In order to obtain information about building daylighting performance earlier in the design stage, recent studies have begun to develop daylight prediction models using machine learning methods. Most studies have adopted design variation parameters as the input parameters for model training, but this method greatly limits the scope of application of the daylight model. By extending application to different design possibilities, this research proposes a novel daylight model, which includes a pre-processing procedure to convert the design model into “Intermediary features” as input parameters representing daylight penetration performance. By changing geometry of one kind of parametric façade and performing daylighting simulation, we were able to generate the data for training, including Intermediary features and sDA/ASE values. The daylight model was then trained using an artificial neural network. Finally, the proposed daylight model was tested by predicting the daylight performance of other kinds of façades. The results indicated that the value of DA and ASE hours per grid were well predicted and that the daylight distribution was reproduced even with different kinds of façades. The deviations of sDA and ASE from simulation results ranged from 1.7 to 6.1% and 0.3–2.1 h respectively. The reproducibility, the predictive capability, and most importantly the extension of model applicability were all demonstrated for the proposed model. Furthermore, comparing to the daylighting simulation, the method using the proposed daylight model is estimated to save 9/10 daylighting evaluation time. This is critical for implementing the evaluation in the early design stage.
All Science Journal Classification (ASJC) codes