A comprehensive review of machine learning techniques for optimizing daylight in high-performance buildings

This paper presents a comprehensive review of the current research on the application of machine learning (ML) techniques to optimize daylighting performance in buildings, encompassing both the early design stage and the operational phase. The authors systematically analyze 50 recent studies in this domain, providing valuable insights into the state-of-the-art and future research directions.

The review reveals that the majority of the existing work has employed supervised artificial neural network (ANN) models, often trained on data generated from detailed daylight simulations. These ML-based frameworks have demonstrated exceptional predictive accuracy in forecasting daylighting and visual comfort metrics, including useful daylight illuminance (UDI), spatial daylight autonomy (sDA), and visual discomfort indices. By integrating parametric building modeling with GPU-accelerated daylight simulations, researchers have been able to generate large synthetic datasets to effectively train high-fidelity ANN models. These trained models can then be deployed as web services, enabling designers to access real-time daylighting analysis within online design platforms, significantly reducing analysis times compared to traditional simulation-based methods.

However, the review also identifies several key limitations of the current research. The studies have predominantly focused on a narrow range of building types, primarily offices, and educational facilities, and there is a lack of consistency in the input variables considered across different works. Additionally, there is limited discussion on the deployment and scalability of these ML-based approaches for real-world applications. The review highlights the potential of more advanced ML algorithms, such as reinforcement learning, to enable adaptive daylight control systems that can optimize the indoor luminous environment based on occupant preferences and contextual factors. It also emphasizes the need to leverage emerging techniques like transfer learning and incremental learning to improve the generalizability of ML-based daylight models, allowing them to be applied to a wider range of building typologies.

In conclusion, this review underscores the transformative potential of machine learning in enhancing daylighting design and control. By addressing the identified limitations and exploring advanced ML techniques, future research can pave the way for more versatile and scalable daylighting solutions. This will ultimately contribute to creating more sustainable and occupant-friendly built environments.