Atmospheric Research 144 (2014) 141150

Contents lists available at ScienceDirect

Atmospheric Research
journal homepage: www.elsevier.com/locate/atmos

Articial neural networks modeling for forecasting the

maximum daily total precipitation at Athens, Greece
P.T. Nastos a,, A.G. Paliatsos b, K.V. Koukouletsos b, I.K. Larissi c, K.P. Moustris d
a
b
c
d

Laboratory of Climatology & Atmospheric Environment, Faculty of Geology & Geoenvironment, University of Athens, Greece
Department of Mathematics, Technological & Education Institute of Piraeus, Greece
Laboratory of Environmental Technology, Department of Electronic Computer Systems, Technological & Education Institute of Piraeus, Greece
Department of Mechanical Engineering, Technological & Education Institute of Piraeus, Greece

a r t i c l e

i n f o

Article history:
Received 31 December 2012
Received in revised form 12 November 2013
Accepted 17 November 2013
Available online 27 November 2013
Keywords:
Artificial neural networks
Maximum daily precipitation
Athens

a b s t r a c t
Extreme daily precipitation events are involved in significant environmental damages, even in
life loss, because of causing adverse impacts, such as flash floods, in urban and sometimes in
rural areas. Thus, long-term forecast of such events is of great importance for the preparation
of local authorities in order to confront and mitigate the adverse consequences. The objective
of this study is to estimate the possibility of forecasting the maximum daily precipitation for
the next coming year. For this reason, appropriate prognostic models, such as Artificial Neural
Networks (ANNs) were developed and applied. The data used for the analysis concern annual
maximum daily precipitation totals, which have been recorded at the National Observatory of
Athens (NOA), during the long term period 18912009. To evaluate the potential of daily
extreme precipitation forecast by the applied ANNs, a different period for validation was
considered than the one used for the ANNs training. Thus, the datasets of the period 1891
1980 were used as training datasets, while the datasets of the period 19812009 as validation
datasets. Appropriate statistical indices, such as the coefficient of determination (R2), the index
of agreement (IA), the Root Mean Square Error (RMSE) and the Mean Bias Error (MBE), were
applied to test the reliability of the models. The findings of the analysis showed that, a quite
satisfactory relationship (R2 = 0.482, IA = 0.817, RMSE = 16.4 mm and MBE = +5.2 mm)
appears between the forecasted and the respective observed maximum daily precipitation
totals one year ahead. The developed ANN seems to overestimate the maximum daily
precipitation totals appeared in 1988 while underestimate the maximum in 1999, which could
be attributed to the relatively low frequency of occurrence of these extreme events within GAA
having impact on the optimum training of ANN.
2013 Elsevier B.V. All rights reserved.

1. Introduction
Extreme precipitation events, which are heavily affected by
the local scale, have adverse impacts on ecosystems, agriculture, infrastructure and even loss of life, while these events are

Corresponding author at: Laboratory of Climatology and Atmospheric

Environment, Department of Geography and Climatology, Faculty of Geology
and Geoenvironment, University of Athens, Panepistimiopolis GR 157 84
Athens, Greece. Tel./fax: +30 210 7274191.
E-mail address: nastos@geol.uoa.gr (P.T. Nastos).
0169-8095/$ see front matter 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.atmosres.2013.11.013

involved in land use changes. During the 20th century, floods

caused more loss of life and property damage than any other
natural disaster in the United States (Easterling et al., 2000).
Barlow (2011) found important contributions of hurricanerelated activity to extreme precipitation over North America.
Min et al. (2011) showed that human-induced increases in
greenhouse gases have contributed to the observed intensification of heavy precipitation events found over approximately
two-thirds of data-covered parts of Northern Hemisphere land
areas. These results are based on a comparison of observed and
multi-model simulated changes in extreme precipitation over

142

P.T. Nastos et al. / Atmospheric Research 144 (2014) 141150

the latter half of the twentieth century. Pall et al. (2011)

