1 - 15 Day Weather Forecast Guidance at The Weather Company

Over the past twenty years, significant advancements in the post processing of numerical weather prediction output have driven substantial improvements in the skill of consumer forecasts. Machine learning techniques to optimally blend ensembles of model and model output statistics output have been operational for over a decade. The principal method used at The Weather Company is based on the NCAR (National Center for Atmospheric Research) DICast system. First developed in the late 1990's, the DICast (Dynamically Integrated ForeCast) system is an automated forecast system designed to maximize objective forecast skill by blending many inputs and adjusting the ensemble through continuous and near real-time comparison against observations. Using a gradient-descent nudged machine learning approach, the DICast system combines these inputs to generate optimized forecasts with skill generally superior to that of any single input forecast.

A customized version of the NCAR DICast system, has been in use for over a decade at The Weather Company and has provided first-guess 1 -15 day forecast guidance to the organization's forecasting infrastructure. Recently, however, a significantly updated and modernized version of the DICast system was developed to drive the Forecast on Demand (FOD) system (Neilley, et. al 2015).

The Weather Company's new 1-15 day guidance employs a global gridded architecture with hourly output timesteps to support the spatial and temporal resolution requirements of FOD. Significantly, the gridded architecture of the new system required the creation of a global gridded observation dataset to drive machine learning adjustments to forecast parameters in the new system. Further, significant improvements were made to the DICast subsystems responsible for deriving consumer forecast parameters from input numerical weather prediction output. Notably, the DICast system's forecasts of sensible weather parameters such as precipitation type and probability were modernized with new techniques, which have resulted in reducing the frequency of subsequent modification of the automated forecasts by human forecasters. Finally, and perhaps most importantly, the runtime architecture was modified to produce optimized forecasts as soon as new numerical weather prediction input data arrives, removing the lag imposed by a pre-defined schedule as in the legacy DICast system.

This presentation will provide a brief background on automated 1 -15 day weather forecasting at The Weather Company, and a description of the recent transition from the legacy DICast system to the new system. Examples will be presented, as will challenges introduced by the modern system's architecture. Finally, some insights on future directions for The Weather Company's 1 -15 day weather forecasting infrastructure will be provided.