Probabilistic Prediction of Tropical Cyclone Intensification with an Analog Ensemble

The Analog Ensemble (AnEn) technique has been extensively tested for the probabilistic prediction of meteorological variables (e.g., wind and temperature), air quality (ground-level ozone and particulate matter), and for renewable energy applications (both solar and power). We will present a novel application of AnEn to tropical cyclone (TC) intensity forecasting. The AnEn technique is used here to create a naturally calibrated ensemble prediction of TC intensity from a training dataset composed of the deterministic predictions from the Hurricane Weather Research and Forecasting (HWRF) model (2015 version) and observed tropical storm intensity. In the AnEn, a set of analog forecasts is created by searching archived HWRF forecasts that share key features with a current forecast from the same configuration of HWRF. The meteorological variables used to identify a past forecast similar to the current one are called analog predictors. The 1-min Best Track maximum surface wind speed observations associated with the best analog forecasts are then used to produce an ensemble forecast.

The AnEn is developed here using HWRF reforecast data and corresponding Best Track intensity data from 2011-2014. There are 83 Atlantic TCs and 105 Eastern Pacific TCs in this dataset. Over 60 predictors (based on environmental and storm inner-core characteristics) were derived from all of the 1110 Atlantic and 1316 Eastern Pacific Ocean HWRF runs available (with 0-126 h lead-times, at 3 h increments). The analogs are selected from a training period defined by HWRF reforecast data covering the period June 2011 – August 2013. This period is also used to select a subset of the best predictors and define their weights optimizing the AnEn's performance in term of Mean Absolute Error (MAE). The predictor weights are kept unchanged during the testing period (August 2013 – November 2014) used to evaluate the AnEn performance. To mimic real-time operations, for each forecast, the training goes from the first TC forecast in the dataset to the last one before which the current forecast is issued.

The AnEn TC intensity deterministic forecasts have been obtained as the mean of the 20 ensemble members at each lead-time. In this presentation, the improvements compared to the raw TC intensity forecast from HWRF will be shown in terms of MAE and other commonly used metrics. An in-depth analysis of important attributes of probabilistic predictions of TC intensity generated with the AnEn will also be shown. These attributes include statistical consistency, reliability, resolution, sharpness, and the spread-skill relationship. The improvements to TC intensity forecasting produced by this inexpensive technique applied to a single deterministic HWRF model, which itself has continually improved over the last decade, are quite promising given NOAA's longstanding goal to improve intensity change forecasting.