Next Article in Journal
Integrated Environmental–Economic Assessment of CO2 Storage in Chinese Saline Formations
Previous Article in Journal
Groundwater Level Estimation Using Improved Transformer Model: A Case Study of the Yellow River Basin
Previous Article in Special Issue
A Review of Low-Cost Point-of-Use Water Treatment Solutions Addressing Water Access and Quality in Resource-Limited Settings
 
 
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Artificial Neural Network-Based Prediction of Clogging Duration to Support Backwashing Requirement in a Horizontal Roughing Filter: Enhancing Maintenance Efficiency

by
Sphesihle Mtsweni
1,*,
Babatunde Femi Bakare
1 and
Sudesh Rathilal
2
1
Environmental Pollution and Remediation Research Group, Department of Chemical Engineering, Faculty of Engineering, Mangosuthu University of Technology, P.O. Box 12363, Durban 4026, South Africa
2
Green Engineering Research Group, Department of Chemical Engineering, Faculty of Engineering and the Built Environment, Durban University of Technology, Steve Campus, S3 L3, P.O. Box 1334, Durban 4000, South Africa
*
Author to whom correspondence should be addressed.
Water 2025, 17(15), 2319; https://doi.org/10.3390/w17152319
Submission received: 14 June 2025 / Revised: 19 July 2025 / Accepted: 22 July 2025 / Published: 4 August 2025
(This article belongs to the Special Issue Advanced Technologies on Water and Wastewater Treatment)

Abstract

While horizontal roughing filters (HRFs) remain widely acclaimed for their exceptional efficiency in water treatment, especially in developing countries, they are inherently susceptible to clogging, which necessitates timely maintenance interventions. Conventional methods for managing clogging in HRFs typically involve evaluating filter head loss coefficients against established water quality standards. This study utilizes artificial neural network (ANN) for the prediction of clogging duration and effluent turbidity in HRF equipment. The ANN was configured with two outputs, the clogging duration and effluent turbidity, which were predicted concurrently. Effluent turbidity was modeled to enhance the network’s learning process and improve the accuracy of clogging prediction. The network steps of the iterative training process of ANN used different types of input parameters, such as influent turbidity, filtration rate, pH, conductivity, and effluent turbidity. The training, in addition, optimized network parameters such as learning rate, momentum, and calibration of neurons in the hidden layer. The quantities of the dataset accounted for up to 70% for training and 30% for testing and validation. The optimized structure of ANN configured in a 4-8-2 topology and trained using the Levenberg–Marquardt (LM) algorithm achieved a mean square error (MSE) of less than 0.001 and R-coefficients exceeding 0.999 across training, validation, testing, and the entire dataset. This ANN surpassed models of scaled conjugate gradient (SCG) and obtained a percentage of average absolute deviation (%AAD) of 9.5. This optimal structure of ANN proved to be a robust tool for tracking the filter clogging duration in HRF equipment. This approach supports proactive maintenance and operational planning in HRFs, including data-driven scheduling of backwashing based on predicted clogging trends.

