The meteorological time series were then converted into wind power generation in Openwind, taking into account plant losses on an hourly basis. All plant losses were tracked separately and at the turbine level, providing the ability to make detailed comparisons with actual operation. Strong agreement is observed between the actual and modeled net power generation even though icing-related losses add to the complexity. The validation of the Openwind simulations indicated that the net wind power generation was well aligned with the actual generation, derived from the SCADA data, where the average hourly coefficient of determination (R2) was 0.80, while the mean daily R2 was 0.88. Our analyses also indicated that the monthly/seasonal trends in net power are well captured by the simulation system. The graphs below provide examples of the net power generation at three different wind farms and shows evidence of good agreement between our simulations and the reality.
We believe that simulating realistic meteorological and wind power generation time series that mimic the dynamic behaviors of wind plants is the logical next step in wind resource assessment. It will have implications not only in energy production estimates but also for curtailment studies (bats, birds, noise, and sector management) and grid integration studies. Compared to conventional fuel combustion systems, nuclear power stations and large hydroelectric plants, wind power generation is more variable. Understanding and predicting variations in atmospheric conditions (including wind) and the resulting wind power generation, especially during periods of high electrical demand, is critical for optimal grid planning and operations. It allows grid operators and planners to address sensitivities to operating reserves, storage options, carbon reduction plans, market scenarios, and other factors. We’re in the early ages of time series energy modeling but as we analyze more and more operational wind farms, our methods for simulating time-varying wind plant losses will only improve. Time series of wind power generation will provide the level of details necessary to learn why wind farms produce more (or less) energy under certain atmospheric conditions. As many of you know already, one must look at more than the tennis player statistics to predict the winner of the Grand Slam tournaments.