Short-term electrical load demand forecasting with deep learning techniques

Abstract

With the advent of smart grid systems enabling efficient allocation of electrical power, the topic of short-term electrical load demand forecasting has gained attention in academic literature. However, despite crucial findings in this area, the topic of forecasting electrical load 30-minutes ahead has seldom been discussed, with extant research focusing on one-day ahead forecasting. Considering the advancements in smart grid technologies to adapt to forecasts shorter than one-day ahead of time, this study focuses on electrical load demand forecasting for 30-minutes and one-day ahead of time. To achieve this, deep learning techniques were used on electrical load datasets from 2013-2015 for each state of Australia- Queensland (QLD), New South Wales (NSW), South Australia (SA), Tasmania (TAS) and Victoria (VIC), obtained from the Australian Energy Market Operator (AEMO). Forecasts were made using Convolutional Neural Networks (CNNs), Vanilla, Stacked and Bidirectional Long Short-Term Memory Networks (LSTMs), Gated Recurrent Units (GRUs) and an ensemble method composed of Maximal Overlap Discrete Wavelet Transform (MODWT), Empirical Mode Decomposition (EMD) and Random Vector Functional Link (RVFL) Models. Results indicate the strong potential for forecasts made 30-minutes ahead of time and the consistency of forecasts made one-day ahead with extant research, with forecasting performance improved by the use of ensemble methods. Implications of the research in this study and future directions are discussed.