System Design for Operational Ocean Forecasting

Gary B. Brassington

Abstract The scientific and technical advances in ocean modelling, ocean data assimilation and the ocean observing systems over the past decade have made the grand challenge of ocean forecasting an achievable goal with the implementation of the first generation systems (Dombrowsky et al. 2009). Implementation of these components into a truly operational forecasting system introduces a number of unique constraints that can lead to reduced performance. These practical constraints, such us the limitations in the coverage and quality of critical components of the ocean observing systems in real-time as well as the constraints of completing forecast integrations within a fixed schedule are unavoidable components for any forecast system and require additional strategies to achieve robustness and maximise performance. We begin by defining commonly used terms such as operational and forecasting in this context. We then review the design choices that can be taken with each component of an ocean prediction system when implemented as an operational system to achieve the most reliable performance.

18.1 Introduction

Operational ocean forecasting systems have been established over the past decade by several agencies and institutions (Dombrowsky et al. 2009). Hurlburt et al. 2009 provides an appraisal of key developments over this period. These systems employ a wide variety of techniques (Kamachi et al. 2004; Cummings 2005; Brasseur et al. 2005; Martin et al. 2007; Oke et al. 2005, 2008) largely due to the maturing state of the science. None of these techniques are theoretically optimal as defined by the use of a 4D variational scheme (Lorenc 2003) or an ensemble Kalman Filter (Evensen 2003). However, the computational cost of eddy resolving models which preclude the use of 4DVar and EnKF approaches, together with the poor knowledge of the

Centre for Australian Weather and Climate Research, Bureau of Meteorology, Melbourne, Australia e-mail: [email protected]

A. Schiller, G. B. Brassington (eds.), Operational Oceanography in the 21st Century, 441

DOI 10.1007/978-94-007-0332-2_18, © Springer Science+Business Media B.V. 2011

background error covariances that apply in the ocean has led to a wide variety of sub-optimal approaches being employed.

Guiding principles for good design can be found in many quotations of which we cite three. The first of these is referred to as the Law of the instrument and is attributed to Abraham Maslow, "When the only tool you have is a hammer, it is tempting to treat everything as if it were a nail". The law of the instrument is a warning to new scientists and engineers that need to work on improving existing systems that many of the design choices are based on the known methods and techniques at that time. All design choices are constrained by those methods and should be regularly questioned and reviewed.

The second quotation is a warning against reductionism and attributed to Albert Einstein, "Make things as simple as possible, but not simpler". All components of the ocean prediction system contain assumptions that reduce the problem into simpler elements that offer advantages e.g., methods of solution. All assumptions that reduce the parameter space of the system are true under defined conditions e.g., Boussinesq, hydrostatic and incompressible assumptions. A thorough knowledge of these assumptions and the conditions under which they hold is critical when re-applying methods or systems for new applications. Alternatively, all the advantages of new efficient method are of no use if they do not solve the target problem to within a required precision.

The third and final quote is the antithesis of the previous quote and again is attributed to Albert Einstein, "Any intelligentfool can make things bigger, more complex, and more violent. It takes a touch of genius—and a lot of courage—to move in the opposite direction". This quote is particularly relevant with present systems as the trend is toward higher model resolutions, more complex data assimilation methods, ensemble forecasting and coupled physical models. It serves to pause and justify before automatically introducing greater system complexity. This trend is scaling with the improvement in computing system performance and is likely to continue.

A good analogy for operational ocean forecasting design today is that of the chronometer invented by John Harrison (Sobel 1995). Take a visit to the museum in Greenwich, London and you will see an incredible piece of design/art called H1 (see Fig. 18.1a). This was designed by John Harrison to solve the Longitude problem by

Fig. 18.1 a The H1 clock and b chronometer designed by John Harrison to solve the longitude problem

producing a clock that could perform accurately at sea and claim the significant monetary prize. Anyone cannot help but admire the quality of the design and the achievement of its clock. However, this particular clock was abandoned by John Harrison after 17 years of development as he realised how he could improve to eventually arrive at a pocket sized device called the chronometer (see Fig. 18.1b). Operational oceanography today is analogous to the H1 where it functions as it was designed to, contains many novel and elegant solutions but remains far from where it will be over the coming decades in terms of its techniques and importantly its reliable performance.

