Creating a Universal Depth-to-Load Conversion Technique for the Conterminous United States Using Random Forests

Abstract

As part of an ongoing effort to update the ground snow load maps in the US, this paper presents an investigation into snow densities for the purpose of predicting ground snow loads for structural engineering design with ASCE 7. Despite their importance, direct measurements of snow load are sparse, compared with measurements of snow depth. As a result, it is often necessary to estimate snow load using snow depth and other readily accessible climate variables. Existing depth-to-load conversion methods, each of varying complexity, are well suited for snow load estimation for a particular region or station network, but none is consistently effective across regions and station networks. In this paper, a random forest regression model is proposed for estimating annual maximum snow loads in the conterminous US that makes use of climate reanalysis data and overcomes the limitations of existing methods. The effectiveness of the random forest model is demonstrated through accuracy comparisons of existing depth-to-load conversion techniques using a compilation of national and state-level data sources. The accuracy comparisons show that the random forest model is competitive for all regions and station networks, whereas other methods are competitive only for certain regions or station networks. These results highlight the feasibility of developing a single depth-to-load conversion method that appropriately characterizes region and climate specific differences in the snow depth–load relationship across the conterminous US. Such universal models are an essential component for creating a unified set of national snow load requirements that eliminate the case study regions currently defined in current national standards.