Over the past year, Weather & Radar has worked diligently to implement a series of new processes and cutting-edge technologies to improve the temperature forecast for North America. One such improvement was the integration of machine learning (ML) into the forecast.

In weather forecasting, only the most advanced models integrate ML into their forecasts and are often used to fix systematic errors in weather parameters such as temperature, wind, and clouds. For the contiguous United States, some of the largest systematic errors usually occur when forecasting temperatures over the rugged terrain of the Rockies and extreme desert regions of the Southwest.

In mountainous areas, these large deviations between expected and observed temperature are partly due to the weather model’s inability to resolve geographical features accurately. This happens when the distance between mountain peaks and troughs is smaller than the model’s highest resolution.

Essentially, the model cannot “see” the mountain and as a result, temperatures are calculated at incorrect altitudes, which leads to these reoccurring, systematic errors. This is a common problem with even some of the most sophisticated and accurate weather models on the planet.

Over the past year, Weather & Radar has trained their ML algorithm to identify these systematic errors and automatically correct them when necessary. However, the training process must be done over at least a year and for the entire globe so the system can become familiar with the most common climate patterns particular to the season.

Despite the distinctive topography of the Rocky Mountains and the extreme weather conditions of the western deserts, early results from applying ML techniques look promising. Across the globe, temperature forecasts have been improved by several tenths of a degree. In more problematic regions, where systematic errors are larger (more than 5 degrees Celsius), the results show even more significant improvements in the temperature forecast.