Gas consumption demand forecasting with empirical wavelet transform based machine learning model: A case study

Abstract

Dispatchable, reliable, and clean energy is essential for sustained economic growth and a better future. This study develops a novel technique of empirical wavelet transform (EWT) to reveal intrinsic patterns in daily gas consumption demand data and ultimately forecast the daily gas consumption demand patterns for the city of Melbourne, Australia, via a hybrid decision tree (M5 model tree) model. EWT algorithm decomposes the explored dataset into the respective intrinsic mode functions (IMFs) and a residual component representing historical demand behavior overcoming non-linearity and non-stationary issues. These sub-series components reflect the frequency types, stochastic behaviors, and several other relevant patterns that are embedded into the model's input variable matrix representing historical changes in demand to forecast future gas consumption demand. An evaluation comparison was conducted between the hybrid EWT-M5 model tree algorithm, the hybrid EWT-multivariate adaptive regression spline (MARS) and the hybrid EWT-autoregressive integrated moving average (ARIMA) models, including the traditional models based on M5 tree, MARS, and ARIMA algorithms. The results show that EWT-models improved significantly compared to traditionally adopted (ie, the non-EWT) methods. The performance metrics showed that the EWT-M5 model tree generated a RRMSE of 29.19% compared to 33.29% for EWT-MARS, 47.76% for EWT-ARIMA, 71.77% M5 model tree, 45.47% for MARS, and 65.25% for ARIMA. Accordingly, the results show the potential utility of EWT modelling approach in assisting the monitoring of energy usage and decision-making tasks for national energy markets.