presented a multi-step, physically based probabilistic event
attribution framework showing that it is very likely that global
anthropogenic greenhouse gas emissions substantially increased the risk of flood occurrence in England and Wales in
autumn 2000. The frequency of heavy precipitation events has
increased over most land areas, consistent with warming and
observed increases of atmospheric water vapor (IPCC, 2007).
More specifically, Groisman et al. (2005) found that changes
in heavy precipitation frequencies are always higher than
changes in precipitation totals and, in some regions, an increase
in heavy and/or very heavy precipitation occurred while no
change or even a decrease in precipitation totals was observed.
Significant positive trend in the frequency of extreme rainfalls
(greater than 50 mm per day) was revealed over the last few
decades in the USA (Karl et al., 1995; Karl and Knight, 1998). In
Australia, Suppiah and Hennessy (1998) found a significant
increase in the 90th and 95th percentiles of rainfall, while in
Japan, Iwashima and Yamamoto (1993) revealed that, the
highest precipitation events were recorded in recent decades.
Similar results were presented by Brunetti et al. (2001) and
Alpert et al. (2002) for the central and western parts of the
Mediterranean basin. Nastos and Zerefos (2008), analyzing the
extreme daily precipitation totals in Greece, concluded that
positive trends, not statistically significant (95% confidence
level.), observed in daily precipitation totals exceeding 30 and
50 mm at the eastern and southeastern regions of the country.
The extreme precipitation events appear during winter months
in the western and eastern regions, and mainly during autumn
in the rest of the country. Similar results have been reported by
Kambezidis et al. (2010). Besides, recent studies have concluded that, heavy storms of convective nature in the developed
mega-cities could be attributed to the urban heat island (UHI)
(Bornstein and Lin, 2000; Cicek and Turkoglu, 2005; Paliatsos
et al., 2005; Guo et al., 2006; Nastos and Zerefos, 2007;
Philandras et al., 2010; Goswami et al., 2010), resulting in more
rapid run off.
However, extreme precipitation events address not only
increased precipitation totals but prolonged droughts over
many regions (Kundzewicz et al., 2006; Lehner et al., 2006;
Sheffield and Wood, 2007; Beniston et al., 2007; Diodato and
Bellocchi, 2008; Kostopoulou and Jones, 2005; Nastos and
Zerefos, 2009).
The simulation of extreme precipitation events was
carried out by using either stochastic weather generators
(i.e., a hybrid technique with a gamma distribution for low
to moderate intensities and a generalized Pareto distribution
for high intensities) (Furrer and Katz, 2008) or appropriate
distributions such as the extreme value distribution of
type II (EV2) for modeling annual maximum rainfall series
(Koutsoyiannis, 2004).
During the last decade, Artificial Neuron Networks
(ANNs) have been applied to rainfall forecasting (Bodri and
Cermak, 2000; Luck et al., 2000; Silverman and Dracup, 2000;
Sakellariou and Kambezidis, 2004; Cigizoglou and Alp, 2004;
Moustris et al., 2011). Additionally, ANNs have been used in
forecasting extreme precipitation events. Sahai et al. (2000)
applied ANN models in order to forecast total precipitation
during the summer monsoon period across India (Root Mean
Square Error (RMSE) equals 54.2 mm), using as input data
the rainfall during JuneJulyAugust and September months

for the period 18711994. Freiwan and Cigizoglu (2005)

developed a number of different multilayer perceptron ANN
models in order to forecast precipitation for the next month
resulting in RMSE between 25.8 and 33.6 mm, depending on
the used ANN model type. Marzano et al. (2006) used neuralnetwork approach in order to estimate precipitation intensity and extinction from ground. They found that the ANN
retrieval algorithm tends to provide a better accuracy and a
reduced error bias, especially for low-to-moderate rain rates.
Manzat (2007) developed ANNs for the prediction of both
the likelihood of occurrence, and intensity of storms over the
region of Friuli Venezia Giulia in Italy, with satisfactory
results. Wardah et al. (2008) used meteorological satellite
data and developed back-propagation ANNs for the estimation of rainfalls caused massive damages and flooding in the
Klang river basin in Malaysia. Mar and Naing (2008) used
ANN models for the prediction of the monthly rainfall
amount in Yangon (MyanmarSouth East Asia). The major
aim was to evaluate a suitable neural network model for
monthly precipitation mapping of Myanmar. In measuring
network performance using RMSE, experimental results
significantly show that 3 inputs-10 hiddens-1 output architecture model gives the best prediction result for monthly
precipitation in Myanmar.
Extreme precipitation is difficult to reproduce, especially for
the intensities and patterns of extreme events, which are heavily
affected by the local scale. For this reason, it is necessary to study
these events by analyzing long time series of observations (IPCC,
2001). The goal of this study is to estimate the capacity of ANNs
in forecasting the maximum daily precipitation for the next
coming year, using historical daily precipitation totals, recorded
at the National Observatory of Athens (NOA), during the period
18912009.
2. Data and methodology
The dataset used for ANN's training and evaluation concerns
the observed annual daily maximum precipitation recorded at
the National Observatory of Athens (NOA), during the time
period 18912009, the longest reliable precipitation time series
for Greece available, due to unchanged position of the
meteorological station. NOA (longitude: 23 43E, latitude: 37
58N, h = 107 m above sea level) is located on the Hill of
Nymphs near the center of Athens within the greater Athens
area (GAA).
ANNs modeling is a branch of artificial intelligence
developed in 1950s aiming at imitating the biological brain
architecture. ANNs are parallel-distributed systems made of
many interconnected nonlinear processing elements (PEs),
called neurons (Hect-Nielsen, 1990). A renewal of interest has
been grown exponentially in the last decade, mainly concerning the availability of suitable hardware that has made
them convenient for fast data analysis and information processing. MultiLayer Perceptron model (MLP) is the most
commonly used type of ANNs. Its structure consists of PEs and
connections. The PEs called neurons are arranged in layers: the
input layer, one or more hidden layers, and the output layer. An
input layer serves a buffer that distributes input signals to the
next layer, which is a hidden layer. Each unit in the hidden layer
sums its input, processes it with a transfer function, and
distributes the result to the output layer. Also, several hidden

P.T. Nastos et al. / Atmospheric Research 144 (2014) 141150

143

Fig. 1. Typical artificial neural network architecture (Caudill and Butler, 1922).

layers are possibly connected in the same fashion. The units in

the output layer operate in a similar manner. Such kinds of
models are defined as feed-forward ANNs, since data flow
within the network, from one layer to the next without any
return path. The structure of a feed-forward MLP artificial
neural network (Caudill and Butler, 1922) can be represented
as in Fig. 1.
In the present work a special kind of MLP known as Time
Lagged Recurrent Network (TLRN) was used (Fig. 2). TLRNs
are MLP which have been extended with short term memory
structures. Most real-world data contains information in such

time structure. Besides, most ANNs are purely static classifiers. TLRNs are the state of the art in nonlinear time series
forecasting, system identification and temporal pattern classification (Wang and Traore, 2009).
The training algorithm used with TLRNs is more advanced
even than the standard back-propagation. The TLRNs topology is found to be very suitable to deal with flood forecasting
problem (Xue and Dibike, 2001). The performance of TLRNs
over the standard back-propagation algorithm for time series
data forecasting is evident in the recent studies of Badjate
and Dudul (2009) and Kale and Dudul (2009). Fig. 2 shows

Fig. 2. Time-lagged recurrent network structure (Wang and Traore, 2009).