In this paper, we begin by offering a definition for commonly used terms related to ocean forecasting specifically identify properties unique to operational forecasting. We then provide a short overview of applications for ocean forecasting and common servicing requirements influencing design. Section 18.4 introduces the system elements of an ocean forecasting system which is followed by an expanded discussion on each of these elements with particular emphasis on the properties of each component that influence the system design. This includes, Sect. 18.5 real-time observing system, Sect. 18.6 real-time forcing system, Sect. 18.7 modelling, Sect. 18.8 data assimilation, Sect. 18.9 initialization, Sect. 18.10 forecasting cycle, Sect. 18.11 system performance. Throughout we have highlighted aspects of an operational system that require design choices to be made and are of a general interest to system design. By way of demonstration, examples are drawn from specific systems with the cautionary note that these may or may not be general practice. A majority of the examples are drawn from the BLUElink Ocean Model, Analysis and Prediction System (OceanMAPS) which is noted throughout. We then end with a short conclusion.

18.2 Definitions

The initial development of all forecasting systems is performed under hindcast conditions (see Table 18.1). In many respects hindcasts frequently attempt to mimic the forecast environment however, many of the conditions that occur in real-time are difficult to reproduce and are not necessarily normally distributed e.g., drop outs in satellite products (see Fig. 18.2). Alternatively, it is often desirable to determine the statistical performance of a system operating under ideal conditions which sets the upper bound in performance. In practice, this level of performance only oc-

Table 18.1 Definitions of terms frequently used in reference to assimilated ocean model states Forecasting terminology

Hind-analysis Best estimation using optimal methods and maximum information Hindcast Behind real-time simulation of forecasts i.e., model initialisation from a hind-

analysis and model projection Hindcasts are typically performed under ideal conditions and represent an upper bound in forecast performance Nowcast Estimation of the state and circulation at real-time that can be used as a persistence forecast Forecast Prediction of the state and circulation beyond real-time


2 105


01/04 2009

11/04 21/04 01/05 11/05 21/05 31/05 10/06 20/06 Date (dd/mm)

Fig. 18.2 Observation retrievals from AMSR-E (-) ascending and (...) descending swaths obtained at the Bureau of Meteorology between 4th January 2009 and 30 June 2009

curs when the forecast conditions approach the ideal. In the design of a forecast system the performance of the system under less than ideal conditions is of equal importance. This frequently introduces additional strategies to minimise impact to achieve the highest lower bound. For this reason it is critical to use terminology of forecast and hindcast systems appropriately and to define the conditions of the system accurately.

The term operational is frequently used with a wide variety of working definitions but interestingly also has a specific philosophical heritage (see http://plato. A useful working definition was outlined during the develoment of EuroGOOS (Prandle and Flemming 1998). The term as it applies to operational forecasting is summarised here in Table 18.2 as relating

Table 18.2 Definitions for the meaning of operational as they apply to world meteorological agencies Operational

Real-time System and products targeting nowcast and forecasts Routine Performs to a regular schedule

Robust Technological: High-end computing and communications with designed fail-overs and fit-for purpose scheduling Scientific: Detect and mitigate changes in system state to ensure minimum impact to performance

Consistent Consistently achieving the designed performance to: real-time services delivered routinely and robustly. Many operational centres measure success against the delivery of services 24/7. A considerable amount of resources are expended in order to achieve the level of servicing of 99.99% up time typical of a WMO agency. Consistency of the quality of the services is also critical to design choices.

