Efficient optimisation of building design using a genetic algorithm
Several conventions on climate change, such as the Kyoto Protocol, the European Union Protocol, and the UK White Paper, have been issued to control and reduce greenhouse gas emissions. Research has been conducted to find friendly environment sources of energy such as renewable energy, and to reduce greenhouses gas emissions by reducing the use of conventional energy. Analysis of the energy consumption from a perspective of end-use indicates that buildings are one of the main energy consumers. Optimization of building design has the potential to save 22% to 32% of building energy consumption [Caldas and Norford, 2001; EU, 2002; and Wetter and Wright, 2003]. There are several optimization algorithms that have been developed to solve engineering problems. However, in this research a probabilistic optimization algorithm (a binary encoded Genetic Algorithms, GA), has been implemented to optimize building design with the aim of finding nearoptimum design solutions with the minimum number of new function calls. The main aim of this research is to identify a GA structure and control parameters that is effective in solving whole building optimization problems, including large scale constrained problems having many design variables. The research is restricted to the single objective, minimising building energy. The performance of the GA was evaluated for two building optimization problems, both based on an example five zone air-conditioned building located in Chicago, USA. The first example is for an unconstrained minimization of building energy use, the optimization of the building construction design. The second problem extends this to include the HVAC system control variables and as a result, includes constraints on the occupant thermal comfort. In each experiment, the performance of the GA was examined for different population sizes, crossover probability, and the mutation rate. The maximum number of new function calls (and building simulations) was restricted in each experiment set (this being the GA stopping criterion). The number of new function calls was selected to allow the optimization problem to be solved in a practical time. For the unconstrained problem, 12 GA control parameter sets were evaluated (with a total of 60,000 building simulations). Whereas for the constrained problem, eight sets of parameters were evaluated. Again the experiments were requiring a further 60,000 trial simulations. The results showed that GA performance was insensitive to most GA control parameter values, such as crossover probability and mutation rate. However, the control parameter that had the most significant effect was the population size. The small population sizes (5 individuals) gave better results on the unconstrained problem, whereas the mid-size population size (15 individuals) showed better result with the constrained problem. It can be concluded from this research that a binary encoded GA with small population sizes can be used to solve unconstrained building optimization problems with 500 or less building simulation calls. However, large scale constrained building optimization problems require in the order of 2000-3000 simulations.