Energy Forecasting in Buildings Using Deep Neural Networks

The building sector is responsible for the largest portion of the total energy consumption in the United States. Conventional physics-based building energy models (BEMs) consider all of the building characteristics in order to accurately simulate their energy usage, requiring an extensive, complex, and costly process, particularly for existing buildings. In recent years, data-driven models have emerged as an additional path toward the prediction of energy consumption in buildings. The purpose of this work is to present a methodology for predicting the energy consumption of buildings using deep neural networks (NNs). Three machine learning algorithms, including a linear regression model, a multilayer perceptron NN, and a convolutional NN (CNN) model, are proposed to solve an energy consumption regression problem using outside dry-bulb temperature as the only input. To assess these methods, a building in Melbourne, FL, is considered and modeled in EnergyPlus. Ten years of data were used as inputs to the EnergyPlus model, and the energy consumption was calculated accordingly. The input to the machine learning algorithm (average daily dry-bulb temperature) and the output (daily total energy consumption) are used for training. Cross-validation was performed on the trained model using actual weather data measured onsite at the building location. The results showed that all three proposed machine learning algorithms were trained successfully and were able to solve the regression problem with high accuracy. However, the CNN model provided the best results when compared with the other two methods. This work also investigates different data filtering techniques that provide the best positive correlation between inputs and outputs for a similar type of problem. Results from this work aim to be used toward accurate energy forecasting that facilitates achieving higher energy efficiency in the building sector. The presented framework provides a readily simple model that allows accurate prediction of outputs when supplied with new inputs and can be used by a wide range of end users.