18.3 Applications

Prior to designing any system it is important to define the applications to be targeted for the system and to define the service requirements that need to be met. This is critical to both the design of the observing system and forecasting systems. However, operational oceanography has not rigorously followed this idealised approach. Operational oceanography was initiated as an experiment, Global Ocean Data Assimilation Experiment (Smith and Lefebvre 1997) motivated by the opportunity presented by the new and expanding global ocean observing system particularly with the introduction of satellite altimetry. The many sectors that could potentially benefit from ocean forecasting services were more or less known at that time. However, the specific applications and the forecast skill requirements were not known. There are several properties of the applications that will influence the design of ocean forecast systems and the impact of those services are summarised in Table 18.3. These include the type of application, its social or economic value, the sophistication of the user community and the service requirements.

A subset of the potential applications are represented in Fig. 18.3. Figure 18.3a, b represents an upwelling event that took place off the Bonney coast in South Australia on the 10th February 2008. Upwelling frequently impacts local marine ecosystems bringing nutrient rich water into the photic zone resulting in a chlorophyll bloom that is observable by ocean colour (Fig. 18.3b) on the 30th March 2008. Upwelling can also have a stabilising effect on the local atmospheric boundary layer reducing the transfer of momentum to the surface. Upwelling can occur very rapidly and therefore can be absent from atmospheric forecasts that persist SST boundary conditions. This specific event resulted in a forecast failure where strong winds were forecast but local observers experienced weak winds. This lead to a complaint by local tourist operators who had cancelled ocean cruises. Upwelling events can also be associated with sea fog which is difficult to observe using either infrared radiation (e.g., Advanced Very High Resolution Radiometer (AVHRR)) due to the presence of fog or microwave (e.g., Advanced Microwave Scanning Radiometer— EOS (AMSR-E)) due to the coarse resolution (~25 km/pixel) and interference near coastlines. A dynamical forecast is required to generate the cool SST's in response to the wind which can be observed however, the precision of the forecast SST's is difficult to validate.

Marine accident and emergency services from ships (Fig. 18.3c) and from oil wells (Fig. 18.3f) including the airborne and ship-based salvage operations (Fig. 18.3g) are an obvious application of ocean forecasting services. However, the

Table 18.3 Properties of the applications and user communities that impact the design choices in ocean forecasting systems



Ad hoc time and space

(e.g., Search and rescue, Marine accident and emergencies, defense)

Planning and management

(e.g., fisheries by-catch, marine park management)


(e.g., offshore oil and gas, ship routing, renewable energy)

Global and continuous

(e.g., weather, wave, ecosystem forecasting)

Coastal shelf

(e.g., ports management, bilge discharge, coastal surge)

Public good

(e.g., recreational fishing, diving, swimming, sailing)

Social and economic

Life, safety or security threatening


Property damage

Marine health

Economic value and energy

User community

Are the users structured and coordinated?


Are the service needs well-defined?

Are the impacts of ocean services understood?

Capacity to interpret ocean products and add value

Capacity to monitor and assess impacts

Capacity to engage in a relationship and provide usable feedback

Service requirements

Hindcasts, Short-, medium-, long-range forecasts

Performance thresholds

Sensitivity to error

Sensitivity to extremes

Observational requirements

Timeliness and frequency of forecast products

requirements for skilful Lagrangian trajectories has been difficult to achieve. The present and future global ocean observing system over the next decade is unlikely to be sufficient to meet the needs of these applications (Hackett et al. 2009; Davidson et al. 2009; Rixen et al. 2009; Brassington et al. 2010a). A characteristic of these events is that they occur infrequently at ad hoc locations and are localised making them suitable for short term, intense observing deployments through the use of gliders, AUV's and drifting buoys etc.

An atmospheric feature that is common to Australia, Brazil and the United States on their respective eastern coastlines is the formation of rapidly intensifying extratropical cyclones (see Fig. 18.3d). These storms are sometimes referred to as bombs due to their severity and impacts. The event in June 2007 made famous by the grounding of the cargo ship Pasha Bulka also resulted in loss of life due to flooding in Newcastle. On the east coast of Australia these storms form when a cut-off low of cold dry air moves over a warm moist marine boundary leading to vertical convection and a positive feedback in the atmosphere of convergence of the up-

