A Comparison Between Indoor Daylight Prediction Methods: Daysim Simulation and LSTM Models

The simulation software is widely used in many research and professional fields. Certainly, they are very useful tools, but it could be a difference between the expected results and the simulated ones. This paper aims to validate methods for calculating natural lighting contributions comparing the results of simulated data calculated by using the climatic file, the results of simulated data calculated by using real climate data, and results calculated by using machine learning methods. The study was applied for a case study located at the Campus of the University of Palermo. The measurement campaign ran from April to September, collecting more than 24,000 measurements of solar radiation and illuminance. The results showed that the machine learning approach provided the most accurate predictions, with the LSTM model achieving an $\mathbf{R}^{\mathbf{2}}$ of 0.99, validated using approximately 700 measurements. In contrast, the Daysim simulation software reached an $\mathbf{R}^{\mathbf{2}}$ of $\mathbf{0. 7 8}$ over the same validation period as LSTM. However, machine learning methods require extensive datasets for training, unlike Daysim. Nevertheless, they can be a valuable tool for estimating indoor daylight distributions using only external solar radiation data.