Keywords: artificial neural networks, building energy consumption, building life-cycle costs, coarse grain modelling, modular model development, non-linear functions, recursive models, simulation modelling, thermodynamic simulation.

Improving the thermal performance of a building can significantly reduce its life cycle costs. However, design decisions aimed at reducing energy consumption (such as increasing the overhang of eaves to reduce solar loading, or increasing the thickness of thermal insulation) tend to increase initial construction costs. Determining an optimal combination of values for all design variables is not straightforward. While the impact of design decisions on construction costs can be easily computed, the impact on energy consumption requires a sophisticated simulation of the structure for dynamic occupant usage and thermal loading profiles. The most accurate mechanism for achieving this is the finite element method (FEM). Unfortunately, FEM models are cumbersome in that they require a lot of effort to reconfigure for each new version of the model and, moreover, they are computationally expensive taking several hours to complete a simulation. Consequently, it is not feasible to test more than a handful of alternative design decisions, and thus the ability to seek an optimal solution is severely limited.

The work presented here is concerned with developing an alternative approach to simulating the thermal behaviour of residential buildings, based on artificial neural networks. The primary objective is to capitalize on the modelling versatility of neural networks to allow design variables to be treated as simple inputs to the model. This way, in contrast to FEM approaches, alternative designs can be evaluated without having to rebuild the model. Such a tool will enable a large number of alternative design decisions to be evaluated within a short period of time (compared to more traditional Finite Element Methods of analysis), thus enabling an architectural design to be fine tuned to minimize life cycle costs.

Specifically, the approach entails the development of a modelling system that operates in three spatial dimensions and time, that can simulate the thermal behaviour (and ultimately energy consumption) of any residential structure subject to various occupant usage profiles and environmental conditions. A modular neural network approach has been adopted to provide the versatility necessary to model the indefinite variety of alternative building configurations possible [1]. Each neural network module will be dedicated to modelling, at a localized position within a model, a given element type (such as a section of a given type of wall), or an interface between element types (such as a corner joint). The modules are assembled by a user into a spatially distributed configuration that represents the building design under investigation. This assemblage is then run in a recursive manner, simulating the time-wise thermal behaviour of the building, under chosen thermal loading and occupant usage profiles. Design variables such as eaves overhang are represented as input variables to the appropriate neural modules, thus circumventing the need to reconfigure the model to test alternative designs.

The paper reports on progress in the development of this concept, ultimately producing a selection of neural network modules to allow modelling a broad range of building elements. Data for training these modules are obtained from an appropriate range of FEM models of each element. Validation of the approach is undertaken by simulating the thermal behaviour of a range of building configurations and comparing with actual performance measured in the CertainTeed™ Building Products Test Facility at the University of Florida.

1

Flood, I., "Modeling Dynamic Engineering Processes Using Radial-Gaussian Neural Networks" in "Journal of Intelligent and Fuzzy Systems", 7, 373-385, 1999.

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £85 +P&P)