Building characterization through smart meter data analytics: Determination of the most influential temporal and importance-in-prediction based features.

• Most effective temporal features from smart meter data for building characterization are found. • State-of-the-art feature selection methods and a proposed customized approach are utilized. • Importance-in-prediction based features that improve building classification are determined. • Number of utilized features is substantially reduced while the accuracy is also enhanced. The present paper aims at determining the most influential features to be extracted from smart meter data to facilitate machine learning-based classification of non-residential buildings. Smart meter-driven remote estimation of the chosen characteristics (the buildings' performance class, use type, and operation group) is significantly helpful in buildings' commissioning, benchmarking, and diagnostics applications. As the first step, state-of-the-art feature selection methods and a proposed customized approach are utilized for determining the most influential parameters in the pool of temporal features, proposed in a previous study. Next, importance-in-prediction based features, generated from an hour-ahead load prediction pipeline, that can improve the classification accuracy are proposed and added as additional input parameters. Finally, interpretations about some of the most influential features for different classification targets are provided. The obtained results demonstrate that, while aiming at estimating the buildings' use type, through performing feature selection and adding importance-in-prediction based features, the number of utilized features is reduced from 290 (initial pool of features proposed in a previous study) to 29, while also increasing the accuracy from 71% to 74%. Similarly, number of employed features for estimating the performance class is decreased from 224 to 17 and the achieved accuracy is improved from 56% to 62%. Finally, using only 6 selected features, compared to 287 features in the initial set, the obtained accuracy for the classification of operation group is increased from 98% to 100%. It is thus demonstrated that the proposed methodology, through selecting and utilizing notably fewer features, results in a notable simplification of the feature extraction procedures, improves the achieved accuracy, and facilitates providing interpretations about the reason behind the influence of some of the most important features. [ABSTRACT FROM AUTHOR]