Minimising Error: Artificial Neural Network Configurations for a User Overridable Dynamic Shading System

Abstract

This research explores the possibilities of integrating environmental and human inputs to achieve precise architectural goals. Specifically, the aim is to create an adaptive façade, trained on historical data relating to human (an override capability) and environmental inputs to maintain optimal internal lighting conditions for inhabitants. The study was conducted using a physical louvered shading system constructed in the Bartlett School of Architecture, University College London.

The historical data collected by the system provided a sample data set to train the Artificial Neural Network (ANN) for which the system would operate. A multi-layer perceptron was the neural network used in the study and a series of experiments allowed for the optimal network architecture to be ascertained. Based on the trained network, further testing was carried out to assess the accuracy of the results with regards to the louver angle suggested during system recall. It was found that the complexity derived from receiving both environmental and human data provided some confusion when recalling, however the system displayed a high level of accuracy, correctly recalling the desired blade angle over 70% of the time. Further testing found that the remaining recall error could be accounted for through environmental input data similarities. By physically building and testing the system this research suggests that a trained physical system based on computational principles can provide an adaptive architectural entity that considers building occupants behaviour and wants as well as the external environments natural imposition.

Title: Minimising Error: Artificial Neural Network Configurations for a User Overridable Dynamic Shading System

Authors: Amir NabilProf Michael PittDr Sean Hanna, Martha Tsigkari

Publication: Proceedings of the International Conference on Constructions in a Changing World (German); 12 pages | full text (PDF)

Year: 2014