Real-time storm tide forecasting for emergency management in Queensland, Australia Dr Joanna M. Burston 1, Dr Andrew M. Symonds 1,2, and Prof Rodger Tomlinson 1 1 Griffith Centre for Coastal Management, Griffith University, Southport, Qld, AUSTRALIA 2 SMEC Gold Coast, Nerang, Qld, AUSTRALIA email: j.burston@griffith.edu.au Abstract Storm surge presents the greatest hazard to life in tropical cyclone (TC) events, and the Queensland (Qld) coastline has particular vulnerability to this hazard. Griffith University has initiated a Qld State Governmentfunded research project in partnership with Emergency Management Qld (EMQ), Qld Cyber Infrastructure Foundation (QCIF) and DHI Australia investigating the potential for real-time forecasting of storm tide during TC events for emergency planning purposes. Here we present a feasible probabilistic real-time storm tide forecasting system using as input existing TC track forecast products available from the Australian Bureau of Meteorology (BOM) and outputting high resolution probabilistic maps of storm tide inundation along the Qld coastline. A method to generate an ensemble of TC tracks suitable for storm surge simulation capturing the range of spatial, temporal and intensity uncertainties in the TC forecasts is presented. The system then uses this track ensemble to generate wind and pressure field which then force a calibrated and validated hydrodynamic model of the Coral Sea. The Coral Sea model was established using DHI s MIKE21 model for high performance computing (HPC) implemented on Griffith University and QCIF s HPC facilities and has been optimised for realtime performance. The proposed real-time system is found to be feasible from a technical perspective and would add value to Qld emergency management decision making. Keywords: storm surge, tropical cyclone, numerical modelling, high performance computing, emergency management. investigating the potential for real-time forecasting of storm surge during TC events using numerical modelling. The aim of this pilot project is to develop a system capable of real-time storm surge forecasting based on an ensemble approach using dynamic wind and ocean models outputting high resolution probabilistic estimates of storm surge water levels along the eastern Qld coast. 1. Introduction During a tropical cyclone (TC), storm surge water levels in combination with the astronomical tide can rise higher than the roof level of a single storey home, and in combination with high flow velocities and wave effects, create a high risk of drowning for people not evacuated from low-lying areas. The Qld coastline has particular exposure to this hazard, and while significant storm surges impacting major population centres have been rare in recent years, they have occurred historically, for example in Townsville (1971) and Mackay (1918). The current emergency management response to an impending storm tide hazard in Qld is evacuation to higher ground. The lead time for effective evacuation is 24 to 48 hours prior to the predicted onset of gale force winds, depending on the local setting. Forecasting storm surge during a TC event is complicated by the wide uncertainties in forecast TC track and intensity, particularly the timing of the peak storm surge relative to the phase of the tide. The current storm tide warning system operated by BoM is deterministic, and the storm tide estimates tend to adopt a worst-case scenario until there is greater certainty about the time and position of landfall. Emergency managers do not currently have a view of the range of scenarios that could result at different points in the forecast and a probabilistic forecasting approach is desirable. Griffith University has initiated a research project in partnership with EMQ, QCIF and DHI Australia 2. Background 2.1 Qld s Current Storm Tide Warning System The existing storm tide warning system operated by BOM consists of matching the deterministic consensus forecast TC track to a library of pre-run TC simulations made using the MMU-SURGE hydrodynamic model forced with single Holland (1980) parametric cyclonic wind fields (Harper et al. 2001). The resultant storm surge height is then added to the highest tide of the day until ~12-24 hours prior to landfall (when there is more certainty in landfall time and location), when the appropriate predicted tide magnitude is used. Potential wave set-up levels are considered by applying offshore forecast wave height and period estimates to the empirical equations of Hanslow and Nielsen (1993) (pers. comm.. B. Gunn, BOM, 2013). 2.2 Identifying Potential Improvements While the current system has the advantage of timeefficiency, workshopping with disaster managers at EMQ and coastal local councils identified a range of potential improvements that would be useful to the end-users of an updated storm tide warning system:
