Australia, Projects, Renewables

Intelligent systems: Smart answers to hard problems

The hunt for ways to push clean energy forward is well underway, with Australian research defining what intelligent systems look like, writes Jeremy Chunn.

It will take plenty of smarts to replace what is essentially a pretty dumb existing energy system, where stuff is burned to make heat to create steam to push turbines. The sunshine and wind that fuel new generation technologies are just as elemental as coal and gas, but mankind now understands why it’s better to power economies with fuels that leave no polluting byproduct.

The power plants that are replacing fossil-fuelled stations are simple in concept – they work when the sun’s shining and the wind is blowing – but it will require vast intellect to weave together many hundreds or thousands of assets, big and small, so output is deemed reliable.

Research teams and brave entrepreneurs have been making progress during the past 10 years, with the result being that the cost of technology is falling as a future where renewable energy is efficiently generated and distributed via connected systems becomes more possible.

In some ways it’s hard to tell who is working on the next big breakthrough because all pieces in the puzzle will impact the others. There has been a lot of talk about the “Internet of Things” and solutions that rely on artificial intelligence, where the energy system learns from every squeeze in supply or surge in demand and adapts itself to improve.

It sounds exciting, but is it real?

EcoGeneration asked some of the forward thinkers around Australia to describe where their work fits in and what they imagine is possible.

Fleet appeal

At the Australian National University (ANU) in Canberra, Johannes Hendriks and colleagues are working on a system that will use machine learning of electric bus energy expenditure based on terrain and weather so fleet owners can decide how to upgrade to emissions-free vehicles.

As electric buses replace the fuming, noisy diesel machines that have ferried people around for the past century, owners of fleets will want to know how expensive batteries can be most efficiently charged so they last as long as possible.

Hendriks, a research associate in the Battery Storage and Grid Integration Program at ANU, has been looking at the operations of the Leichhardt Bus Depot in Sydney, where 40 electric buses will be helping out on inner-city routes. He’s noticed some interesting patterns from the activities of 20 buses deployed so far.

Johannes Hendriks, a research associate in the Battery Storage and Grid Integration Program at ANU. Photo: Supplied.

Passenger loading has a significant impact on the operation of the air-conditioning, which uses a lot of battery energy. Passengers might account for up to 15 per cent of a vehicle’s mass, but the air-con will be running hard in a packed bus as the doors repeatedly open and close.

Also, the optimal operating temperature for a battery in the trial buses is around 10°C to 15°C, which is about standard for winter in Sydney. As you go hotter or colder, energy consumption picks up in a parabolic profile. In hot weather, as air-con units are running hard, any method to minimise drain on batteries is valuable.

“None of this is a problem to converting a large number of our buses to electric,” says Hendriks. “It comes down to making good planning decisions.”

Schedules may be slightly different for electric buses, for example, but it’s an adaptation that smart planners should be able to pull off.

With the data flowing from the trial, Hendriks aims to develop a model that can predict how much energy a bus would need to service a route, knowing the terrain, average speed, number of stops, anticipated passengers and weather. The output will be used to work out an optimal charging schedule.

“You need a way to factor [hot weather and hilly terrain] into your planning,” he says. “There is no point operating on good assumptions – you need to know what the worst case is.”

Operators toying with an investment in electric buses will use the model to understand which routes best suit the technology. Output from the model will also be used to work out what charging gear is needed, if an onsite battery is required, and whether a current grid connection is enough.

When you work on one intelligent solution, you are effectively working on many others, says Hendriks.

“It’s about the ability to take a large amount of data and build up a model of something that is complex and very difficult to write down a mathematical expression for, but instead to use the data in an algorithm and have a model that can be applied elsewhere,” he says.

The limits of machine learning are defined by the inputs, and as the quality and frequency of data increases so will the possibilities.

Weather watchers

“The biggest improvement for operational use of machine learning has been the price drop of cloud computing costs,” says Solcast chief technology officer and co-founder Nick Engerer. Affordable storage means the solar forecasting company can more easily create “instances” – customised configurations of computers with vast storage, memory and processing power.

Solcast downloads several terabytes of satellite imagery a month along with gigabytes of fresh files every 10 minutes, on average, to run algorithms that predict cloud positions based on the recent past. As updated imagery comes in, the prediction that looks best is used to make a new prediction.

“When you make that prediction, machine learning is happening,” says Engerer.

Solcast chief technology officer and co-founder Nick Engerer. Photo: Supplied.

As forecasting is refined, Engerer imagines a future “where we’re running our electricity system on the weather; the weather is our fuel”. In that scenario, the grid is balanced based on short-term weather forecasts enabled through machine learning, performed on the cloud.

“It becomes about wind energy, large-scale solar, rooftop solar, demand, temperature – all these factors come together,” he says.

