Post-Processing NWP Using Deep Learning with a Custom Loss Function to Create Ensemble Forecasts to Quantify Forecast Uncertainty

We present a machine learning model that post-processes high-resolution, deterministic weather forecasts to produce probabilistic forecasts for seven core weather variables over the contiguous United States. Our approach combines the strengths of global Numerical Weather Prediction (NWP) modeling with machine learning to generate ensemble forecasts, thus adding substantial predictive and actionable information to stakeholders. We use a multi-task neural network with a Continuous Ranked Probability Score (CRPS) custom loss function applied to the ECMWF HRES, using station observations from NOAA Global Integrated Surface Database (ISD) and ECMWF reanalysis version 5 (ERA5) as targets. Instead of addressing uncertainty in machine learning model forecasts directly, our custom loss function allows the neural network to learn the predictive distribution and output a full ensemble with a flexible number of ensemble members to best quantify the uncertainty.