More continuous spatial output: The sparsely dispersed warning sites along the Qld coastline makes interpretation of potential onshore effects difficult. Currently, EMQ Spatial Science officers create surfaces corresponding to the estimated storm tide levels and intersect these with the terrestrial digital elevation model (DEM) in a bath-tub inundation approach to identify pre-defined evacuation zones in 0.5 m increments above highest astronomical tide (HAT). Probabilistic output: Real-time decision-making would benefit from an understanding of the worst-case and most-likely scenarios, especially with regard to timing of storm surge with high tide; and Dynamic time lines: emergency managers would find a view of when locally-defined trigger points would be expected to be inundated useful. Other system requirements included: Timing: Initial output is required earlier than the evacuation timeframe of 48 hours prior to the predicted onset of destructive winds; Forecast turn-around within 1 hour of TC track updates; and Consistency in use of vertical datums and terminology, especially clarity over the application of wave set-up. Extremes: potential storm surges > 0.5 m HAT should be identified. While a conservative approach of adding a storm surge height to the predicted high tide may be warranted given the high degree of uncertainty in the timing of TC landfall, arguments made by disaster managers against over conservatism in warnings include: Limited availability of resources to be deployed in an emergency (e.g., emergency personnel to conduct house-to-house evacuations); A significant increase in the number of properties predicted to be affected by inundation when increasing warning levels by a small magnitude; The Boy who cried wolf effect where a proportion of people fail to evacuate having previously experienced a false warning; and The risks to the general population associated with evacuation (heavy vehicle traffic, panic, etc.) that may outweigh the risk posed by a hazard with low certainty. Hence, this project aims to investigate whether some of these needs can be met by extending the current storm tide warning system towards probabilistic real-time output. 3. Proposed Storm Tide Forecasting System Here we present the framework of a dynamic storm tide modelling system capable of generating realtime probabilistic forecasts using TC track information available to Australian forecasters. It consists of: (1) a method to select a probabilistic ensemble of TC tracks suitable for simulating storm tide based on TC intensity, scale, and landfall location; (2) an extension of existing wind field generation techniques using the available TC parameter information; (3) a calibrated and validated hydrodynamic model optimised for realtime forecasting; (4) a method to combine storm surge water levels with probabilistic tide estimates and seasonal water levels; and (5) a probabilistic inundation mapping technique enabling delivery of inundation forecasts to emergency managers. Each of these components is discussed in detail below. 3.1 TC Ensemble Selection The size of the TC ensemble necessary to capture the range of storm surge possibilities at a given point in a forecast informs subsequent development of the wind field and hydrodynamic modelling framework by providing guidance on the model runtime and computational resources necessary to complete such simulations within the desired timeframe. Ascertaining the size and composition of this TC ensemble requires an understanding of the sensitivities of the predicted storm surge to the various input parameters defining the TC wind fields. In terms of available probabilistic TC information in forecast mode, BoM operationally generate a 1000- member stochastic set of tropical cyclone tracks for the purposes of generating wind hazard probability maps (pers. comm.., Dr M. Foley 2013). This set of tracks is based on the consensus forecast track, and created using the method of DeMaria et al. (2009) [4] by sampling the errors in cyclone parameters from distributions computed from the errors between the consensus and ultimate best track at different forecast periods for cyclone events in the Australian region over a 5-year period. The cyclone parameters considered include: track location (calculated in along-track and cross-track components), intensity (Maximum Sustained Winds (MSW)) and scale (given as Radius of Maximum Winds (RMW) and radius to 34, 48, and 64 knot winds (R34, R48, R64) for four quadrants). A similar method is used operationally at NOAA in the USA and used as input for their p-surge model [9]. A sample of this ensemble for TC Yasi at 48 hours from landfall is shown in Figure 1.
