Improving Cloud Forecasts using AI-enhanced WRF

Accurate, long-term cloud forecasts are critical for today's growing solar energy industry and the need for fast, secure, large data transfers made possible by Free Space Optical Communications (FSOC). Traditional Numerical Weather Prediction (NWP) models such as the Weather Research and Forecasting model (WRF) are frequently used for weather forecasts. Different versions of WRF are often designed for more specific forecasting tasks, such as WRF Solar for solar energy forecasts. Traditionally, WRF is modified to optimize a specific task through applying different physics schemes. But when tasked with forecasting clouds out to 48 hours, even the best possible physics schemes have significant limitations. NWP models often have difficulty generating clouds since the processes for cloud formation are microscale. Most schemes rely almost entirely on relative humidity as a primary indicator of clouds even though the physics and dynamics can be much more complicated over terrain, time of day, and for different seasons.

To improve WRF cloud forecasts, we developed an AI-enhanced cloud forecast using WRF output parameters as predictors and satellite-derived clouds as the labeled data. Starting with a 3km resolution WRF model already optimized for cloud forecasting and centered over southern New Mexico, we trained a U-net convolutional neural network (CNN) model with 109 two-dimensional variables from WRF model output as predictors. These predictors include a targeted set of variables from terrain and atmospheric pressure to CAPE, planetary boundary layer height, divergence/convergence, dewpoint, relative humidity, and vertically integrated liquid water content. The results show that the AI-enhanced cloud forecast outperforms persistence and improves the critical success index (CSI) of the WRF cloud fraction output by 11.8% on average. The most significant improvements are seen during daytime forecast hours when convective clouds are likely to form. Additionally, the AI-enhanced cloud forecast demonstrates CSI improvement of up to 40% during the monsoon season. These improvements in cloud forecasting capabilities are especially important as we move towards more usage of solar energy and with increased interest in space to ground FSOC.

This talk will provide an overview of the WRF model configuration and the U-net CNN implementation including data preparation, determination of the predictor set, best-performing U-net hyperparameters, and feature importance. The AI-enhanced WRF cloud forecast is validated against independent satellite data and compared with persistence and standard WRF cloud fraction for forecast hours 13-48 by time of day and by season.

Approved for Public Release: NG24-1533. © 2024 Northrop Grumman Systems Corporation