Semi-Empirical Prediction of the Carrying Capacity of Rack Structure Elements: An ML-Driven Approach

Abstract

Reliable prediction of carrying capacity is essential for efficient design of all carrying structures, including pallet racks. Precise or simplified analytical models are surely indispensable in engineering practice, but when considering racking structures, they still remain limited primarily by specific beam-to-column connections. This study proposes a semi-empirical, machine learning-driven approach for predicting the behavior of rack elements, i.e., upright frames, as a function of three structurally distinct connection types and appropriate racking element properties. Random Forest was selected for feature screening based on comparative analysis of six ML algorithms, and the retained parameters were used to develop interpretable power-law models. Within this research, upright cross-section area and beam-level height were retained as main influential parameters for the first two considered connection types, whereas upright cross-section moment of inertia, beam-level height and beam cross-section height governed the third type. These differences indicate that the contribution of each input parameter to capacity prediction varied among connection types. Grouped correction factors were introduced to account for deviations associated with upright profile and beam-level height, with additional differences specifically for the third connection type. The adopted models achieved near-exact agreement with the upright frame carrying capacities within the calibration domain. Although the resulting coefficients are specific to the analyzed rack structures, the proposed sequence provides a transferable framework for converting large structural datasets into simplified and interpretable engineering models.