Ensemble precipitation forecasts made with quantile regression forests and deterministic harmonie-arome inputs

A gridded 51-member ensemble of precipitation forecasts that are created using a tree-based machine learning method, quantile regression forests (QRF), and inputs from the deterministic Harmonie-Arome (HA) Cy43 forecasts. The target data set is rain-gauge-adjusted radar data that is upscaled by taking 3x3 km means and then a rolling maximum is taken in a 9 x 9 km box. Inputs to the machine learning model include HA precipitation, and indices of atmospheric instability. Spatial and temporal dependencies are restored using the minimum divergence Schaake Shuffle (SSh). Hourly forecasts are issued 8 times per day (00, 03, 06, 09, 12, 15, 18 and 21 UTC) for 60-hours into the future.