Skip to Content

The purpose of this summary is to share insights related to the Short Term Forecasting Trial supported by ARENA.

Short-Term Forecasting

ARENA projects

ARENA’s Short-Term Forecasting (STF) trial is exploring the potential for wind and solar farms to provide their own, more accurate, forecasts of their output into the Australian Energy Market Operator’s (AEMO) dispatch system. This is a summary of the progress and insights gained between October 2019 and April 2020.

Read the STF trial background
Read STF reports

Trial-wide insights

  • On average, the trial’s self-forecasts accredited by AEMO are more accurate than AEMO’s current forecasting system. Trial participants have confidence that onsite, high-dimensional weather forecasting models can outperform AEMO’s current system.
  • Machine learning algorithms that produce the self-forecasts benefit from 12 months of training to account for seasonal changes at site locations. (Windlab & Vestas)
  • Site maintenance and communications outages known as ‘edge cases’ can affect forecast accuracy. (Windlab)
  • Self-forecast models should account for AEMO constraints to avoid higher causer pays charges (Meridian, Fulcrum 3D). E.g. Advisian developed two models, one using Wind Power during non-constrained periods and another that relies on external factors such as wind speed and direction. These models run in parallel and the algorithm switches depending on if a constraint is active. This is essential for accurate forecasts when constraints are lifted. (Advisian)
  • The AEMO Causer Pays Factor methodology and data accessibility was designed for AEMO internal use and static publication. Participants found it difficult to navigate (IMC, Vestas). Some participants have replicated the AEMO model, which will allow the commercial benefits of self-forecasting to be assessed. (Fulcrum 3D, Proa Analytics, Solcast)
  • Self-forecasting is not a “set and forget” process. Both the site asset manager and self-forecast provider need data access to remain valid and signals correct to support the ongoing optimisation of the self-forecast models. (Advisian)

Solar focused insights

  • Skycam and CloudCAM system hardware used to forecast solar patterns are essential for short-term 0-30 minute ahead forecasting. Past the 30 minute mark, other data sources such as satellite imagery become more valuable. (Solcast & Proa Analytics)
  • The CloudCAM system hardware has proven to be effective in the high temperatures and often heavy rainfalls of Northern Queensland (Fulcrum 3D – Solar). In dusty areas, the Skycam lenses need to be regularly cleaned to collect accurate data. (Proa Analytics)

Wind focused insights

  • The wind forecasting machine learning models are performing well. Ongoing adjustments to the models will improve the forecasting of wind generation ramps to account for sudden changes in wind direction (DNV GL, Vestas, Meridian, Aeolius Wind Systems & Advisian).
  • Meteorological masts 10-20 metres high collect sufficient measurements at roughly a quarter of the cost of a 40 metre mast (Windlab).
  • Forecasts should check high temperatures against the cut-out temperature for wind turbines in the forecast model to avoid instances of over forecasting generation. (Meridian)

Experience with AEMO’s MP5F API

  • The AEMO Market Participant 5-minute Forecast API (MP5F) password reset was an ongoing challenge. AEMO notes that they are close to releasing a solution. (Proa Analytics)
  • The MP5F is an evolving platform with room for business improvements. (Advisian & Fulcrum 3D)
  • AEMO servers experience more self-forecast submissions near gate closure time. This challenge of reduced performance could increase as more farms and forecast providers submit self-forecasts. (IMC & Advisian)

Report references

Further reading

Last updated 19 July 2020
Last updated
19 July 2020
Print Friendly, PDF & Email
Back to top