Residential Energy Use Prediction across different Time Scales with Advanced Machine Learning Techniques

As a significant part of total energy consumption, predictive modeling of residential energy use is critically important and highly desired. Previous efforts have proposed a number of statistical models for the prediction of residential energy consumption, while the accuracy and predictability of different models are still highly uncertain. In this study, we explore the effective temporal scale of residential energy use prediction, using the-state-of-the-art machine learning techniques: a fully connected Artificial Neutral Network (ANN) and a Recurrent Neural Network (RNN). For RNN modeling, the Long-Short-Term-Memory (LSTM) realization is employed. We find that ANN model in general has higher predictability than LSTM. Specifically, neither ANN nor LSTM is able to well predict high frequency fluctuation of residential energy use (~10 minutes) due to short-term random error. While, across a relatively longer time frame (from 24 hours to 48 hours), ANN model performs reasonably well and works much better than LSTM. From the perspective of dominating factors, room temperature and humidity are the most relevant ones to predict the building residential energy use. This work will facilitate the energy use prediction and decision-making within the framework of smart grid.