Figure 1 Example of the DeMaria 1000-member TC track ensemble for TC Yasi at 48 hours from landfall, showing considerable spatial spread in potential landfall locations along the Qld coast (black) (courtesy M. Foley, CAWCR). There are several disadvantages with this stochastic track ensemble for storm surge simulation given that it was not developed for this purpose, including: The error distributions are climatological and there is no account for current physical processes that may constrain the range of track possibilities, e.g., sea surface temperature (SST) and upper level steering; The independence of the along-track/crosstrack error in forecast cyclone position means that a change in direction in the forecast TC track leads to unrealistic changes in the forward speed. While having little influence on wind probability calculations, this would prove problematic for storm surge simulation; The size of the ensemble, while found sufficient to ensure convergence of the Monte Carlo approach for wind probability estimation [4], is likely insufficient to generate convergence of probability estimates for the parameters of interest in storm surge simulation, given the sensitivity to intensity, landfall location and timing. Despite these disadvantages, this operational product was identified as the most suitable for application to this project. If an expanded or modified ensemble becomes available, the ensemble selection approach outlined below would still be applicable. Now we develop a method to reconstitute a set of representative tracks from the probability distributions of the parameters important for storm surge derived from this 1000-member ensemble with associated probabilities of occurrence for use in real-time probabilistic storm surge simulation: 1. For each ensemble member in the forecast period, the track is intersected with the eastern Qld coastline and the scale (RMW) and the time and position of landfall (crossing time and position) are calculated. 2. The proportion of tracks making landfall in each coastal segment spanning 0.5 of latitude along the eastern Qld coastline from 15 S to 29 S is calculated. This length of coastal segment was chosen as a balance between the length of the coastline and the scale of cyclone s RMW (typically 20-50 km). 3. For each coastal segment, the empirical probability distributions of crossing time and RMW are determined. 4. To make the new TC track ensemble for storm surge simulation, a set of tracks comprising those representing the range of variability present in the forecast is assembled. The size of the resulting ensemble, n, is: n = n cl x n i x n RMW, (1) where n cl, n i, and n RMW are the number of realisations necessary to appropriately represent the empirical probability distribution of crossing location (cl), intensity (i), and scale (RMW), respectively. Crossing time (ct) is considered offline via linear addition with the estimated tide (see Section 3.4). In more detail, for each coastal segment, we select n i intensities where n i = (i max i min )/ i where i max and i min are the maximum and minimum intensities and i the interval in intensity necessary to capture the sensitivity in response to that parameter in each coastal segment. Since there is no relationship between MSW and the track location over the ocean in the DeMaria ensemble formulation, intensities at landfall are computed from the probability distribution of MSW in all tracks over the ocean at the mean time of crossing for that coastal segment regardless of landfall location. The intensity is linearly adjusted along the track from its estimated intensity in its current position to the required value at landfall. We then generate n i TC tracks passing through the landfall location. Ensemble members with maximum wind speeds lower than gale force are discarded. This process is then repeated to determine a set of n RMW RMW values using the RMW value at landfall. The probabilities associated with each ensemble member are also calculated. The ensemble tracks are recomputed every forecast output period. As the cyclone progresses closer to the coastline and the range of landfall locations narrows, the spatial resolution of the coastal segments is decreased. This approach could be further extended by considering asymmetry in scale as characterised by the R34 values given in the DeMaria set. This option is not considered further in this paper although remains under investigation.
As part of this study, an extensive set of sensitivity analyses will be tested using the TC wind field and hydrodynamic models described in the following sections in order to identify the values of i and RMW necessary to capture the range of possible storm surge heights for each coastal segment. The robustness of the proposed ensemble selection for probabilistic storm surge forecasting would be best tested through comparison of the resulting probability storm surge surface with that generated using a very large (~1 million) set of simulations based on an expanded DeMaria set. As an example, the resulting TC ensemble for TC Yasi selected using this method with the DeMaria stochastic ensemble 48 hours prior to landfall as input contains 282 members assuming i=10 m/s and RMW=5 km. In terms of operational run-time, each simulation would be required to run in (X*45*60)/n seconds per core for X computing cores (e.g., 282 simulations on 32 cores of computing resources = 306 sec per run per core) to achieve a one-hour turn-around from each TC track forecast update assuming 45 min for simulation and 15 min for pre- and post-processing. Such an example is achievable with the hydrodynamic model established for this study (see Table 2). Approaches for ensemble selection for more complicated cases, including non-landfalling cyclones, cyclones crossing from the land out to sea and offshore islands, are under development. 