Deep learning for post-processing global probabilistic forecasts on sub-seasonal time-scales

Abstract

Even though sub-seasonal weather forecasts are crucial for the decision making in many sectors including agriculture, public health and renewable energy production, sub-seasonal predictions are hardly used due to their low skill. We propose several post-processing approaches based on convolutional neural networks (CNNs) to improve and calibrate sub-seasonal forecasts from numerical weather prediction (NWP) models. All proposed methods work directly with the spatial NWP forecasts and are therefore able to retain spatial dependencies in the forecasts. Moreover, these methods have the potential to exploit the predictive information in the spatial structure of the NWP forecasts. The proposed post-processing models use forecast fields of multiple meteorological variables as input, and produce global probabilistic tercile forecasts for biweekly aggregates of temperature and precipitation for weeks 3-4 and 5-6, as commonly done in sub-seasonal to seasonal prediction. The model architectures and the training strategy are optimized to deal with the low signal-to-noise ratio in sub-seasonal forecasts and the limited amount of training data. Half of the tested architectures use the well-known UNet architecture specifically designed for image segmentation. The remaining architectures are based on a standard CNN as typically used for image classification, with the difference that it estimates coefficient values for a set of basis functions to provide spatial predictions. All post-processing methods improve precipitation and temperature forecasts for both lead times. Improvements increase with lead time and are larger for precipitation, for which the NWP forecasts show no skill. Our best performing model is based on the UNet architecture and trained directly on the global NWP forecasts. The post-processed forecasts are substantially less sharp than the respective probabilistic forecasts from ECMWF for all tested methods. We demonstrate that the post-processed forecasts are well calibrated in contrast to the ECMWF benchmark using a calibration simplex, a reliability diagram for probabilistic tercile forecasts. Thus, our proposed post-processing methods are able to derive a reliable uncertainty estimate based only on ensemble mean NWP forecasts.