Applicability of radial-based artificial neural networks RBNN on coliform calculation A case of stud

Abstract

Introduction

Deep-sea discharge (DSD) is a method of disposal to take advantage of the sea’s dilution capacity. The main purpose of DSD systems is to make the wastewater collected with the city wastewater network harmless with very high dilution rates by giving them to the marine environment after being treated at a level determined according to the need. In our country and the world, the discharge of

domestic and industrial wastewaters to coastal waters constitute the main causes of pollution in seas and rivers. It is widely used because wastewater is a reliable and relatively inexpensive waste removal technology with DSD. Today, DSD systems will continue to be used until a better alternative is available [1]. Especially without primary treatment, widespread discharge of sewage is of great importance,

*Corresponding author. *E-mail address: yuksel.ardali@omu.edu.tr This paper was recommended for publication in revised form by Regional Editor Osman Nuri UÇAN Published by Yıldız Technical University Press, İstanbul, Turkey Copyright 2021, Yıldız Technical University. This is an open access article under the CC BY-NC license (http://creativecommons.org/licenses/by-nc/4.0/).

Sigma J Eng Nat Sci, Vol. 40, No. 4, pp. 724–731, December, 2022

Figure 1. DSD Lines in the Black Sea Region (Google Earth) (Ardali, 2015).

because these wastes contain only high concentrations of suspended solids and nutrients, they also contain a significant amount of organic matter and coliform [2]. The Black Sea is located between 41.0° and 46.5° northern latitudes, 27.5° and 41.5° eastern longitudes in an area where the European and Asian continents converge. When analyzed based on hydrographic data, the Black Sea, which has a total area of 413490 km² and a water volume of 537000 km3, can be defined as the largest anoxic basin in the world [1]. In Turkey, the Black Sea region should have greater geographical factors of population density in coastal areas due to both land distribution. There are a total of 36 deep-sea discharge points, 33 domestic and 3 industrial, which are used as wastewater discharge by the municipalities on the Black Sea coast in Figure 1. Sea discharge studies in Turkey, Water Pollution Control Regulations and Regulation of Urban Wastewater Treatment are carried out within the framework of this regulation [1]. Quality standards are defined to protect the beneficial use of the sea and its product. Setting water standards is very complicated and following these standards is strongly linked to water use [3]. Therefore, pathogens are a serious concern for water resource managers. Because it is known that an excessive amount of fecal bacteria in sewage and urban flow indicates an increased risk of pathogenic illness in humans [4]. A small amount of coliform is known to indicate the presence of other harmful bacteria or viruses in the stomachs. Coliforms are among the bacterial indicators to be monitored [3]. Total coliform (TC) is chosen as target marker organisms as it is present in the feces of human / warm-blooded animals and high concentrations in wastewater [1]. Tay and Zhang (1999) modeled the complex process of anaerobic biological treatment of wastewater using neuralfuzzy techniques. Scarlatos made coliform calculations using an artificial neural network (ANN) model using samples of river mouth systems. He stated that the graphical results of

coliform prediction make it difficult to express model performances [5]. In a study, the peak coliform values of the Delaware River were estimated using feed forward artificial neural network (FFNN) models. It is stated in the study that artificial neural network models can be successfully applied in coliform prediction [6,7,8]. In another study, using the FFNN model, a coliform estimation was made using 7 point values from the Southwest Scottish coast. At the end of the study, they stated that the correlation values of the models they prepared approached to 0.50 [9]. In another study, FFNN and regression models were used in coliform prediction. They made the performance evaluation of using the models and as a result of the evaluation, it was stated that although the artificial neural models were more successful, they performed poorly in predicting the peak values [10,11]. Another study investigated the use of regression models to estimate fecal coliform levels in the Charles River basin in Massachusetts [12]. In another study, they tried to estimate the amount of free chlorine using FFNN based on a statistical model such as flow, pH and temperature of a sample drinking water network. They stated that flow and temperature variables are effective in the amount of chlorine [13]. In another study, they estimated the daily coliform amount using 6 different models in which neural network-based sedimentation and sedimentation-based artificial variables were preferred, and they stated that the models containing precipitation parameters gave successful results in coliform estimation [14]. Another work from the Iznik lake basin from turkey has been to develop fecal pollution model structures with FFNN for cost-effective lake water quality management studies. The study was indicated that multilayer FFNN models could be used to predict microbial pollution in deep lakes [15]. In this study, the parameters affecting the total coliform were evaluated using Radial Based Artificial Neural Networks, which is a different artificial neural network model than those used in studies in the literature. Samsun

Sigma J Eng Nat Sci, Vol. 40, No. 4, pp. 724–731, December, 2022

Tekkekoy DSD system was chosen as the study area, taking into account the high population density in the Black Sea region.

