LSTM Forecasts for Smart Home Electricity Usage
Rosemary E. Alden, Huangjie Gong, Cristinel Ababei, Dan M. Ionel
Abstract
With increasing of distributed energy resources deployment behind-the-meter and of the power system levels, more attention is being placed on electric load and generation forecasting or prediction for individual residences. While prediction with machine learning based approaches of aggregated power load, at the substation or community levels, has been relatively successful, the problem of prediction of power of individual houses remains a largely open problem. This problem is harder due to the increased variability and uncertainty in user consumption behavior, which make individual residence power traces be more erratic and less predictable. In this paper, we present an investigation of the effectiveness of long short-term memory (LSTM) models to predict individual house power. The investigation looks at hourly (24 h, 6 h, 1 h) and daily (7 days, 1 day) prediction horizons for four different recent datasets. We find that while LSTM models can potentially offer good prediction accuracy for 7 and 1 days ahead for some data sets, these models fail to provide satisfactory prediction accuracies for individual 24 h, 6 h, 1 h horizons.