144

P.T. Nastos et al. / Atmospheric Research 144 (2014) 141150

a typical time-lagged recurrent network used in this study

where (Z1) represents the delay operator (Wang and Traore,
2009).
Due to the fact that the examined period 18912009 (119
annual maximum daily precipitation data) is not enough long
for an appropriate ANN training concerning maximum daily
precipitation, the following method was applied. Firstly, a
data matrix was constructed consisting of 119 rows and 7
columns. Each line of the matrix represents the first, second,
third and seventh daily maximum precipitation totals of a
specific year, ranking in decreasing order. Thus, for example
the first line of the matrix seems like that: 26.1; 21.8; 14.4;
10.2; 9.1; 8.1; 7.2 representing the seven daily maximum
precipitation totals (mm) for the year 1891 in decreasing
order, respectively. Thereafter, factor analysis (FA) was applied
on the previous matrix, in order to reduce the dimensionality of
the matrix (Nastos et al., 2002; Bartzokas et al., 2003).
The main applications of FA are to reduce the number of
variables and to detect the structure in the relationships
between variables that is to classify variables. Therefore, FA is
applied as a data reduction or structure detection method. The
data should have a bivariate normal distribution for each pair
of variables, and observations should be independent. The
applied Factor analysis is based on the Principal Components
Analysis (PCA) technique that rewrites the original data matrix
into a new set of components that are linearly independent and
ordered by the amount of variance they explain and results
in limiting the number of variables used. Therefore, each of
the p initial variables X1, X2, , Xp can be expressed as a
linear function of m (m b p) uncorrelated factors: Xi =
ai1F1 + ai2F2 + + aimFm where F1, F2, , Fm are the factors
and ai1, ai2, , aim are the factor loadings which express
the correlation between the factors and the initial variables.
The values of each factor are called factor scores and they are
presented in standardized form, having zero mean and unit
variance (Jolliffe, 1986; Manly, 1986). The number m of the
retained factors has to be decided, by using various rules
(eigenvalue 1, scree plot) and considering the physical
interpretation of the results. Another important point of the
analysis is the rotation of the axes, which maximizes some
factor loadings and minimizes some others and in that way a
better separation among the initial variables is succeeded.
Varimax rotation is generally accepted as the most accurate
orthogonal rotation, which maximizes the sum of the variances
of the square factor loadings, keeping the factors uncorrelated
(Richman, 1986).
The depth of the memory in the developed ANN was
setup to 10. The learning rule for each layer in this study
was the momentum. The activation function for the hidden
units as well as the outputs units is the following function
(TanhAxon function) (Rumelhart et al., 1986):

f x

ex ex
e2x 1
:
x
x 2x
e e
e 1

All the above have been chosen after a trial and error
method. The numbers of nodes, so-called processing elements (PEs) in the hidden layer, the step size, as well as the
momentum rate were determined by the use of genetic

algorithms (GA). Genetic Algorithms are one of the most

successful and optimistic techniques of a new generation of
soft computing which includes fuzzy-logic, ANN and supportvector machines (Kamp and Savenije, 2006). Genetic Algorithms use these strategies to find an optimum solution for
any multi-dimensional problem (Goldberg, 1989). Generally
the GA approach applied in this study in order to select
the best network architecture-structure, which finally consists of one input layer, one hidden layer and one output layer
(1, 1, 1).
For the evaluation of the capability and the predicted
ability of the developed ANN four statistical indices, the coefficient of determination (R2), the index of agreement (IA),
the Root Mean Square Error (RMSE) and the Mean Bias Error
(MBE) were used.
The coefficient of determination is calculated according to
the equation (Krause et al., 2005):
2

32

n
X
Oi Oave  P i P ave

6
7
6
7
2
i1
s7
R 6
6s
7
n
n
X
4 X
2
25
Oi Oave 
P i P ave
i1

i1

where n is the number of observations, Oi is the observed

value, Pi is the predicted value, Oave is the average value of all
the observed values and Pave is the average value of all the
predicted values. The coefficient of determination represents
the percentage of the data closest to the line (y = b x + a)
of best fit (Kolehmainen et al., 2001). For example, R2 =
0.850 means that 85% of the total variation in y can be
explained by the linear relationship between x and y. If R2 is
used for model validation, it is therefore advisable to take
into account additional information, which can cope with
that problem. Such information is provided by the gradient
(b) and the intercept (a) of the regression on which R2 is
based. For a good agreement the intercept (a) should be close
to zero, which means that an observed runoff of zero would
also result in a prediction near zero and the gradient (b)
should be close to one. For a proper model assessment the
gradient (b) should always be discussed together with R2
(Krause et al., 2005).
The Index of Agreement is calculated according the
equation (Willmott et al., 1985):
n
X

IA 1

n
X

P i Oi

i1

jP i Oave j jOi Oave j

i1

The index of agreement is a dimensionless index within

the range of 01. IA = 0 means no agreement between
prediction and observation and IA = 1 means perfect
agreement between prediction and observation (Willmott,
1982; Walker et al., 1999).
The Root Mean Square Error is a commonly used measure
of the differences between the values extracted by a forecasting model or an estimator and the real-observed values.

