Optimizing energy predictions: a generalized surrogate modeling approach for diverse building designs and climates

This paper introduces and evaluates a state-of-the-art surrogate modeling approach for predicting heating and cooling load for office buildings, designed to accommodate diverse climates and architectural complexities. The model incorporates weather features from 249 strategically selected locations worldwide, capturing the global climate variations.

Building on previous modeling approaches, the paper introduces additional levels of flexibility to ensure model adaptability to any design scenario. Given the vast 66-dimensional design space of a zone model, a specialized ensemble learning approach of Mixture of Experts (MoE) is applied, achieving an excellent performance on the test set with an R2=0.9993 and 0.9972 for heating and cooling EUI, respectively.

The zone MoE model is successfully applied to two test buildings, with varying complexity, demonstrating its modularity to complex designs. Future work will focus on refining the model to incorporate localized experts from decentralized datasets and to consider various building types, other performance indicators, and annual temporal interactions.