Machine Learning with Numerical Weather Prediction Cloud Forecasts for Optical Communications Support

Weather related interference with space-based optical communications systems is a major challenge as many agencies move away from traditional radio frequency systems. Regional Numerical Weather Prediction (NWP) models offer the capability of providing high-resolution, dynamically-driven, realistic 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 operational decision aids which predict link outages due to clouds over the Haleakala summit of the Hawaiian island of Maui. However, NWP models have limits in terms of their ability to represent the precise timing and location of meteorological phenomena. Although WRF offers many benefits over lower resolution global forecasts and statistical forecasts using course satellite data, more than a year of very high-resolution runs of the WRF model have shown deficiencies in the standard cloud fraction predictions.

Due to the sub-optimal performance of the WRF-simulated cloud cover parameter (CLDFRA) and the need to accurately forecast atmospheric transmission for imaging and optical communication applications, it is important that we work towards improving the very short-range and short range (from hours to days) simulation of clouds. One way to transform NWP forecasts is by taking advantage of Artificial Intelligence (AI) techniques. 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 order to improve upon the WRF cloud forecasts over Haleakala summit, we employ a Random Forest (RF) statistical approach that trains upon a large database of WRF forecasts and validates against in-situ observations. The capability of the RF approach to predict cloud cover at a single location is examined under two configurations. The first uses only local WRF forecast data at the summit, while the second uses WRF forecast data from neighboring locations, including spatial domain statistics. Initial results demonstrate that RF improves the probability of detecting low clouds by about 40% while keeping the false alarm ratio stable and reducing the overall bias. We examine the sensitivity of the RF predictions to several factors including predictor variables, forecast offset, validation dataset, and data sampling protocol. We conclude with a discussion of the merits of this approach and future plans.