Game Climate Models
Game Climate Models
329–357
Abstract: This article reviews the historical development of statistical weather models, from
simple analyses of runs of consecutive rainy and dry days at single sites, through to multisite
models of daily precipitation. Weather generators have been used extensively in water
engineering design and in agricultural, ecosystem and hydrological impact studies as a means
of in-filling missing data or for producing indefinitely long synthetic weather series from finite
station records. We begin by describing the statistical properties of the rainfall occurrence and
amount processes which are necessary precursors to the simulation of other (dependent) mete-
orological variables. The relationship between these daily weather models and lower-frequency
variations in climate statistics is considered next, noting that conventional weather generator
techniques often fail to capture wholly interannual variability. Possible solutions to this
deficiency – such as the use of mixtures of slowly and rapidly varying conditioning variables –
are discussed. Common applications of weather generators are then described. These include the
modelling of climate-sensitive systems, the simulation of missing weather data and statistical
downscaling of regional climate change scenarios. Finally, we conclude by considering ongoing
advances in the simulation of spatially correlated weather series at multiple sites, the
downscaling of interannual climate variability and the scope for using nonparametric
techniques to synthesize weather series.
Key words: climate change, impact assessment, stochastic model, time series, weather
generator.
I Introduction
Models of observed daily weather sequences are frequently used in water engineering
design, and agricultural, ecosystem or climate change simulations because observed
ground-based meteorological data are often inadequate in terms of their length, com-
pleteness or spatial coverage. These statistical models are also known as ‘weather
generators’ since they can in-fill missing data or produce indefinitely long synthetic
positive serial (i.e., auto-) correlation, so that wet and dry runs tend to clump together
in time more strongly than could be expected by chance. The precipitation intensity
process pertains to the modelling and simulation of the nonzero precipitation amounts.
These are typically strongly skewed to the right, with many small values and few but
quite important large precipitation amounts. Although these concepts appear straight-
forward it has taken more than a century to formalize many of the processes within
stochastic precipitation models.
for Tel Aviv, Israel, the authors recognized that the frequency distributions for wet and
dry spell length of the types identified by Williams (1952) and Longley (1953) may arise
from a simple Markov chain model. In particular, Gabriel and Neumann (1962)
proposed the use of a first-order Markov chain for precipitation occurrence, assuming
that the probability of rainfall on any day depends only on whether the previous day
was wet or dry. This model can be fully defined by the two conditional probabilities
p01 = Pr {precipitation on day t | no precipitation on day t – l} (1a)
and
p11 = Pr {precipitation on day t | precipitation on day t – l} (1b)
which are called transition probabilities. Here the vertical bar symbol ‘|’ is read as
‘given’ or ‘conditional on’. Since there are only two possible states on a given day, the
two complementary transition probabilities are p00 = 1 – p01 (dry day following a dry
day) and p10 = 1 – p11 (dry day following a wet day).
It was noted by Gabriel and Neumann (1962) that this simple model for rainfall
occurrence was able to describe closely the persistent nature of daily precipitation
occurrence patterns, and that certain other properties of the occurrence series could be
derived from the two transition probabilities. Of particular importance are the long-run
(i.e., climatological) relative frequency of precipitation days
p01
π= (2a)
1 + p01 – p11
and the first-lag autocorrelation of the precipitation occurrence series
r1 = p11 – p01 (2b)
Because of the persistent nature of daily rainfall occurrence one finds in practice that the
conditional probability in (1b) is larger than that in (1a), so that p01 < π < p11 and r1 > 0.
Furthermore, the lengths of the alternating wet and dry spells produced by the first-
order Markov model are independent, with those lengths distributed according to the
geometric distribution
Pr{X = x} = p(1 – p)x – 1, x = 1, 2, 3, . . . (3)
where p = p01 for dry spells and p = 1 – p11 for wet spells. These authors pointed out
further that the average number of wet days, and the variance of the number of wet
days, within a string of T consecutive days under the first-order Markov model can be
computed as
E[N(T)] = π T (4a)
and
1 + r1
Var[N(T)] ≈ π (1 – π) T (4b)
1 – r1
respectively. Equations (4a) and (4b) can thus pertain to rainfall occurrence statistics for
monthly (T ≈ 30) or seasonal (T ≈ 90) periods. Usually the parameters of the precipita-
tion generation models are fit separately for the 12 calendar months, in order to allow
for seasonal cycles in the precipitation statistics. Equation (4b) is an approximation that
holds for large T, although this approximation was found to be very close for monthly
totals in the Tel Aviv data, and subsequently for other locations as well (e.g., Gregory et
al., 1993; Wilks, 1999a).
