Using Dynamic Ensembles and Machine Learning to Determine Tropical Cyclone Forecast Position Confidence Areas

Operational tropical cyclone products typically include an assessment of uncertainty, or confidence, in the forecast position. The most common objective approaches to produce forecast confidence information have been statistical methods based on historical forecast position errors. This gives regular, circular areas centred on the forecast track. However, these statistical solutions do not account for the environmental influences at the time and therefore do not represent the true spread of possible outcomes.

The many forecast scenarios provided by dynamic ensembles are based on the environmental influences over the forecast period. They provide a situation dependent and more complete representation of the spread of potential outcomes.

A Forecast Confidence Area is the area in which the centre of the tropical cyclone is forecast to be located a certain percentage of instances and is for a single forecast time. A Forecast Confidence Area based on ensemble guidance can be found using a Gaussian Mixture Model, which is a machine learning method.

Verification of Forecast Confidence Areas has shown a multi-model ensemble outperforms any single model ensemble, but the multi-model ensemble is overspread. We account for this overspread by calibrating the size of the Forecast Confidence Areas using a global dataset from the past 3 years. The multi-model ensemble provides a better forecast confidence area than the historical errors method. Better is defined as covering a smaller area while achieving the target observed frequency.

A strength of the process is that Forecast Confidence Areas can be created independently of producing an official forecast track and prior to a tropical low forming. This makes it suitable for cyclogenesis products and forms the basis of the new Bureau of Meteorology Tropical Cyclone 7-Day Forecast. Forecast Confidence Areas for multiple successive timesteps can be amalgamated to create Forecast Confidence Cones, which have replaced the previously statistical, or sometimes subjective, uncertainty areas on Bureau official forecast tracks and ensures consistency between the two products.