A machine learning-based surrogate model to approximate optimal building retrofit solutions. (1st January 2021)

• Record Type: Journal Article
• Authors: Thrampoulidis, Emmanouil; Mavromatidis, Georgios; Lucchi, Aurelien; Orehounig, Kristina
• Is Part Of: Applied energy. Volume 281(2021)
• Publication Date: 2021-01-01
• Subjects: Building retrofit -- Energy efficiency -- Surrogate model -- Machine learning -- Multi-objective optimization -- Pareto-optimal; Power (Mechanics) -- Periodicals; Energy conservation -- Periodicals; Energy conversion -- Periodicals; 621.042
• DOI: 10.1016/j.apenergy.2020.116024
• Languages: English
• ISSNs: 0306-2619

Abstract

Highlights: Machine learning-based surrogate model to predict near-optimal retrofit solutions. Validated with a conventional building simulation-optimization model. Case study reveals good accuracy, ease of application and computational efficiency. Convenient for non-expert decision makers due to small set of input data. The model is scalable and applicable for retrofit analyses in wide-areas. Abstract: The building sector has the highest share of operational energy consumption and greenhouse gas emissions among all sectors. Environmental targets set by many countries impose the need to improve the environmental footprint of the existing building stock. Building retrofit is considered one of the most promising solutions towards this direction. In this paper, a surrogate model for evaluating the necessary building envelope and energy system measures for building retrofit is presented. Artificial neural networks are exploited to build up this model in order to provide a good balance between accuracy and computational cost. The proposed model is trained and tested for the case study of the city of Zurich, in Switzerland, and is compared with one of the most advanced models for building retrofit that uses building simulation and optimization tools. The surrogate model operates on a smaller input set and the time required to derive retrofit solutions is reduced from 3.5 min to 16.4 μsec. Results show that the proposed model can provide significantly reduced computational cost without compromising accuracy for most of the retrofit dimensions. For instance, the retrofit costs and the energy system selections are approximated with an average accuracy of R 2 = 0.9408 and f 1 s c o r e = 0.9450, respectively. Finally, yet importantly, such surrogate retrofit models may effectively be used for bottom-up retrofit analyses for wide areas and can contribute towards accelerating the adoption of retrofit measures.