Application of a machine learning-based prediction model for the design of general lighting systems in offices

• Integration of machine learning with indoor lighting design. • Development of an machine learning model for luminaire recommendation. • General lighting layouts prediction by machine learning model. • A design tool that provides preliminary information for non-expert users in lighting. • The methodology of the prediction model is detailed in the paper with the user interface design and application examples. Lighting design is an interdisciplinary practice that aims to create a specific atmosphere through light, while addressing the functional, physiological, and psychological needs of users. The main objective of interior lighting design involves creating an ideal lighting system which fulfills all necessary requirements of a particular space. Designers focus on examining the interactions between various criteria to produce the most suitable lighting solution. Over time, lighting designers develop an intuitive understanding of how to optimize lighting parameters based on accumulated experience. Machine learning is an area of artificial intelligence that enables computers to make predictions about new data through sample data and past experiences. The aim of this study is to develop a prediction model that reflects the knowledge and intuitive expertise of lighting designers over time. The study involves training a Deep Neural Network (DNN) model to recommend luminaire properties and positions for achieving target illuminance levels and uniformity ratios in office spaces. The model accepts spatial parameters, average illuminance and uniformity ratio as input data while it generates luminaire position, quantity, luminous flux and beam angle as output results. A prediction model interface was developed to provide users with access to the system. With the developed user interface, luminaire recommendations suitable for the luminaire technical specifications obtained as output from the prediction model are presented. Training and validation losses were closely aligned, indicating stable learning. Test-set evaluation yielded a mean R² of 0.5531 across luminaire types and 0.8277 for downlights. Validation against DIALux Evo simulations showed an average agreement of approximately 84 % in illuminance level and 88 % in uniformity. These findings suggest that the proposed model, combined with its user interface, has the potential to serve as a useful tool in the lighting design process. The study is expected to contribute to the literature, particularly in filling the gap in luminaire selection, and support the lighting design industry with a user-friendly, data-driven approach.