Anticipatory modelling of hurricane damage
(1) Blandine L’Hévéder - RiskWeatherTech
(2) Gilles André - RiskWeatherTech
(3) Léa Boittin - Modelling, R&D Department, CCR
(4) Roxane Marchal - Modelling, R&D Department, CCR
INTRODUCTION
In September 2017, Hurricane Irma hit the northern Lesser Antilles, causing massive damage to Saint-Martin and Saint-Barthélemy. It falls into the category of the strongest hurricanes recorded in the North Atlantic. The insured damage on these two islands following Hurricane Irma is estimated at over €2 billion.
An initial estimate of the insured damage was made in the days following the hurricane, using models developed and calibrated by CCR on past events. In view of the amount of damage generated by this type of event, better forecasting would make it possible to alert the State and the insurance companies of the extent of the coming catastrophe. CCR therefore wants to set up an automated cyclone watch capable of generating alerts several days ahead of a cyclone approaching ultra-marine territory. This warning system should make it possible, when the most likely trajectory crosses an area covered by the Nat Cat scheme, to implement fine spatial scale simulations and to simulate wind speeds and precipitation totals. These fine-scale weather forecasts will allow the modelling of insured flood and wind damage up to 4 days ahead of a catastrophe, with a high confidence interval incorporating trajectory uncertainty.
The operational chain is divided into two phases: the first is an early warning system (EWS) which alerts CCR of a hurricane event where one of the possible tracks crosses an ultra-marine territory; the second, triggered if the warning is confirmed, is a downscaling of the hazard and a forecast of the damage insured at D-4 and then at D-1. This second phase is divided into two parts: the atmospheric simulation and the hazard simulation with the CCR models. Fine-scale atmospheric forecasts are made by RiskWeatherTech (RWT) via downscaling with the WRF model [1] from US weather model forecasts (GFS[2]). They provide the input data for the hazard models (runoff and wind), which are then cross-referenced with vulnerability to estimate damage. The territories covered are the islands of the Antilles (Saint-Martin, Saint-Barthélemy, Guadeloupe and Martinique), Mayotte, Réunion Island and Europe (to deal with any medicanes).
METHODOLOGY
The architecture of the operational chain for monitoring and modelling hurricane activity is composed of three units: (1) an Early Warning System (EWS); (2) in the case of a prolonged hurricane warning, a medium-resolution downscaling is carried out for two hurricane tracks, namely the most likely and the riskiest; (3) in the case of a reinforced warning, a high-resolution downscaling is carried out 24 hours before the hurricane hits the island.
The EWS runs daily in each ocean basin during hurricane season (from May to November for the Antilles and from October to March for Réunion Island). It is based on the use of ensemble forecasts[3] from the US atmospheric model, GEFS[4], consisting of a set of 30 weather forecasts.
Each morning, the forecasts of the 30 members of the GEFS are retrieved for the next five days and global risk maps are developed for flooding and high winds over the 5-day period (Figure 1). The wind risk threshold is set at 90 km/h. The flood risk threshold is set at a daily rainfall of over 120mm. Regions in hurricane season are selected and a wind or flood alert is issued if at least 10% of the members exceed the alert threshold in the territory. An email is automatically
Figure 1 – Description of the SAP
In case of prolonged cyclone warning, corresponding to 3 consecutive days of wind warning, phase 2 is triggered. Two of the 30 possible hurricane tracks in the GEFS forecast set are selected: the track with the highest risk for the territory and the most likely track. For each of these two tracks, downscaling is performed with the WRF atmospheric model with 4km resolution, over a 5-day period to cover the hurricane passing over the island (Figure 2).
WRF simulations are performed on a computing cluster and require the joint use of 288 CPUs. Rainfall and wind data output from the WRF simulation with 4km resolution are stored on an hourly basis and used by CCR to assess potential hurricane risk for the territory at D-4 (Example of Cyclone Batsirai on Réunion Island in Figure 3).
