Leveraging deterministic weather forecasts for in-situ probabilistic predictions via deep learning

We propose a neural network approach to produce probabilistic weather forecasts from a deterministic numerical weather prediction. The developed method is applicable to any gridded forecast including the recent machine learning weather prediction model outputs. To postprocess multiple lead times using a single model, we introduce a lead time embedding that encodes the shift in biases as the forecast progresses. We apply our approach to operational outputs from the Global Deterministic Prediction System up to ten-day lead times. The model is trained to predict METAR in-situ surface temperature observations in Canada and the United States. The resulting forecasts have a mean CRPS below 2.5 K at 10 days lead time while maintaining a spread-error ratio of approximately 0.9, suggesting appropriate calibration. For extreme temperatures, the model’s biases are comparable to that of the underlying deterministic forecast. Our approach increases the utility of a deterministic forecast by adding information about the uncertainty, without incurring the cost of simulating multiple trajectories. It requires no information regarding forecast spread and can be used to generate probabilistic predictions from any deterministic forecast.