Short-Term Forecasting Trial: Summary of Key Insights Apr to Oct 2020

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 April 2020 and October 2020.

Read STF reports

• To confidently realise benefits from improved forecast accuracy, it may be necessary to also consider development of forecast models which are optimised to minimise FCAS charges. The complexity and dynamic nature of the FCAS market and charges make this a challenging problem. (DNV GL, Advisian)
• A number of forecasters in the trial found it is possible to satisfy AEMO’s accuracy requirements and achieve as low as zero causer pays charges in some months for both wind and solar farms via MP5F (Market Participant 5-Minute forecast) under normal operations. This was achieved by replicating the Causer Pays Factor procedure based on publicly available data from AEMO. Close consultation with AEMO is likely to be necessary to understand and deploy the required methodology. (Fulcrum 3D, Solcast, Proa Analytics)
• Recovering real-time data from sites was unique for each project. It will be important to allocate adequate resources to this activity for future projects. Extensive discussions with all parties involved in data acquisition and management for the site are necessary in order to fully understand the environment and establish the most appropriate method for data retrieval. (DNV GL, Fulcrum 3D, Solcast)

• Future projects will need to develop robust monitoring systems to pick up and respond to abnormalities quickly. The ability to quickly suppress a forecast is also required, particularly in the case of spurious input data such as inconsistent, missing, stuck and old measurements, as well as dust and wildlife which requires intervention. (Solcast)

– Note: AEMO has in the past and will continue to suppress farms causing gross error (materially effects that dispatch period) for NEM dispatch and instead used the AWEFS or ASEFS forecast. (AEMO)

• Establishing self-forecasting at a new site will present unique challenges and require unique training of machine learning models with bespoke settings to provide accurate results. Depending on how strongly a farm’s generation is affected by seasonality, it is noted that multiple months of input data are likely required, and a full twelve months of data is desirable to capture variations in production due to seasonality. (DNV GL)

• Wind farms that sit on complex topography require additional extrapolation of the wind speed measurements at turbine ‘hub-height’ to ensure that the changes in wind speed through the wind farm are accurately accounted for. (Meridian)
• Rainy conditions affect the visibility of LIDARs, the reduction of visibility can be used as a forecasting indicator that there is likely an incoming wind speed ramp event which could cause a fluctuation of the power generation of the farm. (Meridian)
• Remote sensing forecasting is a relatively new topic with standards and guidelines still under development. As such, forecasters using LIDAR should develop and test various methods to identify best-performing approaches. When choosing a forecasting strategy, particular attention should be given to the associated operational computation costs, as these may differ greatly. Time taken to compute is also an important factor to consider when forecasting in short, 5 minute time periods. (Meridian).

• CloudCAM system hardware is robust with the ability to detect issues such as water ingress in one of Fulcrum 3D’s nine CloudCAMs. The water was detected and corrected for in software by Fulcrum3D remote from the site. The cause of the water ingress was identified and rectified for subsequent builds. The ability to detect and correct such issues in a rapid manner demonstrates the power of modern “smart sensors” and the upside of being the CloudCAM designer, integrator and OEM. (Fulcrum 3D)
• Proa has designed and installed an innovative automatic self-cleaning system able to clean the lenses of skycams with pressurised water on a periodic basis. The project has demonstrated very good results of the self-cleaning system to remove dust or other atmospheric contaminants, but traces or droppings from birds still required manual intervention. For this reason, the functionality of the automatic self-cleaning has been extended to actively deter wildlife (using audio and mechanical deterrents) from approaching the instruments. (Proa Analytics)
• Modelling cloud dynamics is important to unlock the full potential of self-forecasts for solar farms. Accounting for cloud dynamics employs relatively complex models making use of skycam, satellite and weather models data. While these models successfully describe cloud cover, their complexity has presented additional challenges including additional fine calibration testing and monitoring and the required increased computing time within 5 minutes ahead. (Proa Analytics).

• Projects should consider sending both a forecast that will be safely accepted by AEMO as well as one close to the end of the submission window in order to minimise their forecasting error. (Meridian)
• A setting in AEMO’s system appears to round the plant capacity down to the nearest MW in the dispatch engine. This means that it is impossible to forecast accurately when the plant is permitted to produce at full capacity and is doing so. (Fulcrum 3D)

– Note: AEMO’s market systems can only store whole numbers. In the case of the maximum capacity field, this is rounded-down to the nearest whole MW. This can cause up to a 0.9MW error between actual output and forecasts/targets due to the difference between the maximum capacity value defined in the GPS vs max capacity in AEMO’s market systems. (AEMO)

• It is necessary to ensure that the process for interacting with the API is well understood, and that processes are in place for dealing with issues such as credential expiry and outages. (DNV GL, Fulcrum 3D)

– The MP5F API is stable and has been materially streamlined in the last 12 months. The consistent requirement for password resets every 90 days is cumbersome and reliant on the participant. This process could be improved via a secure automated (or semi-automated) system for resetting passwords. (Fulcrum 3D)
– Note: AEMO has published the ChangePassword API that provides a direct secure system for self-forecast providers to update passwords without coordination from the market participants (owner/operator of the solar or windfarm). (AEMO)

• Future projects should keep audit logs of AEMO API response codes for each dispatch interval of each DUID (unique identifier for each farm). This is important to understand how forecasts are classified by AEMO. (Solcast)

Wind

• Fulcrum3D
• Meridian
• DNV GL

Solar

• Proa Analytics
• Solcast
• Fulcrum3D