Application and characterization of metamodels based on artificial neural networks for building performance simulation: A systematic review

In most of the countries, buildings are often one of the major energy consumers, leading to the necessity of achieving sustainable building designs, and to the mandatory use of building performance simulation (BPS) tools in order to retrofit or design new energy efficient buildings. In the last years, the use of artificial neural networks (ANNs) metamodels has increased and gained confidence in BPS applications thanks to their favorable trade-off between accuracy and computational cost. This paper presents a comprehensive and in-depth systematic review of the up-to-date literature related to the application and characterization of ANN-based metamodels for BPS. First, a general insight into the methodology of metamodel generation and ANN theory is presented. The ANN metamodels are classified according to the type of building they are addressed to, screening them by their inputs (building design variables or indicators to take a certain decision) and outputs (energy consumption, comfort index, climatic condition, environment performance). Then, all the stages for the generation of ANN-based metamodels (sampling methods, data pre-processing, architectures, activations functions, the process of training and testing, and the platforms and frameworks for their implementation) are presented giving a brief theoretical introduction and making a critical review of the literature linked to each stage. For each of these analyzed stages, summary tables and graphs are presented showing the distributions of different alternatives and trends. Finally, the current limitations and areas for further investigation are discussed. (C) 2020 Elsevier B.V. All rights reserved.