P.T. Nastos et al. / Atmospheric Research 144 (2014) 141150

RMSE is calculated according the equation (Moustris et al.,

2010):
RMSE

n
1X
2
P Oi
n i1 i

!1
:

n
1X
P Oi :
n i1 i

The MBE provides information in the long term performance of the correlations by allowing a comparison of the
actual deviation between predicted and measured values
term by term. The MBE is used to describe whether a model
over-(positive value) or under-(negative value) predicts the
observation and has the same units as the measured variableparameter, which is forecasted by the model. The ideal value
of MBE is zero (Moustris et al., 2010).
3. Results and discussion
Two factors, which explain 87% of the total variance, were
extracted (Table 1). Thus, for each year the seven maximum
daily precipitation totals in decreasing order were replaced
by two factors (F1 and F2), and this new matrix (119 2) is
used for training and validation of the developed ANN.
Table 2 presents the input and the output data used/
extracted in the developed ANN. The data were divided into
two subsets. The first subset concerns the training data (training
set) and includes the years 18911980. The 80% of the first
subset of data was used for the ANN training and the rest 20% for
the ANN cross validation test during the phase of training. The
validation data (20% of the years 18911980) were randomly
chosen by the model during the phase of training and were used
in order to check the progress of the ANN training, optimizing
its parameters (Moustris et al., 2010). The second subset
(testing set) includes the years 19812009 and was used to
evaluate the forecasting ability of the developed ANN model. It
should be noted here that the testing set is completely unknown
to the ANN and has not been presented-disclosed at any stage
during its training. The shape of training data-input data file
(Table 2) was chosen after a trial-and-error method. This shape
Table 1
Results of the applied factor analysis.

1.
2.

Table 2
Input and output data for the ANN training.
Input data (input layer)

Output data (output layer)

The values of factors F1 and F2

from the five previous years

The maximum daily total

precipitation (mm) for the
next year

These individual differences are also called residuals. The

RMSE has the same units as the measured variable-parameter,
which is forecasted by the model. The value of RMSE provides
information on the short term performance. The value of RMSE
is always positive, represented as zero in the ideal case
(Gunhan et al., 2005). The smaller the RMSE is, the closer the
forecasted values are with the real-observed values.
The Mean Bias Error is calculated according to the
equation (Gunhan et al., 2005):
MBE

145

Eigenvalue

Total
variance
(%)

Cumulative
eigenvalue

Cumulative
total variance
(%)

5.018502
1.051178

71.69289
15.01683

5.018502
6.069681

71.69289
86.70972

was given the best training of ANN according to the available

data.
Fig. 3 depicts the Box & Whiskers plots for the first seven
maximum daily total precipitation values within the year for
the period 18912009. The mean maximum daily precipitation totals (middle point), standard deviation of the mean
(box) and the absolute maximum and minimum value of
precipitation (whiskers) are presented in ranking order. It is
obvious that the mean estimates (middle points) between
the first and the seventh maximum daily precipitation value
within the year shows a range of 33.2 mm. As far as the
absolute maximum values are concerned (whiskers), a difference of 144.8 mm is observed between the first and the
seventh maximum daily precipitation value, against 12.8 mm
with respect to the absolute minimum values. Looking at the
standard deviation of the mean value, an abrupt decrease
appears from the first to the second maximum daily
precipitation, and thereafter the variability is lower among
the remaining ranked values.
The relationship between the seven mean daily maximum
precipitation totals and the ranking order during the period
18912009 is depicted in Fig. 4. A power function (Eq. (6))
express satisfactorily this relationship (p b 0.01) with a
coefficient of determination R2 = 0.999.
0:562

y 49:452x

;1 x 7

where y is the mean daily maximum precipitation totals

during the period 18912009 and x is the ranking order of
occurrence during the year. The above analysis showed the
existed strong relationship between the seven consecutive
mean maximum daily precipitation totals and the ranking
order, meaning that the use of these consecutive figures for
ANNs training improved the forecasting capacity of the
developed models.
Fig. 5 presents the 5-year moving average of the maximum
daily precipitation totals per year recorded at NOA, during the
examined time period 18912009. The highest value of the
maximum daily precipitation totals appear during the period
18951900 (189.6 mm/day in 1895), and thereafter an abrupt
decrease is obvious in the next five years. Then, the maximum
daily precipitation oscillates around 50 mm with secondary
peaks appearing within the years 19101915 (114.9 mm/day
in 1912), 19491954 (142.9 mm/day in 1949) and 20002005
(91 mm/day in 2002). There is not a statistically significant
trend (at 95% confidence level) during the examined period.
Nastos and Zerefos (2007) showed a statistically significant (at
95% confidence level) increasing trend of the number of days
with daily precipitation totals exceeding specific thresholds,
especially 50 mm. This is particularly important as these events
(N 50 mm per day) may cause considerable damage and loss of
life. These increasing trends are more evident and statistically

146

P.T. Nastos et al. / Atmospheric Research 144 (2014) 141150

Fig. 3. Box & Whiskers plots. Mean maximum daily precipitation totals (middle point), standard deviation of the mean (box) and the absolute maximum and
minimum value of precipitation (whiskers) vs. ranking order at National Observatory of Athens, Greece, within the period 18912009.