3.2 Wind field generation In order to generate wind and pressure field inputs for storm surge modelling for each TC ensemble member, an extension of the Holland et al. (2010) single vortex wind profile has been implemented. The radial pressure/wind speed model is extended to time and azimuth dimensions, and the radius to gales is fixed using the average of the R34 quadrant values given in the DeMaria ensemble. The Holland et al. (2010) wind profile model [8] has an advantage over the presently used Holland (1980) [7] model by reducing underestimation of the wind speeds in the tails of the wind profile and also reducing sensitivity to input RMW. It also has the potential to use dynamic environmental information such as SST. Since the DeMaria ensemble contains information on the scale of the TC ensemble members via estimates of the radius to 34, 48, and 64 knot winds for four quadrants, there is potential to extend the wind field model to account for radial asymmetry. This has been neglected at this stage due to the multiplicative effect on the number of ensemble members and the wide range of R34 values at landfall in each coastal segment. 3.3 Hydrodynamic model MIKE21 FM A calibrated and validated hydrodynamic model optimised for real-time forecasting of storm surge due to tropical cyclones has been established using DHI s MIKE21 HD Flexible Mesh model for high performance computing (HPC) and implemented on Griffith University and QCIF s HPC facilities. The model domain is the Coral Sea (Figure 2), calibrated for the astronomical tide for 13 measurement sites from the Gold Coast Seaway (27.95 S) to Cooktown (15.45 S). The most recent bathymetry was sourced from multiple local, state, federal agencies and industry, and adjusted to AHD. Figure 2 Extent of the Coral Sea model and bathymetry. Table 1 Sample comparison of modelled and measured tidal constituent amplitude and phase for the Coral Sea model (extracted from [1]). Location Tidal Cairns Mackay Const Mod Meas Diff Mod Meas Diff (a) Amplitude (m) Q 1 0.02 0.03 0.00 0.03 0.04-0.01 O 1 0.13 0.15-0.02 0.17 0.20-0.02 P 1 0.09 0.09 0.00 0.11 0.12 0.00 K 1 0.28 0.31-0.03 0.36 0.39-0.03 N 2 0.20 0.19 0.01 0.40 0.41-0.02 M 2 0.64 0.58 0.06 1.63 1.67-0.04 S 2 0.33 0.34-0.01 0.57 0.61-0.04 K 2 0.10 0.09 0.01 0.10 0.09 0.02 (b) Phase ( ) Q 1 132 130 2 131 124 7 O 1 150 154-3 146 147-2 P 1 190 187 3 184 183 1 K 1 190 191-1 184 187-2 N 2 260 256 4 304 305-1 M 2 280 276 3 322 322 0 S 2 258 248 11 327 321 6 K 2 254 245 9 323 319 4
The model consists of ~87 000 elements with arc lengths varying between 150 m and 50 km. Sample tidal calibration plots are shown in Figure 3 (additional calibration data provided in [2]). A comparison of modelled to measured tidal constituents shows good agreement in both amplitude and phasing (sample in Table 1). The model has been validated against several historical storm tide events [3] and shown to be suitable for modelling storm surge magnitudes (sample in Figure 4). (Figure 5). The model run-time required to accurately simulate the storm surge in this event is 130 sec/core (Table 2), which is sufficient to meet the required turn-around time using only 16 computational cores. Table 2 Example model run-time in seconds (speed-up factor) across for MIKE 21HD FM on Griffith University s HPC facilities. Run No. Cores Domain Length 1 2 4 8 16 32 Coral Sea TC Yasi 28 days 24 hrs 81717 (1) 130 (1) 26468 (3.1) 68 (1.91) 13827 (5.9) 40 (3.25) 7468 (10.9) 24 (5.42) 3660 (22.3) 16 (8.13) 2018 (40.5) 18 (7.22) Figure 3 Modelled (black) and predicted (red) water levels at four model calibration sites for Dec 2010 to Jan 2011. Figure 5 Sample model domain for TC Yasi from Townsville to Cairns showing the storm surge water level at landfall using Holland et al. (2010) wind/pressure field. Figure 4 Measured tidal residual (blue) vs modelled storm surge (red) (Holland et al. 2010 winds) for (a) Clump Point, (b) Cardwell, (c) Lucinda, and (d) Townsville for TC Yasi. Times in UTC. Model run-time is substantially reduced by selecting an appropriate domain extent. A pre-defined library of meshes has been created and a tool to select the appropriate simulation domain based on the forcing TC area of influence is under development. An example of such a model domain is shown for the best track for TC Yasi containing 6742 elements 3.4 Storm tide estimation The output of the storm surge model simulations of TC events is the storm surge water level at discrete output points located at 1-2 km spacing along the eastern coastline of Qld in populated areas. Simulations are made without a tide given the uncertainty in forecast TC track timing, so n ct tidal water levels are added onto the each member of the simulated set of storm surge water levels. This makes the assumption that storm surge and tide are linearly additive. Using the calibrated Coral Sea tidal model, a predicted tide time series for each of the model output points is generated. For the standard port sites, the measured tidal constituents are used to predict the tide. For each member of the DeMaria ensemble, the TC crossing time is used to select the corresponding tidal water level at each output point. The empirical probability distribution of the tidal water level at each output point is then calculated from which n ct realisations are drawn and combined with the storm surge ensemble results.
