The problem starts simply: when clouds block the sun, solar panels produce less power.
Whether it’s a solar farm or a residential rooftop, the amount of energy generated by photovoltaic panels can drop quickly when clouds shade the sun.
As renewable energy provides increasing levels of energy to the grid in Australia, grid operators must carefully manage the balance between generation and consumption. By managing this energy variation, they can make the best use of abundant renewable energy resources.
For a solar farm, cloud cover can cause fast changes in energy output to the grid. Solar farm operators are now responsible for providing the grid operator—the Australian Energy Market Operator (AEMO)—with forecasts of their output into the grid, and into AEMO’s central dispatch system that controls how much energy each generator outputs every five minutes.
AEMO needs these forecasts to balance energy supply and energy use. If these don’t balance, then the grid voltage and frequency will fluctuate. Once these fluctuations exceed set voltage or frequency limits, then loads have to be shed or supplies isolated.
To keep the grid frequency stable, it helps if AEMO has forecasts of future changes in the solar energy output from each solar farm on the grid. AEMO can then make up any energy shortfalls by increasing or decreasing the energy output of other slower-responding and often more expensive, but controllable, generators (like gas or coal fired generators).
AEMO already has an ageing centralised forecasting system, but there is lots of room for improvement. Improvements can be made by customising forecasts for individual farms by using extra data gathered from local sensors, and by combining several forecast models.
This is where the Solar Power Ensemble Forecaster steps in.
Predicting the future, five minutes at a time
By having ground-based cloud-sensing equipment on site, solar farms can provide a more accurate prediction of future cloud cover.
We’re working on this challenge, through an ARENA-funded project called the Solar Power Ensemble Forecaster. The Newcastle company Industrial Monitoring and Control is leading the project, along with CSIRO Energy, the University of New South Wales and the University of South Australia.
Together, we’re working to produce a solar forecast from multiple different models, an ‘ensemble’ of forecasting models. Combining models into an ensemble allows selection of the best forecast in a range of weather conditions, seasons and times of day, increasing the accuracy beyond what any single model can achieve by itself.
We have installed the forecasting system at five solar farms in New South Wales, Queensland and Victoria. They are providing five-minute ahead forecasts that significantly outperform the older forecasting system.
“This project will help move people away from the thought that renewable energy is unreliable,” explained CSIRO’s Chris Knight. “Renewable energy is not unreliable, but it can be unpredictable. Models like this will help us improve the prediction of how much renewable energy will be produced. Solar farms are reliable as any other infrastructure. Basically, this technology allows us to predict the future.”
For a solar farm, this provides greater grid stability, increased revenue and better utilisation of the available solar resources at these sites.
And for the rest of Australia, better short-term solar energy forecasts mean lower-emissions, cheaper energy and a more stable electricity grid.
Better together: Why we combine predictive models
“An ideal forecasting system will make use of multiple techniques, as each technique has strengths at different timeframes,” says Dr Merlinde Kay, from the University of NSW.
Our ensemble model learns which model has historically operated well in certain conditions. Then, it can intelligently weight and blend these predictions together to give a final forecast that is almost always more accurate and robust than any individual model.
The individual models include:
Sky Imaging Camera Model
The CSIRO Sky Imaging Camera takes photos of the entire sky every 10 seconds. Using the captured images, clouds can be tracked and their path can be predicted, enabling the forecasting of solar panel shading events. This, coupled with historical records of the solar farm’s output, can accurately predict how future electricity output will be affected by clouds. Focusing on individual clouds in this way is useful for predictions up to 20 minutes ahead, while lower resolution images from satellites or data from numerical weather predictions are better for forecasting events that are farther away.
Numerical Weather Prediction Model
Weather bureaus, like the Bureau of Meteorology, use sophisticated weather and climate models run on supercomputers to provide long-term weather forecasts for hours and days into the future. Those models include irradiance forecasts, which we combine with information from a Power Conversion Model (detailed information about the solar panels and inverters at a farm) to predict how much power a solar farm can generate. This forecasting method is less accurate than other models for short timeframes, but it has several advantages: it operates independently of the farm’s historical data and is not affected by outages of local communications systems or production limits imposed by the solar farm or market operators.
Statistical Model
We also use statistical time series techniques to forecast the electrical power output from solar farms. There are two factors that we input into this model: the average seasonal weather patterns and the maximum power output available from the farm during optimal (clear sky) conditions based on historical data. By using this information, we can determine the best input criteria for current weather conditions, to generate other data.
Satellite Model
Finally, we use raw satellite imagery to produce Cloud Motion Vectors. Forecasting the motion of clouds in this way is most useful on timescales of one to six hours. If we can accurately determine the movement of a cloud over a solar farm, we can determine the impact the cloud will have on irradiance.
No single type of forecasting model outperforms all the others in all conditions. For example, Sky Imaging Cameras often work very well in intermittent conditions with fast-moving clouds. Statistical models excel on clear-sky days. And satellites can see clouds approaching from further away, so (when combined with Numerical Weather Prediction models) will give forecasts for many hours ahead.
Some models also rely on recent power output from the farms to predict future output, whereas others can predict power using only external data sources—which is more accurate when communications systems fail, or a farm’s output is capped.
“The ensemble model intelligently learns from historical predictions made from several forecasting models, favouring forecasting models that perform better than the rest,” says Dr Merlinde Kay, from the University of NSW.
We have been trialling the models for about six months. For far, the combined ensemble forecast is significantly outperforming AEMO’s existing forecast system. The results from our forecasting trial will be available in late 2020, in the final report.