Equations (2)–(4) illustrate the use of the first-order Markov model to characterize
important aspects of the precipitation occurrence climate. This model also provides a
very convenient and efficient means of generating sequences of random numbers that
resemble (i.e., simulate) the corresponding real weather data. For each simulated day, a
random number u is drawn from the interval [0,1] in a way that any real number (up to
the precision of the computer) in that interval is equally likely to be picked. In practice
these are usually produced by widely available computer algorithms called uniform
pseudo-random number generators (e.g., Press et al., 1986; Bratley et al., 1987). Note,
however, that there are a number of surprising and subtle pitfalls in the generation of
uniform random numbers, and in particular that the simplest of these algorithms (as
might be implemented as part of a computer’s operating system) can have serious defi-
ciencies that may compromise the results of stochastic simulations. See, for example,
Press et al. (1986) for a brief exposition and guide to improved algorithms, or Knuth
(1997), which is the standard reference on this subject.
Once the random number u has been generated, whether the next day in the sequence
is wet or dry is determined using (1). If the previous day (t – 1) was dry, then day t is
simulated to be wet if u ≤ p01, and otherwise it is also dry. If the previous day was wet,
then the current day is simulated to be wet if u ≤ p11, and is dry otherwise (cf. Figure
2a). Because first-order Markov models fit to daily precipitation data yield p01 < π < p11,
the simulations yield sequences of wet and dry days that are more persistent than
independent draws according to the climatological probability π.
For some climates it has been found that the simple first-order Markov model
generates synthetic rainfall series with too few long dry spells (e.g., Buishand, 1977;
1978; Racsko et al., 1991; Guttorp, 1995). Dennett et al. (1983), Singh and Kripalani
(1986), Jones and Thornton (1997) and Wilks (1999a) addressed this deficiency by
considering Markov chains of higher order. These techniques increase the length of the
Markov model’s ‘memory’ of antecedent wet and dry days. For example, second-order
Markov chains use the wet/dry state on both the preceding day, and two days prior,
such that eight transition probabilities pijk must be defined. Here each of the indices i,
j, k may be either one (wet) or zero (dry). Hence, p101 would be the probability of a wet
day given that the previous day was dry, and the day before that was wet. Third and
higher-order Markov chains can be similarly defined, although the number of
parameters (i.e., transition probabilities) required increases exponentially as the order
increases, being 2k for a kth-order chain. When only the dry spells are not adequately
modelled by the first-order Markov model it is possible to improve the statistics of the
simulated dry spells using ‘hybrid-order’ Markov models, in which the Markov
‘memory’ extends further back in time for the dry spells only (Stern and Coe, 1984;
Wilks, 1999a).
When deciding among models having different degrees of complexity, one must
judge how elaborate a model is justified by the data. Gabriel and Neumann (1962)
compared the first-order Markov model for precipitation occurrence at Tel Aviv with
the next simplest model, namely, independent Bernoulli (i.e., binomial) occurrences,
(a)
(b)
using a Chi-square goodness-of-fit test. While this is a reasonable approach when only
two alternatives are being considered, ambiguities in such statistical tests arise when
multiple comparisons are made, for example when choosing among Markov models of
zeroth (Bernoulli distribution), first and second orders. The usual approach in circum-
stances like this is to employ an objective order-selection criterion such as Akaike’s
information criterion (AIC – Akaike, 1974) or the Bayesian information criterion (BIC –
Schwarz, 1978). Both AIC and BIC are likelihood-based criteria, in that they choose the
model having the largest maximized likelihood, after application of a penalty that
increases with the number of free parameters allowed by each of the models
considered. (The likelihood function is notationally analogous to the probability distri-
bution function or the probability density function; but the data are considered as fixed,
while values of the parameters associated with the global maximum of this function are
the fitted maximum likelihood estimates.) The AIC and BIC differ only in the forms of
their penalty functions. Gates and Tong (1976) concluded that second-order Markov
dependence was justified according to the AIC for the Tel Aviv precipitation data
considered by Gabriel and Neumann (1962). However, Katz (1981) concluded first-
order dependence for these data was adequate on the basis of the BIC, which he also
showed to be asymptotically consistent (i.e., the BIC is correct on average for suffi-
ciently large data samples). Since many years of meteorological data are typically
available for fitting stochastic weather models, this consistency result is a strong
argument for use of the BIC in this context.
An alternative to Markov chain models for simulating precipitation occurrences is
the use of spell-length models. Rather than simulating rainfall occurrences day by day,
spell-length models operate by fitting probability distributions such as (3) to observed
relative frequencies of wet and dry-spell lengths. This kind of model is sometimes
called an ‘alternating renewal process’ (Buishand, 1977; 1978; Roldan and Woolhiser,
1982), in that random numbers are generated alternately from the wet and dry spell-
length distributions (Figure 2b). That is, a new spell length L is generated only when a
run of consecutive wet or dry days has come to an end, at which point a new spell of
the opposite type is simulated. Of course, if geometric distributions (3) are used to
model the lengths of wet and dry spells the resulting synthetic series will exhibit the
same characteristics as the equivalent first-order Markov process (1). Higher-order
Markov chains have spell-length distributions associated with them that are general-
izations of the geometric distribution. Precipitation occurrence sequences with different
statistical characteristics can be obtained using different distributions for the
frequencies of spell lengths. Such distributions include the truncated negative binomial
(Buishand, 1977; 1978; Roldan and Woolhiser, 1982), the negative binomial distribution
(Wilby et al., 1998; Wilks, 1999a), and the mixed geometric distribution (Racsko et al.,
1991). For climates where (1) or (3) yield very long dry spells with insufficient
frequency, these more elaborate choices for modelling precipitation occurrence have
been found to yield more realistic results in this regard (Buishand, 1977; 1978; Racsko et
al., 1991; Wilks, 1999a). However, this method can be susceptible to poor parameter
estimates in arid regions or wherever less than 25 years of observations is available
(Roldan and Woolhiser, 1982).
