Гібридна модель штучного інтелекту інтегрована в ГІС для прогнозування аварій на мережах водопостачання
The search for an effective and reliable model for predicting accidents on water supply networks by determining their exact locations has always been important for effectively managing water distribution systems. This study, based on the adaptive neuro-fuzzy logical inference system (ANFIS) model, w...
Saved in:
| Date: | 2024 |
|---|---|
| Main Authors: | , |
| Format: | Article |
| Language: | English |
| Published: |
The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute"
2024
|
| Subjects: | |
| Online Access: | https://journal.iasa.kpi.ua/article/view/280665 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Journal Title: | System research and information technologies |
| Download file: |  |
Institution
System research and information technologies| _version_ | 1866302909483843584 |
|---|---|
| author | Zaychenko, Yuriy Starovoit, Tetiana |
| author_facet | Zaychenko, Yuriy Starovoit, Tetiana |
| author_sort | Zaychenko, Yuriy |
| baseUrl_str | http://journal.iasa.kpi.ua/oai |
| collection | OJS |
| datestamp_date | 2024-08-11T01:12:49Z |
| description | The search for an effective and reliable model for predicting accidents on water supply networks by determining their exact locations has always been important for effectively managing water distribution systems. This study, based on the adaptive neuro-fuzzy logical inference system (ANFIS) model, was developed to predict accidents in the city of Kyiv (Ukraine) water supply network. The ANFIS model was combined with genetic algorithms and swarm optimization (ACO) methods and integrated into a GIS to visualize results and determine locations. Forecasts were evaluated according to the following criteria: mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R2). Depending on the amount and type of input data, ANFIS optimization with genetic algorithms and swarm optimization (ACO) can, on average, increase the accuracy of ANFIS predictions by 10.1% to 11%. The obtained results indicate that the developed hybrid model may be successfully applied to predict accidents on water supply networks. |
| doi_str_mv | 10.20535/SRIT.2308-8893.2024.2.04 |
| first_indexed | 2025-07-17T10:28:11Z |
| format | Article |
| fulltext |
Yu. Zaychenko, T. Starovoit, 2024
52 ISSN 1681–6048 System Research & Information Technologies, 2024, № 2
TIДC
ТЕОРЕТИЧНІ ТА ПРИКЛАДНІ ПРОБЛЕМИ
ІНТЕЛЕКТУАЛЬНИХ СИСТЕМ ПІДТРИМАННЯ
ПРИЙНЯТТЯ РІШЕНЬ
UDC 004.02, 004.67, 004.891.3
DOI: 10.20535/SRIT.2308-8893.2024.2.04
A HYBRID MODEL OF ARTIFICIAL INTELLIGENCE
INTEGRATED INTO GIS FOR PREDICTING ACCIDENTS
IN WATER SUPPLY NETWORKS
Yu. ZAYCHENKO, T. STAROVOIT
Abstract. The search for an effective and reliable model for predicting accidents on
water supply networks by determining their exact locations has always been impor-
tant for effectively managing water distribution systems. This study, based on the
adaptive neuro-fuzzy logical inference system (ANFIS) model, was developed to
predict accidents in the city of Kyiv (Ukraine) water supply network. The ANFIS
model was combined with genetic algorithms and swarm optimization (ACO) meth-
ods and integrated into a GIS to visualize results and determine locations. Forecasts
were evaluated according to the following criteria: mean absolute error (MAE), root
mean square error (RMSE), and coefficient of determination (R2). Depending on the
amount and type of input data, ANFIS optimization with genetic algorithms and
swarm optimization (ACO) can, on average, increase the accuracy of ANFIS predic-
tions by 10.1% to 11%. The obtained results indicate that the developed hybrid
model may be successfully applied to predict accidents on water supply networks.
Keywords: ANFIS, ACO, GA, spatial objects, geodatabase, metaheuristics, spatio-
temporal analysis, water loss.
INTRODUCTION
Forecasting accidents in water distribution systems is important in the manage-
ment of water resources, as it makes it possible to identify problem areas in the
network and eliminate them in advance. Intelligent predictive systems are models
and algorithms that provide valuable information about the future performance of
a system as a decision support system. With the development of supervisory con-
trol and data acquisition (SCADA) systems, real-time monitoring of pressure and
data flows is commonly used to detect pipe bursts. Machine learning [5] and clus-
ter analysis models were developed for optimal assessment. Failures in the net-
work can also be analyzed using hydraulic models [6].
The techniques mentioned above were successful in detecting accidents, but
not in pinpointing their exact locations [7]. The model-based approach relies
heavily on the accuracy of hydraulic models [8] and may not be suitable for larger
water supply systems. Other methods that utilize pressure/flow measurements and
GIS have also been proposed. For instance, [9] utilized triangle-based cubic inter-
polation to establish a pressure drop surface during network breaks to locate the
A hybrid model of artificial intelligence integrated into GIS for predicting accidents…
Системні дослідження та інформаційні технології, 2024, № 2 53
source of the problem. In [10] the measuring zone’s rupture location in the water
supply network was identified by assessing the sensitivity of various pres-
sure/flow measurements in relation to emergency leaks. [11] employed a multi-
variate graphical model that utilized data from multiple pressure gauges to iden-
tify potential accident locations, employing a combination of Gaussian and
geostatistical methods. Typically, fluctuations in demand can make it difficult to
detect hydraulic indicators resulting from accidents. Therefore, these methods can
only provide a general idea of where network breaches may occur, with an error
range of hundreds of meters and several pipes. Unfortunately, this is not precise
enough to quickly locate and fix network issues, resulting in delayed system res-
toration.
