Neural Network Used In Thunderstorm Prediction
Viviana-Mihaela Bostan (Damian), Brandusa Pantelimon
Polytechnic University of Bucharest
Faculty of Electrical Engineering
Measurements Department, Electrical and Static Converters
Bucharest, Romania
damianviviana@gmail.com, brandusa.pantelimon@upb.ro
Abstract- Extreme phenomena such as thunderstorm, heavy
snowfall, flooding or tornados are no longer strange in our daily
forecast. The thunderstorm represents one of the global
phenomena that can appear anywhere in the world at any time. It
is also known as lightning storm or hailstorm associated with
tornados. This storm is a type of weather characteristic
containing strong wind, lightning, tornados, heavy rain, and
sometimes hail. Nevertheless, even if thunderstorm is generally
very short-lived phenomena it has a great potential to produce
serious damage to human life and property. Lightning damaging,
tornados, large sized hail, heavy precipitation, and flooding can
occur anytime during a thunderstorm. Many regions in our
country experience thunderstorms at higher frequency during
spring and summer months (April–September). Severe
thunderstorms form and move generally from northwest to
southeast over the south-eastern regions of Romania during the
summer season. This powerful thunderstorms associated other
elements such as thunder lightning, tornados and hail usually
cause extensive loss in agriculture, damage to property, flooding
and also loss of life. The casualties reported due to lightning
associated with thunderstorms in the field region are the highest
in our country. These severe thunderstorms have significant
socioeconomic impact in the eastern and south eastern parts of
the country. An accurate location specific and timely prediction
are required to avoid loss of lives and property due to strong
winds and heavy precipitation associated with these severe local
storms. As such, the uncertainty of both the future occurrence
and amount of precipitation can have a negative impact on many
sectors of our economy, like agriculture. There is, therefore, a
need to use innovative computer technologies such as artificial
intelligence to improve the accuracy of precipitation predictions.
Artificial neural networks have been shown to be useful as an aid
for the prediction of weather variables.
Keywords-component; ANN; thunderstorms; hail and rainfall
prediction
I. INTRODUCTION
Global warming is occurring and is scientifically
proven by recording and studying climatic history from the
19th century until now: researchers show that the trend of the
last 30 years is global heating with extreme phenomena such
as thunderstorms, rainfall asociated with hail, heavy snowfall,
floodings or tornados [1]. Like an example the first tornado
was noticed in Romania in Bucharest in June 1886. After the
first one some others followed.
The most powerful tornado encountered in Romania
was the one from Făcăieni in Ialomița on the 12 of August
2002. This was an F3 tornado on the Fujita scale, which means
a severe tornado with wind speed from 252 km/h up to 330
km/h. This was caused by high thermal difference between
two air masses a cold polar one and a tropical one, which
intersected the country.
The effects of tornadoes in the commune Făcăieni,
over 2 minutes were disastrous: 33 houses completely
destroyed and 395 houses partially damaged, 14 people
seriously injured, 2 people dead, 1,000 people stricken, 120 ha
of acacia forest destroyed by breaking the trees to about 1 m
from the surface.
Other extreme phenomena occurred in our country on
the 28 of May 2005 in the commune Pielești from Dolj where
falling rain was accopmanied by windstorm and hail. The hail
size was between 5 mm and up to the size of a pigeon’s egg
(about 4-5cm), and the amount of hail on the ground was 20 to
30 cm that formed a layer which resisted until the next day at
noon. Hail duration was of 20-25 minutes and the rainfall was
about 2 hours.
The most intense hail falling occurred in Oltenia, in
Băileşti in 1995 when the hail dimentions were the size of a
goose egg. Precipitation, in meteorology, is defined as any
product, liquid or solid, of atmospheric water vapor that is
accumulated onto the earth’s surface. Water, and thus
precipitation, has a major impact on our daily livelihood. [2]
In our country the thunderstorms depends upon various
parameters such as pressure level in the given region as
compared to surrounding region, humidity, vapor pressure etc.
If we look at the trends and the pattern of
thunderstorms in the central part of Romania we find that the
rainfall on one day has its effect on the rainfall for the next
few days. The thunderstorms appear for number of days
together and then they disappear for number of days.
There is a correlation between the thunderstorms on a
given day with the thunderstorms on previous days. There are
two possible methods for thunderstorms prediction.
The first one is based on the study of the rainfall
processes and its dependence on other meteorological
parameter such as pressure, humidity, vapor pressure,
temperature and the impact between two different air masses
etc.