Material And Methods

Study Area This study was carried out in Tekkekoy DSD of Samsun in Turkey’s northern coast. The location of tekkeköy DSD satellite image, which is a domestic discharge system, is given in Figure 2. Field studies were carried out between July 2015 and 2016 under the contractor of the Ministry of Environment and Urbanization and the direction of Ondokuz Mayıs University Environmental Engineering Department. Sampling from various points, dissolved oxygen, pH values were determined at the time of sampling by

5. L Niskin water sampler, seawater samples were collected

at three different water depths (surface, middle, and bottom) at each point. Laboratory studies were performed according to standard methods and then tested for three parameters (total suspended solids (TSS, APHA 2540D), BOD5 (APHA, 5210B), total coliform (TS EN ISO 9308-1). Detailed information about the sampling points and their geographical locations are presented in Table 1. Satellite image of the deep sea discharge sampling points is given in Figure 3.

Modeling Study

Radial based artificial neural networks Radial Based Artificial Neural Networks (RBNN) was developed in 1988 inspired by the effect response behaviors

Sigma J Eng Nat Sci, Vol. 40, No. 4, pp. 724–731, December, 2022

seen in biological nerve cells and entered the history of ANN by applying it to the filtering problem [16]. It is possible to view the training of RBNN models as a curve-fitting approach in multidimensional space [17]. For this reason, the training performance of the RBNN model turns into an interpolation problem, finding the most suitable surface for the data in the output vector space. RBNN models are defined in three layers as the input layer, hidden layer and output layer, similar to general ANN architecture

(Figure 4). However, unlike conventional ANN structures, RBNN s use radial-based activation functions in the transition from the input layer to the hidden layer. The structure between the hidden layer and the output layer continues to function as in other ANN types, and the actual training is carried out here. Radial-based artificial neural networks (RBNN) are networks that have radial-based activation functions in the transition to the hidden layer, unlike other networks [19]. There are three components for the radially symmetrical middle layer processor element. The first is a center vector in the input space. This vector is stored as the weight vector between the input and hidden layers. The second is the distance measure to determine how far an input vector is from the center. Typically this criterion is taken as the standard Euclidean distance. The last one is an activation function structure that determines the output value of the processor element, which is one variable and takes the distance function output as input. The processor elements in the first layer do not use the weighted shape of the inputs. The outputs of the processor elements in the first layer are determined according to the distance between the ANN inputs and the center of the basic function. The last layer of the RBNN structures is linear and the total output weighted from the outputs of the first layer is produced [20]. The output (y) produced by the network in RBNN models can be calculated with the help of equation 1. N

yi = ∑ wikϕk(x,ck) = ∑ wikϕk||x – ck||2, i = 1,2,… ,m Figure 3. Sampling points of Tekkekoy DSD (Google Earth).

In this equation, x∈Rn×1 is the input vector of the network; ϕk∈R+ radial based activation function; ck∈Rn×1

Table 1. Geographical locations of sampling points (Ardali, 2015). Station

Sigma J Eng Nat Sci, Vol. 40, No. 4, pp. 724–731, December, 2022

radial-based centers selected from a subset of the input vector space; ||.||2 is the Euclidean norm which is a measure of how far the input vector is from the center; wik weights in the output layer; N indicates the number of cells in the hidden layer. Many types of functions can be used as activation functions in RBNN models. Linear, Cubic, Gauss, MultiQuadratic, Inverse Multi-Quadratic functions are some of them and the Gauss function was preferred in this study. The mathematical structure of the Gauss function is shown in equation 2. ϕk(x) = exp

In this equation, x represents the input vector, CK centers. σ symbolizes the standard deviation value. In ANN terminology, it is also referred to as the scatter parameter that significantly affects the performance of the RBNN model [17]. The scattering parameter is usually taken as a constant for all cells. Although there are approximate equivalents for the dispersion parameter in RBNN models, this parameter can also be determined by the trial-and-error method (Ham & Kostanic 2001). In this study, 10 different parameters have been tried with the values of the dispersion parameter between 0.1-1 and step size of 0.1.