A more accurate method is needed to locate pipe bursts, which involves
gathering detailed information about the water system’s behaviour in potential
locations to detect anomalies. This can be achieved by placing accelerometer sen-
sors and analyzing acoustic signals, which can automatically determine the loca-
tion of the rupture or leak [12]. However, the reliability of this method depends
on the characteristics of the leakage conditions, such as pressure and flow rate,
and the detection range is limited by the clarity and correlation of the acoustic
signals. Another approach is based on transient processes [13], which analyzes
characteristic transient waves to determine the location of accidents. However,
background noise or other activities in the system can interfere with transient sig-
nals caused by discontinuity, especially if the number of channels to be analyzed
increases [7]. Hence, methods based on transient processes may not be suitable
for locating pipe breaks with exact precision.
Many researchers have explored the use of machine learning in water re-
sources research [14], but there is no consensus on the best model for predicting
water supply network emergencies. To address this, a forecasting model was de-
veloped that can pinpoint the exact location of potential emergencies. Artificial
neural networks are commonly used in water resource assessment due to their
computational efficiency [15–17], but they may produce errors in some cases due
to poor prediction or overtraining [15]. Therefore, it is necessary to optimize the
ANN and look for new approaches and new classes of neural networks.
Studies [18–20] have proposed a high-precision hybrid model called ANFIS,
which combines artificial neural networks (ANN) and fuzzy logic. The hybrid
ANFIS model has better performance than the two separate models, but it has cer-
tain limitations in finding the best weight parameters, which greatly affect the
prediction performance [15]. Furthermore, different optimization algorithms yield
varying results based on the geoenvironmental factors of the area being studied.
Therefore, developing new hybrid algorithms to determine the best weights and
produce reliable results is fundamental for flow modeling processes.
The purpose of this work is the development of a new model of artificial in-
telligence and the study of its effectiveness in the tasks of predicting accidents on
water supply networks with the determination of exact locations. This research is
conducted for the first time on the water distribution system of the city of Kyiv.
MODEL DEVELOPMENT AND TRAINING METODOLOGY
Data set collection for spatial modeling
The proposed modeling method is applied to the GIS water supply system of the
city of Kyiv (Ukraine). The length of the water supply networks in the city is in-
Yu. Zaychenko, T. Starovoit
ISSN 1681–6048 System Research & Information Technologies, 2024, № 2 54
creasing due to the inclusion of street and intra-quarter networks from enterprises.
As of 2019, the total length of the networks was 4,284.8 km.
In the structure of the city’s water supply networks, the main part is street
networks — 2614.8 km or 61% of the total length of pipes; intra-quarter
networks — 1275.1 km or 29.8%; water pipes — 394.9 km or 9.2%. The vast
majority of pipelines, namely 65.9%, are made of cast iron; 30.5% — from steel
and only 3.6% — from plastic materials.
21.4% of the pipes of the water supply network have been operated for more
than 50 years and another 33.2% — about 50 years; the service life of 27.1% is up
to 35 years, 12.3% — up to 25 years, 4.6% — up to 15 years, and only 1.4% —
up to 5 years. According to the degree of wear, 50.4% of the pipes are worn by
more than 90%; 24.3% of pipes — by 50–75%; 15.5% — by 75–90%; 6.3% —
by 25–50%; 3.5% — less than 25%.
Pipelines made of cast iron have the longest average age — 46.8 years,
pipelines made of steel — 45.4 years, the smallest — made of plastic — 15 years.
According to the pipe depreciation indicator, the water distribution system is
characterized as follows: the average degree of wear of steel pipes is 90%, cast
iron pipes are 75%, and plastic pipes are 23%.
The accident rate, which is determined by the number of accidents per unit
length of the network, has fluctuated in the range of 2.0–2.2 accidents/km in re-
cent years, and the tendency to increase the number of accidents was observed
specifically for water pipes.
The methodology of this study is shown in Fig. 1, and includes the following
stages:
1) preparation of input data;
2) separation of data into training (70%) and test (30%) sets;
3) training of ANFIS neuro-fuzzy network;
4) optimization of the ANFIS model by genetic algorithms and the swarm
optimization algorithm (ACO);
5) checking the accuracy of ANFIS, ANFIS-GA and ANFIS-ACO models.
ANFIS
Elevation
Pipe diameter
Pressure
Demand
Year
Type of
accident
Pipe length
Flow
Soil type
Soil
corrosion
index
Pipe material
Pipe stiffness
Distance to
railway
Calculate weight of each class of factors
Factors that cause emergencies
on the water supply network
Final Product
GA ACO
Model is OK Divide flood to
train
an test data
Calculate
Historical Floods
Train data
70 percent
Test data
30 percent
Optimize parametr of ANFIS by
modeling optimization problems
Fig. 1. Structural diagram of the development and optimization of the ANFIS model
A hybrid model of artificial intelligence integrated into GIS for predicting accidents…
Системні дослідження та інформаційні технології, 2024, № 2 55
It is important to consider how the problem occurs in relation to other factors
to make accurate spatial predictions. Table 1 shows the data used in our predictive
model, with some entered into the GIS and the rest determined through hydraulic
modeling based on the GIS model.
T a b l e 1 . Factors and conditions used in the model that impact the emergence
of issues in the water supply network
Factors/conditions Units Description
The degree of proximity
of the location to railway tracks/
m
When trains are in motion, the ground
vibrates, causing pipes to crack and gate
valves to be damaged.
Age year Year of laying the pipe
Length m Length of a pipe
Diameter mm Size of a pipe
Soil type index NA Soil type
Geoposition NA Geospatial location
Accident date year Accident date on network
Pressure bar Pressure from hydraulic calculation results
Volume of consumption m3/hour Volume of water consumption per hour
Volume of consumption m3/month Volume of water consumption per month
Demand NA Water demand
flow rate NA Flow rate according to hydraulic calculation
Pipe materials rigidity NA rigidity coefficient
Consumers NA Individuals and legal persons
It is probable that certain factors may affect the occurrence of pipe ruptures
or damages in specific parts of the network, while leaving other areas unaffected.
