Potter T.D., Colman B.R. (co-chief editors). The handbook of weather, climate, and water: dynamics, climate physical meteorology, weather systems, and measurements

Подождите немного. Документ загружается.

CHAPTER 32

MODERN WEATHER FORECASTING

LAWRENCE B. DUNN

1 INTRODUCTION

It has often been said that weather forecasting is part science and part art. Essentially,

weather forecasting is the application of science to a very practical problem that

affects the lives and livelihoods of people and nations. How the science is applied

varies somewhat from forecaster to forecaster, and this subjectivity is the aspect of

weather forecasting that is still a bit of an art form. Many meteorologists and atmos-

pheric scientists are uncomfortable with the subjectivity of weather forecasting since

it inevitably leads to inconsistency, and there is a never-ending attempt to provide

forecasters with tools to make the process as objective as possible. Given the preced-

ing discussion, it should be clear that, like forecasting itself, any description of

modern weather forecasting will vary, at least slightly, according to the experiences

of the author.

Modern weather forecasting is applied to a broad spectrum of scales and applica-

tions. Forecasting is done by government and private meteorologists for a variety of

users. Forecasts of general conditions for each of the next 7 days are widely distri-

buted and are familiar to virtually everyone. In the United States, forecasts and

warnings to protect life and property are provided by government meteorologists,

although private-sector forecasters will often add value to these products, especially

through dissemination of the products. Speciﬁc forecasts for commercial interests

are done by private-sector meteorological ﬁrms. Forecast applications are numerous.

Examples include weather predictions for aviation, the marine community, manage-

ment of wildﬁres and prescribed burns, ﬂood control, reservoir management, utility

power demand, road surfa ce condit ions, agriculture, snowmaking at ski resorts, the

retail industry, commodity traders and brokers, the ﬁlm industry, special outdoor

Handbook of Weather, Climate, and Water: Dynamics, Climate, Physical Meteorology, Weather Systems,

and Measurements, Edited by Thomas D. Potter and Bradley R. Colman.

ISBN 0-471-21490-6 # 2003 John Wiley & Sons, Inc.

677

sporting events, and many others. The increased skill of modern weather forecasts

has resulted in more and more decision makers in all ﬁelds basing high resource

commitments on weather predictions.

Early in the twentieth century weather forecasting was based on the collection of

surface observations of wind, temperature, pressure, dew-point temperature, clouds,

and precipitation. The observations were plotted and analyzed by hand. By the 1930s

rawinsonde observations of the upper atmosphere began to supplement the surface

observation network that was the backbone of operational meteorology. By mid-

century the front al theory of the time, often referred to as the Norwegian Frontal

Model because of its origins, was being applied as a unifying concept to interpret the

observations and to guide analysis. Numerous empirical rules existed that related

clouds, precipitation, and frontal evolution to the observations. To a great extent

these rules described how to prognosticate weather systems through extrapolation.

The nonlinear nature of atmospheric processes limits the utility of extrapolation

techniques, and the problem of cyclone development and decay were addressed in

the 1940s and 1950s by application of theoretical ‘‘development’’ equations that

related upper-level vorticity and thermal advection to divergence and ultimately to

changes in surface pressure. The quasi-geostrophic (QG) approximation, basically

stating that the wind velocity remains in approximate balance with the pressure

gradient, was invoked as part of these development equations. To apply these

concepts to weather forecasting, vorticity charts were constructed via manual graphi-

cal methods that were quite laborious. The goal of all these efforts was to predict the

future position of cyclone centers and the attendant fronts as accurately as possible

and then to derive the sensible weather that was characteristically associated with

each weather system.

By the late 1950s and 1960s numerical weather prediction using recently devel-

oped computers became skillful enough that humans were no longer preparing

prognostic charts by hand, and forecasting instead focused on interpreting the

output from the various computer models of the atmosphere. This continues to

this day and will be discussed in detail later in this chapter.

