On the use of artificial neural networks to model household energy consumptions. (September 2017)

• Record Type: Journal Article
• Authors: Martellotta, F.; Ayr, U.; Stefanizzi, P.; Sacchetti, A.; Riganti, G.
• Is Part Of: Energy procedia. Volume 126(2017)
• Page Start: 250
• Page End: 257
• Publication Date: 2017-09
• Subjects: Artificial Neural Network (ANN) -- Energy consumptions prediction -- Smart houses -- Energy saving; Power resources -- Congresses; Power resources -- Periodicals; Power resources; Conference proceedings; Periodicals; 333.7905
• DOI: 10.1016/j.egypro.2017.08.149
• Languages: English
• ISSNs: 1876-6102

Abstract

Abstract: Modern houses are more and more frequently characterized by the presence of "smart" metering devices, capable of measuring air temperature, relative humidity, air quality, and in the more sophisticated cases, even electric equipment consumptions. In addition, other relevant parameters such as illuminance may often be determined and they can be used as proxy variables to account for other important aspects (such as solar irradiance) influencing the energy balance of a building. Such information, in combination with weather data which can be retrieved by other sources (or by additional sensors), may conveniently contribute to the creation of a "black box" model in which, given a few input variables it is possible to output a variable which would result from otherwise complex calculations (e.g. an energy balance) requiring many data. The availability of such a "black box" could be helpful under many points of view, such as benchmarking energy consumptions and stimulating virtuous behavior from the occupants. To test whether such approach can be feasible, an EnergyPlus model of a real house was made, trying to accurately reproduce building features, systems set-points, and occupant behaviors. The overall simulated energy consumptions were compared with the real ones resulting from energy bills, thus ensuring a good agreement with reality. The dataset resulting from EnergyPlus was then used to train an artificial neural network (ANN) capable of yielding hourly energy consumptions based on limited input data. Finally, the relative importance of the different input variables was analyzed to understand which might influence prediction accuracy most.