One such factor could be the presence of railway tracks. The vibrations caused by
freight trains passing by can lead to frequent failures in the water supply network,
resulting in pipe ruptures or damage to fittings. Additionally, the type of pipe ma-
terial used also plays a significant role in determining its lifespan. Steel pipes typ-
ically last for 25 years, while plastic or cast iron pipes can last up to 50 years.
PREPARATION OF DATA SET FOR TRAINING AND TESTING
In order to check if the model is practical, the data set for analysis should be split
into two groups: one for building the model (called the training data set) and the
other for testing it (called the test data set) [21]. To create the training data set,
70% of locations with and without previous accidents on the network (a total of
313 locations) were randomly chosen and combined.
The remaining 30% were then used to create the test dataset. Both data sets were
originally in vector format but were converted to csv format for further analysis.
For both data sets, the value 1 was assigned to indicate the presence of an accident
on the network, while 0 was assigned to indicate the absence of accidents.
We conducted a statistical analysis to thoroughly examine the data and im-
prove the intelligent model.
We performed a statistical analysis of spatial data in order to determine the
parameters of the membership function for training the ANFIS network and its
optimization (Fig. 2).
Yu. Zaychenko, T. Starovoit
ISSN 1681–6048 System Research & Information Technologies, 2024, № 2 56
1
2
1 — low
2 — med
3 — high
3
1
2
1 — Q1
2 — mean
3 — Q3
3
Fig. 2. Statistical analysis of the spatial data
1
2 1 — Q1
2 — mean
3 — Q33
1 2
1 — low
2 — high
1
2
1 — Q1
2 — mean
3 — Q3 3
1
21 — low
2 — med
3 — high
3
1 2
1 —Q1
2 — mean
3 — Q3
3
1
2
1 — low
2 — medium
3 — high
1
2
1 —Q1
2 — mean
3 — Q3
3
1
2
1 —low
2 — med
3 — high
3
1
2
1 — Q1
2 — mean
3 — Q3
3
1 2
1 — low
2 — high
A hybrid model of artificial intelligence integrated into GIS for predicting accidents…
Системні дослідження та інформаційні технології, 2024, № 2 57
DEVELOPMENT OF THE ANFIS MODEL AND ALGORITHMS FOR ITS
OPTIMIZATION
Adaptive neuro-fuzzy logical inference system
ANFIS (Adaptive Network Based Fuzzy Inference System) is an adaptive fuzzy
logical inference system proposed by Sugeno based on IF-THEN rules. It is a
method that combines artificial neural networks (ANNs) with fuzzy ones. This
neural network is used for membership function tuning and rule base tuning in a
fuzzy expert system. Below is the Sugeno model of fuzzy logic inference (Fig. 3).
The layers of this fuzzy neural network perform the following functions.
Layer 1. Membership Function Layer
In this layer, each neuron uses a membership function (fuzzifier) to trans-
form the input signal x or y. The most commonly used functions are the bell-
shaped function or the Gaussian function:
2
1
1
)(
i
i
A
a
cx
x
i
;
2
exp)(
i
i
A a
cx
x
i
.
Layer 2. Antecedent Layer
Each neuron is represented by the symbol . It performs an intersection be-
tween sets of input signals, which simulates a logical AND operation. The neuron
then sends an output:
niiyxw
ii BAi ,..,2,1),()( .
In fact, any T-norm operator that generalizes the AND operation can be used
in these neurons.
2 2 2 2f a x b y r
1 1 2 2
1 1 2 2
1 2
w f w f
f w f w f
w w
1111 rybxaf
Fig. 3. Sugeno’s fuzzy logic model
Yu. Zaychenko, T. Starovoit
ISSN 1681–6048 System Research & Information Technologies, 2024, № 2 58
Layer 3. Normalization Layer
Each neuron in this layer calculates the normalized strength of the rule:
ni
w
w
w
ii
i
i ,..,2,1,
.
Layer 4. Consequent Layer
The values of output variables are formed in neurons:
)(4
iiiiiii rybxawfwO .
Layer 5. Aggregation Layer
We receive the output signal of the neural network and perform defuzzifica-
tion of the results:
i i
i ii
ii w
fw
fwO outputoverall5 .
The neural network of the ANFIS architecture is trained using the gradient
descent method.
OPTIMIZATION OF ANFIS WEIGHTING COEFFICIENTS AND OFFSETS BY
THE ANT COLONY ALGORITHM
Ant Colony Optimization (ACO) is a probabilistic method for solving complex
computational problems that find optimal parameters in a search environment.
This algorithm, which was proposed by Marcus Dorigo in 1996, imitates the be-
haviour of ants in finding the optimal path from their nest to a food source. In
[22; 23], the author optimizes the weighting coefficients of an artificial neural
network using ACO and investigates the performance of the network. In the
search space, a population of weights is created which is considered as an objec-
tive function and is found according to the formula:
SEP = 100
ponn
OO minmax 2
11
) ( p
i
p
i
n
i
n
p
ot
op
,
where p
it and p
io are the expected and actual value of the output neuron and for
the template p.
The terms maxO , and minO represent the highest and lowest values of the
output signal from a specific neuron, while on and pn refer to the number of
output neurons.
The ACO algorithm is a tool for optimizing neural network parameters such
as synaptic weights, number of layers, and number of hidden neurons. It begins
by randomly selecting decisions from a predefined set of data, which are then
evaluated and assigned to the decision space based on their fitness values. New
solutions are created using information from previous iterations, with a higher
likelihood of selecting values with a greater concentration of pheromones [23].