f(x) =
1
µ
exp [ ]–x
µ
(5)
The exponential distribution is probably the simplest reasonable model for daily pre-
cipitation amounts, as it requires specification of only one parameter, µ, yet reproduces
qualitatively the strong positive skewness exhibited by daily precipitation data. The
average nonzero precipitation amount according to (5) is µ, and the corresponding
variance σ2 = µ2. Exponential distributions have also been used by Richardson (1981)
and Wilby (1994), among others.
A number of more elaborate models have also been proposed for the distribution of
daily precipitation amounts given the occurrence of a wet day. The two-parameter
gamma distribution has been the most popular choice (e.g., Thom, 1958; Katz, 1977;
Buishand, 1977; 1978; Stern and Coe, 1984; Wilks, 1989; 1992), and has the probability
density function
(x/β)α – 1 exp[–x/β]
f(x) = (6)
β Γ(α)
This distribution involves two parameters: the shape parameter α and the scale
parameter β. The factor Γ(α) is the gamma function (see, e.g., Abramowitz and Stegun,
1984, or Wilks, 1995) evaluated at α. This distribution has mean µ = αβ, and variance
σ2 = αβ2. For α ≤ 1 gamma distributions are qualitatively similar to exponential distrib-
utions (5) in concentrating most of the probability near zero and producing large pre-
cipitation amounts only rarely. For α = 1 gamma distributions (6) reduce to exponential
distributions (5), but in general the additional parameter in (6) allows more flexible
accommodation of rainfall amount frequencies, and so improves the realism of
stochastic precipitation models.
Another natural generalization of the exponential distribution (5) is the mixed
exponential distribution, which is simply a probability mixture of two one-parameter
exponential distributions. Its probability density function is
α –x 1–α –x
f(x) =
µ1
exp
[ ] µ1
+
µ2
exp
[ ]µ2
(7)
tically) significantly different from zero, but is typically quite small and usually of little
practical importance (Katz, 1977; Buishand, 1977; 1978; Foufoula-Georgiou and
Lettenmaier, 1987). In contrast, accounting for serial correlation of nonzero precipitation
amounts is essential if the precipitation model has an hourly (or smaller) rather than a
daily time step (Katz and Parlange, 1995).
Precipitation (mm)
where each Tk is any of the nonprecipitation variables, µk,0 and σk,0 are its mean and
standard deviation for dry days, and µk,1 and σk,1 are its mean and standard deviation
for wet days. The seasonal dependence of the means and standard deviations in (10) is
usually achieved through Fourier harmonics (i.e., sine and cosine functions) that vary
smoothly through the year, simulating the annual cycle for each of these parameters.
Note that even though the z(t) as produced by (8) have Gaussian distributions, the
resulting nonprecipitation weather variables T are not necessarily also Gaussian
because of the random application of dry or wet-day means and variances in (10)
(Katz 1996), although in practice the simulated distributions are often close to
Gaussian. The means and variances of the Ts in (10) depend not only on the dry and
wet-day means and variances on the right-hand side of (10) but also on the uncondi-
tional rain-day probability (Katz, 1996), which for first-order Markov dependence is
given by (2a).
A number of formulations for producing more than the basic set of three nonprecip-
itation variables in the WGEN framework have been proposed. In addition to
maximum temperature, minimum temperature and solar radiation, the model of Wallis
and Griffiths (1997) also simulates daytime winds, night-time winds and daily dew
point through a straightforward extension of (8) to include K > 3 nonprecipitation
variables. Parlange and Katz (1999) followed a similar approach to simulate wind speed
and dew point in addition to temperatures and radiation, and furthermore subjected
the wind speed data to a square-root transformation in order to simulate better its
skewed distribution. Less comprehensive algorithms were used by Bruhn et al. (1980)
to add simulation of daily minimum relative humidity; and by Richardson and Nicks
(1990) also to simulate wind speed and wind direction.
Figure 2 shows flowcharts for the basic WGEN structure when 1) the precipitation
submodel is based on a first-order Markov chain; and 2) when it is based on direct
simulations of spell lengths. In this figure, numbers in square brackets refer to
equation numbers in the foregoing text. Regardless of how the precipitation
occurrences are handled, the nonprecipitation submodels (right-hand sides of the
panels) are identical.
