Artificial neural networks to predict energy performance and retrofit scenarios for any member of a building category: A novel approach. (1st January 2017)

• Record Type: Journal Article
• Authors: Ascione, Fabrizio; Bianco, Nicola; De Stasio, Claudio; Mauro, Gerardo Maria; Vanoli, Giuseppe Peter
• Is Part Of: Energy. Volume 118(2017)
• Page Start: 999
• Page End: 1017
• Publication Date: 2017-01-01
• Subjects: Building energy retrofit -- Building category -- Surrogate models -- Artificial neural networks -- Sensitivity analysis -- Office buildings; Power resources -- Periodicals; Power (Mechanics) -- Periodicals; Energy consumption -- Periodicals; 333.7905
• DOI: 10.1016/j.energy.2016.10.126
• Languages: English
• ISSNs: 0360-5442

Abstract

Abstract: How to predict building energy performance with low computational times and good reliability? The study answers this question by employing artificial neural networks (ANNs) to assess energy consumption and occupants' thermal comfort for any member of a building category. Two families of ANNs are generated: the first one addresses the existing building stock (as is), the second one addresses the renovated stock in presence of energy retrofit measures (ERMs). The ANNs are generated in MATLAB ® by using the outcomes of EnergyPlus simulations as targets for networks' training and testing. A preliminary 'Simulation-based Large-scale sensitivity/uncertainty Analysis of Building Energy performance' (SLABE) is conducted to optimize the ANNs' generation. It allows to identify the networks' inputs and to properly select the ERMs. The developed ANNs can replace standard building performance simulation tools, thereby producing a substantial reduction of computational efforts and times. This can allow a wide diffusion of rigorous approaches for retrofit design, which are currently hampered by the excessive computational burden. As case study, office buildings built in South Italy during 1920–1970 are investigated. Comparing the ANNs' predictions with EnergyPlus targets, the regression coefficient is between 0.960 and 0.995 and the average relative error is between 2.0% and 11%. Highlights: High computational costs hinder to assess energy behavior of whole building stocks. Artificial Neural Networks are used to perform it reliably with a novel approach. ANNs' generation is optimized by means of uncertainty and sensitivity analyses. A first family of ANNs predicts building energy consumption and thermal comfort. A second family of ANNs predicts energy retrofit scenarios.