 However, this method is complex because
thunderstorm is the result of a number of many atmospheric
parameters which vary both in space and time, and the
parameters are limited in both the spatial and temporal
dimensions [3].
Another method to forecast thunderstorms is based on
the pattern recognition methodology. In these method relevant
spatial and temporal features of thunderstorms series in past
are extracted. These features are then utilized to predict the
thunderstorms in future.
In this technique only rainfall data is considered and
therefore this approach is considered appropriate for short-
duration thunderstorms prediction. If thunderstorm pattern can
be estimate then it can help in predicting the thunderstorm for
the next day [4, 5].
II. DATA ANALYSING

Therefore in this paper an attempt has been made to

estimate thunderstorms pattern using neural network. Neural
network is the best tool for pattern recognition. All
computational works were carried out using the programming
code of MATLAB.
This thunderstorms prediction model developed in
the paper is based on the use of Artificial Neural Networks to
understand the thunderstorms pattern and also to predict the
future thunderstorms.
The ANN tries to replicate the parallel distributed
processing of the human nervous system and have proven to
be a very powerful tool in dealing with complicated problems,
such as pattern recognition and approximation function.
Hornik [6] has shown that an ANN with adequate
complexity is capable of approximating any function to a
reasonable accuracy. In addition, ANN is computationally
robust, in the way that they have the ability to learn and give
correct output even if the input contains noise.
There have been a number of reported studies that
have used ANN to solve problems in hydrology. For example,
French [7] used an ANN to forecast thunderstorms for
catchments with artificial thunderstorm inputs, while Hsu [8]
applied an ANN to model the thunderstorm-runoff process.
Thunderstorm process is nonlinear as well as random
process and therefore most previous studies have used a
Multilayer perceptron (MLP) or generalized feed forward
neural network (GFNN), without considering the other types
of ANN.
Although the GFNN is a powerful nonlinear model,
there are many other types of ANN which may have been
more appropriate to the problem. ANN appropriate for rainfall
forecasting includes the time-delay neural networks (TDNN).
Historical rainfall data was collected from Brașov
stations from 1961 to 1990 [8]. The graphical meteorological
data of 1961–1990 shows that the average annual relative
humidity (RH) is about 49, 51%.
Figure 1. Annual relative humidity in Brașov
The equation can be described as:
DNorm=(D0-Dmin)/ (Dmax-Dmin)
Where: D0 is for the observed sdata, Dmax and Dmin are
maximum and minimum of the observed data, respectively.
The scope of this paper is to develop a model to
predict thunderstorm one day ahead. As it is known that
thunderstorm is random process it has been modeled as a
process where the probability of a kth output depends upon the
finite previous output in sequence. This class of random
process is called as Markov process.
The rainfall value at a given location and time is a
function of a finite set of previous occurrence. With this
assumption, a model structure can be expressed as:
R(t + 1) = F (R(t),R(t - l),R(t – 2), . . . ,R(t - k + 1)) --(1)
Where: R (t+1) represent the next day rainfall
R (t) represents rainfall today
R (t-1) represents rainfall one day before and so on
F ( ) is a nonlinear mapping function, which will be
approximated using an ANN, k is the (unknown to be solved)
number of days in past usually, k refers to the lag of the
network; if k = 1, the rainfall at the next time-step is related
only to the present rainfall, thus giving a lag-l network.
There is a number of alternative ANN that can be
used in order to represent the Markovian model expressed in
the equation, where the continuous process of rainfall is being
represented by a discrete process. Based on the review of these
alternatives, two suitable ANN model were identified and
adopted in this study for comparison.
 Analyzing these alternatives, two suitable ANN
model were identified. They are: 1. Generalized feed forward
network (GFNN), 2. Time lag focused neural network
(FTLNN).
One of those models is: 1. Generalized feed forward
network (GFNN). In a neural network if the interconnection is
restricted to feed forwarding activations (neither feedback nor
self connections), the neural network is defined as feed
forward. Feed forward networks are instantaneous mappers;
i.e. the output is valid immediately after the presentation of an
input.
A special class of feed forward networks is the
layered class, which is called the multilayer perceptron
(MLP). This name comes from the fact that Rosenblatt’s
network, which was called the perceptron, consisted of a
single layer of nonlinear PE s without feedback connections.
Multilayer perceptrons shown in Fig.2 have an input
layer of source nodes and an output layer of neurons (i.e.,
computation nodes); these two layers connect the network to
the outside world. Layers that receive the input from the
external world are called the input layers, layers in contact
with the outside world are called output layers. In addition to
these two layers, the multilayer perceptrons usually has one or
more layers of hidden neurons.