significant (p b 0.05) since 1970 to present. Besides, Philandras

et al. (2010) found that the annual and seasonal mean and
maximum rain intensity time series at Athens (NOA) present
significant increasing trends, starting in 1990's, during fall and
winter and become more intense within the last decade, while
the increasing trends begin earlier in 1980's during spring and
summer. The observed increase in the frequency of extreme
events could be attributed to the urban heat island (UHI) in the
wider area of the National Observatory of Athens (NOA). The
urbanization effect is attributed to the extensive building of
Athens around NOA after the Second World War and the rapid
increase of the population and the number of vehicles mainly
after 1970. The urbanization effect in NOA refers mainly to

maximum air temperature (an increase ~2 C) and to the

warmer seasons of the year (Philandras et al., 1999). A lot of
studies have been carried out providing evidence that UHI is
associated with convective precipitation in Atlanta (Bornstein
and Lin, 2000), in Mexico City (Jauregui and Romales, 1996), in
Tel Aviv (Goldreich and Manes, 1979), in Beijing City (Guo et
al., 2006), in Tokyo (Yonetani, 1982), in London (Atkinson,
1971) and in Ankara (Cicek and Turkoglu, 2005).
The best training for the developed ANN was succeeded
using data from the previous five years (Table 2), after a
trial-and-error methodology. After the appropriate training
of ANN, the second subset of data concerning the period
19812009, which was absolutely unknown to the trained

Mean maximum daily precipitation totals

(mm)

55
50
45

y = 49.452x -0.562

40

R2 = 0.999

35
30
25
20
15
10
5
0
0

Ranking order
Fig. 4. Mean maximum daily precipitation totals vs. ranking order at National Observatory of Athens, Greece, within the period 18912009.

Maximum daily total precipitation (mm)

P.T. Nastos et al. / Atmospheric Research 144 (2014) 141150

147

200
180
160
140
120
100
80
60
40
20

18
90
18
95
19
00
19
05
19
10
19
15
19
20
19
25
19
30
19
35
19
40
19
45
19
50
19
55
19
60
19
65
19
70
19
75
19
80
19
85
19
90
19
95
20
00
20
05
20
10

Year
Fig. 5. Five year moving average of the maximum daily total precipitation totals at National Observatory of Athens, Greece, within the period 18912009.

ANN, was used to evaluate the forecasting ability of the

developed ANN model. The results of the evaluation are
shown in Table 3. It seems that the forecasting ability of the
developed ANN model is satisfactory enough with RMSE =
16.4 mm (total rainfall = 50 mm), which is within acceptable limits, taking into consideration the results from other
relative studies (RMSE = 4.7 mm-total rainfall = 43 mm,
Wardah et al., 2008; RMSE = 54.2 mm-total rainfall =
840 mm, Mar and Naing, 2008), where RMSE was divided
by the respective mean annual rainfall in order to be
normalized and thereafter compared. Thus, the normalized
RMSE values equal to 0.3, 0.1 and 0.1 for the present and the
aforementioned studies respectively.
The prognosis of such a random and extreme event, such
as the maximum daily precipitation totals for the next year it
is a very difficult matter. However, the developed ANN shows
significant and substantial forecasting ability. Fig. 6 depicts
the time series of observed (blue line) and forecasted (red
line) maximum daily precipitation totals at NOA for the test
period 19812009.
The developed ANN seems to overestimate the maximum
daily precipitation totals appeared in 1988, while underestimate the maximum in 1999. As far as the rest of the years
are concerned, the forecasting of maximum daily precipitation
totals per year is satisfactory enough with relatively small
differences between forecasted and observed values (Table 3).
The index of agreement (IA = 0.817) indicates that in general
terms the forecasted values are very close to the observed
values. The ANN's weakness in forecasting successfully the
extreme peaks could be attributed to the relatively low
frequency of occurrence of these events within GAA, which
are very rare. This limitation may be due to the fact that other

factors-input data are required for a better training of ANN. We

tried to improve better the forecasting ability of ANN taking
into account the annual absolute maximum and minimum
temperatures and the annual temperature range as inputs for
the appropriate ANN's training. The results were much worse
compared to those without the use of the three temperature
parameters (R2 equals to 0.081, 0.092 and 0.112 respectively),
indicating that the examined temperature parameters do not
improve the ANN's ability to forecast the annual maximum
daily precipitation. Despite that, another important issue is that
longer historical datasets are necessary for a better training,
taking into consideration that the examined time series
concerns 119 values (annual maximum daily precipitation
totals during the period 18912009). We acknowledge the
limitations of our effort, but we consider the extracted results
of the performed methodology quite satisfactory in forecasting
such extreme values for the next year, based only on historical
daily precipitation time series.
The relationship between the observed and the forecasted
values of maximum daily precipitation totals per year at
National Observatory of Athens for the examined period
19812009 is depicted in Fig. 7. The black dashed line
corresponds to perfect fit (y = x), while the red line to the
least-square fit. The 48.2% (R2 = 0.482) of the total variation
in y during the examined period 19812009 can be explained
by the linear relationship between x and y:
y 0:673x 21:089

where, (x) represents the observed values of the maximum

daily total precipitation and (y) represents the forecasted
values by the developed ANN model.
4. Conclusions

Table 3
Values of statistical indices for the evaluation of ANN forecasting ability.
MBE (mm)

RMSE (mm)

R2

IA

+5.2

16.4

0.482

0.817

The forecasting of maximum daily precipitation totals for

the next coming year using traditional statistical models and
methods is a very difficult issue. In this work an effort was
carried out in order to forecast the maximum daily precipitation totals for the next year with the application of the

P.T. Nastos et al. / Atmospheric Research 144 (2014) 141150

Maximum daily total precipitation (mm)

148

140
OBSERVED VALUES (mm)

PREDICTED VALUES (mm)

