Deep learning versus traditional machine learning methods for aggregated energy demand prediction

In this paper the more advanced, in comparison with traditional machine learning approaches, deep learning methods are explored with the purpose of accurately predicting the aggregated energy consumption. Despite the fact that a wide range of machine learning methods have been applied to probabilistic energy prediction, the deep learning ones certainly represent the state-of-the-art artificial intelligence methods with remarkable success in a spectrum of practical applications. In particular, the use of Multi Layer Perceptrons, recently enhanced with deep learning capabilities, is proposed. Furthermore, its performance is compared with the most commonly used machine learning methods, such as Support Vector Machines, Gaussian Processes, Regression Trees, Ensemble Boosting and Linear Regression. The analysis of the day-ahead energy prediction demonstrates that different prediction methods present significantly different levels of accuracy in the case of a challenging dataset that comprises an interesting mix of consumers, wind and solar generation. The results show that Multi Layer Perceptrons outperform all the eight methods used as a benchmark in this study.

I. Introduction

Many approaches aiming to accurately and robustly predict the energy consumption have been proposed. In general, at the building level, two types of approaches are discerned. Models of the first type are based on physical principles to express their thermal dynamics and energy behavior. Depending on the type of building and the number of the considered parameters, these models might include space heating systems, natural ventilation, air conditioning systems, passive solar heating, photovoltaic systems, financial considerations, occupants' behavioral characteristics, etc. The second type of approaches is based on statistical methods. Such methods are used to predict energy consumption by correlating it with influencing variables related to weather and energy costs. Interested readers are referred to [1] and [2] for a more comprehensive discussion on the applications of artificial intelligence in building energy systems, and the more recent reviews [3], [4]. Moreover, to account for the evolution of future building energy systems, hybrid approaches which combine some of the aforementioned models to optimize predictive performance have been developed [5]–[8].