The 1950s and 1960s were also the time when remote sensing of the atmosphere

via radar and satellite began to have an impact on weather forecasting. In particular,

these types of observations becam e primary tools along with the conventional

surface observations to make forecasts and warnings of weather phenomena on a

short temporal scale (0 to 6 h) and a spatial scale of approximately towns and

counties. Remote-sensing technology continued to evolve through the remainder

of the twentieth century, and short-term warnings and forecasts of severe and signiﬁ-

cant weather have become one of the primary activities of operational weather

forecasters.

Modern weather forecasting has evolved as a result of changes in observational,

computational, and communications technology. Increases in computational

resources have made a signiﬁcant impact on the complexity of numerical weather

prediction, but, as importantly, also on display capabilities where forecasters view

model output and observational data. Perhaps more important to the evolution of

weather forecasting than the changes in technology are the application of new

678 MODERN WEATHER FORECASTING

knowledge from research results that examine the fundamental structures and

processes of atmospheric phenomena on all scales. Research has led to the devel-

opment and application of numerous conceptual models of atmospheric phenomena

that provide a framework for forecasters as they attempt to interpret the large volume

of model output and observational data now available.

The rest of this chapter will be divided into three parts. A discussion of the

forecast process will be followed by sections on long-term forecasting and short-

term forecasting. The distinction between short-term and long-term will be made

based on the time scale when numerical weather prediction begins to become the

dominant tool used by the forecaster. At the time of this writing, this transition takes

place between 6 and 12 h. For the purposes of this chapter, long-term forecasting

covers the period from 12 h through 7 days. Forecasting at ranges longer than 7 days

is beyond the scope of this chapter.

2 FORECAST PROCESS

A prediction of the future state of the atmosphere requires that a forecaster must

understand what is taking place in the present. This fundamental ﬁrst step is required

regardless of the spatial or temporal scale of the prediction. This step is required

even if the basis for the prediction is output from a numerical weather prediction

model for reasons that will be descri bed in the long-term section below.

As part of the process of understanding the present state of the atmosphere, a

forecaster must pose and then answer a series of questions. Examples of some of the

more obvious questions might be: Where is it raining? Where is it not raining? Why

is it raining in one location but not another? How are the clouds distributed horizon-

tally and vertically? How has this been changing with time? Where are the centers of

high and low pressure? Where are the fronts? How are the troughs and ridges aloft

distributed and how have they been evolving? Where are the temperature gradients at

the surface and aloft and how are they changing? A quick look at an animated

sequence of satellite or radar images will bring these and dozens of other questions

quickly to the forefront of the problem to be solved by the forecaster.

This part of the forecast process is called analysis. During analysis, a forecaster

will examine both direct and remotely sensed observations and attempt to answer the

many questions that, in the aggregate, lead to an understanding of the present state of

the atmosphere. Analysis of directly observed parameters is often done objectively

by software, with graphical output displayed on a computer. Analysis is also done by

hand, with a forecaster carefully scrutinizing a plot of observations with a pencil in

hand, drawing contours and making annotations of signiﬁcant features. Skillful hand

analysis can provide a forecaster with insight that is no t likely to be brought out by

objective analysis routines, but the utility of hand analysis is dependent on the skill

of the analyst.

Analysis of remotely sensed data is also performed to ascertain the current state

of the atmosphere. The most common example of this type of analysis is the

examination of satellite imagery. The three most commonly available satellite data

2 FORECAST PROCESS 679

types are infrared (IR), visible (VIS), and water vapor (WV) imagery. Other types of

satellite imagery exist that detect precipitable water, rainfall rates, ozone, and other

parameters, and these are seeing increased usage in forecasting applications. Anima-

tion of WV imagery quickly shows a forecaster the positions of mid and upper-level

circulation features such as troughs, ridges, closed cyclones, and anticyclones, as

well as areas of ascent and subsidence aloft. Animation of IR imagery shows devel-

opment and decay of cloud systems by changes in cloud top temperature with time.