The layers without direct access to the external world,
i.e. not connected to the input or output, are called hidden
layers, so called because these neurons are not directly
accessible. The hidden neurons extract important features
contained in the input data. The training of an MLP is usually
accomplished by using a back propagation (BP) algorithm [9].
Figure 2. Three layers FFNN model
The other model is:
2. Focused time lag neural network (FTLNN). To
extract temporal information, time is an essential dimension of
learning. First the temporal information from the rainfall series
is extracted and then this information can used to predict the
rainfall one unit ahead in time. Time parameter may be
incorporate into the design of a neural network implicitly or
explicitly.
A straight forward method of implicit representation
of time is to add a short-term memory structure in the input
layer of a static neural network. This memory structure is used
to store the past information, which can be used to analyze the
data in a more efficient way. The element in the memory
structure can be varied as per requirement [10].
The resulting network is called as focused time
lagged feed forward network (TLFN). The short-term memory
structure may be implemented in different form one of Forms
is Tapped-Delay-Line (TDL) Memory.
This is the most commonly used form of short-term
memory. It consists of p unit delays with (p + 1) terminals.
This may be viewed as a single input–multiple output
network.
The figure shows a focused TLFN network using the
combination of a TDL memory and multilayer perceptron. In
this figure the unit-delay is denoted by z-1. The memory depth
of a TDL memory is fixed at p, and its memory resolution is
fixed at unity, giving a depth resolution constant of p.
Figure 3. TLFN model
The two-column paper length is strictly limited to a
Two different models were developed for prediction of rainfall
one day ahead. In the first model a generalized feed forward
neural network (1-42-1) (GFNN) with one hidden layer was
designed [11].
The back propagation (BP) algorithm used for
learning was Levenberg-Marquardt (LM). The Levenberg-
Marquardt (LM) algorithm is one of the most appropriate
higher-order adaptive algorithms known for minimizing the
mean square error of a neural network. A TAHN activation
function was adopted for the hidden nodes, whereas a linear
activation function was used for the output nodes.
The use of a TAHN function was to enable
nonlinearity of the network. The TAHN function, however,
was not used for the output nodes because it would force the
output to be bounded between 0.0 and 1.0
 Input data applied to this network was the rainfall Focused TLFN model was considered for this
time series with delay. A delay means data from previous day. simulation. Input data given to the network was for last seven
Number of delays for which the model was tested was taken days and the rainfall prediction was done for two days ahead
from 1 to 7. It means data applied to the network was for last to seven days a head. The results are as shown in Table 2.
seven days to last one day. Network model was also tested for
different number of neurons in the hidden layer. TABLE II. CORRELATION WITH NUMBERS OF DAYS AHEAD PREDICTION
Optimum numbers of neurons in the hidden layer No. of Normalizes mean square error Correlation
were found to be 42. At this number the mean square error days coefficient
was found to be at minimum. ahead
Second network model developed was a Focused 2 0,0025 0,87
time lagged feed forward neural network (1-23-1) with one 3 0,0033 0,82
hidden layer and learning algorithm LM. This model was 4 0,0046 0,78
tested again for different number of lag and different number 5 0,0108 0,75
of neurons at hidden layer. 6 0,0135 0,68
Number of neurons in the hidden layer at which the 7 0,0150 0,62
model performance is optimum was found to be 23. This is
less as compare to that required in GFNN network model. From the result it is seen that if the prediction is done
Various network models have been tried to predict the more in advance the accuracy of prediction decreases or the
thunderstorms and therefore rainfall one day ahead. error of prediction increase. Therefore it is concluded that the
Thunderstorm prediction was done based on the temporal model designed is capable of short term rainfall prediction
pattern learning using neural network. Two neural network with better accuracy.
models are found to be best for thunderstorm and rainfall Also it is found that FTLNN model is better than
prediction. GFFN network for learning temporal pattern. So we can make
Both the models have been tested with different lags a remarkable conclusion that the networks with a lesser lag
and different number of neurons in the hidden layer. From the were better like performance than those of a higher lag. We
result shown in the table below, it is observed that when the conclude that the rainfall series has a short term dependent
number of lag are three for GFFN model and two for FTLNN structure. Also it can be concluded that as the number of lags
the normalized mean square error is low as well the are increased the network become more complex and therefore
correlation between the actual rainfall and predicted rainfall learning becomes more difficult. Another Simulation was
series is high. done for prediction of rainfall number of days ahead from two
days to seven days.