While most weather generators operate on a daily time step, their output nevertheless
exhibits longer-term variations that are the counterparts of lower-frequency variations
in real weather data. For example, both weather generator output and observed
weather data will exhibit different monthly mean temperature or monthly total precip-
itation in different years. A natural way to characterize these variations is to compute
the variance of these monthly quantities, which is often referred to as the interannual
variance. Of course one would like the interannual variability in weather generator
output to match closely the corresponding variability in the observed data, and it has
been suggested (Gregory et al., 1993) that this comparison provides a crucial test of the
similarity of synthetic and real climates.
Comparison of the precipitation statistics is particularly informative. Let S(T) be the
sum of T consecutive daily precipitation amounts, which will correspond to a monthly
precipitation total if T ≈ 30. The statistics of the synthetic S(T) climate can be expressed
in terms of its long-term mean E[S(T)] and interannual variance Var[S(T)] according to
(e.g., Gregory et al., 1993; Katz and Parlange, 1998):
stationarity to climatic change. A more common view currently is that the nonstation-
arity results from interannual climate variations. For example, Figure 4 shows
interannual variations in three climatic quantities that are also stochastic rainfall
parameters, namely: the conditional dry and wet-day probabilities 1 – p01 and p11, and
the mean wet-day amount µ. For this site, it is evident that year-to-year precipitation
variability includes contributions both from random within-year fluctuations, and from
contributions deriving from more systematic longer-period variations in these three
parameters.
(a)
(b)
(c)
It has been found that a relatively simple and straightforward approach to simulating
interannual variability better is to condition the weather generator parameter set on a
covariate. That is, different parameter sets are chosen for simulation in different years,
depending on the value of some external variable. For example, much of the variation
evident in Figure 4 is strongly correlated with large-scale, free-atmosphere vorticity
over Britain (Wilby, 1997). Simple covariates that have been used successfully to
condition the choice of stochastic model parameters include the monthly statistics
themselves (Wilks, 1989), long-range forecasts of the monthly statistics (Briggs and
Wilks, 1996), or even random numbers (Jones and Thornton, 1997). Katz and Zheng
(1999) used a ‘hidden’ mixture approach to capture some interannual variability not
exhibited by a stationary weather generator formulation.
More commonly, the stochastic model parameters are conditioned on some aspect of
large-scale atmospheric circulation (e.g., Hay et al., 1991; Bardossy and Plate, 1992;
Woolhiser et al., 1993; Hughes and Guttorp 1994a; Katz and Parlange, 1993; 1996; Wallis
and Griffiths, 1997; Kiely et al., 1998) or by a hierarchy of precipitation mechanisms (e.g.,
Sansom and Thomson, 1992). For example, Wilby (1995) simulated daily precipitation
for sites in the UK using the Lamb Weather Type (Lamb, 1972) (LWT) categories, sub-
classified by the presence or absence of weather fronts. Table 1 compares the probability
of a frontal system under anticyclonic, cyclonic and purely directional airflows, as well
as the corresponding likelihood of precipitation occurrence at Kempsford, UK. From
such information LWTs may be used to condition both the probability of frontal
weather (over the British Isles as a whole) and the occurrence of precipitation at
individual stations. However, the veracity of weather pattern models depends upon the
chosen weather classification system and on the observed links between model
parameters and circulation classes remaining constant through time (Wilby, 1994; 1997).
More recent examples of the circulation-based weather generators have favoured the
use of continuous atmospheric variables as an alternative to conditioning by discrete
weather categories (e.g., Conway et al., 1996).
In a comparable study of interannual precipitation variability at selected sites in the
UK, Wilby (1998) employed the North Atlantic Oscillation Index (NAOI) and North
Atlantic sea surface temperature (SST) anomalies as conditioning variables. For
example, Figure 5 demonstrates that weak but statistically significant correlations may
be detected between p11, p00 and µ, and spring (March–May) North Atlantic SST
anomalies, 1901–90. It is noteworthy that for this site – with the exception of the spring
p11 parameter – the correlation strengths were not generally improved by lagging the
SST anomalies (cf. Colman, 1997). In this unusual case, spring wet-spell persistence p11
at Kempsford, Cotswolds, UK was most strongly correlated with SST anomalies in the
preceding winter (November–December). However, the over-riding implication of the
Table 1 The probability of frontal weather (F > 0) and of precipitation occurrence (π)
by circulation type at Kempsford in the Cotswolds, UK, 1970–90
Circulation type Probability (F > 0) Probability (π) (F > 0) Probability (π) (F = 0)
study was that the realism of stochastic rainfall models may be enhanced using
mixtures of slowly and rapidly varying conditioning variables (in this case, monthly
SST anomalies and daily vorticity, respectively).
