Enhancing Optimization in Building Performance: Simplified Models, Parallel Computing, and Neural Networks

Abstract

Accurate and computationally efficient building energy models are essential for retrofit decision-making under climate uncertainty. This study evaluates the use of simplified building energy models within simulation-based multi-objective optimization frameworks. Detailed models provide high predictive accuracy but impose substantial computational costs when optimization requires large numbers of simulations. Simplified models reduce computational effort but may introduce errors that affect the identification and ranking of energy conservation measures. This research quantifies these trade-offs, evaluates climate sensitivity, and proposes methods to improve predictive accuracy without modifying physical structure.

The study is organized into three phases. The first phase assesses the effects of common simplification strategies, including thermal zone aggregation, HVAC abstraction, material property reduction, and geometry simplification, on model accuracy and optimization performance. A commercial building case study shows that simplification reduces simulation and optimization time by more than an order of magnitude. Error magnitude and direction depend strongly on building typology and abstraction type. HVAC idealization and excessive zoning aggregation distort Pareto-optimal solutions and lead to suboptimal retrofit recommendations. Parallel computing further improves efficiency, enabling large-scale optimization using both detailed and simplified models.

The second phase examines the sensitivity of simplified models to climate inputs, including typical and actual meteorological years, future climate projections, and extreme hot and cold conditions. Using a mid-rise dormitory as a case study, results indicate that simplification-induced errors vary across climate scenarios and may intensify during extreme events or seasonal transitions, highlighting the importance of climate-aware model selection.

The third phase introduces an external artificial neural network adjustment that improves the predictive accuracy of simplified models without altering their physical structure. The approach restores hourly load and objective-function accuracy across all case studies, particularly for complex buildings where simplification effects are most pronounced. Embedded within a prNSGA-III framework, the method preserves Pareto-front structure while maintaining computational efficiency.

Overall, the results demonstrate that simplified models can support efficient and reliable retrofit optimization when their limitations are explicitly evaluated and addressed. The proposed methodology offers a structured framework for assessing simplified models under climate uncertainty and enhancing accuracy through data-driven correction, supporting scalable retrofit planning for existing buildings.