Artificial neural networks to estimate daylighting in residential environments with surrounding obstruction

Abstract

Many countries adopt regulatory instruments to improve the performance of buildings and luminous quality is often addressed in their scopes. Simplified methods facilitate the application of such instruments and artificial intelligence has been shown to be useful for this purpose. Thus, the objective of this work is to propose a metamodel, using artificial neural networks, to verify the luminous performance of residential buildings, considering the impact of the built environment in the context of the revision of the Brazilian standard “ABNT NBR 15.575-1 Housing buildings — Performance". For this, the simulated database was adopted for the proposed revision of the standard, containing 36,000 cases that relate the influence of external obstructions to the building to its performance regarding the sufficiency and uniformity of natural light. Thus, metamodels of artificial neural networks Multilayer Perceptron were trained with data from the cities of Curitiba, Brasília and Belém. The architecture of the networks consisted of 3 layers, the input, a hidden and the output. Aspects of its architecture and the grouping of input parameters, building variables, and output parameters, ALNE200lx,50% and ALNE60lx,50%, were tested. Its overall performance was considered acceptable, with an average percentage error of less than 10%, requiring its refinement to reduce outliers. It was concluded that the ANN can be an alternative as a simplified method for application in the standard, pointing out as options for refining the metamodel the variation of the learning algorithm, the partition of the training and test sets, and the expansion of its scope with other proportions and visible transmissions.