(a)
(b)
(c)
statistics are available for a site, and weather generator parameters are needed for
simulation at the daily timescale. Since the monthly statistics (Equation 11, for precipi-
tation) depend on the daily parameters, one can exploit these relationships to develop
specifications for the daily parameters in terms of the known monthly statistics. The
accuracy of this process is improved for locations at which the average number of wet
days per month (and thus the unconditional probability of a wet day, π) is available, in
addition to the monthly precipitation totals from which the monthly precipitation mean
and variance can be computed. Hershfield (1970) found approximately linear relation-
ships for selected USA locations between π and the two Markov transition probabilities
(1), and Hutchinson (1986) confirmed (different) linear relationships for these variables
in Australian data. Geng et al. (1986) extended this kind of empirical linear ‘inverting’
of (11) to include the gamma distribution (6) parameters α and β, and for varied
climates in different parts of the world. However, Hutchinson (1995) has pointed out
that such linear, empirical specification equations are valid only over the range of their
calibration data, and thus can lead to serious errors or even nonsense (e.g., negative
probabilities) if extrapolated too far. A more theoretically satisfying approach to this
problem is to use (11) directly to define the relationships between the daily parameters
on the right-hand sides, in terms of the monthly or seasonal statistics on the left-hand
sides (e.g., Katz, 1996; Wilks, 1992; 1999b).
V Future directions
between winter and spring (MAM), and then from spring to summer (JJA), there were
progressive retreats of the area of significant positive correlations and an expansion of
the negative correlation field. For summers there was a breakdown of the dipole pattern
with statistically insignificant correlations over all the British Isles. However, by
autumn (SON) significant positive correlations emerge once again in northern and
western Scotland.
The correlation fields shown in Figure 6 have two implications for weather
generation. First, it is clear that the modest gains in the simulation of seasonal precipi-
tation reported for Wilby’s (1998) model reflect the location of the two test sites (i.e.,
Kempsford, central England, and Durham, northeast England) with respect to the
regions of strongest forcing. Had sites been chosen in highland Scotland, for example,
it is probable that the incorporation of the slowly varying predictor (i.e., the NAOI)
would have yielded more significant gains in model performance relative to the
vorticity-only model, particularly in winter. Secondly, it is evident that the strength of
NAOI forcing of WGEN parameters is highly seasonal, accounting for as much as 40%
of the interannual variability in wintertime p11 over parts of western Scotland, but
having little or no influence in summer at the same sites. Therefore improved repre-
sentations of interannual precipitation variability in downscaling schemes by means of
low-frequency predictor variables (such as the NAOI) will be spatially and seasonally
specific.
The strength of correlations between slowly varying predictors and WGEN
parameters at individual sites may also be increased by optimizing the predictor
domain. This is accomplished by inverting the correlation search, using a fixed target
site (e.g., Kempsford) and searching a global predictor set for the spatial domain of
strongest correlation(s). For example, the weak positive correlation (shown in Figure 5a)
between spring p00 and North Atlantic SST anomalies was obtained using sea-surface
temperature anomalies averaged over the domain 40° N – 60° N, 35° W – 5° E.
However, as Figure 7 indicates, this domain (shown by the grey shading) only partially
samples the region of significant correlations. Given the correlation field for spring, it
would be legitimate to employ a predictor domain that extends further south, even as
far as 20° N. Alternatively, the shape of the predictor domain might be modified in the
case of autumn (when weak and/or negative correlations at the western edge of the
domain ‘dilute’ the strongly positive correlations to the east). In the case of summer and
autumn it might even be reasonable to employ an additional predictor domain such as
the ocean region off the eastern seaboard of the USA.
The preceding examples suggest that slowly varying predictors have explanatory
power for certain meteorological variables, in certain regions and seasons, under
present climate conditions. However, the validity of using such empirical relationships
for statistically downscaling future precipitation will depend on 1) the ability of GCMs
faithfully to reproduce large-scale forcing patterns (such as the NAO) under current
climate conditions (see, for example, Davies et al., 1997; Osborn et al., 1999); 2) the extent
to which global warming will affect future interdecadal climate variability (e.g.,
Trenberth and Hoar, 1997); and 3) the stationarity of the empirical relationships
between mesoscale forcing and local-scale meteorological responses (Wilby, 1997).
351
Wilby’s (1998) downscaling model is shown in grey
352 The weather generation game: a review of stochastic weather models
3 Nonparametric approaches
Finally, we conclude with a brief description of nonparametric (i.e., not requiring that
particular theoretical probability distributions be assumed) approaches to stochastic
simulation of weather series. Many modern nonparametric statistical methods involve
‘resampling’ (e.g., Efron, 1982), which means that large ersatz samples are synthesized
by repeatedly but randomly copying data values from an existing finite record. For the
case of weather series, an important complication that usually invalidates simple
resampling schemes is the prominent time correlation evident in these data, the
modelling of which is the subject of much of this review. That is, simply concatenating
a random sample of daily weather data (even if restricted to a single month or season)
will simulate the original series poorly because the prominent time correlation will be
destroyed by the resampling process.
More complex procedures to resample weather series in ways that also capture the
important time correlations have been described by Young (1994), Lall and Sharma
(1996), Lall et al. (1996) and Rajagopalan et al. (1997). Recently, Bardossy (1998) has
proposed the use of an algorithm known as simulated annealing to reshuffle simple
resampled data in a way that important statistical properties of the original series
(especially, but not limited to, the time correlation) are reconstituted also. The simulated
annealing approach is quite intensive computationally, but may prove to be especially
useful for simulating series with very short (minutes or hours) time steps.
