Predicting daylight autonomy metrics using machine learning

This study analyses the efficacy of using machine learning though artificial neural networks (ANN) to predict daylight autonomy metrics in typical office spaces. Based on a literature review of the use of ANN for non-linear problems, the chosen approach was deemed promising for its use in predicting daylight performance with the assumption that previous training data can be provided. The ANN approach, while empirical, has advantages when compared to conducting full simulations in the areas of speed and computing resources. In this study, several network architectures were analysed against several test cases. The accuracy of the obtained results mirror those in other studies when applied to daylight autonomy metrics. In addition, accuracy improved with the addition of a larger set of training data as well as the enhancement of the network architecture itself.