Results And Discussion

Various combinations were used as input to make coliform calculations using a radial-based artificial neural network. The 17 elements that make up these combinations are, pH, suspended solids, dissolved oxygen, crude oil, and its derivatives, organic pollutants, chlorophyll, phenols, ammonia, Cu, Cd, Cr, Pb, Ni, Zn, Hg, As and light transmittance values. The Models which have five to ten neurons were prepared to apply with the RBNN model were given in Table 2.

In the prepared models, all parameters are divided into their maximums and normalized to values between 0-1 and size homogeneity is provided. The normalized data were inserted into the RBNN structure with dispersion parameters ranging from 0.1-1, and the results were denormalized and evaluated according to the selected error evaluation criteria. In the prepared models randomly selected 70% of

Sigma J Eng Nat Sci, Vol. 40, No. 4, pp. 724–731, December, 2022

data are used for training, 15% for validating and 15% for testing [21]. Root mean square error (RMSE), correlation (R) and determination (R2) coefficients were used as error evaluation criteria. All data sets by using RBNN models have been also estimated.

Coliform estimation was made for all 30 samples via prepared models and statistical results of the models are given in Table 3a. and Table 3b, the scatter diagrams of the best results of the best models of two secenarios are presented in Figure 5 and Figure 6. In the scatter diagrams, the

Sigma J Eng Nat Sci, Vol. 40, No. 4, pp. 724–731, December, 2022

Table 4. Comparison of RBNN results and FFNN (Ayeri et al., 2018) results Model

x-axis formed measured coliform values determined as a result of laboratory studies, and the y-axis formed the predicted coliform values as a result of RBNN. The best results determined in the prepared models were obtained when the dispersion parameter was chosen as 0.1 the ten-neurons M1 model is determined to be the model that gives the best results. The models prepared by Ayeri et al. (2018) using the FFNN structures and the same models results of this study using the RBNN structures were compared. Values for comparing studies are presented in Table 4. It has been determined that the models using the RBNN structures give better results than the models used in FFNN.

except heavy metal and ammonia has a superior performance compared to other models. In coliform estimation, it was determined that RBNN structures are more successful than FFNN structures and Artificial Neural Network structures can be used successfully in coliform prediction. More successful models can be developed by using various model structures and artificial neural network architectures in large work areas. In coliform estimation, it was determined that RBNN structures are more successful than FFNN structures and Artificial Neural Network structures can be used successfully in coliform prediction. More successful models can be developed by using various model structures and artificial neural network architectures in large work areas.

Acknowledgements

This work was supported by the Ministry of Environment and Urbanization, Project name; Determination of Deep Sea Discharge Design Criteria. The authors would like to thank for this support.

Authorship Contributions

Aydın Er conducted experimental and statistical studies. Sisman supported the map drawings and writing. Ardali supported writing, experimental studies and statistical studies.

Data Availability Statement

It is a fact that sea discharges, which have a very important place in reducing environmental pollution, are used extensively in the Black Sea region. Coliforms with indicator parameters can be determined in the laboratory environment or can be estimated using artificial intelligence methods in the light of laboratory measurements. In this study, the coliform values of Tekkekoy deep sea discharge were made to determine using a radial-based artificial neural network. Samples taken from the study area were analyzed in a laboratory environment and presented as input to RBNN architecture. The best model was determined by comparing the amount of coliform estimated using RBNN with the amount of coliform determined as a result of laboratory studies. In the study where the dispersion parameter, which is one of the parameters of the RBNN structure, varies between 0.1–1.12 different model structures were prepared. Although heavy metals and ammonia were added as parameters in 6 of the 12 models prepared, these parameters were not used in the other models. The correlation value in RBNN structures where the dispersion parameter of the best results was determined as 0.1 was 95.7%–98.67%. It has been determined that the model with 10 neurons expressed as M3 and all parameters

The authors confirm that the data that supports the findings of this study are available within the article. Raw data that support the finding of this study are available from the corresponding author, upon reasonable request.

