An Intercomparison of Machine Learning Models to Determine the Planetary Boundary Layer Height Using Surface Level Measurements

The planetary boundary layer height (zi) is a key parameter in meteorology and climatology for applications such as weather prediction, cloud formation modelling and understanding the vertical transport of scalars (such as aerosols and greenhouse gases) and energy near the surface of the planet. This study presents an intercomparison of Machine Learning (ML) models that predict the value of zi from surface data that include fluxes of traces gases and energy, meteorological measurements and surface energy balances measurements. Data from the ARM data repository (cleaned and upsampled to 30 minute frequency) are used for training and testing. In testing the performance of the models multiple entire months amounting to approximately 10% of the available data are withheld from the model at training time and used to test model predictions. The predictions are evaluated using RMSE, nRMSE and R2 as metrics with ERA-5 predictions from the nearest grid point as baseline. A modified recursive feature elimination algorithm utilising feature importance as a heuristic is also introduced to obtain optimal feature sets with the minimal number of features needed to obtain maximal model performance. This study demonstrates the applicability of machine learning models to model zi and identify optimal features for model building.