The capacity factor of wind

Guest Post by John Morgan. John is Chief Scientist at a Sydney startup developing smart grid and grid scale energy storage technologies.  You can follow John on twitter at @JohnDPMorgan.

A lot of ink is spilled on wind intermittency, and not necessarily based in data.  So I have extracted and analyzed a high resolution dataset of a year’s worth of Australian wind power for a number of interesting properties.  I previously wrote about the capacity factor as a limit to the share of electricity that wind and solar can acquire, so I also ask how wind capacity factor changes with time, place and season.  In particular, how does it change during sunlight hours and what does it mean for the capacity factor limit on renewable energy penetration?

Australian wind fleet data

The Australian Energy Market Operator (AEMO) publishes data on all generators connected to the National Electricity Market (NEM) grid, which covers the eastern states including Tasmania, but excludes Western Australia and the Northern Territory.  The data includes power generation every five minutes for every generator for the last year, their capacities as registered with the grid operator, and more.  It is not very accessible, being in the form of thousands of SCADA data files, many of which contain errors.  But with a bit of work the data can be extracted.  Here, for instance, is the output of all grid-connected wind farms at five minute resolution over one year:

Wind capacity factor

Here is the top level summary of the Australian wind farm fleet over the last year:

The nameplate capacity is the total capacity of all wind farms – 3753 MW.  But the whole fleet only manages 3238 MW at peak.The whole is less than the sum of its parts – half a gigawatt less in this case. Why is this?

The fleet is comprised of wind farms distributed over a large area of eastern Australia.  To achieve maximum theoretical power the wind would have to be blowing at the optimum speed for each wind farm, at all wind farms, simultaneously.  This is a statistical improbability and quite possibly a hydrodynamic impossibility, as it would require a high velocity correlated flow field over very large distances.

So while we often hear that the wind is always blowing somewhere, it is equally true that it is always not blowing somewhere else, and the fleet output never achieves full capacity.  Australia in theory has 3753 MW of wind capacity, but this will never be realized in practice.  Similarly, statements like: “The US added x GW of wind capacity last year”, overstate the capacity addition because the new wind build is unlikely to ever produce its maximum power in full correlation with the rest of the fleet.

In other words, national capacity statistics overstate the potential output of wind.  The flip side of this is that the capacity factor limit underestimates the potential for wind penetration.  We can push the penetration of wind a bit higher than the capacity factor – generation would start to exceed demand at 33% share, rather than 29% share.

Wind correlation times and the synoptic scale

Over what distances are wind farm outputs correlated?  Its actually easier to ask, over what period in time is wind power correlated?  This information is contained in the wind power autocorrelation function, which we can calculate from the dataset:

The autocorrelation function tells us how long the influence of a particular state of wind persists.  If its windy now, for how long will it remain windy?  Surely for the next five minutes.  But will it still be windy tomorrow?  The autocorrelation function is a “memory” function that tells us how long wind “remembers” how hard it was blowing.

The autocorrelation function decays in about 40 hours (since we have 5 minute data the x-axis is in units of 5 minute “lag”s).  This means wind, on average, bears no relation to its output 40 or more hours ago.  Wind has a “correlation time” of about 40 hours.

Its interesting to interpret this as the time for a body of moving air to pass over a windfarm.  If we knew how fast it was moving we’d know how big it was.  Wind resource maps suggest a velocity of about 7-8 ms^-1, so that interpretation suggests a wind correlation distance of about 1200 km – the “synoptic scale”.  This seems a pretty reasonable estimate of the size of weather systems and can in fact be done for a single wind farm with a similar result.  Its remarkable to be able to pull large scale geographic information out of just the power fluctuations of a single wind farm!

http://www.seabreeze.com.au/Photos/View/2871107/Weather/Australia-wind-energy-map/

If we wanted to cover intermittency we would need to ensure our wind fleet is dispersed over distances of 1200 km and more, so that the output of at least some of its wind farms will be uncorrelated.  This smooths the output of the wind fleet, reducing maximum output below the nameplate capacity, but increasing the amount of energy that can be integrated without having to spill, store or manage excess generation.

