The Problem

This study was to determine if Artificial Intelligence and Machine Learning can improve management of the network by forecasting supply for time periods of up to 72hrs in advance for network operators.

The major difference between this pilot project and some popular energy market forecasting platform e.g. AEMO was looking at the data from generation (supply) perspective and not the consumption (demand).

 

The Solution

VROC used two popular ML algorithms with the goal to forecast the network extreme (min and max) total megawatts generated as the output parameters. VROC trained the models on 2 years of historical major substations total megawatts PV generation as well as weather data.

power lines

The Outcome

VROC's forecasting performance based on Mean Absolute Error was exceptional with a 5.5% MAE average (acceptable <8% MAE error threshold defined by the client and industry). VROC was able to automatically forecast results 7 days ahead with the data provided to us.
However, we recommend additional modelling with abnormal conditions (heat waves and extreme cloud cover) to improve upon the existing model.

Accurate forecasting of the energy marker supply and demand is crucial, with National Electricity Market (NEM) supplying about 200 terawatt hours of electricity to businesses and households each year. Improving the electricity supply and demand forecasting accuracy by only 1% would result in an improvement of over-supply or undersupplying electricity to consumers with an estimated 800M AUD annual savings for both the government and energy businesses.