A parallel decomposition approach for building design optimization

Integrated building design can promote high-performance, energy-efficient, and sustainable building designs by holistically considering design variables of simulation models from different disciplines. However, the complexity of the design increases drastically with the increasing dimension of the design problem. In some cases, solving high-dimension problems is not technically feasible nor time-efficient. Parallel decomposition can solve this problem by dividing the original problem into several smaller subproblems to be solved separately. Variable grouping is the first step of parallel decomposition. In order to obtain global optimal solutions, the existing variable grouping criterion focuses on whether there is an interaction between variables. Since many variables in building simulation interact, the building design optimization problems could be inseparable using this criterion. This study proposes a new variable grouping criterion, which can achieve the global optimum and decompose variables with strong interactions. This proposed criterion is combined with the existing one and presented as a dual-criteria variable grouping method for parallel decomposition. Sensitivity analysis is used to assess the existing criterion, and interaction plots are used to evaluate the proposed criterion. To demonstrate the applicability and runtime efficiency of the proposed parallel decomposition, the approach is applied to solve the single-objective optimization problems of a benchmark function and a low-rise office building. The results show that the proposed approach finds the global optimal solutions and takes less computation time than optimization without decomposition. In this study, each subproblem is solved while the variables in other subproblems are set as default values. With the results of the reference case of a full factorial design, it is found that the subproblem might achieve local optimum depending on the default values of variables in the other subproblems. This potential issue can be overcome by optimizing the subproblem with the strongest main effect first and using the results for subsequent subproblems with decreasingly weaker main effects.