This process generates a matrix of size M × N, where M represents the decision
population size and N represents the number of decision variables.
A hybrid model of artificial intelligence integrated into GIS for predicting accidents…
Системні дослідження та інформаційні технології, 2024, № 2 59
M
j
X
X
X
X
2
1
Population =
MNMiMM
jNjijj
Ni
Ni
xxxx
xxxx
xxxx
xxxx
21
21
222221
111211
,
where jX j -th solution, xji – i-th solution variable for the j-th solution, and M is
the size of the number of solutions. The value xji is chosen randomly from the set iV :
1 2{ , ,..., , }i i i id i iV v v v v D , 1,2,...,i N ,
where iV set of predefined values for the i-th decision variable, idv d -th pos-
sible value for the i-th decision variable, and iD total number of possible values
for the i-th decision variable [23].
GENETIC ALGORITHMS
Genetic algorithms develop optimal solutions by sampling from all possible solu-
tions. The best of these solutions are then combined using the genetic operators of
crossover and mutation to generate new solutions. This process continues until a
certain termination condition is met [4]. The diagram of the GA process is shown
in Fig. 4. The first step is the initial state in which we want to find the Hamilto-
nian cycle with the smallest sum of weights. In the second step, the fitness func-
tion estimates the Hamiltonian cycles, with lower cost functions indicating the
best individuals. Finally, in the third step, the most adapted individual is
identified.
GA can be used to optimize various parameters in water distribution sys-
tems. It uses the following mechanisms: crossover, mutation, selection. The goal
of training is to minimize the root mean square error:
M
k
kk wyd
M
WE
1
2))((
1
)( ;
],[ OI WWW ;
I
ijI wW ;
O
ijO wW .
Step 1
Step 2
Step 3
Fig. 4. Scheme of the process of genetic algorithms
Yu. Zaychenko, T. Starovoit
ISSN 1681–6048 System Research & Information Technologies, 2024, № 2 60
We set the initial population in which any individual is represented by the
corresponding weights of N individuals: )]0(),...,0(),...,0([ NiI WWW .
We calculate the fitness index (Fitness Index) and evaluate the quality of
forecasting:
max)()( ii WECWFI ,
where С — constant.
IMPLEMENTATION OF THE MODEL
The technique of forecasting with a combination of GIS and artificial intelligence
methods were applied to predict accidents on the water supply network of the city
of Kyiv. The Sugeno method was used, as it shows better accuracy. The optimal
membership function was chosen by trial and error. The ANFIS method was op-
timized by GA and ACO to improve accuracy.
The performance of the ANFIS, ANFIS-GA, ANFIS-ACO models was de-
termined from the resulting mean absolute error (MAE), which indicates a risk
metric corresponding to the expected value of the absolute error loss or the loss
rate:
)ˆ,( yyMAE
1
samplesn
1
ˆ
samplesn
i i
i
y y
.
The mean squared error indicates the risk indicator corresponding to the ex-
pected value of the squared error or loss:
)ˆ,( yyRMSE
samplesn
1 2
0
)ˆ( ii
n
i
yy
samples
.
The 2R function calculates the coefficient of determination, which repre-
sents the proportion of variance (y) that was explained by the independent vari-
ables in the model:
2
1
2
12
)(
)ˆ(
1)ˆ,(
yy
yy
yyR
i
n
i
ii
n
i .
The function explained variance calculates the estimate of the explained
variance:
explained_variance
}{var
} ˆ ,{var
1)ˆ,(
y
yy
yy
.
RESULTS
Spatial-temporal assessments and prediction of the occurrence of accidents
on the water supply network of the city of Kyiv
Spatiotemporal GIS analysis and modeling are essential for studying and predict-
ing future events. For modeling, we used the ESRI GIS package: ArcGIS Pro 2.7.
The first step was data acquisition and preparation. The obtained information was
A hybrid model of artificial intelligence integrated into GIS for predicting accidents…
Системні дослідження та інформаційні технології, 2024, № 2 61
summarized in the netCDF data structure, which was used for spatial statistical
analysis and creation of a space-time cube (Fig. 5) [24] .
A space-time cube is a well-known model in ArcGIS that combines spatial
data and time into a three-dimensional data structure of the netCDF (total network
shape) format, containing an array of bins with absolute location and absolute
time [24]. So, we aggregated incidents of accidents on the water supply network
within a grid size of 500 × 500 m2 (distance interval) with an absolute step inter-
val of 1 month. This approach made it possible to investigate cases of accidents
on the water supply network of the city of Kyiv (Ukraine).
We applied the space-time cube to a forest-based prediction model, which
generated a 2D object class indicating the predicted locations within the original
space-time cube. Each location is predicted individually (as shown in Fig. 5) and
has its own schedule (as seen in Fig. 6).
In Fig. 6, the graph displays the input, data gaps restored as a result of calcu-
lations, predicted values and confidence intervals. Confidence intervals are cre-
ated for each predicted time step, which are presented as fields of output objects.
The upper and lower bounds of the confidence intervals for the first pre-
dicted time step are calculated using quantile random forest regression. To predict
values for a future time, observations from each leaf of the tree are averaged to-
gether. The confidence interval of the second forecast is calculated in a similar
way, but is adjusted taking into account the confidence interval of the first fore-
Fig. 5. The result of spatial forecasting
Fig. 6. Graph of the values of the locations of the space-time cube
Yu. Zaychenko, T. Starovoit
ISSN 1681–6048 System Research & Information Technologies, 2024, № 2 62
cast [25]. The real confidence interval of the second forecast is calculated by add-
ing the lengths of the limits of the confidence interval of the two forecasts. Subse-
quent time steps are calculated by adding previous predictions. The real confi-
dence level of these intervals is at least 90%, but in reality the accuracy may be
higher [25].