These nonparametric methods are attractive in that subjective judgements about
most appropriate model forms and probability distributions are avoided, and
furthermore as data-based methods they can capture deviations from theoretical
probability distributions for the individual variables, and nonlinearities in the relation-
ships among variables. On the other hand, being based wholly on the observed data
they are somewhat limited in the range of extreme values that can be generated,
although ‘smoothed’ resampling schemes (e.g., Rajagopalan et al., 1997) ameliorate this
problem somewhat. Another difficulty with the nonparametric methods for some appli-
cations is that future climate regimes cannot be easily constructed through simple
parameter adjustments.
VI Concluding remarks
From the preceding review it is evident that there have been considerable advances in
weather generation research in the 150 years since Quetelet’s pioneering study. As in
many other branches of meteorology there is an ongoing need for critical assessment of
emergent techniques as well as for comparisons between conventional and refined
model formulations. The widespread and growing use of regionally specific, environ-
mental impacts models implies a corresponding demand for high integrity weather
data. For as long as there remains a mismatch between the data that our limited obser-
vational networks can supply and that which the impacts community requires, there
will be a continuing need for statistical weather generation.
Acknowledgements
We thank Rick Katz and an anonymous reviewer for constructive suggestions that have
improved this review. This work was sponsored in part by ACACIA (A Consortium for
the Application of Climate Impact Assessments), the National Center for Atmospheric
Research and by ACCORD (Atmospheric Circulation Classification and Regional
Downscaling) under the European Community Environment Research Programme
(contract: ENV4-CT97-0530). NCAR is sponsored by the National Science Foundation.
References
Abramowitz, M. and Stegun, I.A., editors, 1984: Royal Meteorological Society 64, 631–34.
Pocketbook of mathematical functions. Frankfurt: Colman, A. 1997: Prediction of summer Central
Verlag Harri Deutsch. England Temperature from preceding North
Akaike, K. 1974: A new look at the statistical Atlantic winter sea surface temperature. Inter-
model identification. IEEE Transactions on national Journal of Climatology 17, 1285–1300.
Automatic Control 19, 716–23. Conway, D., Wilby, R.L. and Jones, P.D. 1996:
Precipitation and air flow indices over the
Bardossy, A. 1998: Generating precipitation time British Isles. Climate Research 7, 169–83.
series using simulated annealing. Water Cox, D.R. and Isham, V. 1988: A simple spatial-
Resources Research 34, 1737–44. temporal model of rainfall. Proceedings of the
Bardossy, A. and Plate, E.J. 1992: Space-time Royal Society London A415, 317–28.
model for daily rainfall using atmospheric
circulation patterns. Water Resources Research Davies, J.R., Rowell, D.P. and Folland, C.K. 1997:
28, 1247–59. North Atlantic and European seasonal pre-
Besson, L. 1924: On the probability of rain. dictability using an ensemble of multi-decadal
Monthly Weather Review 52, 308. AGCM simulations. International Journal of
Bras, R. and Rodriguez-Iturbe, I. 1976: Rainfall Climatology 17, 1263–84
generation: a nonstationary time varying multi- Dennett, M.D.J., Rodgers, J.A. and Keatinge,
dimensional model. Water Resources Research 12, J.D.H. 1983: Simulation of a rainfall record for a
450–56. new site of a new agricultural development: an
Bratley, P., Fox, B.L. and Schrage, L.E. 1987: A example from northern Syria. Agricultural
guide to simulation. New York: Springer-Verlag. Meteorology 29, 247–58.
Briggs, W.M. and Wilks, D.S. 1996: Extension of Devroye, L. 1986: Non-uniform random variate
the CPC long-lead temperature and precipita- generation. New York: Springer.
tion outlooks to general weather statistics. Dubrovsky, M. 1997: Creating daily weather
Journal of Climate 9, 3496–504. series with use of the weather generator.
Bruhn, J.A., Fry, W.E. and Fick, G.W. 1980: Environmetrics 8, 409–24.
Simulation of daily weather data using
theoretical probability distributions. Journal of Efron, B. 1982: The Jackknife, the Bootstrap, and
Applied Meteorology 19, 1029–36. other resampling plans. Philadelphia, PA: Society
Buishand, T.A. 1977: Stochastic modeling of for Industrial and Applied Mathematics.
daily rainfall sequences. Mededlingen Land-
bouwhogeschool Wageningen 77–3. Favis-Mortlock, D.T. and Boardman, J. 1995:
––––– 1978: Some remarks on the use of daily Nonlinear responses of soil erosion to climate
rainfall models. Journal of Hydrology 36, change: a modelling study on the UK South
295–308. Downs. Catena 25, 365–87.
Favis-Mortlock, D.T., Boardman, J. and Bell, M.
Chin, E.H. and Miller, J.F. 1980: On the 1997: Modelling long-term anthropogenic
conditional distribution of daily precipitation erosion of a loess cover: South Downs, UK. The
amounts. Monthly Weather Review 108, 1462–64. Holocene 7, 79–89.