Visible imagery, which is available only during sunlit hours of the day, shows

regions of low and mid-level clouds, as well as regions of fog and snow cover. Visible

imagery can also be used to identify areas of convection, especially at times of low

sun angle, when convective clouds cast shadows. Forecasters use animation of the

different types of satellite imager y to see movement and evolution of features.

Similar analysis techniques of radar data can quickly show a forecaster the loca-

tion and intensity of both stratiform and convective precipitation. The evolution and

vertical structure of individual convective cells can be discerned from careful analy-

sis of radar data. Doppler radar offers forecasters the opportunity to remotely sense

the wind speed and direction on a continuous basis at a variety of levels in the

vertical if there are adequate reﬂectors. Wind proﬁlers, which are a type of vertically

pointing Doppler radar, also provide remo tely sensed information about the three-

dimensional wind ﬁeld.

The various directly and remotely sensed observations described above, along

with other parameters not included in this discussion, are often best analyzed in

combination. Computational resources are commonly available that allow for the

combination of the various observed data sets. To a great extent, this was not

possible prior to the 1990s, and many of these data were often analyzed separately

from each other. A graphical overlay of rawinsonde and surface observations on a

satellite image can make it very easy to answer a fundamental question such as

‘‘where is the front at the surface and aloft,’’ but this was not always the case and is

still quite difﬁcult over oceanic regions.

To make sense of the observations, forecasters will often attempt to ﬁt a concep-

tual model to the observed data. A conceptual model is a mental picture of an

atmospheric structure or process. They are usually the result of research where

intensive observations have been made on a temporal and spatial scale that is far

greater in resolution than the operational observational network. The Norwegian

Frontal Model is one of the best-known examples of a conceptual model. The

structure and evolution of a so-called supercell thunde rstorm is another example

of a conceptual model. There are hundreds, and perhaps thousands, of different

conceptual models in use to describe virtually every type of phenomena on virtually

every spatial and temporal scale. Conceptual models are convenient structures for

organizing the huge volume of data with which forecasters are faced; however, they

are inevitably simpliﬁcations of actual atmospheric phenomena and processes.

Research results lead to constant reﬁnement of existing conceptual models,

proposals of new models, and even the abandonment of some conceptual models.

Models of phenomena are often applicable in some geographic regions, but not

others. The Norwegian Frontal Model is a very old paradigm, and it has been revised

680 MODERN WEATHER FORECASTING

many times, and its complete abandonment has been recommended on numerous

occasions. However, a satisfactory replacement for this conceptual model has not

been found, so at least parts of this model continue to be used by forecasters. Most

conceptual models have similar life cycles.

Another method employed by forecasters to organize the forecast process is to

start their analysis of the data on the largest scale and then move progressively

downscale. This method has been called the ‘‘forecast funnel’’ in that a forecaster

starts at the top, or wide por tion, of the funnel, which cor responds to the planetary

scale, and then descends into the synoptic-scale in the middle of the funnel, and

ﬁnally the narrow portion of the funnel focuses on the forecast problem of a speciﬁc

place and time, which is determined by conditions on the mesoscale. The forecast

funnel is itself a conceptual model in that it is based on the idea that larger scale

atmospheric conditions drive smaller scale phenomena, and that it will be difﬁcult or

impossible to accurately predict conditions on the smallest scales without taking into

account conditions on the next larger scale, and so on.

A simple example of scale interaction and the application of the forecast funnel

might be to consider the potential for a snowstorm along the mid-Atlantic states. On

the planetary scale, a longwave trough position must be present over eastern North

America for there to be a chance of cyclogenesis, and the track of the storm will, to

some extent, be determined by the planetary-scale conditions. A migratory short-

wave trough(s) on the synoptic-scale is required to initiate cyclogenesis. The inter-

action of the synoptic-scale troughs and ridges with the local topography will

determine whether cold air remains trapped on the east side of the Appalachian

Mountains. This synoptic-scale interaction with the terrain drives the meso scale

conditions and modulates the precipitation type and intensity.