The result of the assessment of the total accuracy of the forecast in different
locations, using the forest-based method, is shown in Table 2.
T a b l e 2 . The result of the overall assessment of forecast accuracy in different
locations
Category Min Max Mean Median Mean sq. dev.
RMSE of the prediction 0.00 1.25 0.26 0.24 0.15
RMSE of validation 0.00 2.89 0.56 0.48 0.45
This forecasting method is best used for time series with a complex shape
and trends that are difficult to model using simple mathematical functions. The
correct selection of time steps during model validation is important. The more
time steps that are excluded, the less time it takes to evaluate the validation mod-
el. However, if too few time steps are included, the RMSE value will be estimated
using less data and may be misleading. Also, this tool can produce unstable and
unreliable forecast results if the same value is repeated too often in time series
[25]. To optimize and improve the accuracy of the predictive model, we com-
bined GIS methods with hybrid artificial intelligence methods.
Configuration of hybrid models
In this study, we integrated the ANFIS model with GA and ACO algorithms, and
compared the performance of the models. The algorithms are implemented in the
Spyder environment (Anaconda 3). In order to test the model with different opti-
mization algorithms, the data were organized into separate training and test data-
sets, which were divided into 70% and 30% (Fig. 7).
Fig. 7. Results of model training
ANFIS ANFIS-GA ANFIS-ACO
A hybrid model of artificial intelligence integrated into GIS for predicting accidents…
Системні дослідження та інформаційні технології, 2024, № 2 63
Fig. 7 shows the result of model training: change in error frequency during
training; comparing predicted values with actual data on the training set and com-
paring predicted values with actual data on the test set.
The first step was to import the training data into the ANFIS, ANFIS-GA,
ANFIS-ACO models to reveal the hidden relationships between the factors affect-
ing the emergency of the water supply network. As a next step, the validation data
were used to test the performance and predictive capabilities of the models. MAE,
RMSE, R2, and explained_variance were used to measure accuracy. Table 3
shows the result of learning hybrid models (the first 5 iterations in GA and ACO).
T a b l e 3 . Comparison and performance testing of models
Test data Train data
Model
MAE RMSE R2 Cov MAE RMSE R2 Cov
0.043 0.094 0.613 0.613 0.062 0.125 0.576 0.578
ANFIS
Study time: 0:00:06.19
0.041 0.097 0.599 0.600 0.061 0.124 0.575 0.574
Study time: 0:00:08.31
0.042 0.098 0.583 0.585 0.061 0.124 0.575 0.577
Study time: 0:00:08.98
0.041 0.098 0.587 0.589 0.061 0.125 0.575 0.577
Study time: 0:00:09.09
0.044 0.098 0.585 0.587 0.062 0.124 0.573 0.575
Study time: 0:00:08.41
0.042 0.098 0.584 0.586 0.061 0.125 0.575 0.576
ANFIS-GA
Study time: 0:00:08.24
0.041 0.096 0.593 0.595 0.061 0.124 0.573 0.575
Study time: 0:00:11.96
0.042 0.097 0.585 0.587 0.061 0.124 0.574 0.576
Study time: 0:00:11.86
0.043 0.098 0.585 0.586 0.062 0.125 0.572 0.575
Study time: 0:00:12.22
0.041 0.098 0.586 0.588 0.061 0.125 0.576 0.577
Study time: 0:00:11.92
0.042 0.097 0.585 0.587 0.062 0.125 0.573 0.575
ANFIS-ACO
Study time: 0:00:12.21
The MAE values for the ANFIS, ANFIS-GA, and ANFIS-ACO models were
calculated for both the test and training data. The results show that the ANFIS-
GA model had the best performance with a MAE value of 0.042 for the test data
and 0.061 for the training data. The GA algorithm was found to be more efficient
than the ACO algorithm, which had a similar performance but required twice as
much training time. It’s important to note that these results may vary based on the
input data. Overall, the ANFIS-GA model is stable, efficient, and has a fast con-
vergence rate.
CHECKING AND COMPARING MODELS
We used three different optimization models, namely ANFIS, ANFIS-GA, and
ANFIS-ACO, which were developed and implemented in Spyder (Anaconda3).
Yu. Zaychenko, T. Starovoit
ISSN 1681–6048 System Research & Information Technologies, 2024, № 2 64
The results obtained from these models were then visualized in ArcGIS Pro 2.7.
To train these models, we divided the pointed objects of accidents into two cate-
gories: 30% for training and 70% for testing. We used the training data set to es-
tablish relationships between the occurrence of accidents (1) and the absence of
accidents (0).
We checked the accuracy and performance of hybrid intelligent models by
calculating the mean absolute error of MAE. Fig. 8 shows the membership func-
tions of the input variables of the ANFIS model. Fig. 9 illustrates the graph of the
change in the loss function depending on the number of iterations. Membership
functions indicate the fuzziness of the inputs. A comparison of the accuracy
scores in Fig. 8 shows that the ANFIS network performs well.
As a result, the accuracy of the ANFIS model was 95.49%. The accuracy de-
creases when the number of inputs increases, so to increase the accuracy, it is
necessary to improve the network with optimization algorithms.
The result of training the ANFIS-GA and ANFIS-ACO models was not
much better than the classic ANFIS, moreover, the ANFIS-ACO model required
much more time. In ANFIS-GA, the training time was the same as in ANFIS (one
iteration on average 0:00:06.24), while in ANFIS-ACO the total training time
took 0:58:19.69 (0:00:12.38 one iteration). Overall, the predictions aligned well
Fig. 8. Membership functions of the used input variables
Fig. 9. The graph of the change of the loss function depending on the number of
iterations
3
1
4
2
1 — mae
2 — mse
3 — val mae
4 — val mse
1 — Training loss
2 — Validation loss
1
2
A hybrid model of artificial intelligence integrated into GIS for predicting accidents…
Системні дослідження та інформаційні технології, 2024, № 2 65
and matched the experimental data accurately. It’s worth mentioning that the test
results demonstrate the developed models’ proficiency in forecasting data beyond
the training range.