Cochran, W.G. 1938: An extension of Gold’s Favis-Mortlock, D.T., Evans, R., Boardman, J.
method of examining the apparent persistence and Harris, T.M. 1991: Climate change,
of one type of weather. Quarterly Journal of the winter wheat yield and soil erosion on the
Katz, R.W. and Zheng, X. 1999: Mixture model simulation models. Agricultural Systems 44,
for overdispersion of precipitation. Journal of 377–96.
Climate, in press.
Kiely, G., Albertson, J.D., Parlange, M.B. and Osborn, T.J., Briffa, K.R., Tett, S.F.B., Jones, P.D.
Katz, R.W . 1998: Conditioning stochastic and Trigo, R.M. 1999: Evaluation of the North
properties of daily precipitation on indices Atlantic Oscillation as simulated by a climate
of atmospheric circulation. Meteorological model. Climate Dynamics, in press.
Applications 5, 75–87.
Kittel, T.G.F., Rosenbloom, N.A., Painter, T.H., Parlange, M.B. and Katz, R.W . 1999: An
Schimel, D.S. and VEMAP Modeling extended version of the Richardson model for
Participants 1995: The VEMAP integrated data- simulating daily weather variables. Journal of
base for modeling United States ecosystem/ Applied Meteorology, in press.
vegetation sensitivity to climate change. Journal Pickering, N.B., Stedinger, J.R. and Haith, D.A.
of Biogeography 22, 857–62. 1988: Weather input for nonpoint source
Klugman, M.R. and Klugman, S.A. 1981: A pollution models. Journal of Irrigation and
method for determining change in precipita- Drainage Engineering 114, 674–90.
tion data. Journal of Applied Meteorology 20, Press, W.H., Flannery, B.P., Teukolsky, S.A. and
1506–509. Vetterling, W.T. 1986: Numerical recipes, the art
Knuth, D.E. 1997: The art of computer programming, of scientific computing. Cambridge: Cambridge
v 2, semi-numerical algorithms. Reading, MA: University Press.
Addison-Wesley.
Racsko, P., Szeidl, L. and Semenov, M. 1991: A
Lall, U., Rajagopalan, B. and Tarboton, D.G. serial approach to local stochastic weather
1996: A nonparametric wet/dry spell model for models. Ecological Modelling 57, 27–41.
resampling daily precipitation. Water Resources Rajagopalan, B., Lall, U., Tarboton, D.G. and
Research 32, 2803–23. Bowles, D.S. 1997: Multivariate nonparametric
Lall, U. and Sharma, A. 1996: A nearest-neighbor resampling scheme for generation of daily
bootstrap for resampling hydrological time weather variables. Stochastic Hydrology and
series. Water Resources Research 32, 679–93. Hydraulics 11, 65–93.
Lamb, H.H. 1972: British Isles weather types and a Richardson, C.W. 1981: Stochastic simulation of
register of the daily sequence of circulation patterns, daily precipitation, temperature, and solar
1861–1971. Meteorological Office, Geophysical radiation. Water Resources Research 17, 182–90.
Memoir 116. London: HMSO. ––––– 1985: Weather simulation for crop manage-
Longley, R.W . 1953: The length of dry and ment models. Transactions of the American
wet periods. Quarterly Journal of the Royal Society of Agricultural Engineers 28, 1602–606.
Meteorological Society 79, 520–27. Richardson, C.W. and Nicks, A.D. 1990: Weather
generator description. In Sharpley, A.N. and
Mearns, L.O., Katz, R.W. and Schneider, S.H. Williams, J.R., editors, Erosion/productivity
1984: Extreme high-temperature events: impact calculator. 1. Model documentation.
changes in their probabilities with changes Agricultural Research Service Technical Bulletin
in mean temperature. Journal of Applied 1768, Washington, DC: US Department of
Meteorology 23, 1601–13. Agriculture, 93–104.
Mearns, L.O., Rosenzweig, C. and Goldberg, R. Richardson, C.W. and Wright, D.A. 1984: WGEN:
1996: The effect of changes in daily and a model for generating daily weather variables
interannual climatic variability on CERES- (ARS-8). US Department of Agriculture.
wheat yields. A sensitivity study. Climatic Riha, S.J., Wilks, D.S. and Simoens, P. 1996:
Change 32, 257–92. Impact of temperature and precipitation
––––– 1997: Mean and variance changes in climate variability on crop model predictions. Climatic
scenarios: methods, agricultural applications, Change 32, 293–311.
and measures of uncertainty. Climatic Change Roldan, J. and Woolhiser, D.A. 1982: Stochastic
35, 367–96. daily precipitation models. 1. A comparison of
occurrence processes. Water Resources Research
Newnham, E.V. 1916: The persistence of wet and 18, 1451–59.
dry weather. Quarterly Journal of the Royal
Meteorological Society 42, 153–62. Sansom, J. and Thomson, P.J. 1992: Rainfall
Nonhebel, S. 1994: The effects of use of average classification using breakpoint pluviograph
instead of daily weather data in crop growth data. Journal of Climate 5, 755–64.
