Optimum window design with machine learning for effective use of daylight in living spaces

In this thesis study, various models have been developed for three different purposes for the effective use of daylight in spaces and the inclusion of machine learning in the window design process. First of all, unlike previous studies, spatial daylight autonomy and daylight factor prediction models have been developed to evaluate the effect of window design on daylight intake into spaces and to help designers in decision making. Linear regression, polynomial regression, support vector regression, decision tree regression, random forest regression, deep neural networks, convolutional neural networks and capsule networks were used in the development of the models. For this purpose, space dimensions, window direction, window dimensions, window parapet height, window shape, ratio of window area to wall area and ratio of window area to floor area were used as input. Secondly, window dimensions and parapet height were obtained by considering the space dimensions, the desired spatial daylight autonomy and the daylight factor, in order to obtain a window design suitable for the desired daylight performance and according to the spaces. For this purpose, a model was not developed with support vector regression because the outputs were different from the first purpose: window sizes and parapet height. Finally, using visual data, models were developed with convolutional neural networks and capsule networks that classify whether the desired daylight autonomy and daylight factor value are suitable or not. This thesis study not only aims to develop a model with machine learning methods for the effective use of daylight in spaces, but also to select the appropriate model for this problem by developing various models. As a result of the developed models, in line with the first aim, random forest algorithm for spatial daylight autonomy estimation, random forest regression and decision tree regression in the models developed for daylight factor estimation were equally successful. However, since the daylight factor has little or no variation depending on the direction, new models have been developed for one direction due to the possibility of memorizing the data. As the data decreased, the most successful model became the polynomial regression model. In the models developed for the second purpose, the random forest regression model was more successful. In the classification models developed for the third purpose, convolutional neural networks were more successful. The success of the methods has varied depending on the purposes and data types. This situation emphasizes the importance of choosing the appropriate method for the problem. The successful models obtained will provide a support tool for daylight analyzes of living spaces at the early design stage, providing cost, time and ease of application, unlike simulations. These developed models will guide designers in the window design process and it will be possible to decide on window sizes suitable for the desired daylight performance. For designers who check before the implementation of projects, the window will help decide whether the design is suitable or not. Integrating these developed models into the architectural design process will enable window designs that allow daylight into the space efficiently and effectively. Thus, the physiological and psychological needs of the users are met; efficiency and productivity increases. By providing sufficient natural lighting, sustainable designs are achieved and energy savings are achieved. An unbalanced data set was used in the thesis study. The success of the models can be increased by adding different daylight performance evaluation metrics and using a balanced data set. Additionally, using and comparing various methods can help optimize the performance of models to be developed for future studies. Keywords: Daylight, optimal window design, decision making, machine learning, deep learning