Development of data intelligent models for electricity demand forecasting: case studies in the state of Queensland, Australia

Development of data intelligent models for electricity demand forecasting: case studies in the state of Queensland, Australia

Electricity demand (G) forecasting is a sustainability management and evaluation task for all energy industries, required to implement effective energy security measures and determine forward planning processes in electricity production and management of consumer demands. Predictive models for G forecasting are utilized as scientific stratagems for such decision-making. The information generated from forecast models can be used to provide the right decisions regarding the operation of National Electricity Markets (NEMs) through a more sustainable electricity pricing system, energy policy, and an evaluation of the feasibility of future energy distribution networks. Data intelligent models are considered as potential forecasting tools, although challenges related to issues of non-stationarity, periodicity, trends, stochastic behaviours in G data and selecting the most relevant model inputs remain a key challenge.

This doctoral thesis presents a novel study on the development of G forecasting models implemented at multiple lead-time forecast horizons utilizing data-intelligent techniques. The study develops predictive models using real G data from Queensland (second largest State in Australia) where the electricity demand continues to elevate. This research is therefore, divided into four primary objectives designed to produce a G forecasting system with data-intelligent models.

In first objective, the development and evaluation of a multivariate adaptive regression splines (MARS), support vector regression (SVR) and autoregressive integrated moving average (ARIMA) model was presented for short-term (30 minutes, hourly and daily) forecasting using Queensland’s aggregated G data. MARS outperformed SVR and ARIMA models at 30-minute and hourly horizon, while SVR was the best model for daily G forecasting.

The second objective reported the successful design of SVR model for daily period, including short-term periods (e.g., weekends, working days, and public holidays), and the long-term (monthly) period. Subsequently, the hybrid SVR, with particle swarm optimization (i.e., PSO-SVR) integrated with improved empirical mode decomposition with adaptive noise (ICEEMDAN) tool was constructed where PSO is adopted to optimize SVR parameters and ICEEMDAN was adopted to address non-linearity and non-stationary in G data. The capability of ICEEMDAN-PSO-SVR to forecast G was benchmarked against ICEEMDAN-MARS and ICEEMDAN-M5 Tree, including traditional PSO-SVR, MARS and M5 model tree methods.

As G is subjected to the influence of exogenous factors (e.g., climate variables), the third objective established a G forecasting model utilizing atmospheric inputs from the Scientific Information for Land Owners (SILO) observed data fields and the European Centre for Medium Range Weather Forecasting outputs. These models were developed using G extracted from the Energex database for eight stations in southeast Queensland for an artificial neural network (ANN) model over 6-hourly and daily forecast horizons.

The final objective was to advance the methods in previous objectives, by applying wavelet transformation (WT) as a decomposition tool to model daily G. Using real data from the University of Sothern Queensland (Toowoomba, Ipswich, and Springfield), the maximum overlap discrete wavelet transform (MODWT) was adopted to construct the MODWT-PACF-online sequential extreme learning machine (OS-ELM) model. The results revealed that newly developed MODWT-PACF-OSELM (MPOE) model attained superior performance compared to the models without the WT algorithm.

In synopsis, the predictive models developed in this doctoral thesis will to provide significant benefits to National Electricity Markets in respect to energy distribution and security, through new and improved energy demand forecasting tools. Energy forecasters can therefore adopt these novel methods, to address the issues of nonlinearity and non-stationary in energy usage whilst constructing a real-time forecasting system tailored for energy industries, consumers, governments and other stakeholders.