Schimel, D.S., VEMAP Participants and simulation for regional climate change impact
Braswell, B.H. 1997: Spatial variability in assessment. Water Resources Research 30,
ecosystem processes at the continental scale: 3395–403.
models, data and the role of disturbance. ––––– 1995: Simulation of precipitation by
Ecological Monographs 67, 251–71. weather pattern and frontal analysis. Journal of
Schwarz, G. 1978: Estimating the dimension of a Hydrology 173, 91–109.
model. Annals of Statistics 6, 461–64. ––––– 1997: Nonstationarity in daily precipitation
Semenov, M.A. and Barrow, E.M. 1997: Use of a series: implications for GCM downscaling
stochastic weather generator in the using atmospheric circulation indices.
development of climate change scenarios. International Journal of Climatology 17, 439–54.
Climatic Change 35, 397–414. ––––– 1998: Statistical downscaling of daily pre-
Semenov, M.A., Brooks, R.J., Barrow, E.M. and cipitation using daily airflow and seasonal
Richarrdson, C.W. 1998: Comparison of the teleconnection indices. Climate Research 10,
WGEN and LARS-WG stochastic weather 163–178.
generators for diverse climates. Climate Research Wilby, R.L., Conway, D. and Jones, P.D. 1999:
10, 95–107. Prospects for downscaling seasonal precipita-
Semenov, M.A. and Porter, J.R. 1995: Climatic tion variability using conditioned weather
variability and the modelling of crop yields. generator parameters. Hydrological Processes,
Agricultural and Forest Meteorology 73, 265–83. submitted.
Singh, S.V. and Kripalani, R.H. 1986: Analysis of Wilby, R.L., Greenfield, B. and Glenny, C. 1994:
persistence in daily monsoon rainfall over A coupled synoptic-hydrological model for
India. Journal of Climatology 6, 625–39. climate change impact assessment. Journal of
Stern, R.D. and Coe, R. 1984: A model fitting Hydrology 153, 265–90.
analysis of daily rainfall data. Journal of the Wilby, R.L., Wigley, T.M.L., Conway, D., Jones,
Royal Statistical Society Series A 147, 1–34. P.D., Hewitson, B.C., Main, J. and Wilks, D.S.
Stern, R.D., Dennett, M.D. and Dale, I.C. 1982: 1998: Statistical downscaling of general
Analysing daily rainfall measurements to give circulation model output: a comparison of
agronomically useful results. II. A modelling methods. Water Resources Research 34,
approach. Experimental Agriculture 18, 237–53. 2995–3008.
Strandman, H., Vaisanen, H. and Kellomaki, S. Wilks, D.S. 1988: Estimating the consequences of
1993: A procedure for generating synthetic CO2-induced climatic change on North
weather records in conjunction of climate American grain agriculture using general
scenario for modelling of ecological impacts circulation model information. Climatic Change
of changing climate in boreal conditions. 13, 19–42.
Ecological Modelling 70, 195–220. ––––– 1989: Conditioning stochastic daily precipi-
tation models on total monthly precipitation.
Thom, H.C.S. 1958: A note on the gamma distrib- Water Resources Research 25, 1429–39.
ution. Monthly Weather Review 86, 117–22. ––––– 1992: Adapting stochastic weather
Todorovic, P. and Woolhiser, D.A. 1975: A generation algorithms for climate change
stochastic model of n-day precipitation. Journal studies. Climatic Change 22, 67–84.
of Applied Meteorology 14, 17–24. ––––– 1995: Statistical methods in the atmospheric
Trenberth, K.E. and Hoar, T.J. 1997: El Nino and sciences. San Diego, CA: Academic Press.
climate change. Geophysical Research Letters 24, ––––– 1998: Multisite generalizations of a daily
3057–60. stochastic precipitation generation model.
Journal of Hydrology 210, 178–91.
Valdes, J.B., Seoane, R.S. and North, G.R. 1994: ––––– 1999a: Interannual variability and extreme-
A methodology for the evaluation of global value characteristics of several stochastic daily
warming impact on soil moisture and runoff. precipitation models. Agricultural and Forest
Journal of Hydrology 161, 389–413. Meteorology 93, 153–69.
––––– 1999b: Multisite downscaling of daily pre-
Wallis, T.W.R. and Griffiths, J.F. 1997: Simulated cipitation with a stochastic weather generator.
meteorological input for agricultural models. Climate Research 11, 125–36.
Agricultural and Forest Meteorology 88, 241–58. ––––– 1999c: Simultaneous stochastic simulation
Waymire, E., Gupta, V.K. and Rodriguez-Iturbe, of daily precipitation, temperature, and solar
I. 1984: A spectral theory of rainfall at the meso- radiation at multiple sites in complex terrain.
β scale. Water Resources Research 20, 1453–65. Agricultural and Forest Meteorology, in press.
Wilby, R.L. 1994: Stochastic weather type Williams, C.B. 1952: Sequences of wet and of dry