Deep Learning to Improve Numerical Weather Prediction Cloud Forecasts

Regional Numerical Weather Prediction (NWP) models offer the capability of providing high-resolution, dynamically-driven weather forecasts over any part of the world, which can aid in the mitigation of weather-related impacts. We are using a modified version of the Weather Research and Forecasting (WRF) model to produce cloud forecast decision aids over the Hawaiian Islands. However, NWP models have limits in terms of their ability to represent the precise timing and location of meteorological phenomena, such as clouds. Although WRF offers many benefits over lower resolution global forecasts and statistical forecasts using coarse satellite data, several years of very high-resolution runs of the WRF model have shown deficiencies in its standard cloud cover predictions.

One way to improve upon NWP forecasts is by taking advantage of Artificial Intelligence (AI). AI techniques show promise in closing the gap between the accuracy requirements of decision aids and the performance of traditional NWP and statistical post-processing tools. These techniques use advanced computer science and statistical tools to train models that have high predictive capacity without any prerequisite for a comprehensive understanding of physical processes. In this work, we use U-Net, a Fully Convolutional Neural Network, to train directly from the output of the high-resolution WRF model with the goal of improving the accuracy of total column-integrated cloud cover forecasts. The U-Net model is trained to predict clouds using standard meteorological variables such as temperature, winds, and humidity from a 48-hour WRF forecast as predictors and two-dimensional satellite-derived cloud masks as the labeled data. The U-Net model is trained on a compute cluster equipped with three 16GB NVIDIA P100 GPUs. Results show significant improvements in the cloud forecasts when compared to the WRF cloud cover parameter, with increases in forecast accuracy of around 20% and a large reduction in the WRF cloud forecast bias. This presentation describes the data preparation, work performed to determine the best set of predictors and the optimal U-Net model configuration, and the prediction accuracy of the U-Net cloud forecasts.