Operational scheme for predicting tropical cyclone wind radius based on a statistical–dynamical approach and track pattern clustering

Abstract

AbstractAn operational scheme for predicting the symmetric R30 and R50 of tropical cyclones (TCs) in the western North Pacific was developed using a statistical regression method and track pattern clustering (four clusters). The statistical–dynamical model employs multiple linear regressions of two to eight variables at each cluster and forecast lead time. The dependent variable for prediction was the change in the 5‐kt wind radius (R5)—a proxy of TC size—relative to the initial time. The performance of the model was compared for the training (2008–2016) and testing (2017–2018) periods. The effect of clustering on TC size prediction was evaluated by comparing the performance of the non‐clustering and clustering models. The clustering model improved the prediction of TC size by 3%–24% at all lead times during the training period, especially with a significant improvement of up to 43% in Cluster 2. In Cluster 2, because most TCs tend to develop strongly and continue to increase in size, it greatly reduced the variability in TC size through clustering, allowed for smarter predictor selection, and ultimately improved TC size prediction. In the real‐time R30 and R50 predictions for the 2017 and 2018 TCs, the error of the clustered model was 18%–19% less than that of the non‐clustered model. The analysis results revealed that the real‐time prediction errors of the current model increase when the TC tracks are difficult to classify into specific clusters, the predicted environments and TC tracks are inaccurate, and the size and intensity of a TC rapidly increase.