C Gouy-Pailler, H Najmeddine, A Mouraud, F Suard, C Spitz, A Jay, P Maréchal
DISTANCE AND SIMILARITY MEASURES FOR SENSORSSELECTION IN HEAVILY INSTRUMENTED BUILDINGS:APPLICATION TO THE INCAS PLATFORM
Abstract: Energy management in residential buildings is taking an increasing role in the construction workflows.It entails understanding thermal processes at stake in the buildings and quantifying energyconsumption, which meets inhabitants comfort requirements. Experimental platforms such as INCASaim at providing experts with a practical way to study such problems in real conditions. These heavilyequipped buildings yield huge amounts of real-time data (sampling rates, number and types of sensors)for which new automatic approaches could be useful to thermal experts. Generic similarity measuresfrom data-mining could therefore provide comprehensive analysis tools to thermal experts. This paper focuses on the ability of some distance and similarity measures to organize millions ofdata from homogeneous and heterogeneous sensors into coherent clusters. Simplifying datainterpretations to thermal experts in highly equipped buildings, this approach could also stand as abasis for studying smart grids of less equipped domestic houses studies. Two types of similarity measures are explored. The first one consists of a set of three distances,and accounts for differences in terms of amplitude scaling and shifting between pairs ofmeasurements. It relies on the comparison of homogeneous sensors by quantifying the relativeproximity of their amplitude in terms of mean value, variance and time shift. The second type ofsimilarity measure employs a pre-processing step transforming continuous signals into binary events.The resulting spike trains are then compared by quantifying the amount of unitary transformations(events moves or events deletions/additions) needed to align pairs of events sequences. These proximity measures are computed on real data from experimental buildings of the INCASplatform. It comprises three experimental buildings (with different construction types) dedicated totesting various approaches regarding systems, control and energy-saving policies. These geometricallyidentical buildings are equipped with hundreds of sensors measuring temperature, humidity,differential pressure, and others data at various positions of the structures with sampling rates of onemeasurement per minute. Simulation-based temperatures are integrated in the sensors set providing acomparison between real and simulated data. Results illustrate the contribution of the applied methods when dealing with large amounts ofmeasurements related to instrumented buildings behaviors. Actually results show that coherent clustersregarding distinct signal properties are automatically generated. These clusters can be used fordimensionality reduction (clusters of sensors could be summarized by a single virtual measurement),or relative comparisons between sensors or between real and simulated datasets.
Keywords: INCAS, low-energy consumption, sensor selection, multivariate data mining
Full text: content.pdf (803,535 bytes) (available to registered users only)