Coupling convolutional neural networks with gated recurrent units to model illuminance distribution from light pipe systems

• Published in: Building and environment Vol. 237; p. 110276
• Main Authors: Ngarambe, Jack, Duhirwe, Patrick Nzivugira, Van Quang, Tran, Nzarigema, Jean d’Amour, Yun, Geun Young
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01-06-2023
• Subjects: Convolutional neural networks; Daylighting modelling; Gated recurrent units; Light pipe systems; Machine learning; Light pipe systems; Daylighting modelling; Convolutional neural networks; Gated recurrent units; Machine learning
• ISSN: 0360-1323; 1873-684X
• DOI: 10.1016/j.buildenv.2023.110276

Existing methods used to predict the performance of light pipe systems depend on empirical models based on specific conditions (e.g., climate, configuration). As a result, they are not readily applicable to conditions outside the initially considered boundaries. We propose machine learning (ML) models as an alternative to existing empirical models. 147 ML models were trained to predict illuminance distribution from a light pipe. Three ML algorithms were considered – convolutional neural networks (CNN), gated recurrent units (GRU) and an ensemble of CNN + GRU. The CNN + GRU model (R2 = 0.987) showed a higher predictive performance than the GRU model (R2 = 0.981). Additionally, the CNN + GRU model required less time to train and was significantly lighter than the GRU model; these attributes make the ensemble CNN + GRU model preferable during deployment. The developed models also offer better generalization capabilities than existing empirical models and are easily subjected to continued learning as more data becomes available.
•Developed 143 ML models to predict the performance of light pipe systems.•176 different configurations of light pipe systems are considered.•The developed models can accurately predict illuminance distribution (R2 > 0.90).•The ensemble CNN + GRU model outperforms and is lighter the GRU model.