A robust automated machine learning predictive model for natural period of buildings

The fundamental natural period of a building is its most important dynamic characteristic parameter, influenced by many factors. Various building codes provide empirical formulas for period prediction, however, these typically consider only one or two factors, such as building height and material type, while ignoring the effects of others. Data-driven Automated Machine Learning (AutoML) offers a novel approach to quickly develop powerful predictive models while avoiding the tedious and time consuming iterative tasks involved in traditional machine learning model development. This study establishes a database comprising full-scale measured period samples from more than 3,000 existing buildings, obtained through a rigorous literature search and data filtering process. The AutoGluon Python package is employed to develop a robust predictive model with ten influencing factors, including five numerical features and five categorical features, as inputs. Compared to empirical formulas in building codes and those proposed by researchers, the proposed AutoML model demonstrates better accuracy across a wider range of building types. A coefficient of determination of 0.93 on the test set is achieved, and the model's generalization capability is validated using independent third-party measurement data. Furthermore, the proposed model is deployed online and made openly accessible for quick, reliable predictions.