Compared to GIS forecasting methods, developed artificial intelligence
models provide an opportunity to expand and increase forecast accuracy, and in-
dicate specific problematic pipes. Also, the developed models can be easily inte-
grated into ArcGIS Pro in the form of geoprocessing tools, and published on cor-
porate geoportals.
CONCLUSIONS
Adaptive neural fuzzy logic inference system (ANFIS) and its hybrid learning
methods: ANFIS-GA, ANFIS-ACO were used to predict water supply network
accidents. This model was integrated into GIS (ArcGIS Pro) to visualize and de-
termine the exact locations of possible accidents and was verified in practice (all
predicted accident locations for the next three days coincided with accidents that
occurred on the Kyiv water supply network). The following conclusions can be
drawn from the forecasting model described above:
Performance evaluation and model validation results of selected metrics:
R2, RMSE, and MAE for both training and testing on a small amount of data
showed that the hybrid models did not outperform ANFIS model.
When the amount of input data increased, the accuracy of the ANFIS
model decreased and it became necessary to optimize the ANFIS with genetic
algorithms and the swarm optimization algorithm (ACO). This optimization
increased the accuracy of ANFIS prediction by 10.1%, 11%.
The results of ANFIS, ANFIS-GA, and ANFIS-ACO intelligent models
combined with GIS indicate a large information potential that can support real-
time operational control of water supply systems. Fuzzy models of emergency
forecasts have a significant advantage as they require less information about water
supply systems than conventional probabilistic models. In addition, this informa-
tion may be vague and inaccurate. The ANFIS model is suitable for modeling
complex problems, especially when the relationship between factors is unknown.
It is especially useful for identifying threats and providing advance warnings
about the likelihood of an accident.
REFERENCES
1. C.M. Fontanazza, V. Notaro, V. Puleo, P. Nicolosi, and G. Freni, “Contaminant in-
trusion through leaks in water distribution system: experimental Analysis,” Procedia
Eng., 119, pp. 426–433, 2015. doi: 10.1016/j.proeng.2015.08.904.
2. “Urban Water Statistics Yearbook 2017,” (in Chinese), CUWA.
3. Dung H. Tran, A.W.M. Ng, and B.J.C. Perera, “Neural networks deterioration
models for serviceability condition of buried stormwater pipes,” Engineering
Applications of Artificial Intelligence, 20(8), pp. 1144–1151, 2007. doi:
10.1016/j.engappai.2007.02.005.
4. M. van Dijk, S.J. van Vuuren, J.E. van Zyl, “Optimising water distribution systems
using a weighted penalty in a genetic algorithm”, Water SA, 35(5), pp. 537–548,
2008. doi: 10.4314/wsa.v34i5.180651.
5. M. Romano, Z. Kapelan, and D.A. Savic, “Evolutionary algorithm and expectation
maximization strategies for improved detection of pipe bursts and other events in
Yu. Zaychenko, T. Starovoit
ISSN 1681–6048 System Research & Information Technologies, 2024, № 2 66
water distribution systems,” J. Water Resour. Plan. Manag., 140 (5), pp. 572–584,
2014. doi: 10.1061/(ASCE)WR.1943-5452.0000347.
6. J. Meseguer et al., “A decision support system for on-line leakage localization Envi-
ron,” Environmental Modelling & Software, vol. 60, pp. 331–345, 2014. doi:
10.1016/j.envsoft.2014.06.025.
7. Y. Wu and S. Liu “A review of data-driven approaches for burst detection in wa-
ter distribution systems,” Urban Water J., 14 (9), pp. 972–983, 2017. doi:
10.1080/1573062X.2017.1279191.
8. C.J. Hutton, Z. Kapelan, L. Vamvakeridou-Lyroudia, and D.A. Savic, “Dealing with
uncertainty in water distribution system models: a framework for real-time modeling
and data assimilation,” J. Water Resour. Plan. Manag., 140 (2), pp. 169–183, 2014.
doi: 10.1061/(ASCE)WR.1943-5452.0000325.
9. S.R. Mounce, A. Khan, A.S. Wood, A.J. Day, P.D. Widdop, and J. Machell, “Sen-
sor-fusion of hydraulic data for burst detection and location in a treated water distri-
bution system,” Inf. Fusion, 4 (3), pp. 217–229, 2003. doi: 10.1016/S1566-
2535(03)00034-4.
10. B. Farley, S.R. Mounce, and J.B. Boxall, “Development and field validation of a
burst localization methodology,” J. Water Resour. Plan. Manag., 139 (6), pp. 604–
613, 2013. doi: 10.1061/(ASCE)WR.1943-5452.0000290.
11. M. Romano, Z. Kapelan, and D.A. Savic “Geostatistical techniques for approximate
location of pipe burst events in water distribution systems” J. Hydroinf., 15 (3),
pp. 634–651, 2013. doi: 10.2166/hydro.2013.094.
12. J. Kang, Y.J. Park, J. Lee, S.H. Wang, and D.S. Eom, “Novel leakage detection by
ensemble cnn-svm and graph-based localization in water distribution systems,” IEEE
Trans. Ind. Electron., 65 (5), pp. 4279–4289, 2018. doi: 10.1109/TIE.2017.2764861.
13. A.F. Colombo, P. Lee, and B.W. Karney, “A selective literature review of transient-
based leak detection methods,” J. Hydro Environ. Res., 2 (4), pp. 212–227, 2009.
doi: 10.1016/j.jher.2009.02.003.