Wind power ramp rates

Wind output is constantly changing and requires the rest of the grid to be flexible enough to ramp up power or shed load to balance wind fluctuation.  This rate of change is just the time derivative of wind power.  The plot below shows this “acceleration rate” throughout the year.  It’s a normal distribution, the symmetry showing that wind power picks up as fast as it drops off, and that the grid needs to be responsive at a rate of 20 MW per minute, in both directions, to cover most conditions.

As more wind is added, the flexibility of the rest of the grid will have to increase proportionately – double the wind energy would require about 40 MW/min ramp rate.  But this additional ramping ability must be delivered by the shrinking dispatcheable generator sector.  So the intrinsic flexibility of the rest of the grid must increase, and faster than in simple proportion to wind penetration.  Practically that means increasingly strong pressure to shift from coal generation to gas as wind share grows.

Low wind days

Lets look at the distribution of low wind days.  We can ask, for how many days was wind output below some level?  For instance, we can find 29 days in which output was below 10% of capacity, and 127 days below 20% capacity.  127 days is, incidentally, pretty close to the number of weekends and annual leave of most Australians – Australian wind is obviously governed by Australian workplace awards!

The plot below shows the number of days below a particular output level.  Interestingly, the daily average power output never exceeded 75% of capacity, or 2810 MW, almost 1 gigawatt less than installed.  The fleet was never totally becalmed, but the lowest recorded day in the year saw output of just 2.7%.

Also of interest is the number of consecutive low wind days.  This affects the strategies we might use to cover wind outages – whether we store energy in batteries, or with pumped hydroelectricity, or shed load, cut in gas generators, or coal.  For instance, of the 29 days of wind output below 10% of capacity, 15 are single isolated low wind days, and then there are 7 pairs of 2 day long low wind runs.  If we look for sequences of days with less than 20% output, we find 2 runs of 5 low wind days.  The full distribution is shown below.

The number and distribution of low wind days show that while wind contributes energy, it does not provide capacity.  Alternative generation capacity must be available to meet the near absence of wind about one day in ten, and for two or more days in sequence.  But many of these low wind events are of just a single day duration.  This is a difficult timescale for coal plants, so again, increasing wind penetration drives the residual mix towards gas.

Wind capacity factor by month and day

To get a handle on wind seasonality we can look at monthly output and capacity factors.  Each point in the plot below is the average power output for a day in the year.  The coloured blobs show the distribution of power output in each month.  Also shown is the capacity factor for each month.

The nameplate capacity factor varies from month to month (30%±5% covers it). Every month has low wind days and high wind days and everything in between with little seasonal structure.  The winter months have more high wind days, but they also have more low wind days, and one cannot confidently assert the monthly CF variation is greater than noise, in this year.

We can see this in a box plot of daily capacity factor – the data distribution is very wide and a capacity factor of 25% is consistent with the data for every month.  The highest daily capacity factor for the whole fleet was 75%, and the lowest was 2.7%.  These maximum and minimum output days both occurred in winter, when solar power is at a minimum.

Wind capacity factor by hour

We can push this through to a still finer grain by looking at wind output by hour of the day.  The following plot shows the average capacity factor by hour of the day, for each month.  We can see if wind picks up or drops off in any consistent way through the course of a day.  The summer months show some daily pattern, perhaps, but the rest of the year does not:

Does the wind output drop when the sun is shining?  It would be convenient if it did, as it would allow more solar  power on the grid.  Lets nominate solar hours as 10 am – 4 pm, and compare solar hours with non-solar hours.

