Cluster analysis and artificial neural networks in predicting energy efficiency of public buildings as a cost-saving approach

Abstract

Although energy efficiency is a hot topic in the context of global climate change, in the European Union directives and in national energy policies, methodology for estimating energy efficiency still relies on standard techniques defined by experts in the field. Recent research shows a potential of machine learning methods that can produce models to assess energy efficiency based on available previous data. In this paper, we analyse a real dataset of public buildings in Croatia, extract their most important features based on the correlation analysis and chi-square tests, cluster the buildings based on three selected features, and create a prediction model of energy efficiency for each cluster of buildings using the artificial neural network (ANN) methodology. The main objective of this research was to investigate whether a clustering procedure improves the accuracy of a neural network prediction model or not. For that purpose, the symmetric mean average percentage error (SMAPE) was used to compare the accuracy of the initial prediction model obtained on the whole dataset and the separate models obtained on each cluster. The results show that the clustering procedure has not increased the prediction accuracy of the models. Those preliminary findings can be used to set goals for future research, which can be focused on estimating clusters using more features, conducted more extensive variable reduction, and testing more machine learning algorithms to obtain more accurate models which will enable reducing costs in the public sector.

Machine Learning Techniques Focusing on the Energy Performance of Buildings: A Dimensions and Methods Analysis

“…Regarding the first question (RQ1), 13 studies used the EPC dataset [30,[32][33][34][35]52,58] as explained in Sections 4.4 and 4.5. This kind of data is multi-dimensional, given that each energy certificate has many attributes.…”

“…As for data classification and clustering techniques, most studies applied the K-means clustering algorithm to characterize the cluster sets with given energy performance, as explained in Sections 4.4 and 4.5. Some studies used a density-based spatial clustering of application with noise algorithm (DBSCAN) to handle outliers and correlation analysis to identify the best input demission for their clustering analysis [32][33][34]52]. A few studies referring to RQ1 used GIS and geospatial maps to visualize their clustering results [30,58].…”

“…Some of them combine different types of buildings. Five papers combined residential and commercial buildings-offices and schools [25,32,44,52,58], two papers addressed offices [31,49] and six analyzed schools and universities [27,38,40,50,51,59]. Only one addressed a hospital [23] and one a care home [48].…”

“…Furthermore, only the most recent papers utilized building EPC data for their analysis [28,29,31,34,[50][51][52][53][54]56,57,60,61]. This aspect is surprising since the first directive on building energy performance, "the Energy Performance of Building Directive (EPBD)," was introduced by the European Parliament in 2002.…”

“…The most-used techniques for predicting EP and retrofitting were energy performance simulation techniques, statistical-based approaches, genetic algorithms, and ANN. Few studies use only ML methods, namely ( 13) studies [28,29,31,[33][34][35]40,45,48,50,52,56,59]. The most common clustering and classification techniques were K-means (7), statistical methods (6), Latin hypercube sampling (2), other manual groupings (2), decision tree (2), and probability density function (1) [23,25,[28][29][30][31][32][33][34]47,[50][51][52]54,55,57,59].…”