3.5 Probabilistic inundation mapping Probabilistic storm tide heights at each model output point are calculated using the simulated storm tide heights and probabilities associated with each ensemble member. These probabilities can be expressed either as the probability of storm surge greater than X m, or surge magnitude (m) that is exceeded by Y % of ensemble members (as for NOAA s p-surge model [9]). Both outputs were identified as useful for disaster managers by EMQ. However, EMQ wish to consider the total ocean water level wave set-up where applicable. While wave set-up is outside the scope of this study, functionality to include a wave set-up height estimate is included. This could be used to map inundation in a predefined open coast zone. The resultant total ocean water level heights would then be delivered to EMQ to map inundation using a bath-tub approach in GIS software. A storm tide surface for each probability level is created by interpolating between the model output points, and then intersected with the terrestrial DEM (1 m resolution derived from LiDAR supplied by the Qld Department of Community Services) to calculate inundation depths. The resultant probabilistic inundation depth extents and surfaces can then be intersected with the pre-defined evacuation zone maps by EMQ and local emergency managers to identify areas requiring evacuation orders. 4. Summary This paper describes a workable system for probabilistic real-time forecasting of storm tide heights and mapping of inundation. An approach for building a TC ensemble suitable for storm surge simulation based on available forecast information is introduced, as are a wind/pressure model, a hydrodynamic model and a probabilistic inundation mapping technique. The optimised hydrodynamic model has been established on high performance computing and yields run-times for TC events suitable for this ensemble modelling approach. A range of sensitivity experiments to test the required resolution of the various parameters making up the ensemble are in progress. Such a probabilistic system will enable disaster managers to have an understanding of the likelihoods of possible outcomes from an impending event. While the high uncertainty in current TC forecasts may not lead to a decrease in the conservative nature of storm tide predictions (for instance, by refining the TC crossing time with the phase of the tide), an improved understanding of the spatial variability in landfall locations is useful. Any improvements to TC forecasting will necessarily refine the output of this storm tide modelling system. The uncertainty present in the accuracy of TC forecasts would not warrant efforts to refine the storm surge modelling for other uncertainties or neglected processes such as storm surge and tide dynamics, the influence of wave-induced radiation stress or wave-dependent variations in wind drag coefficients etc. Future directions include extension into storm surge inundation modelling in populated areas (currently underway) to test whether the accuracy derived from such high resolution modelling is warranted against the computational expense. The wave setup inundation component also remains an outstanding issue for future work. 5. References [1] Burston, J.M. and Symonds, A.M. (2013a). Coral Sea Hydrodynamanic Model Set-up and Calibration. Technical Report 140. Griffith Centre for Coastal Management. [2] Burston, J.M. and Gee, D. (2013b). Coral Sea Wind and Hydrodynamanic Model: Event Validation. Technical Report 142. Griffith Centre for Coastal Management. [3] Courtney, J. and Knaff, J. (2009) Adapting the Knaff and Zehr wind-pressure relationship for operational use in Tropical Cyclone Warning Centres. Australian Meteorological and Oceanographic Journal 58.3:167-179. [4] DeMaria, M., Knaff, J.A., Knabb R., Lauer, C., Sampson, C.R., DeMaria, R.T. (2009). A new method for estimating tropical cyclone wind speed probabilities. Weather and Forecasting 24.6:1573-1591. [5] Hanslow, D. & Nielsen, P. (1993) Shoreline set-up on natural beaches, Journal of Coastal Research SI 1-10. [6] Harper, B.A., Hardy, T.A., Mason, L.B., Bode, L., Young, I.R. and Nielsen, P. (2001). Queensland climate change and community vulnerability to tropical cyclones, ocean hazards assessment. Stage 1 report. Department of Natural Resources and Mines, Queensland, Brisbane, Australia, 368 pg. [7] Holland, G.J., 1980. An analytical model of the wind and pressure profiles in hurricanes. Mon. Wea. Rev. 180, 1212 1218. [8] Holland, G.J., Belanger, J.I., and Fritz, A. 2010. A revised model for radial profiles of hurricane winds. Mon. Wea. Rev. 138: 4393-4401. [9] Taylor, A.A., and Glahn, B. (2008). Probabilistic guidance for hurricane storm surge. In: 19th Conference on probability and statistics. Pp. 7-4.