Input variable selection for thermal load predictive models of commercial buildings

• Published in: Energy and buildings Vol. 137; pp. 13 - 26
• Main Authors: Kapetanakis, Dimitrios-Stavros, Mangina, Eleni, Finn, Donal P.
• Format: Journal Article
• Language: English
• Published: Lausanne Elsevier B.V 15.02.2017; Elsevier BV
• Subjects: Building thermal loads; Buildings; Climate; Climatology; Commercial buildings; Commercial real estate; Computer simulation; Data analysis; Data processing; Datasets; Energy consumption; Energy management; Forecasting; Historical buildings; Input selection; Performance prediction; Prediction models; Predictive model; Simulation; Thermal analysis; Weather; Input selection; Building thermal loads; Predictive model
• ISSN: 0378-7788; 1872-6178
• DOI: 10.1016/j.enbuild.2016.12.016

•Selection of input variables performing linear and monotonic correlation analysis.•Accuracy of predictive models maintained at the same level with selected variables.•Complexity of predictive models reduced with selected variables as inputs. Forecasting of commercial building thermal loads can be achieved using data from Building Energy Management (BEM) systems. Experience in building load prediction using historical data has shown that data analysis is a key factor in order to produce accurate results. This paper examines the selection of appropriate input variables, for data-driven predictive models, from wider datasets obtained from BEM systems sensors, as well as from weather data. To address the lack of available complete datasets from actual commercial buildings BEM systems, detailed representation of reference buildings using EnergyPlus were implemented. Different types of commercial buildings in various climates are examined to investigate the existence of patterns in the selection of input variables. Data analysis of the simulated results is used to detect the correlation between thermal loads and possible input variables. The selection process is validated by comparing the performance of predictive models when the full or the pre-selected set of variables is introduced as inputs.