A Probabilistic, Large-Ensemble Approach to Tropical Cyclone Forecasting

Over the past decade, improvements in forecasts of the track, intensity, and wind field of tropical cyclones (TCs) have been largely driven by the improvements in high-resolution numerical weather prediction (NWP) models. While these deterministic models can resolve the 3D structure of a TC, they are computationally expensive, making it infeasible to sample the internal variability of the TC and its environment. To remedy this issue, numerical weather prediction centers have been running ensemble prediction systems. Unfortunately, these models do not possess the resolution to resolve strong gradients in dynamical fields or explicitly resolve cumulus processes, and thus have large negative biases in predicting the intensity of TCs.

In light of these biases, we developed a physics-based, large-ensemble tropical cyclone model that can sample the internal variability in the large-scale environment and quantify its effect on the track, intensity, and wind field of TCs. The model consists of three components: (1) a track model that generates synthetic tracks from the track covariance of an ensemble NWP, (2) a simplified intensity model that is primarily driven by potential intensity and mid-level ventilation, and (3) a physically-based TC wind field model. The TC's intensity and wind field evolution in this model are determined by a time-evolving environmental field, which is stochastically generated and derived from ensemble NWP models. The system is computationally inexpensive, such that it can sample the inherent uncertainty in the forecast by generating thousands of realizations.

We evaluated this large-ensemble model over four years (2015-2018) of TC forecasts in the Atlantic and East Pacific basins. We find that the model produces competitive track and intensity forecasts, and shows significant improvements in the continuous ranked probability score for TC intensity forecasts, as compared with the deterministic Hurricane Weather Research and Forecasting Model (HWRF). Probabilistic, point-wise forecasts of wind speeds exceeding a given threshold are also evaluated and compared with the operational Hurricane Wind Speed Probability Product. The relative success in forecasting TCs, using a simple model that simulates TCs embedded in realistic, stochastically varying environmental fields, indicates the importance in quantifying the internal variability in the forecast system.