Fig. 18.3 A collage of applications that require real-time forecast services a forecast SST's for a coastal upwelling event of South Austrlai. b The ocean colour response for the same vent, c Oil washed on shore off Queensland due to a leak from the Pacific Adventurer, d Modelled 10 m winds of an East coast cyclone off the NSW coast, e The modelled SST conditions associated with the event, f Oil discharge from the

Montara oil well, g Ship based salvage operations for the same event, h Surge along the Derwent river in -t-

Tasmania and i the forecast sea level for this event "-1

Fig. 18.3 A collage of applications that require real-time forecast services a forecast SST's for a coastal upwelling event of South Austrlai. b The ocean colour response for the same vent, c Oil washed on shore off Queensland due to a leak from the Pacific Adventurer, d Modelled 10 m winds of an East coast cyclone off the NSW coast, e The modelled SST conditions associated with the event, f Oil discharge from the

Montara oil well, g Ship based salvage operations for the same event, h Surge along the Derwent river in -t-

Tasmania and i the forecast sea level for this event "-1

per layer potential vorticity (Mclnnes et al. 1992). The ocean heat content along these coastlines is highly variable due to the turbulent western boundary current that transports warm/fresh water from the tropics to higher latitudes. The modelled SST shown in Fig. 18.3e exhibits a temperature front in the same position as the storm. The warm SST's were maintained by a warm-core anticyclonic eddy (Brassington 2010; Brassington et al. 2010b). An ocean forecast system can provide forecasts of the heat content conditions with potential for use in coupled forecasts.

High sea level along the coast is typically associated with the coincidence of tides and storm surge. Forecasting systems are typically based on so-called "storm-surge" models local to the event to estimate risk in combination with tides and sea level pressure. Simulations of non-tidal sea level in ocean forecasting systems can also be impacted by other oceanographic effects of remote coastally trapped waves and impinging warm boundary currents. For example, a high sea level event in the Derwent river (see Fig. 18.3h) resulted from a local storm and a large amplitude coastal trapped wave propagating from South Australia. A characteristic of the coastal trapped wave is the high sea level in the Bass Strait. (see Fig. 18.3i). Regional forecasters did not issue a warning due to their use of traditional methods of computing sea level which did not account for the remote contribution. Ocean forecast systems have the potential to provide total sea level forecasts.

In each of these applications the oceanographic conditions play an important role for which accurate forecasts can provide valuable information. Detailed analysis of these and other similar cases can identify the relevant oceanographic variables and the sensitivity to error to derive the requirements in terms of performance. In these examples, SST, heat content, surface currents and sea level are directly relevant which accounts for four of the five prognostic variables in a hydrostatic ocean general circulation model. Though it is important to note that their forecast are dependent upon the knowledge of all prognostic variables. National agencies and institutions are regularly engaged with local users and have opportunities to acquire this information. The JCOMM Expert-Team on Operational Ocean Forecast Systems (ET-OOFS) is tasked with providing international coordination to generalise this information into observational and service requirements.

18.4 System Elements

All operational ocean forecasting systems available today follow a similar sequential and cyclic structure which involves handling of the latest observational data, performing a model-data fusion, performing a model forecast to generate data products including ocean state estimates, performance diagnostics and error estimates. This sequential procedure is repeated on a regular schedule or performed in an ad hoc basis e.g., triggered by a specific event. The system diagram for the BLUElink OceanMAPS system is shown in Fig. 18.4. This includes retrieval and archival storage of observations, surface fluxes, model and data assimilation dependent data files.

Satellite data archival system

Profile data archival system

Bureau operational NWP products

Bureau operational SST products

Data servicing system

ODAS data retrieval system

OGCM data retrieval system

Satellite data retrieval system

Profile data retrieval system

Surface flux retrieval system

ODAS product archival system

OGCM product archival system

ODAS data retrieval system

OGCM data retrieval system

Satellite data retrieval system

Profile data retrieval system

Surface flux retrieval system

ODAS product archival system

OGCM product archival system

■ 1





CD £ a CO



Fig. 18.4 A schematic diagram of the system elements for an operational ocean forecasting system. (Based on the BLUElink OceanMAPS, Brassington et al. 2007)