14. H.R. Maier, A. Jain, G.C. Dandy, and K.P. Sudheer, “Methods used for the devel-
opment of neural networks for the prediction of water resource variables in river sys-
tems: Current status and future directions. Environ,” Model. Softw., 25, pp. 891–909,
2010. doi: 10.1016/j.envsoft.2010.02.003.
15. D.T. Bui, B. Pradhan, H. Nampak, Q.-T. Bui, Q.-A. Tran, Q.-P. Nguyen, “Hybrid
artificial intelligence approach based on neural fuzzy inference model and metaheu-
ristic optimization for flood susceptibilitgy modeling in a high-frequency tropical
cyclone area using GIS,” J. Hydrol., 540, pp. 317–330, 2016. doi:
10.1016/j.jhydrol.2016.06.027.
16. H. Ghalkhani, S. Golian, B. Saghafian, A. Farokhnia, and A. Shamseldin, “Applica-
tion of surrogate artificial intelligent models for real-time flood routing,” Water En-
viron. J., 27(4), pp. 535–548, 2013. doi: 10.1111/j.1747-6593.2012.00344.x
17. M. Rezaeianzadeh, H. Tabari, A.A. Yazdi, S. Isik, and L. Kalin, “Flood flow fore-
casting using ANN, ANFIS and regression models,” Neural Comput. Appl., 25(1),
pp. 25–37, 2013. doi: 10.1007/s00521-013-1443-6.
18. F.-J. Chang and M.-J. Tsai, “A nonlinear spatio-temporal lumping of radar rainfall
for modeling multi-step-ahead inflow forecasts by data-driven techniques,” J. Hy-
drol., vol. 535, pp. 256–269, 2016. doi: 10.1016/j.jhydrol.2016.01.056.
19. Y.S. Güçlü and Z. Şen, “Hydrograph estimation with fuzzy chain model,” J. Hydrol.,
vol. 538, pp. 587–597, 2016. doi: 10.1016/j.jhydrol.2016.04.057.
20. A. Lohani, R. Kumar, and R. Singh, “Hydrological time series modeling: A compari-
son between adaptive neuro-fuzzy, neural network and autoregressive techniques,”
J. Hydrol., vol. 442–443, pp. 23–35, 2012. doi: 10.1016/j.jhydrol.2012.03.031.
21. C.-J.F. Chung and A.G. Fabbri, “Validation of spatial prediction models for land-
slide hazard mapping,” Nat. Hazards, vol. 30, pp. 451–472, 2003. doi:
10.1023/B:NHAZ.0000007172.62651.2b.
A hybrid model of artificial intelligence integrated into GIS for predicting accidents…
Системні дослідження та інформаційні технології, 2024, № 2 67
22. K. Socha and C. Blum, “An ant colony optimization algorithm for continuous opti-
mization: application to feed-forward neural network training,” Neural Computing
and Applications, vol. 16(3), pp. 235–247, 2007. doi: 10.1007/s00521-007-0084-z.
23. C. Blum and K. Socha, “Training feed-forward neural networks with ant colony op-
timization: An application to pattern classification,” in Fifth International Confer-
ence on Hybrid Intelligent Systems (HIS’05), IEEE, 2005, November. doi:
10.1109/ICHIS.2005.104.
24. Omid Bozorg-Haddad, Mohammad Solgi, and Hugo A. Loáiciga, Meta-heuristic
and Evolutionary Algorithms for Engineering Optimization. 2017. doi:
DOI:10.1002/9781119387053.
25. “ArcGIS analysis workflows for public safety,” ESRI.
Received 28.05.2023
INFORMATION ON THE ARTICLE
Yuriy P. Zaychenko, ORCID: 0000-0001-9662-3269, Educational and Research Institute
for Applied System Analysis of the National Technical University of Ukraine “Igor Sikor-
sky Kyiv Polytechnic Institute”, Ukraine, e-mail: zaychenkoyuri@ukr.net
Tetiana V. Starovoit, ORCID: 0009-0008-6335-7679, Educational and Research In-
stitute for Applied System Analysis of the National Technical University of Ukraine “Igor
Sikorsky Kyiv Polytechnic Institute”, Ukraine, e-mail: starovoyt.tania@lll.kpi.ua
ГІБРИДНА МОДЕЛЬ ШТУЧНОГО ІНТЕЛЕКТУ ІНТЕГРОВАНА В ГІС
ДЛЯ ПРОГНОЗУВАННЯ АВАРІЙ НА МЕРЕЖАХ ВОДОПОСТАЧАННЯ /
Ю.П. Зайченко, Т.В. Старовойт
Анотація. Пошук ефективної та надійної моделі прогнозування аварій на ме-
режах водопостачання з визначенням їх точних розташувань завжди був важ-
ливим для ефективного керування системами розподілу води. Дослідження,
засноване на моделі адаптивної нейронечіткої системи логічного висновку
(ANFIS), розроблено для прогнозування аварій на мережі водопостачання міс-
та Києва (Україна). Модель ANFIS поєднано з генетичними алгоритмами та
методами ройової оптимізації (ACO) та інтегрували в ГІС для візуалізації ре-
зультатів і визначення їх розташування. Прогнози оцінювали за такими крите-
ріями: середньої абсолютної похибки (MAE), середньої квадратичної похибки
(RMSE) та коефіцієнтом детермінації (R2). Залежно від кількості та вигляду
вхідних даних оптимізація ANFIS генетичними алгоритмами та алгоритмом
ройової оптимізації (ACO) може в середньому збільшувати точність передба-
чення ANFIS на 10,1%, 11%. Отримані результати свідчать про те, що розроб-
лена гібридна модель може бути успішно застосована для прогнозування ава-
рій на мережах водопостачання.
