P1.53 Predicting the strength of tropical cyclone seasons using clustering and linear regression

Thursday, 12 November 2009
Caitlin M. Race, University of Minnesota, Minneapolis, MN; and M. Steinbach and V. Kumar

The relationship between August tropical cyclone frequency and reanalysis sea surface temperature (SST) data is analyzed in order to derive a linear model to predict future storm counts based on simulated SST data from the Community Climate System Model 3.0 (CCSM). Kmeans clusters were used to group the data, as using the average of all the coastal data yielded a regression equation with an R2 of only 0.23 (p < 0.05). After clustering the data and regressing the cluster centroids with the storm counts, varying R2 values of up to 0.55 (p < 0.05) were obtained, indicating that the storm counts are better modeled by some subsets of the region than others. Then, both quantile and linear regression were performed as a way to explore the data and find the best predictors possible. Quantile regression showed us that there is a significant (p < 0.05) positive correlation between all of the SST quantiles (not just the centroids) in certain clusters and the concatenated storm counts. While not immediately useful for prediction, these results confirm the significant relationship in that area and leave us further room for exploration in comparing the regression equations of various quantiles to determine if higher SST is better correlated with a faster increase in storm counts than is lower SST. Finally, for prediction, we found that regressing the log of the storm counts against the cluster centroids is most stable as the predictions are non-negative and have a much smaller range (determined experimentally) than do the cluster centroids regressed with the actual storm counts. After informal testing to find the cluster that yielded the best predictions, which involved comparing observed storm counts in 1982-1999 vs. predicted storm counts from the Climate of the 20th Century data (a simulated data set from the CCSM), our results matched what we would have expected from previous results. For the worst case climate change (A1FI data from the CCSM output), storm counts increased exponentially over the decades from 2000-2099, which could be devastating to infrastructures, even discounting the effects the rising global temperature would have on its own. For the best case climate change (B1 data from the CCSM output), storm counts stay relatively stable over the next century, which is indicative that climate change can be slowed and damage minimized.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner