Deep Learning Techniques for Demand Forecasting: Review and Future Research Opportunities

Introduction

Over the last few years, demand forecasting difficulties have piqued the interest of the research community. The studies such as predicting energy consumption for independent buildings (Bedi & Toshniwal, 2019), foreseeing the demands for individual products of a retail company or forecasting water demand of individual households (Xu et al., 2018) are the examples which shows the increase in the interest of research in demand forecasting domain. These forecasting problems are developed in a view to reduce the human time (Salinas et al., 2019). The applications of demand forecasting incudes transport services (Nejadettehad et al., 2020; Chandakas, 2020), natural gas supply (Hribar et al., 2018) and multiple products demand identification (Abbasimehr et al., 2020).

However, identifying these demands using physical models by extracting specific features is a difficult task since they exhibits lower accuracy than its data driven counter parts (Wen et al., 2020). Hence, numerous machine learning methods such as support vector machine (SVM), K-nearest neighbors (KNN), linear regression have been used to predict the future demands (Abbasimehr, 2020). Other than these machine learning techniques, deep learning techniques like recurrent Neural Networks (RNN) (Hribar et al., 2018), Long Short Term Memory (LSTM) (Kim et al, 2020; Bedi et al., 2019), Convolutional Neural Networks (CNN) (Chen et al., 2020 & Ke et al, 2018), Auto encoders (Zhang et al., 2020) or Gated Recurrent Units (GRU) (Wen et al., 2020) are also widely used in the literature. Recent research has focused on deep learning techniques that create a non-statistical relationship between input and output. The benefit of these deep learning techniques is that they do not require feature engineering to extract the relevant attributes of the dataset, instead they capture, memorise, and use inner patterns for predicting (Kantasa-ard et al., 2020). This advantage will eliminate the cost and time required for feature extraction.

A few review articles are found in the literatures which explain the necessity of deep learning techniques in demand forecasting problems. To do a proper analysis of these review papers, their features are analysed. It is found that only one literature review paper is available during the review period, which is restricted to the analysis of water demand forecasting (Ghalehkhondabi et al., 2017) using deep learning technique. Another limitation of the existing review paper is that, it performs an exhaustive analysis of all the soft computing techniques in the demand forecasting problem domain. Hence, the existing review paper has not been given a detailed picture of all the deep learning techniques which can be used for demand forecasting. A comparison with existing review and our work is given in Table 1. This raises a need for systematic review paper which performs an in depth analysis of all the deep learning techniques that can be used for demand forecasting which help the researchers to select an appropriate deep learning technique for their particular demand forecasting problem.