Are deep learning models more effective against traditional models for load demand forecasting?

Consumer level load demand forecasting has caught an eye of many researchers since the data from smart meters has become available with high temporal resolutions. Long-sort-term-memory (LSTM) based deep learning models have gained significant popularity in time series forecasting tasks. While many such models have been proposed in the literature for the job, its difficult to analyze and compare the efficacy of these models due to two main reasons. Firstly, the lack of explainability of these models; and secondly, the proposed models are generally tested on very different datasets. Hence, this paper aims to analyze two of the most commonly used models, i.e. standard LSTM, and Bi-directional LSTM (Bi-LSTM), on a publicly available dataset in order to provide common ground for valid comparison. The paper considers standard baselines, statistical and machine learning models where it notes that random forest can perform better or at par with the considered LSTM-based models. This is the result of efficient data curation and pre-processing techniques that were employed using the information from heuristics.