Evaluation of the causes and impact of outliers on residential building energy use prediction using inverse modeling

Inverse modeling techniques are often used to predict the performance and energy use of buildings. Residential energy use is generally highly dependent on occupant behavior; this can limit a model's accuracy due to the presence of outliers. There has been limited data available to determine the cause of and evaluate the impact of such outliers on model performance, and thus limited guidance on how best to address this in model development. Thus the main objective of this work is to link the use of outlier detection methods to the causes of anomalies in energy use data, and to the determination of whether or not to remove an identified outlier to improve an inverse model's performance. A dataset of 128 U.S. residential buildings with highly-granular, disaggregated energy data is investigated. Using monthly data, change-point modeling was determined to be the best method to predict consumption. Three methods then are used to identify outliers in the data, and the cause and impact of these outliers is evaluated. Approximately 19% of the homes had an outlier. Using the disaggregate data, the causes were found to mostly be due to variations in occupant-dependent use of large appliances, lighting, and electronics. In 20% of homes with outliers, the removal of the outlier improved model performance, in particular all outliers identified with both the standard deviation and quartile methods, or all three methods. These two combinations of outlier detection methods are thus recommended for use in improving the prediction capabilities of inverse change point models.