Conflict Of Interest

The author declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Ethics

There are no ethical issues with the publication of this manuscript.

References

1. Mossa M. Field measurements and monitoring of 2014;70:399–408. [CrossRef] wastewater discharge in sea water. Estuar Coast [14] Choi S-W, Bae H-K. Daily prediction of total coli- Shelf Sci 2006;68:509–514. [CrossRef] form concentrations using artificial neural net-
2. Noble RT, Lee IM, Schiff KC. Inactivation of indica- works. KSCE J Civ Eng 2017;22:467–474. [CrossRef] tor micro-organisms from various sources of faecal [15] Katip A. The usage of artificial neural networks in contamination in seawater and freshwater. J Appl microbial water quality modeling: a case study from Microbiol 2004;96:464–472. [CrossRef] the lake Iznik. Appl Ecol Environ Res 2018;16:3897–
3. Scarlatos PD. Computer modeling of fecal coliform 3917. [CrossRef] contamination of an urban estuarine system. Water [16] Broomhead DS, Lowe D. Multivariable functional Sci Technol 2001;44:9–16. [CrossRef] interpolation and adaptive networks. Complex Syst
4. Brion GM, Neelakantan TR, Lingireddy S. Using 1988;2:321–355. neural networks to predict peak Cryptosporidium [17] Ham FM, Kostanic I. Principles of Neurocomputing concentrations. J Am Water Works Assoc for Science and Engineering. New York: McGraw 2001;93:99–105. [CrossRef] Hill; 2001.
5. Neelakantan TR, Brion GM, Lingireddy S. Neural [18] He H, Yan Y, Chen T, Cheng P. Tree height esti- network modelling of Cryptosporidium and Giardia mation of forest plantation in mountainous ter- concentrations in the Delaware River, USA. Water rain from bare-earth points using a DoG-coupled Sci Technol 2001;43:125–132. [CrossRef] radial basis function neural network Remote Sens
6. Neelakantan TR, Lingireddy S, Brion GM. 2019;11:12–71. [CrossRef] Effectiveness of different artificial neural network [19] Okkan U, Dalkilic HY. Monthly runoff model for training algorithms in predicting protozoa risks in Kemer dam with radial based artificial neural net- surface waters. J Environ Eng 2002;128:533–542. works. Teknik Dergi 2012;23:5957–5966. [CrossRef]
7. Lin B, Kashefipour SM, Falconer RA. Predicting [20] Sagiroglu S, Beşdok E, Erler M. Artificial Intelligence near-shore coliform bacteria concentrations using Applications in Engineering 1: Artificial Neural ANNS. Water Sci Technol 2003;48:225–232. [CrossRef] Networks. 1st ed. Kayseri: Ufuk Publishing; 2003.
8. Mas D, Ahlfeld D. The development and evaluation [Turkish]. of artificial neural networks for modeling indica- [21] Zhang L. Optimizing ANN training performance tor organism concentrations. Proc., 2005 UCOWR for chaotic time series prediction using small data Annual Conf. River and Lake Restoration: Changing size. Int J Mach Learn Comput 2018;8:606–612. Landscapes, Portland, Ore, 2005. [22] Ayeri T, Aydın Er B, Zeybekoglu U, Ertan E, Ardali
9. Mas DML, Ahlfeld DP. Comparing artificial neu- Y. Evaluation of Samsun Tekkekoy deep sea dis- ral networks and regression models for predict- charge system in Turkey’s Black sea coast using arti- ing faecal coliform concentrations. Hydrol Sci J ficial neural networks. International Symposium on 2007;52:713–731. [CrossRef] Urban Water and Wastewater Management, October
10. Eleria A, Vogel RM. Predicting fecal coliform bacte- 25–27, Denizli, 2018. ria levels in the Charles river, Massachusetts, USA. J Am Water Resour Assoc 2005;41:1195–1209. [CrossRef]

Share and Cite