A Supervisor Monitor Scheduler (SMS) developed at the European Centre for Medium-Range Weather Forecasts (ECMWF; see data/software/sms.html) or equivalent software, is implemented at operations centre, to control the job flow monitoring the successful completion of dependent system components. The data and file handling is performed on servers whilst the large memory and computationally intensive tasks for data assimilation and model integration are submitted to high-end super-computing systems. The performance of the computing environment and the level of optimisation that can be achieved with the software is critical to design of ocean forecasting systems. Eddy-resolving ocean forecast systems are at the high-end of supercomputing application both for the prognostic model and the data assimilation inversion. The total wall clock time and the computing resources available in an operations centre are limited and managed among several other forecast systems. The efficiency of the software and the consistency of the completion times for different components has important impacts on design. For example, the computational cost of a data assimilation system will scale with the size of the inversion problem. Targeting a reduction in cost may compromise the number of observations processed through super-observations or thinning, require the implementation of localisation or limit the specification of the background error covariance. Similarly the cost of ocean model design scales with the number of grid points/cells and the timestep constraint for numerical stability. Targeting a specific cost limit will compromise the horizontal/vertical resolution within the model or the area of high resolution.

The system described above accounts for the majority of the science and technical design for an ocean forecasting service. However there are several steps to the provision of a quality service to end users. These include infrastructure for robust data product dissemination, forecaster guidance as well as support services for specifying user requirements and evaluating impacts. These important steps are not discussed further here.

18.5 Real-Time Observing System

The global ocean is now observed by a growing number of instruments and platforms that each have specific properties, some common and some unique, that will impact the design in operational ocean forecasting. These properties are summarised in Table 18.4 and include the timeliness, coverage, expected errors and quality. The relative immaturity of ocean instrumentation and infrastructure leads to more frequent system failures in practice compared with numerical weather prediction. System failures are frequently random and unpredictable though the sensitivity of the forecasting system to failures in the observing system are measurable. Strategies to minimise the impact need to be considered in the system design. For a more detailed discussion on aspects of the ocean observing system refer to Le Traon (2011) and Ravichandran (2011).

18.5.1 In Situ—Profiles

The ocean state is routinely profiled in real-time by Conductivity-Temperature-Depth (CTD) sensors from traditional platforms such as ships and moorings and

Table 18.4 Properties of the real-time ocean observing system that result in unique design choices in ocean forecasting systems Real-time bserving system

Timeliness How close to real-time are the observations received?

Non-normal behaviour


Observation error estimation

Quality control

Are delayed products available with higher quality?

What is the minimum/maximum coverage?

How homogeneous is the coverage?

Instrument error

Representation error

Does the product include quality flags?

Valid tests for the observation error model

Instrument failures, communication and system failures relatively new platforms such as autonomous Argo floats and gliders, from volunteer ships. In addition expendable Bathy-Thermograph (XBT) are operated from volunteer ships and reported in real-time. The sampling by in situ measurements has significantly increased over the past decade and coverage has increased in regions that have been historically poorly sampled such as the Indian Ocean and Southern Ocean.

The Argo array is now the dominant source of in situ sampling having largely achieved its target density of one float per 3° * 3° over the global ocean with > 3000 autonomous floats. Each float profiles the ocean water column from ~2000 m to the surface every 10 days reporting in real-time at the surface via ARGOS or Iridium (Ro-emmich et al. 2010). A user guide to the range of Argo data products available and the server access points are given online ( html). Observations are retrieved by a network of Data Assembly Centres (DAC's) which are responsible for performing an automatic quality control procedure and distributing the observations to both the WMO Global Telecommunication System (GTS) and the two Global DAC's (GDAC's). The DAC's also perform an objective quality control in delayed mode. Profiles that pass the automatic quality control are reported in real-time to the GTS without quality control information in TESAC format. A fast mode product is available from GDAC servers within 3 days in a format that contains the quality control flags and native observations on pressure coordinates.