Finally, in the event of a heightened alert, the forecast of the hurricane’s track becomes relatively reliable 24 to 48 hours before it hits the island. A second downscaling simulation with WRF is performed, this time at a finer resolution of 1km, to be able to assess the insured damage with sufficient accuracy (Figure 4). The precipitation and wind data provided by this simulation are then integrated into the runoff and wind hazard models and then into the damage models to get first estimates of the expected loss amounts.
# cyclone warning
# WRF modelling
# hurricane
# precipitation
# wind
# insured losses
Figure 2 - WRF model configuration for phase 2. Since GEFS simulations are at 50 km resolution, the WRF configuration consists of two nested domains at 12km and 4 km resolution. The simulation on the 12 km resolution domain is used to force the final simulation on the 4 km resolution domain.
Figure 3 - WRF model configuration for phase 2. Since GEFS simulations are at 50 km resolution, the WRF configuration consists of two nested domains at 12km and 4km resolution. The simulation on the 12km resolution domain is used to force the final simulation on the 4km resolution domain. Simulation of the most likely track and the riskiest track at D-4 for Cyclone Batsirai, which came close to Réunion Island in February 2022.
Figure 4 - Simulation of the most probable path at D-1 for Cyclone Batsirai passing close to Réunion Island in February 2022.
Figure 5 - Comparison of the maximum gust values in the forecasts of the various models on D-1 ahead of Batsirai.
RESULTS
In February 2022, the operational chain was tested when Cyclone Batsirai skirted Réunion Island. At D-4, rainfall totals and wind speeds were modelled at 4 km resolution for the most likely trajectory and for the trajectory most at risk for Réunion Island (Figure 3).
On D-1, a second WRF simulation was performed at 1 km resolution (Figure 4), from which the hazard and damage models of CCR estimated damage between €33M and €45M for the most probable trajectory and the associated weather conditions.
On 2 February 2022, Reunion Island was placed under a cyclonic red alert at 4pm local time. This category 4 cyclone passed far enough north of Réunion Island to cause minor damage for a cyclone of this category.
It is however interesting to compare the GFS forecasts used in the operational chain with the forecasts of other available models: ICON is the German model, ECMWF the European model and ARPEGE and AROME the French models. 24 hours before the cyclone skirted Réunion Island, the tracks of the various models and the estimated intensity of strong winds and cumulative rainfall are still quite different (Figure 5).
CONCLUSION
This operational chain of hurricane activity monitoring, implemented in collaboration with RiskWeatherTech, makes it possible to follow the evolution of hurricanes in real time in the various ocean basins covered by the system in place. This process, which is still in the testing phase, should make it possible to anticipate the areas most affected by wind and heavy precipitation (runoff) hazards and the resulting damage. CCR’s objective is to forecast a major event and alert the public authorities and insurers as early as possible. Because of the uncertainty in the tracks modelled several days in advance, this system, once in place, must be tested and validated before being used operationally./
THE PARTNERS
RiskWeatherTech and CCR are partners, namely in projects related to modelling the consequences of climate change on insured losses. As for the independent consulting firm Atmoterra, this project initiated a first collaboration during the set-up of the project ‘Actuarial modelling of financial losses from Flood Risk in Morocco - Selection No. 1267599, World Bank’ and its development.
REFERENCES
1. WRF (Weather Research and Forecasting Model) is a mesoscale model developed by the NCAR.
2. GFS (Global Forecast System) is the weather forecasting model of the US National Weather Service.
3. Ensemble forecasting is a weather forecasting method using numerical forecasting models, which consists of perturbing the initial state of the model with perturbations smaller than the observation errors. Indeed, the accuracy of the calculations made by a numerical model decreases the further away the calculations are made from the time of the observation records: thus the “gross” predictability limit of a model is reached, which is currently about 3 days; it is to forecast the weather beyond this limit that the ensemble forecasting method is used, which assesses a confidence index of the forecast and provides a set of possible weather conditions beyond 3 days.
4. GEFS (Global Ensemble Forecast System) is a weather model created by NCEP that generates 30 different forecasted, called ensemble members, to estimate the uncertainties associated with atmospheric modelling.
CITATION
L’Hévéder, B., et al, Cyclone warnings and forecasting insured damage. In CCR 2022 Scientific Report; CCR, Paris, France, 2022, pp. 54-57