There is no difference between wind capacity factor when the sun is high in the sky and when it is not.  Wind does not cooperate with solar by subsiding in the middle of the day.  Possibly wind blows harder when clouds block sunlight.  Unfortunately I don’t have a dataset for solar PV output and can’t test this.  My guess is that the number of days and the number of sites at which such anticorrelation occurs is not large enough to shift the average output of the total fleet enough to change the overall picture.

What does this mean for the capacity factor threshold?

As explained in detail by Jenkins and Trembath, it is increasingly difficult to build more wind or solar capacity as their market share approaches their capacity factor (CF) because they will then, at times, be producing energy in excess of demand.  The economic drag incurred by dealing with surplus generation by storage, curtailment or demand reduction undermines the economics of building additional capacity.  The capacity of wind and solar is thus limited to be roughly numerically equal to 100% of grid demand.

In “Less than the sum of its parts” I argued that adding solar to the mix actually reduces the combined amount of wind and solar energy that can be added to the grid.  This is because solar competes with wind for share of capacity, but contributes less actual energy due to its lower capacity factor.  Building solar thus reduces the maximum amount of renewable energy we can get onto the grid.

You can get around this if wind and solar generate at different times of the day, or year.  But from the data above we can say that wind does not drop during the day or pick up at night, to any significant degree.  The capacity factor of wind during “solar” hours is the same as during “non-solar” hours.

Turning to the seasonal variation, its possible wind has a higher capacity factor in winter when solar output is low, but the evidence of the last year is not compelling.  The lowest wind capacity factor in the year was actually in the winter month of June, and January in high summer was one of the higher producing wind months.  The winter months have more high wind days, but they also have more low wind days.

If there is some seasonal synergy between wind and solar, its not particularly strong, and the contention that the maximum share of renewable energy is achieved by building wind and not solar seems sound.

But the capacity factor threshold does require an adjustment.  Recall the peak wind output was only 3238 MW, less than the nameplate capacity of 3753 MW.  So we could build more capacity without fear of excess generation.  Instead of spillage or storage being required at 29% wind share, we can accommodate a more generous 33% share.  This greater share of wind energy is possible due to the geographic distribution of wind smoothing out some of the peaks.

Conclusions

There’s a lot of information in noise.  Deducing the size of large scale weather systems from the power fluctuations is pretty cool, as is seeing the signature of spatial distribution of wind farms in a one-dimensional time series.  Notably, a national wind fleet will not achieve full output due to geographical smoothing, but this smoothing also increases the capacity factor threshold for wind share a bit, from about 29% to 33%.

As expected, intermittency means wind contributes energy but not capacity to the grid, meaning wind acts as a fuel saver for fossil plants, which must increasingly shift to gas rather than coal as wind penetration grows, to accommodate higher ramp rates.

The capacity factor does not show strong consistent variation across hours, days or months, and share of renewable energy is limited as Jenkins and Trembath describe.  There is little evidence of a synergy between wind and solar in the Australian grid, supporting my earlier conclusion that a combination of wind and solar can displace less fossil energy than wind alone.  If we really wanted to push towards maximum renewable energy, we would build wind and not solar, and variable renewables share could grow to about 33%.

Data

The Australian Energy Market Operator Generation and Load data can be found at this page.

Five minute data for all generators can be extracted by parsing the SCADA data files in this directory.  Data goes back a little over a year.  Older data is not available – AEMO appear to delete the oldest files on this page on a daily basis.

AEMO lists all generators connected to the grid, by technology here, in the spreadsheet “Registration and Exemption List”, in the tab “generators and Scheduled Loads”.  Each generator has a unique identifier, the DUID, allowing it to be located in the SCADA files.

The SCADA files contain a number of errors – in many files the output of many generators is double counted.  A corrected data set was created by filtering out duplicate generator entries.

The data was extracted and analysed with python code using the excellent SciPy scientific python tools, iPython notebook, pandas data analysis library, with MatPlotLib and Seaborn data visualization libraries for plotting.