Energy prices often increase as the temperature fluctuates, but did you know that homes further inland tend to have higher levels of use regardless of the season? Or that households with higher incomes use more gas? And that when the temperature rises above 13 degrees, gas consumption drops to some of its lowest levels?
These insights are crucial when planning household energy distribution, pricing and infrastructure, and avoiding issues such as power surges and black outs.
However, with the majority of Australian households not equipped with smart metres that measure electricity use, and barely any that record gas usage, obtaining the required amount of data to make accurate inferences is challenging.
By applying a machine learning technique, CSIRO’s Energy and Data61 researchers have developed an energy forecasting model that uses predictive analytics and time series data modelling to accurately hypothesise insights from missing data, a methodology that has outperformed current state-of-the art energy forecasting models.
Using 2014 to 2015 gas consumption datasets provided by Australian Energy Regulator, Data61 researchers were able to predict the average monthly gas usage of a Victorian household during 2016, with an accuracy of over 90%.
“Rather than providing a single estimate for energy usage, which is currently the practice, this new method provides a range of estimates,” explains research lead, Dr Dilusha Weeraddana.
“This allows us to consider a range of possible scenarios, enabling energy authorities to develop flexible and agile planning.”
In Australia, interval or smart metres typically record electrical usage in half hour intervals, offering a rich source of data to develop forecasting models.
This is in contrast to the majority of meters used throughout the country - basic or accumulator metres - which continuously add up a dwelling’s usage until it’s read every month or quarter, making the set of critical data points 4,000 times smaller per quarterly read than that of smart metres. Almost all gas metres are the basic kind.
Because of this, providers face increased difficulties in forecasting energy usage accurately.
But by applying a machine learning algorithm based on a probability theory known as Gaussian process, this new method can effectively capture the unique characteristics of each household and compare this data to that of other homes in the area to identify trends.
This novel solution also integrates seasonal weather, building characteristics and area-specific demographics data to produce a holistic analysis of once complex and unattainable information produced by millions of unique households.
“Any tool, technique, or data that improves the ability to estimate the amount of energy required goes a long way in making sure these aspects of the networks are accurately predicted, and therefore well managed,” explains Lachlan O’Neil, team leader of Energy Data Sciences at CSIRO.
“This research enables a more detailed understanding of our nation’s energy use, especially when there is missing information, and can be used to better inform decision makers in the energy sector and the Australian Government on the future, sustainability and affordability of Australian's energy consumption and billing.”
The application of data and technology science is crucial to developing innovative and effective solutions to some of the nation’s greatest challenges, with machine learning set to play a key role in delivering an energy efficient and cost-effective future to all Australians.
This project has been undertaken as part of the NEAR (National Energy Analytics Research) program (a collaboration between CSIRO, the Australian Government and the Australian Energy Market Operator), real-world data provided by Australian Energy Regulator (AER) was employed in this research.