摘要:
Understanding the underlying dynamics of building energy consumption is the very first step towards energy saving in building sector; as a powerful tool for knowledge discovery, data mining is being applied to this domain more and more frequently. However, most of previous researchers focus on model development during the pipeline of data mining, with feature engineering simply being overlooked. To fill this gap, three different feature engineering approaches, namely exploratory data analysis (EDA) as a feature visualization method, random forest (RF) as a feature selection method and principal component analysis (PCA) as a feature extraction method, are investigated in the paper. These feature engineering methods are tested with a building energy consumption dataset with 124 features, which describe the building physics, weather condition, and occupant behavior. The 124 features are analyzed and ranked in this paper. It is found that although feature importance depends on specific machine learning model, yet certain features will always dominate the feature space. The outcome of this study favors the usage of effective yet computationally cheap feature engineering methods such as EDA; for other building energy data mining problems, the method proposed in this study still holds important implications since it provides a starting point where efficient feature engineering and machine learning models could be further developed.
Website