120
100
80
60
40
20

19
8
19 1
8
19 2
8
19 3
8
19 4
8
19 5
86
19
8
19 7
8
19 8
8
19 9
9
19 0
91
19
9
19 2
9
19 3
9
19 4
9
19 5
9
19 6
9
19 7
9
19 8
99
20
0
20 0
0
20 1
0
20 2
0
20 3
0
20 4
0
20 5
0
20 6
07
20
0
20 8
09

Year
Fig. 6. Time series of observed (blue line) and forecasted (red line) maximum daily precipitation totals at National Observatory of Athens for the forecasting
period 19812009.

artificial neural networks. It is worthy to remark that this

study is a pioneer one concerning the wider area of Athens,
Greece. The results are satisfactory enough and encouraging,
taking into consideration such a random and extreme event.
The index of agreement (IA = 0.817) and the coefficient of
determination (R2 = 0.482) indicate a quite satisfactory relationship between the forecasted and observed values (p b 0.01).
The advantage of the proposed ANN model is that it works
only with historical daily precipitation totals (ex post data).
Further, the knowledge of the maximum daily precipitation
totals from the five previous years is just required. Nevertheless, the disadvantage of the proposed model is that it would
require a great number of data (exemplars) for an appropriate
training and it is not able yet to predict accurately the peaks of
the extreme maximum daily precipitation, as well. The use of

more available daily precipitation time series in the future in

combination with more research in this discipline will improve
ANN's forecasting capacity for such an extreme event, mitigating the adverse impacts in infrastructure of urbanized areas
and even in life loss. The performed methodology, even with
the aforementioned deficiencies, is scheduled to be used within
the research and educational activities of the Laboratory of
Climatology and Atmospheric Environment of the University of
Athens and the Met Service, in an operational program plan
for the estimation of extreme daily precipitation one year
ahead. In a future work we will try to optimize the ANNs ability
by including broad-scale atmospheric circulation indices such
as North Atlantic Oscillation Index, Mediterranean Oscillation
Index and North Caspian Pattern Index having impacts on
Mediterranean precipitation totals.

Forecasted precipitation (mm)

140
120
100
80
60

y = 0.673x + 21.089
R2 = 0.482

40
20
0
0

20

40

60

80

100

120

140

Observed precipitation (mm)

Fig. 7. ANN forecasted vs. observed values of maximum daily precipitation totals per year at National Observatory of Athens for the forecasting period 19812009.

P.T. Nastos et al. / Atmospheric Research 144 (2014) 141150

References
Alpert, P., Ben-Gai, T., Baharad, A., Benjamini, Y., Yekutieli, D., Colacino, M.,
Diodato, L., Homar, V., Ramis, C., Romero, R., Michaelides, S., Manes, A.,
2002. The paradoxical increase of Mediterranean extreme daily rainfall
in spite of decrease in total values. Geophys. Res. Lett. 29 (11), 31-131-4.
http://dx.doi.org/10.1029/2001GL013554.
Badjate, S.L., Dudul, S.V., 2009. Multi step ahead prediction of north and south
hemisphere sun spots chaotic time series using focused time lagged
recurrent neural network model. WSEAS Trans. Inf. Sci. Appl. 6 (4),
684693.
Barlow, M., 2011. Influence of hurricane-related activity on North American
extreme precipitation. Geophys. Res. Lett. 38, L04705. http://dx.doi.org/
10.1029/2010GL046258 (5 pp.).
Bartzokas, A., Lolis, C.J., Metaxas, D.A., 2003. The 850 hPa relative vorticity
centres of action for winter precipitation in the Greek area. Int. J. Climatol.
23, 813828.
Beniston, M., Stephenson, D.B., Christensen, O.B., Ferro, C.A.T., Frei, C.,
Goyette, S., Halsnaes, K., Holt, T., Jylh, K., Koffi, B., Palutikof, J., Schll, R.,
Semmler, T., Woth, K., 2007. Future extreme events in European climate:
an exploration of regional climate model projections. Clim. Chang. 81,
7195.
Bodri, L., Cermak, V., 2000. Prediction of extreme precipitation using a neural
network: application to summer flood occurrence in Moravia. Adv. Eng.
Softw. 31, 211221.
Bornstein, R., Lin, Q., 2000. Urban heat islands and summertime convective
thunderstorms in Atlanta: three cases studies. Atmos. Environ. 34,
507516.
Brunetti, M., Maugeri, M., Nanni, T., 2001. Changes in total precipitation,
rainy days and extreme events in northeastern Italy. Int. J. Climatol. 21,
861871.
Caudill, M., Butler, C., 1922. Understanding Neural Networks. MIT Press.
Cicek, I., Turkoglu, N., 2005. Urban effects on precipitation in Ankara.
Atmoosfera 18 (3), 173187.
Cigizoglou, H.K., Alp, M., 2004. Rainfall-runoff modelling using three neural
network methods. Artificial Intelligence and Soft Computing-ICAISC
2004 Lecture Notes in artificial Intelligence, 3070, pp. 166171.
Diodato, N., Bellocchi, G., 2008. Drought stress patterns in Italy using agroclimatic indicators. Clim. Res. 36 (1), 5363.
Easterling, D.R., Evans, J.L., Groisman, P.Ya., Karl, T.R., Kunkel, K.E., Ambenje,
P., 2000. Observed variability and trends in extreme climate events: a
brief review. Bull. Am. Meteorol. Soc. 81, 417425.
Freiwan, M., Cigizoglu, H.K., 2005. Prediction of total monthly rainfall in
Jordan using feed forward back propagation method. Fresenius Environ.
Bull. 14 (2), 142151.
Furrer, E.M., Katz, R.W., 2008. Improving the simulation of extreme precipitation events by stochastic weather generators. Water Resour. Res. 44,
W12439. http://dx.doi.org/10.1029/2008WR007316.
Goldberg, D.E., 1989. Genetic Algorithms in Search, Optimization, and
Machine Learning. Addison-Wesley Pub. Co.
Goldreich, Y., Manes, A., 1979. Urban effects on precipitation patterns in the
greater Tel-Aviv area. Archiv fr Meteorologie, Geophysik und
Bioklimatologie, Serie B 27 (23), 213224.
Goswami, P., Shivappa, H., Goud, B.S., 2010. Impact of urbanization on
tropical mesoscale events: investigation of three heavy rainfall events.
Meteorol. Z. 19 (4), 385397.
Groisman, P.Ya., Knight, R.W., Easterling, D.R., Karl, T.R., Hegerl, G.C., 2005.
Trends in intense precipitation in the climate record. J. Clim. 18 (9),
13261350.
Gunhan, T., Demir, V., Hancioglu, E., Hepbasli, A., 2005. Mathematical
modelling of drying of bay leaves. Energy Convers. Manag. 46, 16671679.
Guo, X., Fu, D., Wang, J., 2006. Mesoscale convective precipitation system
modifiedd by urbanization in Beijing City. Atmos. Res. 82, 112126.
Hect-Nielsen, R., 1990. Neurocomputing. Addison-Wesley, Reading, M.A.
IPCC, 2001. Climate Change 2001: The Scientific Basis. Contribution of
Working Group 1 to the Third IPCC Scientific Assessment, Chapter 8,
Model Evaluation.
IPCC, 2007. Summary for policymakers. In: Solomon, S., Qin, D., Manning, M.,
Chen, Z., Marquis, M., Averyt, K.B., Tignor, M., Miller, H.L. (Eds.), Climate
Change 2007: The Physical Science Basis. Contribution of Working
Group I to the Fourth Assessment Report of the Intergovernmental Panel
on Climate Change. Cambridge University Press, Cambridge, United
Kingdom/New York, NY, USA.
Iwashima, T., Yamamoto, R., 1993. A statistical analysis of the extreme
events: long-term trend of heavy daily precipitation. J. Meteorol. Soc.
Jpn. 71, 637640.
Jauregui, E., Romales, E., 1996. Urban effects on convective precipitation in
Mexico City. Atmos. Environ. 30 (20), 33833389.
Jolliffe, I.T., 1986. Principal Component Analysis. Springer-Verlag, New York.

