A comparison of approaches with different constraint handling techniques for energy-efficient building form optimization

Building performance optimization (BPO) has been a common method in energy-efficient building design. How to deal with the constraints in the optimization model is critical to optimal solution quality and efficiency, while is usually ignored when selecting optimization approaches. This study introduces three typical categories of constrained optimization approaches that integrate common and advanced constraint handling techniques (CHTs), including optimization with penalty function (SPF and DPF), biased bi-objective optimization (BBO) and reinforcement learning guided optimization (RL-SPF). To better match the high-efficiency requirement of practical problems, some improvements are made to the latter two approaches. An energy-efficient building form optimization problem is described and combined with different spatial and environmental constraints to construct five groups of experiments. The results illustrate that BBO achieves the best performance on average, particularly outstanding in convergence speed. The superiority of RL-SPF is obviously affected by the constraint handling capacity of RL model, and both penalty functions are not robust and efficient enough in solving practical BPO solutions. Additionally, further investigation on different CHTs reveals that preserving good infeasible solutions are more effective than simply improving the proportion of feasible solutions, which provides a promising direction for future work on developing new constraint optimization approaches.

Automated summary

Building performance optimization (BPO) is a common method in energy-efficient building design. This study introduces three categories of constrained optimization approaches and makes improvements to better match practical problems. Results show that biased bi-objective optimization achieves the best performance on average, while reinforcement learning guided optimization is affected by the capacity of the RL model. Penalty functions are not robust and efficient enough in solving practical BPO problems. Further investigation reveals that preserving good infeasible solutions is more effective than simply improving the proportion of feasible solutions.