TABLE I. NETWORK MODEL PERFORMANCE PARAMETER Focused TLFN model was considered for this
No. of
Mean square error Correlation simulation. Input data given to the network was for last seven
lag
coefficient days and the rainfall prediction was done for two days ahead
GFNN FTLNN GFNN FTLNN to seven days a head.
2 0,00131 0,00091 0,925 0,950
3 0,00111 0,00054 0,957 0,969
4 0,00112 0,00053 0,963 0.963
5 0,00438 0,00057 0,946 0.964
6 0,00281 0,00060 0,957 0.958
7 0,00260 0,00068 0,940 0.951

The result shows that the rainfall pattern learning is

better when taken for short duration. As the duration increases
accuracy of prediction decreases because the rainfall on a
given day is dependent on previous day or two. Also it is
found that FTLNN model is better than GFFN network for
learning temporal pattern. From the result mentioned in the
Table 1 another remarkable conclusion can be made that the
networks with a lesser lag have a better performance than
those of a higher lag. This means that the rainfall series has a
short term dependent structure. Figure 4. Actual and predicted rainfall using different ANN model
Also it can be concluded that as the number of lags
are increased the network become more complex and therefore
learning becomes more difficult. Another Simulation was From the result it is seen that if the prediction is done
done for prediction of rainfall number of days ahead from two more in advance the accuracy of prediction decreases or the
days to seven days. error of prediction increase.
 Therefore it is concluded that the model designed is
capable of short term rainfall prediction with better accuracy.
A plot of normalized actual and predicted rainfall using both
the model is shown in the Fig 4.
And scatter plot between actual and predicted rainfall
with residual for TFLN model is shown in the Fig 5.
Figure 5. Actual and predicted rainfall with residual for TFLN model
III. CONCLUSIONS
Rainfall prediction using ANN has been discussed in
the paper. Two different ANN models found suitable for this
task were designed and compared. Following points were
observed:
-For each type of network, there existed an optimal solution,
which was a function of the number of hidden nodes, number
of neurons in the hidden node and the lag of the network.
-The networks with lower lag have the tendency to outperform
the ones with higher lag.
This indicates that the rainfall time series have very
short term memory characteristics. In future ANN Models
designed can be used for rainfall prediction at different
location by properly training the network to estimate the
temporal pattern of rainfall series at given location. Also
different ANN models can be designed for different
geographical area to increase the accuracy of prediction.
REFERENCES
[1] C. L. Wu, K. W. Chau, and Y. S. Li, “Predicting monthly stream- flow
using data-driven model scoupled with data-preprocessing techniques,”
Water Resources Research, vol. 45, no. 8, Article ID W08432, 2009.
[2] L. L´opez, E. Garc´ıa-Ortega, and J. L. S´anchez, “A short-term forecast
model for hail,” Atmospheric Research, vol. 83, no. 2–4, pp. 176–184,
2007.
[3] Kin C. Luk, J. E. Ball AND A. Sharma, 2001, An Application of
Artificial Neural Networks for Rainfall Forecasting, Mathematical and
Computer Modelling 33 pp. 883-699.
[4] Poggio,T., and F. Girosi, 1990, “Networks for approximation and
learning,” Proc. IEEE, vol. 78, pp. 1481–1497.
[5] Sven F. Crone: A Business Forecasting Competition Approach to
Modeling Artificial Neural Networksfor Time Series Prediction. IC-AI
2004, pp. 207-213.
[6] K. Hornik, M. Stinchcombe and H. White, Multilayer feedforward
networks are universal approximators, Neural Networks 1989 2, pp.359-
366.
[7] M.French, W. Krajewski and R.R. Cuykendall, Rainfall forecasting in
Space and time using a neural network, Journal of Hydrology 1992 137,
1-31.
[8] K.L. Hsu, V. Gupta and S. Soroshian, Artificial neural network modeling
of the rainfall-runoff process, Water Resources Research l995 31 (10),
pp.2517-2530
[9] I.N.M.H Webpage information and data regarding Historical rainfall data
was collected from Brașov stations from 1961 to 1990
[10] Irwin W. Sandberg 2001, “Nonlinear Dynamical Systems: Feedforward
Neural Network Perspectives”, pp.11.
[11] Werbos, P. J., 1990, “Backpropagation through time:What it does and
how to do it,” Proc. IEEE, vol. 78, pp. 1550–1560.