Energy Demand Analysis and Forecasting
Australian Energy Market Operator
Australian Energy Marker Operator

Background

Understanding the way Australians consume energy is becoming increasingly complex. Our team aims to develop machine learning algorithms to provide meaningful and accessible information that will unlock the mysteries of our energy behaviour.

This project integrates and enhances information about Australia's energy use, including energy consumption patterns, demographics, building characteristics, battery usage identification, weather sensitivity patterns, and more. The key objective of this work is to better understand the complex nature of Australian consumer's behavior and provide comprehensive information to underpin key decision making, helping to deliver a secure, affordable and sustainable energy future for Australia.

Breakdown of residential energy consumption by end use
Residential energy consumption for households with and without gas services

Major Challenges

Forecasting residential energy usage is critical for informing network infrastructure spending, market decision making and policy development. With concerns not only over the total volume of consumption within a time period, but also the maximum ‘peak’ simultaneous consumption, this becomes a complicated topic to explore. Interval meters in Australia typically record electricity usage in half hour intervals. This provides a rich source of data to develop forecasting models. However, it is estimated that only about one quarter of Australia's near 14 million electricity meters are smart meters. The other three quarters are basic accumulation meters, which simply accumulate the amount of consumed energy and are typically read once every quarter. When it comes to gas metering, almost all meters are basic meters. This low penetration rate of interval meters in electricity and gas metering hinders the industry's ability to forecast accurately.

  1. Successfully forecast energy requirements in our changing world.
  2. Design effective machine learning models that deliver value to Australian households and businesses.

Our Response

To address these challenges, our research aims to develop a method which can be used to successfully predict residential energy consumption for individual households using readings from basic meters (and interval meters) and other available information such as weather or demographics.

  • Short-term demand forecasting: using smart metering and weather data, develop machine learning models that provide half an hour ahead, hour-ahead, day ahead forecasts for individual households as well as aggregate demand for a region as well as.
  • Long-term forecasting: using smart metering and weather data to forecast gas consumption and identify households with gas usage. It will help to paint a clearer picture of Australian residential energy use.
Electricity demand forecasting using weighted support vector regression
Electricity demand forecasting using weighted support vector regression. Predictors = day, time, weather, previous consumptions. Peak samples were penalized when modelling improved the peak forecasts.