Benchmarking predicted and actual production losses

Assessment of production losses
400MW portfolio – Belgium, France & the Netherlands

3E investigated the performance of 57 wind farms across Belgium, France and the Netherlands, representing 184 wind turbines for a total installed capacity of 389MW. The portfolio included different wind turbine types and manufacturers, and consisted of recent wind farms with at least 1 year of data, as well as older wind farms with more than 4 years of operation.

The aim was to better assess the actual wind turbine technical availability, as well as benchmarking a number of production losses that are used in the pre-construction energy yield assessment, either as assumptions or as calculated values, to obtain the P50 figures from the gross energy yield. Most of the production loss values are derived from time-based values and in many cases, the assumption was made that a 1% loss in time represents a 1% loss in production. For example, wind turbine production-based unavailability is usually considered as 3% on the basis that O&M contracts warranty a time-based availability of 97%. However, it is clear that turbine shutdown during low wind conditions does not have the same impact on production than the same shutdown when the wind blows at 15 m/s.

3E developed a code (Python) to handle the various SCADA data formats and store them in a one unique database. For each turbine, the analysis unfolds as follows:

  • From the error log book, each stop is allocated to a single error code by clearly defining the start and end date of the stop event, regardless of the succession of errors that may have been trigged by the initial error.
  • The duration of each stop is then calculated, and the status of the turbine is added to the 10min database. By default, the status is 0 when the turbine is operational, and is the value corresponding to the error code of the error causing the shutdown, when the turbine is not operational.
  • A SCADA-based power curve (called “observed power curve”) is created, by filtering out all the 10min timestamps when the turbine is not producing for 600s (the grid connection counter from the 10min database is used). In many cases, different power curves were detected due to noise, bat or grid curtailment.
  • For every 10min of downtime or curtailment , the lost production is assessed by looking up the average wind speed in the observed power curve. This power is then converted into production that the turbine would have generated, had the turbine been operational. This approach may seem simplistic, however when tested over long periods of 100% technical availability, the difference between the production estimated using the observed power curve and the actual production as provided by the wind turbine counters, was less than 0.8%.

When no nacelle wind speed is available, typically during a grid outage, the wind speed was estimated using MERRA data. To do so, a linear regression was established between the nacelle wind speed and the MERRA wind speed of the nearest MERRA grid point, for timestamps during which the turbine is operational. This was then used to recreate the nacelle wind speed during grid outage.
The production-based technical availability was found to be equal to 97% in average, as can be seen below (50% of the wind turbines operated at more than 97.8%).

WIND_April_2016_600x329_1

The following figure shows a clear improvement in time and production-based availability over the first 3 years of operation.


WIND_April_2016_600x381_2b

The next figure shows that the wind turbine production-based technical unavailability was 2.71%, which is lower than the pre-construction assumption of 3%. The other sources of production-based unavailability break down as follows: curtailment 0.29%, grid 0.28% and ambient errors 0.75%. Unlike the other losses, 3E noted a very wide spread of curtailment losses, which reflects the very changing noise/bats and power curtailment requirements from one site to the other.

WIND_April_2016_600x349_3

Conclusions:

– The pre-construction loss assumptions are broadly in line with actual losses: wind turbine unavailability losses are slightly lower than predicted while grid losses are generally slightly underestimated. All in all, the total amount of losses predicted before construction matches the reality.
– Good O&M practices result in the production-based availability being higher than the time-based availability. In other words, wind turbines are generally stopped for maintenance at low wind.
– The ramping-up effect is not a myth, and a clear availability increase has been identified over the first 3 years of operation, again suggesting good and sound O&M practices.