Probabilistic Forecasting of Cloud-Base Height and Visibility Using Quantile Regression Forests, Based on NWP and Observation Features

Abstract We have applied quantile regression forests (QRFs) to generate probabilistic forecasts of weather conditions associated with low-visibility procedures (LVPs) at Schiphol Airport (Amsterdam, the Netherlands). LVPs are determined by combined thresholds of cloud-base height and (runway) visibility. Forecasts of these conditions are critical for airport operations, as they inform operational planning, with the potential of minimizing meteorologically induced disruptions. Using a dataset of 5 years of hourly data, we have performed a forward feature selection and optimized QRF’s hyperparameters for this specific application, and evaluated the model’s performance for different forecast lead times and different LVP classes. Hereby, LVP forecasts were obtained by combining separate models for cloud-base height and (runway) visibility, applying a Schaake shuffle approach for restoration of the dependencies between these parameters. The verification revealed consistent positive Brier skill scores (BSSs) for the three most common LVP classes: marginal, A, and B. Although the skill was not always positive for the more extreme LVP classes, C and D, we argue that also for these conditions forecasters might derive valuable indications from the forecast system. We demonstrate the operational utility of the system with an example, also illustrating the support of interpretability through the use of Shapley additive explanation (SHAP) values. Our results underscore the potential of QRF for probabilistic forecasts of meteorological conditions, for aviation and other purposes.