Statistical learning versus deep learning: performance comparison for building energy prediction methods

In this paper, deep learning methods are compared with traditional statistical learning approaches for the purpose of accurately predicting the electrical energy consumption at the building level. Despite the fact that a wide range of machine learning methods have already been applied to energy prediction, deep learning methods certainly represent the state-of-the-art in artificial intelligence, and have been used with remarkable success in a wide range of applications. In particular, the use of Deep Belief Network (DBN), Multi Layer Perceptron and Artificial Neural Network methods are considered in this work. Furthermore, deep learning performance is compared with the most commonly used statistical learning methods, such as Support Vector Machines, Hidden Markov Models and Factored Hidden Markov Models. The analysis of the day-ahead and weekahead energy prediction demonstrates that different prediction methods present significantly different levels of accuracy, with the DBN offering the most consistent performance over various lookahead horizons and resolutions. The methods are validated with the Pecan Street large-scale dataset that comprises an interesting mix of consumer behaviors, electrical vehicles and photovoltaic generation.