Every wind plant and solar farm that is 30MW or larger has to send the Australian Energy Market Operator (AEMO) a five-minute projection of its instantaneous power so the operator can issue generators with supply instructions to match demand. However, five minutes can be a long time when the weather has a mind of its own. In South Australia, which has been known to rely totally on rooftop solar at around midday on some days, fast-forming cloud over Adelaide can require a sudden ramp up of gas and diesel generation.

In South Australia, Solcast is working with wind forecaster Weatherzone and Tesla’s demand forecasting arm on the ARENA-backed Gridded Renewables Nowcasting Demonstration project to develop high-quality forecasts. The forecasts will be assessed and used by AEMO, network operator SA Power Networks and generator operators.

“That’s a balance-of-system challenge we’re helping AEMO engage with,” says Engerer. “AEMO is doing some world-class work in integrating forecasts.”

A green brain

At the University of Technology Sydney (UTS), Adam Berry is working on forecasting residential and commercial loads. In a system supplied by unpredictable wind and solar, complemented with storage, any room for flexibility in the shape and seasonal nature of demand will be highly valued. Looking at where self-correcting intelligent systems may offer the most impact, he says the most likely near-term benefit is forecasting.

A coal power station is relatively easy to control. If you want more out of it, you shovel more coal into it. Wind and cloud cover – the variables that dictate output from renewables plants – cannot be dialled up or down. Improved near-term and longer-term forecasting of wind and irradiance will better enable the grid operator to respond to intermittency, with historical load patterns feeding into the decision-making process.

Adam Berry, deputy director of the University of Technology Sydney Data Science Institute. Photo: Supplied.

“This is where AI [artificial intelligence] is potentially beneficial,” says Berry. “Where historical behaviour in household electricity consumption can be married up with local environmental conditions. If you get both sides of that coin right – for generation and for load – the final piece of the puzzle for AI is: now what?”

This is where predictive smarts in software will deploy storage, say, to maximise lifespan, minimise variance in electricity demand and ensure capacity in the future, says Berry. Predictive maintenance scheduling is a long-overlooked opportunity for intelligent systems to be a leap into the future, where work can be timed to better reflect load rather than a rigid calendar strategy.

It’s hard to resist describing a utopian ideal of the energy system when contemplating the possibilities of a totally connected network of self-correcting devices. Instead, the utilisation of air-conditioners – a huge contributor to peak demand – is beyond AEMO’s control.

“It’s a problem because if we have demand response initiatives in place and someone pulls a switch which might turn off air-conditioners, we don’t really know how much that load is going to fall,” says Berry.

If you don’t know enough about today, how can you make decisions about tomorrow?

“If we don’t really know what infrastructure is there, how can we do some scenario modelling of what’s going to happen on a 40-degree day?” says Berry.

An intelligent, connected system that incorporates a virtual brain will be adept at timing cooling in buildings and homes, and the charging or discharging of electric vehicles and large-scale storage to complement renewables generation.

But what exactly is AI? Some people might smirk and say it’s any solution that hasn’t been developed yet – something that’s always on the horizon. Others see it as an umbrella term that can include iterative processes where decisions are improved as more data is fed into them, which can also be called machine learning. Using the analogy of the human brain, artificial intelligence can be expected to operate consistently with the way human intelligence operates, where connections are made between different things so, for example, a red stove element means don’t touch. That’s how we learn, always self-correcting so the optimal path or response is found, with a few mistakes along the way.

“Build a structure, feed it information, give it examples of what we want it to do and, over time, it learns the mapping from those examples to the outputs,” says Berry. “When new things come in that we haven’t seen before, we hope that mapping can be replicated.”

Dr Guandong Xu, a professor of data science at UTS, and director of the Smart Future Research Centre. Photo: Supplied.

It’s how we connect

Also at UTS, the Smart Future Research Centre is working on a Cooperative Research Centre project, with Federal Government funding, which aims to analyse solar farms so output, storage and dispatch can be optimised and maintenance scheduling refined based on calculated likelihood of failure.

“Because solar farms are often built in remote areas, maintenance can be time consuming and costly,” says Dr Guandong Xu, a professor of data science at UTS, and director of the Smart Future Research Centre. “Machine-learning time series-based forecasting is used to make predictions to try to reduce operation costs.”

The researchers have used data from operating solar plants to build a model used to make predictions of functionality and maintenance based on the physical parameters of the solar panel, plus external conditions including temperature and cloud cover. The multivariate model approximates a relationship between the inputs and output, says Dr Xu, who claims 80 to 85 per cent accuracy for the model so far. A scheduling model can be used to decide how much generation will be stored and how much will be dispatched.

“It’s heavily related to the price bidding strategy to help owners make better decisions,” he says of the model, which is still in its testing stage.

Owners can test the model against their maintenance scheduling benchmarks, using drone video surveillance and sensors. The Smart Future Research Centre is working in partnership with clean technology investment firm Providence Asset Group, which is developing a portfolio of community-scale solar farms in NSW.

Send this to a friend