Other important CTD profiles are obtained from the mooring arrays, TAO/TRITON (Pacific; McPhaden et al. 2001), PIRATA (Atlantic; Bourles et al. 2008) and recently RAMA (Indian; McPhaden et al. 2009). These moored arrays report in real-time, multiple times per day and are reported onto the GTS. Increasingly gliders are being used to adaptively sample the ocean however, the data acquisitions are as yet not coordinated internationally in the same way as Argo and lack a common real-time quality control procedure, integration with the GTS and other DAC/GDAC product delivery.

XBT's have been maintained along specific ship routes and sampling is constrained by the frequency of the volunteer ships that occupy the route (Goni et al. 2010). XBT's provide high vertical resolution profiles of temperature and depth at regular spacing along the ship route. Profiles, subsampled in the vertical, are reported on the GTS without quality control flags. The profiles are subsequently subjectively quality controlled by a number of centres using a common set of procedures (Bailey et al. 1994).

As an example the number of profiles that were retrieved at the Bureau of Meteorology each day from the GTS and the two Argo GDAC's (Coriolis and USGO-DAE) between the 15 th January 2010 and 1st March 2010 are shown in Fig. 18.5. The GTS reports consistently ~1200 profiles per day although the most recent retrievals show an increase in the number of observations due to shallow coastal observations in the USA. The GDACS's report on average 300 profiles per day corresponding to the expected number of Argo floats surfacing each day. The number of profiles retrieved from each GDAC do not correlate and are clearly not simply a mirror site. Coriolis also frequently has bursts of profiles which largely contain old profiles that a DAC has subjectively QC'd. The best daily observations available in near real-time are obtained by sorting amongst the three sources. Ideally the three sources should contain a maximum of three duplicates for the same profile that must

In situ Profile

In situ Profile


15/01 20/01 25/01 30/01 04/02 09/02 14/02 19/02 24/02 01/03

Fig. 18.5 The number of ocean profiles retrieved daily between 15 January 2010 and 1st March 2010 from the GTS (purple), Coriolis (blue) and USGODAE (green) and the number of duplicate free profiles (red)


15/01 20/01 25/01 30/01 04/02 09/02 14/02 19/02 24/02 01/03

Fig. 18.5 The number of ocean profiles retrieved daily between 15 January 2010 and 1st March 2010 from the GTS (purple), Coriolis (blue) and USGODAE (green) and the number of duplicate free profiles (red)

be reduced to one. The best profile is determined to be the one that has both the most complete set of observations and the maximum set of quality control tests applied. The number of profiles obtained for each day from the duplicate checking procedure is shown in Fig. 18.5 in red. The best daily observations provides consistently ~1200 profiles per day. The decline in profiles near real-time shows the impact of timeliness of the profiles with a small percentage of the total profiles obtained several days behind real-time.

An algorithm developed at the Bureau of Meteorology (Brassington et al. 2007) to select the best profiles replaces profiles obtained from the GTS with more complete profile information, particularly quality control, obtained from the GDAC's. A typical example of the timeliness, volume and source of the profiles obtained from that system is shown in Fig. 18.6 for the 13th September 2009. Within the first two days of real-time, the number of profiles is dominated by those obtained from the GTS. Within the first day GTS profiles are being replaced by profiles from the GDAC's. In the 3rd and subsequent days profiles from the GDAC's continue to replace those obtained from the GTS. The number of profiles replaced declines as the time behind real-time increases.

Sea surface temperature is the most frequently observed ocean state variable by satellites with multiple sensors and multiple orbits. Microwave sun-synchronous and IR Geostationary platforms provide higher coverage whilst the IR polar orbiting missions provide the highest resolution and accuracy in cloud free conditions. There

Was this article helpful?

0 0
Guide to Alternative Fuels

Guide to Alternative Fuels

Your Alternative Fuel Solution for Saving Money, Reducing Oil Dependency, and Helping the Planet. Ethanol is an alternative to gasoline. The use of ethanol has been demonstrated to reduce greenhouse emissions slightly as compared to gasoline. Through this ebook, you are going to learn what you will need to know why choosing an alternative fuel may benefit you and your future.

Get My Free Ebook

Post a comment