Like all conceptual models, the forecast funnel is a simpliﬁcation. In the example

above, one could also argue that the preexisting conditions on the mesoscale are just

as important, or more so, for the creation of a snowstorm. Where is the cold air

already present? Are there any preexisting temperature gradients or old frontal

boundaries? Is the air in the warm sector of the storm unstable enough to produce

widespread precipitation and convection, which could lead to further intensiﬁcation

of the cyclone. Forecasters generally ﬁnd that they must also ascend the forecast

funnel at times to take into account ‘‘upscale’’ interactions.

3 FORECASTING 12 h TO 7 DAYS

Modern weather forecasting of conditions beyond 12 h in the future is primarily

based on forecaster interpretation of output from numerical weather prediction

models. Forecasters have access to many different models. Through the Internet,

forecasters can access models from the national centers of various countries, military

national centers, and universities around the world. Depending on their employer’s

computational resources, forecasters may even have access to models run locally. At

the time of this writing, model output is typic ally used in a deterministic sense. This

means that given realistic initial conditions and a model with adequate resolution and

3 FORECASTING 12 h TO 7 DAYS 681

accurate representations of the pertinent physical process, the model prediction of

the future state of the atmosphere can be used to explicitly derive the sensible

weather elements at Earth’s surface (or aloft for aviation forecasting). By the end

of the twentieth century considerable effort was being expended on so-called ensem-

ble forecasting, where output from an ensemble of models is used either by a fore-

caster or directly through objective techniques to produce probabilistic forecasts of

sensible weather, with a goal of including information about uncertainty with each

forecast element. A very brief discussion of ensemble techniques is included at the

end of this section, but most of what follows is based on a deterministic approach,

which is still the most widely used method for most forecasting tasks.

Before using any model, forecasters will go through the process of analysis in an

attempt to understand current conditions. This must be done on all spatial scales.

The analysis step is the cornerstone of proper use of deterministic model output. It is

important for at least two reasons. First, the forecaster must decide if the model

initial conditions and the early hours of the forecast capture the essential detail of

what is happening in the real atmosphere. Second, through analysis of current

conditions, the forecaster must grasp the physical processes that are important.

Different models represent the physical processes of the atmosphere in different

ways, and if the forecaster expects certain processes to be particularly important,

then this information will be a factor in how the forecaster uses the output from the

various models.

A critique of the initial conditions of models has become much easier with the use

of modern graphic=image workstations. It is a simple task to overlay various model

ﬁelds with animating satellite imagery. Comparison of model initial ﬁelds with

plotted observations and satellite imagery allows the forecaster to quickly determine

if troughs, ridges, vorticity centers, baroclinic zones, jet stream axes, moisture

plumes, and other features are reasonably well depicted in the model’s initial condi-

tions. This critique is particularly important over data-sparse regions such as the

oceans but is also important over relatively data-rich areas such as North America

since there are times when relatively small errors in initial conditions can lead to

signiﬁcant errors in the forecast.

When the initial conditions of one model are superior to all the others, a fore-

caster may decide to give the solution from this model g reater weight as the forecast

is developed. Often the situation is not clear as to which model is best initialized.

Differences are often subtle, and even with current satellite capabilities, there is

considerable uncertainty about the exact locations in the horizontal and vertical of

speciﬁc gradients and features over oceanic regions. There are times when all of the

available models have a poor grasp of the real atmosphere. In these situations, a

forecaster will have to fall back to some of the empirical rules of the pre-NWP

(Numerical Weather Prediction) era to mentally modify model output in the direction

most likely to minimize the error. Most often in this situ ation the forecaster will have

low conﬁdence in what the ultimate conditions will be, and this uncertainty will be

expressed quantitatively and qualitatively within the allowable format of the forecast

product.

