Forecasting Monthly Tropical Cyclone Activities

1. Forecasting Monthly Tropical Cyclone Activities Lee Sai Ming Hong Kong Observatory, Hong Kong, China ESCAP/WMO Integrated Workshop on Urban Flood Risk Management in a Changing Climate: Sustainable and Adaptation Challenges 6 Sep 2010, Macao, China

2. The need for long-range TC forecast • To support decision-making • To support disaster prevention and preparedness planning

3. Availability of seasonal TC forecast • WMO Bulletin 56 (4) : Seasonal Tropical Cyclone Forecasts – a very comprehensive overview • Examples of forecasts: • No. of TC / named storms / ACE index over an ocean basin [no region-specific info.] • No. of TC landfalls [TC affecting a region/city does not need to make landfall there]

4. Typhoon Chanthu, July 2010 4

5. Typhoon Hagupit, September 2008 5

6. Heavy rain brought by Typhoon Chanthu >= 100 mm of rainfall 6

7. Flooding in Tai O after Typhoon Hagupit (courtesy of TVB)

8. TC activity affecting a region / city • Definition: No. of TC within a certain range and a certain period of time • Hong Kong: N500 [within 500 km of HK] • Long term mean of annual N500 ≈ long term mean of annual Nsig [issuance of warning signals]

9. Monthly N500 of Hong Kong HK TC season ~ June to October

10. July N500 of HK

11. Forecast formulation • N500 is a count parameter • Can be modelled by the Poisson distribution • The Poisson dist. belongs to the family of exponential dist. • Canonical form: where a(y)=y, b(θ)=log θ, c(θ)= θ, d(y)=-logy!

12. Generalized Linear Model (GLM) i = 1, 2, … N Yi = response variable [monthly N500] E = expectation of the dist. g = link function xi = explanatory variables or predictors β = model parameters What are the predictors ?

13. NCEP CFS • Climate Prediction Center, NOAA, USA: a WMO designated Global Producing Centre (GPC) of Long Range Forecasts • CPC provides digital long range forecast and hindcast [generated by the NCEP Climate Forecast System] • Hindcast used in this study: 1981-2008 • Variables: mslp, upper air wind, stream function, geopotential height, SST, vorticity, divergence … [26 variables]

14. Spatial coverage of data used Atmospheric variables and SST 10S – 50N, 90E – 150W Eq. Pacific SST: 15S – 15N, 150E – 80W

15. Data compression • Horizontal resolution of data: • 1 lat. x 1 lon. for SST • 2.5 lat. x 2.5 lon. for others • No. of data grid points >= 1225 [for each variable] • Compress data by EOF analysis • 1225 data points  28 principal components

16. Selection of predictors and combinations • Fit a single predictor GLM, search for skilful single predictor • Fit a multiple predictor GLM [predictors from step 1], filter out redundant predictors by stepwise regression [no. of combinations >= 2 x 108, hence randomly select a limited no. of combinations] Cross-validate the ‘reduced GLM’ [from step 2], search for top performers [a ‘brute force’ approach]

17. Cross-validation • Hide the observation of 1 year • Estimate the GLM parameter from the rest of the observations and the principal components • Verify the GLM forecast against the hidden observation • Rotate the process through 28 years Cross-validation result provides skill estimates for real-time forecasts.

18. A low-cost / no-cost tool

19. Performance comparison against the climatological forecastHindcat period: 1981-2008

20. Hindcast Vs Actual July N500 (1981-2008)

21. Physical Interpretation The 4th EOF of 500 hPa geopotential height of June

22. Physical Interpretation The 1st EOF of 850 hPa zonal wind of October

23. Multi-GLM combination • Weigel et al., 2008: Can Multi-model Combination Really Enhance the Prediction Skill of Probabilistic Ensemble Forecasts? Quarterly Journal of the Royal Meteorological Society • A message in respect of deterministic forecasts: combination of similarly skilful models can enhance prediction skill • Further exploitation of the cross-validation result

24. Performance comparison against the climatological forecastHindcat period: 1981-2008

25. Conclusion • Monthly TC forecast can be formulated in terms of GLM • Dynamical climate model (e.g. NCEP CFS) forecast data contain a lot of predictive information • The ‘brute force’ approach is viable in identifying skilful predictors and combinations • Further skill enhancement is made possible by multi-GLM combination

26. Remarks Too many predictors: inexhaustible combinations. Not all EOF can be easily interpreted. 26

27. Thank you Acknowledgement: The Hong Kong Observatory gratefully acknowledges NOAA/CPC for providing CFS forecast and hindcast data on the web to support research and seasonal forecasting operation conducted by the Observatory. 27

28. Table 6a. Top 20 GLM combinations for June. The number k after the underscore indicates the kth PC. Same convention for Table 6b to 6e. Meaning of the variable is given in Table 1.

29. Table 6b. Top 20 GLM combinations for July.

30. Table 6c. Top 20 GLM combinations for August.

31. Table 6d. Top 20 GLM combinations for September.

32. Table 6e. Top 20 GLM combinations for October.