An Australian university has developed software that by increasing accuracy of five-minute forecasts can push down associated liabilities to the NEM.

The energy sources that will supply increasing amounts of electricity in tomorrow’s grid – sunlight and wind – are abundant in Australia. But that doesn’t mean we can always rely on them. There are plenty of still and cloudy hours.

Owners of solar and wind generation connected to the grid are expected to tell the Australian Energy Market Operator how much electricity they can supply in the following five minutes. If they can’t keep that promise, they’ll be charged by AEMO for any shortfall as the operator then sources that power from somewhere else so it can match supply with demand.

The art of forecasting, then, is taken seriously by wind and solar operators. Or it should be.

Currently, AEMO uses plant owners’ internal supervisory control and data acquisition (SCADA) data feeds to make its assumption of future output, but researchers at Monash University have developed a system they say is 45% more accurate.

The software solution the team has developed, which is being marketed by collaboration partner Worley as PowerPredict, relies on the data supplied by the wind or solar farm to AEMO, so there is no need to invest in extra gear such as sensors or weather masts. “We use the sensors we have and focus on improving the forecasting model using machine learning methods,” says Dr Christoph Bergmeir, a senior research fellow in Monash University’s department of data science and artificial intelligence.

What’s happening up there

Wind farms all send AEMO a list of required data for each turbine, such as wind speed, wind direction and power produced, and some machines supply additional data.

Plant operators are required to supply an unconstrained forecast, “but if constraints apply then obviously … that makes it a bit more complex,” Bergmeir says.

To estimate the output of a plant five minutes into the future AEMO uses “persistence forecasts”, which blithely assume the next five minutes will be the same as the past five minutes. This simple assumption makes a lot of sense but it doesn’t always work.

“Let’s say you have a ramping event, then the persistence forecast is not going to be a very good forecast,” Bergmeir tells EcoGeneration. Another example would be when power production is already running at maximum, and a good hunch would say that in five minutes time it would be more likely to decrease than remain unchanged.

“There are certain dynamics in the signal that you can try to capture and that our solution can capture,” he says.

Wind goes in, watts come out

The output from a wind turbine relies on two things: the wind that goes in and what the turbine does with that wind. Any software that attempts to predict output from a turbine by looking at weather variables alone will miss out on the “non-linear transformation” between what goes in and what comes out. “The system needs to do two things: it needs to forecast the wind but it also needs to adequately figure out this non-linear transformation [of wind to power],” Bergmeir says.

The output of any turbine is constrained, beyond which additional wind energy will not translate into additional power. “You need to look at both the power output and the wind, because if the wind is already well above where the turbine produces maximum power then even if the wind drops it will still stay at maximum output,” he says. “So, the higher the wind the higher the probability you stay at the maximum.”

The researchers had access to the 130.8MW Waterloo Wind Farm in South Australia and 11MW Ross River Solar Farm in Queensland, both operated by project partner Palisade Energy. The Australian Renewables Energy Agency provided close to $1 million in funding for the project.

Ensemble model

The software relies on machine learning, where characteristics of data from the recent past – the past 60 minutes, say – inform an “ensemble model” that predicts the next five minutes.

The research threw up some surprises. For instance, Bergmeir noticed that data up to two hours old could be used to produce quite accurate forecasts. One might expect that what happened two hours in the past should have no influence on what happens in the next five minutes, but there is some evidence to the contrary.

In truth, he says, expected delays in settling the energy market mean that the model is in reality looking about seven minutes out, not five. That stretches the forecasting window by about 40%. “It was quite surprising how big of a difference that is in terms of accuracy.”

Bergmeir concedes more work is needed on developing a solar model. The wind model works on the SCADA data available to AEMO, he says, but a solar model will need additional data before it satisfies the researchers. That may mean installing sensors.

Owners of wind or solar plants may already have taken other options and invested in forecasting technology, where local and international developers claim operational cost savings and a boost to revenue. Bergmeir won’t reveal anecdotal benchmarking against forecasting technology that might compete with PowerPredict, but Worley claims the product has achieved a 45% improvement in its customers’ power output predictions.

Wind picks up

On the face of it, it could be conjectured that plant owners might lose money if they rely on AEMO’s forecasts. Should it be up to them to spend money on a solution like PowerPredict, or is AEMO the obvious buyer?

In Bergmeir’s mind, AEMO is doing plant owners a service by allowing them to supply their own forecasts. If those forecasts are not accurate, there is a cost to be paid to AEMO. “In that sense it is really good that innovation is driven by competition between companies that try to get good forecasts,” he says. “On the other hand, could AEMO provide a better forecast?” The answer is blowing in the wind.

The key benefits of the project include increased renewable energy penetration in the grid due to improved dispatchability of renewable generation and reduction in frequency control ancillary services payments by generators resulting from the failure to meet forecast targets.

“If renewable generators can lower their causer pays factors, they can produce electricity cheaper, and eventually that saving could be passed on to the customers,” Bergmeir says. “It would also make renewables more competitive, which is also a desirable outcome.”

The team worked in collaboration with the Monash Business School’s department of econometrics and business statistics and the project was initiated by the Monash Energy Institute’s Grid Innovation Hub.

Energy consultancy Worley says that in 2020 inaccurate power predictions cost Australian generators $210 million.