Characterising Renewable Generation in the Grid
Conference
Format: IPS Abstract
Keywords: energy, forecasting, time series
Session: Invited Session 2A - Tools for the Energy Transition
Monday 2 December 1:30 p.m. - 3 p.m. (Australia/Adelaide)
Abstract
The transition to 100% renewable based electricity is expected to make use of a diverse range of complementary tools for characterising and forecasting energy data. Our investigations focus on using methods from the AI and Machine Learning domains such as Long Short-Term Memory (LSTM) models, Autoencoders and Transformers, due to the complex interplay and intricate nature of instantaneously matching generation to the load in a market-based energy grid such as the Australian National Energy Market (NEM). We use LSTM models to extract long term, complex sequential patterns from the temporal data of this dynamic system, Transformers to work in a spatial-temporal domain capturing dependencies and global patterns withing the sequential data, and Autoencoders to reduce the high-dimensional data into lower dimension latent spaces extracting the essential features at the same time as reducing noise and redundancy. This modelling aids in monitoring the performance of the system and to identify and assess the impact of deviations in actual vs. forecasted generation and load.
This presentation will detail the current state of our investigations into the modelling, comparing results for characterising both the energy and econometric variables such as energy price used in characterising the supply of renewable energy and possible applications of the methods in effectively delivering reliable energy.