682 MODERN WEATHER FORECASTING

Given a reasonable outcome from the initial condition critique, the forecaster will

often next move onto consideration of the different model characteristics. Typically,

resolution is greater in limited domain models, while lower resolution models have

larger and often global domains. High-resolution models will have more realistic

representations of Earth’s topography, including mountains, valleys, lakes, and coast-

lines. Larger domai n models will often perform best in regimes with strong westerly

ﬂow and fast-moving storm systems.

Similar differences typically exist between models with respect to vertical resolu-

tion, particularly in the boundary layer. Greater vertical resolution generally leads to

more accurate representation of temperature inversions, unstable layers, and the

vertical distribution of moisture. There can be signiﬁcant differences between

models in their treatment of moist processes such as convection and stratif orm

precipitation due to differences in their vertical resolution. It also can d rastically

affect how a model predicts temperature and winds near the surface. Models also

differ in their vertical coordinate system and whether the basic equations assume

hydrostatic balance or not. Some vertical coordinate systems are terrain-following;

others are not. Some dynamically place the greatest resolution in regions of large

potential temperature g radient; others are ﬁxed regardless of the situation. The

vertical coordinate can lead to different model characteristics for phenomena such

as mountain-induced windstorms or cyclogenesis that takes place in the lee of

orographic barriers.

Different methods of representing physical processes are used in the various

models available to forecasters. Processes that cannot be resolved explicitly within

the model’s formulation of the primitive equations are represented through what are

called parameterization schemes. These are basically subroutines that are invoked

throughout the model run that use ‘‘state’’ variables that are explicitly predicted, such

as wind, temperature, pressure, and some form of moisture, to handle such pheno-

mena and conditions as soil moisture, moist and dry convection, shortwave and

longwave radiation, evapotranspirati on from vegetation, turbulence, and so forth.

The parameterization schemes are designed to feedback into the model and alter

the state variables. Although many of these schemes are quite complex, an under-

standing of some of the simpler schemes can be used effectively by forecasters in

their subjective use of model output.

Model output is now widely available in gridded form. Prior to the 1990s, only a

very limited set of graphics at speciﬁc levels were typically available. These were, to

a great extent, the same set of graphics that were produced via manual methods prior

to the NWP era. Gridded model output allows forecasters to examine the model

solution in great detail, and in particular it allows forecasters to gain insight into the

three-dimensional structures of the model’s simulated atmosphere. Software and compu-

tational resources exist that allow the output to be viewed via three-dimensional

renderings, but at the time of this writing these resources are not widely in use yet,

and most model output is still viewed via two-dimensional graphics, such as plan

views, cross sections, time–height sections, and forecast soundings.

Forecasters will apply their knowledge of the important physical processes

gleaned from the analysis step to examination of the model output. Knowledge of

3 FORECASTING 12 h TO 7 DAYS 683

local conditions plays a role in the forecaster’s decision about which processes are

most important in each situation, and this becomes a signiﬁcant factor since the

strengths an d weaknesses of different models are considered. If ﬂow interaction with

terrain is a key consideration for a given event, such as in cases of orographic

precipitation, then a model with high horizontal resolution might be favored. If

lake effect clouds and precipitation are signiﬁcant forecast problems, then a model

that correctly initializes the areal extent of the lake and its water temperature might

be favored. If two models are forecasting a trough to move through a mean longwave

ridge position with the same timing, but with different depths, the larger domain

model that has a better representation of the planetary-scale waves might be

preferred. If the forecaster knows that precipitation has recently occurred and

near-surface conditions are moist, then the model that has most correctly predicted

precipitation in recent model runs might be expected to have the best representation

of soil moisture and boundary layer conditions. Each situation must be evaluated and

forecasters must apply their knowledge of initial conditions and model character-

istics to decide, when possible, which model solution offers the greatest skill and

utility.