149

Kale, S.N., Dudul, S.V., 2009. Intelligent noise removal from EMG signal using
focused time-lagged recurrent neural network. Applied Computational
Intelligence and Soft Computing Article ID 129761, p. 12. http://
dx.doi.org/10.1155/2009/129761.
Kambezidis, H.D., Larissi, I.K., Nastos, P.T., Paliatsos, A.G., 2010. Spatial
variability and trends of the rain intensity over Greece. Adv. Geosci. 26,
6569.
Kamp, R.G., Savenije, H.H.G., 2006. Optimising training data for ANNs with
genetic algorithms. Hydrol. Earth Syst. Sci. Discuss. 3, 285297.
Karl, T.R., Knight, R.W., 1998. Secular trends of precipitation amount frequency
and intensity in the United States. Bull. Am. Meteorol. Soc. 79, 231241.
Karl, T.R., Knight, R.W., Plummer, N., 1995. Trends in high-frequency climate
variability in the twentieth century. Nature 377, 217220.
Kolehmainen, M., Martikainen, H., Ruuskanen, J., 2001. Neural networks and
periodic components used in air quality forecasting. Atmos. Environ. 35,
815825.
Kostopoulou, E., Jones, P.D., 2005. Assessment of climate extremes in the
Eastern Mediterranean. Meteorol. Atmos. Phys. 89, 6985.
Koutsoyiannis, D., 2004. Statistics of extremes and estimation of extreme
rainfall: I. Theoretical investigation. Hydrol. Sci. J. 49 (4), 575590.
Krause, P., Boyle, D.P., Bsel, F., 2005. Comparison of different efficiency
criteria for hydrological model assessment. Adv. Geosci. 5, 8997.
Kundzewicz, Z.W., Radziejewski, M., Pinskwar, I., 2006. Precipitation
extremes in the changing climate of Europe. Clim. Res. 31 (1), 5158.
Lehner, B., Dll, P., Alcamo, J., Henrichs, T., Kaspar, F., 2006. Estimating the
impact of global change on flood and drought risks in Europe: a
continental, integrated analysis. Clim. Chang. 75 (3), 273299.
Luck, K.C., Ball, J.E., Sharma, A., 2000. A study of optimal model lag and
spatial inputs to artificial neural networks for rainfall forecasting.
J. Hydrol. 227 (14), 5665.
Manly, B.F.J., 1986. Multivatiate Statistical Methods: A Primer. Chapman &
Hall, London.
Manzat, ., 2007. Sounding-derived indices for neural network based shortterm thunderstorm and rainfall forecasts. Atmos. Res. 83, 349365.
Mar, K.W., Naing, T.T., 2008. Optimum neural network architecture for
precipitation prediction of Myanmar. World Acad. Sci. Eng. Technol. 48,
130134.
Marzano, F.S., Fionda, E., Ciotti, P., 2006. Neural-network approach to
ground-based passive microwave estimation of precipitation intensity
and extinction. J. Hydrol. 328, 121131.
Min, S.-K., Zhang, X., Zwiers, F.W., Hegerl, G.C., 2011. Human contribution
to more-intense precipitation extremes. Nature 470, 378381. http://
dx.doi.org/10.1038/nature09763.
Moustris, K.P., Ziomas, I.C., Paliatsos, A.G., 2010. 3-day-ahead forecasting of
regional pollution index for the pollutants NO2, CO, SO2 and O3 using
artificial neural networks in Athens, Greece. Water Air Soil Pollut. 209,
2943.
Moustris, K.P., Larissi, I.K., Nastos, P.T., Paliatsos, A.G., 2011. Precipitation
forecast using artificial neural networks in specific regions of Greece.
Water Resour. Manag. 25, 19791993.
Nastos, P.T., Zerefos, C.S., 2007. On extreme daily precipitation totals at
Athens, Greece. Adv. Geosci. 10, 5966.
Nastos, P.T., Zerefos, C.S., 2008. Decadal changes in extreme daily precipitation
in Greece. Adv. Geosci. 16, 5562.
Nastos, P.T., Zerefos, C.S., 2009. Spatial and temporal variability of consecutive
dry and wet days in Greece. Atmos. Res. 94, 616628.