Ключові слова: геоінформаційні системи, ANFIS, ACO, GA, просторові
об’єкти, просторово-часовий аналіз, втрати води.
|
| id | journaliasakpiua-article-280665 |
| institution | System research and information technologies |
| keywords_txt_mv | keywords |
| language | English |
| last_indexed | 2025-07-17T10:28:11Z |
| publishDate | 2024 |
| publisher | The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" |
| record_format | ojs |
| resource_txt_mv | journaliasakpiua/69/1a343fb08ce5ab8718695a463cd62569.pdf |
| spelling | journaliasakpiua-article-2806652024-08-11T01:12:49Z A hybrid model of artificial intelligence integrated into GIS for predicting accidents in water supply networks Гібридна модель штучного інтелекту інтегрована в ГІС для прогнозування аварій на мережах водопостачання Zaychenko, Yuriy Starovoit, Tetiana геоінформаційні системи ANFIS ACO GA просторові об’єкти просторово-часовий аналіз втрати води ANFIS ACO GA spatial objects geodatabase metaheuristics spatiotemporal analysis water loss The search for an effective and reliable model for predicting accidents on water supply networks by determining their exact locations has always been important for effectively managing water distribution systems. This study, based on the adaptive neuro-fuzzy logical inference system (ANFIS) model, was developed to predict accidents in the city of Kyiv (Ukraine) water supply network. The ANFIS model was combined with genetic algorithms and swarm optimization (ACO) methods and integrated into a GIS to visualize results and determine locations. Forecasts were evaluated according to the following criteria: mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R2). Depending on the amount and type of input data, ANFIS optimization with genetic algorithms and swarm optimization (ACO) can, on average, increase the accuracy of ANFIS predictions by 10.1% to 11%. The obtained results indicate that the developed hybrid model may be successfully applied to predict accidents on water supply networks. Пошук ефективної та надійної моделі прогнозування аварій на мережах водопостачання з визначенням їх точних розташувань завжди був важливим для ефективного керування системами розподілу води. Дослідження, засноване на моделі адаптивної нейронечіткої системи логічного висновку (ANFIS), розроблено для прогнозування аварій на мережі водопостачання міста Києва (Україна). Модель ANFIS поєднано з генетичними алгоритмами та методами ройової оптимізації (ACO) та інтегрували в ГІС для візуалізації результатів і визначення їх розташування. Прогнози оцінювали за такими критеріями: середньої абсолютної похибки (MAE), середньої квадратичної похибки (RMSE) та коефіцієнтом детермінації (R2). Залежно від кількості та вигляду вхідних даних оптимізація ANFIS генетичними алгоритмами та алгоритмом ройової оптимізації (ACO) може в середньому збільшувати точність передбачення ANFIS на 10,1%, 11%. Отримані результати свідчать про те, що розроблена гібридна модель може бути успішно застосована для прогнозування аварій на мережах водопостачання. The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" 2024-06-28 Article Article application/pdf https://journal.iasa.kpi.ua/article/view/280665 10.20535/SRIT.2308-8893.2024.2.04 System research and information technologies; No. 2 (2024); 52-67 Системные исследования и информационные технологии; № 2 (2024); 52-67 Системні дослідження та інформаційні технології; № 2 (2024); 52-67 2308-8893 1681-6048 en https://journal.iasa.kpi.ua/article/view/280665/301101 |
| spellingShingle | геоінформаційні системи ANFIS ACO GA просторові об’єкти просторово-часовий аналіз втрати води Zaychenko, Yuriy Starovoit, Tetiana Гібридна модель штучного інтелекту інтегрована в ГІС для прогнозування аварій на мережах водопостачання |
| title | Гібридна модель штучного інтелекту інтегрована в ГІС для прогнозування аварій на мережах водопостачання |
| title_alt | A hybrid model of artificial intelligence integrated into GIS for predicting accidents in water supply networks |
| title_full | Гібридна модель штучного інтелекту інтегрована в ГІС для прогнозування аварій на мережах водопостачання |
| title_fullStr | Гібридна модель штучного інтелекту інтегрована в ГІС для прогнозування аварій на мережах водопостачання |
| title_full_unstemmed | Гібридна модель штучного інтелекту інтегрована в ГІС для прогнозування аварій на мережах водопостачання |
| title_short | Гібридна модель штучного інтелекту інтегрована в ГІС для прогнозування аварій на мережах водопостачання |
| title_sort | гібридна модель штучного інтелекту інтегрована в гіс для прогнозування аварій на мережах водопостачання |
| topic | геоінформаційні системи ANFIS ACO GA просторові об’єкти просторово-часовий аналіз втрати води |
| topic_facet | геоінформаційні системи ANFIS ACO GA просторові об’єкти просторово-часовий аналіз втрати води ANFIS ACO GA spatial objects geodatabase metaheuristics spatiotemporal analysis water loss |
| url | https://journal.iasa.kpi.ua/article/view/280665 |
| work_keys_str_mv | AT zaychenkoyuriy ahybridmodelofartificialintelligenceintegratedintogisforpredictingaccidentsinwatersupplynetworks AT starovoittetiana ahybridmodelofartificialintelligenceintegratedintogisforpredictingaccidentsinwatersupplynetworks AT zaychenkoyuriy gíbridnamodelʹštučnogoíntelektuíntegrovanavgísdlâprognozuvannâavaríjnamerežahvodopostačannâ AT starovoittetiana gíbridnamodelʹštučnogoíntelektuíntegrovanavgísdlâprognozuvannâavaríjnamerežahvodopostačannâ AT zaychenkoyuriy hybridmodelofartificialintelligenceintegratedintogisforpredictingaccidentsinwatersupplynetworks AT starovoittetiana hybridmodelofartificialintelligenceintegratedintogisforpredictingaccidentsinwatersupplynetworks |