The ﬁnal step in the deterministic use of model output for forecasting is the

derivation of sensible weather. This includes parameters such as precipitation

amount, onset, end, type, and areal distribution, potential for severe convection,

lightning activity, sustained wind speed and gusts at the surface, wind shift

timing, cloud cover, cloud layers, visibility, areas of turbulence, icing, blowing

snow, blowing dust, temperature, wind chill, heat index, wave and swell height,

sea-ice accumulation on ships, and so forth, depending on the type of forecasting

that is being done. Statistical postprocessing of model output exists that attempts to

derive sensible weather elements directly from model output. In the United States,

some of these statistical methods are called model output statistics (MOS). The MOS

regression equations are derived by comparing past model forecasts with subsequent

observations at individual locations. This method has the advantage that it can

correct for model biases. It has the disadvantage that in the modern era operational

models are undergoing nearly const ant change, and the developmental data set for

MOS techniques may not be representative of the current incarnation of the model.

There are other objective algorithms that convert model output into sensible weather

elements and more are always being developed, but a complete discussion of this

topic is beyond the scope of this chapter. A popular display of this type of algorithm

output is known as a meteogram, where a group of time series of derived sensible

weather is displayed g raphically in the same way that a graphic of observations

might look. In general, experienced forecasters are quite expert at deriving the

sensible weather from model output, once they have applied the knowledge and

techniques described above to determine which model output is most appropriate

for a given situation.

In deterministic forecasting, a forecaster will consider a number of different

model solutions during the forecast process. In essence the forecaster is examining

a small ensemble of forecasts. The number is typically ﬁve or fewer. The forecaster

may be looking for model-to-model consistency and=or run-to-run consistency from

684 MODERN WEATHER FORECASTING

a given model as an indication of the uncertainty associated with a given forecast.

Ensemble forecasting techniques are based on the idea that there is inherent uncer-

tainty in the initial conditions and perhaps similar uncertainty about which is the

optimal model conﬁguration for a given situation. The goal of ensemble forecast

methods is to quantify the uncertainty by running many versions of NWP models

with either perturbed initial conditions and =or versions of models with different

parameterization schemes, different basic conﬁgurations, or different analysis

schemes. The method of perturbing the initial conditions, or the basis for deciding

the makeup of the ensemble members, is beyond the scope of this discu ssion, but

two key premises are that each member has an equal probability of being correct and

that the more ensemble members included the more skillful the probability distribu-

tion of the output.

Ensemble techniques are being developed, both as an alternative and as an

adjunct to the determin istic forecas t method. Although experienced forecasters are

very skillful at deriving sensible weather from model output, the task of determining

which model output is best in a given situation, and estimating the magnitude and

shape of the model’s most likely error is quite difﬁcult. So me would argue that this

task is essentially impossible. This last statement is a topic of active debate. Another

reason for ensemble techniques is that an increasing number of users of forecast

information, especially those who must make large resource commitments based on

the weather forecast, require a quantiﬁcation of the uncertainty associated with the

forecast. While deterministic forecasts often include uncertainty, such as probability

of precipitation statements, and narrative wording that indicates alternative scena-

rios, the method does not lend itself to quantiﬁcation of uncertainty of all parameters

in time and space, and certainly not with consistency from forecaster to forecaster. A

strength of ensemble techniques is that probabilistic output falls out quite naturally,

offering consistent quantitative assessment of uncertainty. Of course, skillful algo-

rithms that convert model output to sensible weather elements must exist, and this is

still problematic in many situations.

Ensemble forecasting is a rapidly emerging technique. It could become the

primary method of weather forecasting in the future. A description of modern

weather forecasting in 2010 may look back on deterministic forecasting methods

as something from a bygone era, much as today we look back on the pre-NWP era

with nostalgia. More likely, some forecast applications will lend themselves to

ensemble techniques, especially where a sophisticated user community exists,

while other applications will be best accomplished by deterministic methods for

the foreseeable future.

4 FORECASTING 0 TO 12 h

Output from numerical predi ction models is of limited use in short-term forecasting

for a number of reasons. The most obvious reason is that the collection of obser va-

tions, quality control of the obser vations, objective analysis, the execution of the

numerical model, and the distribution of the model output takes time. Usually these

4 FORECASTING 0 TO 12 h 685