Development and evaluation of data-driven models for electricity demand forecasting in Queensland, Australia

Development and evaluation of data-driven models for electricity demand forecasting in Queensland, Australia

Queensland (QLD) is the second largest state in Australia, with a growing demand for electricity, but existing studies appear to lack their ability to accurately model the consumer demand for electricity. In this Master of Science Research (MSCR) thesis, two kinds of hybrid forecasting models were developed by integrating the Extreme Learning Machines (ELM) with a Markov Chain Monte Carlo (MCMC) algorithm based bivariate copula model (ELM-MCMC) and also, a conditional bivariate copula model to probabilistically forecast the electricity demand (D). The study has incorporated statistically significant lagged electricity price (PR) datasets as a non-linear regression covariate into the final D-forecasting model.

In the first objective of the MSCR thesis, the ELM model was trained using statistically significant historical electricity demand at (t–1) timesteps for the state of Queensland used as a predictor variable, derived from Partial Autocorrelation Functions (PACF). This represented historical usage patterns in the electricity demand datasets used to forecast the future usage. It was then tested against current electricity demand (D(t)) to forecast the future D values. The output (i.e., simulated and observed tested D values) from the independent test dataset of the ELM model was used as the input for the MCMC-based copula model to derive the best copula model and to further improve forecasting accuracy. This involved the adoption of twenty-six copulas (e.g., Gaussian, t, Clayton, Gumble, Frank, etc.) and enabled us to also rank the best copulas based on the Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), and Maximum Likelihood (MaxL) to establish the dependence of historical D with the current and future D values. The results for the ELM-MCMC copulabased model outperformed both of its counterpart models (i.e. MCMC copula-based model and the standalone ELM model) based on vigorous statistical performance metrics. For 6 and 12-hours timescales, the MCMC-Fischer-Hinzmann copula yielded the highest Legates and McCabe Index (LM) (0.98 and 0.98), and lowest error terms including root mean square error (RMSE) (285.480 and 534.090), relative root mean square error (RRMSE) (0.348 and 0.320%), mean absolute error (MAE) (262.241 and 490.661 MW), relative mean absolute error (RMAE) (0.336 and 0.309 %), AIC (-63136.102 and -34727.466), BIC (-63125.530 and -34718.279), and MaxL ( 51570.051 and 17365.733), respectively. Similarly, for the daily timescale, the ELM-MCMC-Cuadras-Auge copula outclassed its counterpart models by displaying LM (0.98), MSE ( 482703.8 MW), RMSE (694.769 MW), RRMSE (0.220 %), MAE (638.365 MW), RMAE (0.208 %), AIC (-14514.312), BIC (-14510.412), and MaxL (7258.156).

These present results indicated that the hybrid ELM-MCMC copula-based model had an excellent performance, evidenced by attaining less than 10% RRMSE and RMAE, and Legates McCabe value close to unity. This is further supported by better model fits as denoted by lower AIC and BIC values and small residual error between observed and predicted data as indicated in higher MaxL values for the respective timescales.

In another phase of this study, we explored the ability of both local and global optimization techniques in achieving the best parameter estimate for the 26 copulas. It has shown that the global MCMC optimization method delivers accurate parameter estimates for 6 and 12-hours timescales whilst presenting information on the posterior distribution by computing uncertainty range of parameter values within a Bayesian framework. The local method appeared to provide better estimates of copula parameters for the daily timescale of D-forecasting.

In the second objective of the MSCR thesis, this study has developed a conditional bivariate copula model to probabilistically forecast electricity demand by incorporating the significant lagged electricity price (PR) from the Australian Energy Market Operator (AEMO) as a covariate into the final D-forecasting model. The use of energy price data to predict the energy demand is an important contribution given the relationships between these variables are well established. This objective resulted in the bivariate BB7 and BB8 copulas as being ranked highly for the probabilistic forecasting of D at a timescale of 30 minutes, 1-hour, and daily. The conditional exceedance probability of electricity demand greater than 7000 MW, 14000 MW, and 360000 MW for 30-minutes, 1-hour, and daily timescales given their respective prices greater than AU$25/MWh, AU$60/MWh, and AU$165/MWh predicted to be 20%, 30%, and 50% respectively. Similarly, the conditional non-exceedance probability of electricity demand greater than 7000 MW, 14000 MW, and 360000 MW for 30-minutes, 1-hour, and daily timescales given their respective prices greater than AU$25/MWh, AU$60/MWh, and AU$165/MWh was predicted to be 80%, 72%, and 70% respectively.

When benchmarked with literature, the proposed research methodologies for objective (i.e., projection of demand based on antecedent behaviour) and objective 2 (i.e., projection of demand based on antecedent energy price data) appear to be versatile tools possessing a robust predictive capability for forecasting D in Queensland, Australia. Hence, this research project is the first to develop and test these novel techniques, especially using price as regression covariate to forecast demand to achieve high forecasting accuracy, when the models are applied for multiple forecasting horizons of 30-minutes, 1-hour, 6-hourly, 12-hourly, and daily. It is noted that these timescales are relevant for stakeholders (e.g., energy utilities) to develop decision systems for better energy security, and can potentially be adopted in real power grid operations to ensure stability, cost reduction and improved efficiency whilst granting consumer satisfaction.

In summary, the novel energy demand modelling techniques presented here can help address research gaps in electricity usage monitoring sector by making a significant contribution towards improved forecasting accuracy of energy demand. While this study has currently been limited to Queensland, the research findings are immensely useful for energy experts in the National Energy Markets elsewhere including supporting the work of AEMO, Energex and other companies to enhance their energy forecasting and monitoring skills. These can assist in informed decisions and addressing the growing challenges within electricity
industry, through improving energy demand and price monitoring, consumer satisfaction and maximized profitability endeavours of energy companies.