Applications of neural networks to produce optimum designs

Abstract

An artificial neural network can be considered as an information processing system where the architecture essentially mimics the biological system of the brain. A neural network consists of a number of interconnecting processing units referred to as neurons. Each of these connections has numerical weights associated with them. These weights determine the nature and strength of the influence between the interconnected neurons. The neural network can be trained by using a suitable set of data. The trained network can be tested for its performance. When the performance is adequate, the net can be used to make predictions for new cases. Neural networks have been successfully used in many disciplines on engineering such as in many civil engineering applications (Goh, 1994), prediction of pile capacities (Chow et al., 1995), Multi-objective and multirecourse decision support systems (Wei & Singh, 1995), Control systems (Macnab & D’Elenterio, 1995) etc.

Design is often considered as an intuitive process where the designer has to deal with a lot of uncertainty and complexity. Generally, design process consists of conceptual, preliminary, detailed and design documentation phases. When the designs generated fail to satisfy the objectives, redesign can be carried out at any of the above stages.

In many design tasks, the final aim is to produce optimum solutions, which suits the constraints imposed by various factors such as design objectives and what is practically possible to achieve with the available technology. In design tasks, the production of optimum designs should generally start at the conceptual and preliminary design stages since optimisation of an undesirable concept may even need complete redesign. However, designers face many difficulties at the initial design stages since they have to predict the effects of their decisions on the performance of the final product. This is not an easy task since many of these decisions may influence the performance indirectly rather than directly, and also certain parameters may be interdependent.

In such instances, the development of computer tools that can assist designers by predicting the effects of various decisions would be of immense value. This report concentrates on the development of one such tool that can assist in designing buildings with passive elements.

One of the primary objectives of the building designer is to ensure that the ‘built environment* is thermally comfortable to its occupants throughout the day, around the year. Generally, air conditioning of buildings is considered as one of the options available for achieving the thermal comfort. However, the energy required for air conditioning can be considered as a major contributing factor to the running cost of buildings. This cost can be controlled to a certain extent by incorporating passive elements to the building facade so that the total direct solar gains of the building is minimised. These passive elements include strategic location of windows, use of shading devices, selection of appropriate colour for facades, selection of suitable materials for facade walls etc.

In building designs, the architect handles the conceptual design stage where important decisions on aesthetics, facade types, visual, thermal and acoustic comfort are taken. These decisions can have a very significant effect on the final performance of the building and the energy demand. Thus, the advice of heating, ventilation and air conditioning (HVAC) engineers should be sought at these conceptual stages. However, HVAC engineers will also have problems in using sophisticated computer packages to simulate the thermal performance of buildings since they work from detail to whole building, whereas architects tend to work from whole to detail (Holm, 1993)

Therefore, these computer simulations may lack sufficient data at these initial stages and by the time sufficient data is available, architect and the client may be attached to the design so much, they may not be willing to optimise the design further using the advice given by HVAC engineers. In order to overcome this problem, it is useful to develop Artificial Neural Networks, which can provide approximate answers on energy demand when the orientation, shape, floor area in each floor, facade type, shading devices, opening sizes, surface area to floor area ratios etc. are known.

When such tools are available, the architect will be able to predict the implications of the decisions made at the initial stages on the running cost of the building. Therefore, they will be able to take appropriate action to control the energy demand thus optimising the design while meeting the other criteria like functionality, aesthetics, constructability etc. This approach is possible since the time taken by the ANN is generally independent of the complex nature of the problem. HVAC engineers will also be able to check the results that they obtain with detailed computer simulations with the results of ANNs so that a comparison can be made. The data from detailed simulations can also be used to continuously upgrade the ANN.