Nastos, P.T., Philandras, C.M., Repapis, C.C., 2002. Application of canonical
analysis to air temperature and precipitation regimes over Greece.
Fresenius Environ. Bull. 11 (8), 488493.
Paliatsos, A.G., Nastos, P.T., Tzavelas, G., Panagiotakos, D.B., 2005. Characteristics of precipitation in urban Athens area, from 1891 to 2000. Fresenius
Environ. Bull. 14, 422428.
Pall, P., Aina, T., Stone, D.A., Stott, P.A., Nozawa, T., Hilberts, A.G.J., Lohmann,
D., Allen, M.R., 2011. Anthropogenic greenhouse gas contribution to
flood risk in England and Wales in autumn 2000. Nature 470, 382385.
Philandras, C.M., Metaxas, D.A., Nastos, P.T., 1999. Climate variability and
urbanization in Athens. Theor. Appl. Climatol. 63 (12), 6572.
Philandras, C.M., Nastos, P.T., Paliatsos, A.G., Repapis, C.C., 2010. Study of the
rain intensity in Athens and Thessaloniki, Greece. Adv. Geosci. 23, 3745.
Richman, M.B., 1986. Rotation of principal components. J. Climatol. 6,
293335.
Rumelhart, D.E., Hinton, G.E., Williams, R.J., 1986. In: Rumelhart, D.E.,
McClelland, J.L. (Eds.), Parallel Distributed Processing Explorations in the
Microstructure of Cognition, 1. MIT Press, Bradfords Books, Cambridge,
MA, pp. 318362.
Sahai, A.K., Soman, M.K., Satyan, V., 2000. All India summer monsoon rainfall
prediction using an artificial neural network. Clim. Dyn. 16, 291302.
Sakellariou, N.K., Kambezidis, H.D., 2004. Prediction of the total rainfall
amount during August and November in the Athens area, Greece.
Fresenius Environ. Bull. 13 (3), 289292.

150

P.T. Nastos et al. / Atmospheric Research 144 (2014) 141150

Sheffield, J., Wood, E.F., 2007. Characteristics of global and regional drought,
19502000: analysis of soil moisture data from off-line simulation of the
terrestrial hydrologic cycle. J. Geophys. Res. D Atmos. 112 (17) (No
D17115).
Silverman, D., Dracup, J.A., 2000. Artificial neural networks and long-range
precipitation predictions in California. J. Appl. Meteorol. 39 (1), 5766.
Suppiah, R., Hennessy, K.J., 1998. Trends in total rainfall events and number
of dry events in Australia. 19101990. Int. J. Climatol. 18, 11411164.
Walker, S.-E., Slordal, H.L., Guerreiro, C., Gram, F., Grnskei, E.K., 1999. Air
pollution exposure monitoring and estimation Part II. Model evaluation
and population exposure. J. Environ. Monit. 1, 321326.
Wang, Yu.-M., Traore, S., 2009. Time-lagged recurrent network for forecasting
episodic event suspended sediment load in typhoon prone area. Int.
J. Phys. Sci. 4 (9), 519528.

Wardah, T., Abu Bakar, S.H., Bardossy, A., Maznorizan, M., 2008. Use of
geostationary meteorological satellite images in convective rain estimation for flash-flood forecasting. J. Hydrol. 356, 283298.
Willmott, C.J., 1982. Some comments on the evaluation of model performance. Bull. Am. Meteorol. Soc. 63 (11), 13091313.
Willmott, C.J., Ackleson, S.G., Davis, R.E., Feddema, J.J., Klink, K.M., Legates,
D.R., O'Donnell, J., Rowe, C., 1985. Statistics for the evaluation and
comparison of models. J. Geophys. Res. 90, 89959005.
Xue, Y., Dibike, Y.B., 2001. Flood forecasting mode for Huai River in China using
time delay neural network. Proceedings of the XXIX IAHR congress,
Beijing, China, Theme C, pp. 5966.
Yonetani, T., 1982. Increase in Number of Days with Heavy Precipitation in
Tokyo Urban Area. J. Appl. Meteor. 21, 14661471. http://dx.doi.org/
10.1175/1520-0450(1982)021 b1466:IINODWN2.0.CO;2.