1. Introduction

The broad application of filtration encompassing various methods such as sand filtration and roughing filtration, including horizontal flow roughing filters, up-flow roughing filters, down-flow roughing filters, and direct horizontal roughing filtration, is one of the most indispensable and long-standing technologies in water pre-treatment. Filtration practices have been conducted for many hundreds of years particularly in larger parts of developing communities where pressing water problems remain prevalent [1,2]. Its remarkable degree of cost-effectiveness and exceptional extent of efficiency in removing a wide array of pollutants along with its high degree of flexible design characteristics makes it one of the ideal choices in resource-constrained environments [3,4]. Numerous studies consistently highlight the remarkable degree of effectiveness of roughing filtration as a comprehensive design-to-performance evaluation technology [5,6]. Roughing filtration generally demonstrates a high level of efficiency in terms of removal of suspended solids, turbidity, coliform bacteria, and other various water pollutants [7]. This exceptional degree of efficacy renders it an indispensable pre-treatment technology, particularly in rural low-resource communities where access to advanced water treatment systems is constrained. Therefore, the attainment of proper design and thorough maintenance in roughing filtration systems cannot be underestimated [8].
The focus of previous investigations into roughing filtration has predominantly concentrated on theoretical approaches for detecting clogging duration, fouling, or clogging. These theoretical approaches have provided a foundational understanding of the mechanisms involved, enabling researchers to develop more effective filtration systems and maintenance strategies. Several studies on roughing filtration technology have reported clogging durations ranging from 1 to 3 months for systems with horizontal flow configurations and from 2 to 4 months for configurations of upflow and downflow [1,9]. A schematic representation of different types of roughing filters is presented in Figure 1, which illustrates the influent and effluent flow stream configurations for horizontal flow, upflow, and downflow roughing filter technologies. The construction and flow characteristics in each filter type is indicated in accordance with the design guidelines proposed by [2], ensuring optimal design and filtration performance.
Despite the remarkable efficiency of roughing filtration technology, it remains susceptible to clogging, with increasing amounts of accumulation of solid matter within the layer of filtration medium. This can ultimately pose a performance issue, which can be a limiting factor to the capacity of the filter in terms of the effective area of filtration [10]. With continuous accumulation of solid matter in the filtering space of the filter, this reduces its sustainability, leads to shorter periods of filter run, and compromises standards of water quality [3,8]. In roughing filtration, the mechanism of the filtration process is typically explained through the concept of filtration theory. The filtration theory describes how the concept of particles of specific sizes and the speed or settling rates relates to the filtration process and removal mechanism of a range of sizes of particles in water as they diffuse through a layer of the filtering media [1]. Concentrated particles interact with the filtering media and adhere to it due to electrostatic charge interactions. This process results in the gradual accumulation of solids within the media’s pore spaces, which over time can lead to a point of clogging and reduce the efficiency of the filtration system [11,12,13].
Therefore, effectively managing these accumulations is critically important for maintaining optimal performance in filtration systems. Moreover, to maintain the desired filter efficiency, a backwashing phase is essential for removing the amount of accumulated solid matter. Filtration performance is typically monitored and guided by the availability and analysis of filtration data, including metrics of hydraulic flowrates, pressure drops, particle sizes, contaminant levels, and other relevant filtration parameters. Therefore, the filtration system can then be optimized to ensure that it continues to operate effectively and maintains the required efficiency. Regular backwashing therefore aids in preventing clogging and extends the duration of the filtration system [8].
The steps of tracking and monitoring filter clogging in roughing filters remain crucial for the effective steps of maintenance planning, including the timely periods of initiation of backwashing. One of the common methods of assessing and tracking filter performance in roughing filtration is by measurement of pressure drop across the filter [3,14]. Another method encompasses assessment and interpretation of effluent water data against predefined water quality limits and standards. This method ensures that the treated water meets the required or necessary quality criteria. However, significant fluctuations in operating conditions can highly impact the extent and level of filter performance and fouling coefficient frequencies [10].
Accurate measurement of clogging duration requires a combination of monitoring processes, analytical techniques, and effective methods of data interpretation based on analytical methods of influent and effluent water quality analyses in the filtration system [8,11]. Calibration periods of pressure head instruments then become critical for ensuring accurate measurements of clogging duration in roughing filtration systems [1]. Several studies have rigorously examined the concept of fouling and clogging duration in roughing filtration with much attention on conventional methods using empirical or theoretical techniques and models [3,5,6,11,15]. Responses to conventional methods of clogging duration in roughing filtration are frequently constrained by different types of specific filtration conditions and by a variety of assumptions and may not effectively adapt in the event of rapidly changing operational scenarios [10]. Additionally, higher levels of concentrations of influent solids may result in an increased amount of solids retention per unit volume [1].
Roughing filtration technologies have been increasingly applied as one of the essential pre-treatment units in decentralized and small-scale water treatment systems, especially in low-resource settings where affordability, simplicity, and operational resilience are of greater priority. These systems are particularly suitable in the context of rural and peri-urban communities where centralized wastewater treatment infrastructure is lacking and given the demand thereof for low-cost, low-maintenance solutions [2]. HRFs, in particular, have demonstrated robust performance in terms of treating a wide variety of influent types, including surface water with high turbidity, domestic wastewater, graywater, low-strength industrial effluents, and agricultural runoffs [3,16,17,18,19,20,21]. The primary sources of wastewater typically include households, communal sanitation systems, local food-processing facilities, and irrigation return flows. These sources may vary significantly in terms of quality and quantity depending on geographic location, infrastructure availability, and intended reuse [22,23,24,25]. For instance, domestic graywater sources may be reused for irrigation purposes, whereas pre-treated industrial effluents may require additional steps of filtration before discharge [16]. Of note, a thorough understanding in the context of the origins and characteristics of any water source is also essential for the appropriate selection and design of filtration systems such as HRFs, which must align with specific pollutant loads, flow variability, and final water quality targets [2,22].
A wide range of filter media have been investigated for use in HRFs. These include coarse materials such as gravel, crushed stones, coconut husks, charcoal, and other organic or synthetic materials, and these can significantly enhance the removal of various pollutants while maintaining cost-effectiveness and ease of maintenance [3,26]. Media selection is typically guided by local availability, cost, porosity, and pollutant removal performance. These characteristics make HRFs particularly favorable, mostly in the case of decentralized and resource-constrained settings [27]. Design parameters of HRFs can vary depending on the influent water source and characteristics and the desired level of treatment. The standard configurations often include small hydraulic loading rates ranging from 0.3 to 1.5 m3/m2·h, filter bed depths between 0.2 and 1.2 m, and multi-layered beds arranged in graded particle sizes from coarse to fine to facilitate sequential sedimentation. These parameters directly affect treatment efficiency, hydraulic retention time, clogging frequency, and long-term performance [2].
ANNs offer a unique alternative to conventional methods and are rapidly emerging as a preferred modeling technique [1,13]. In the event of rapid dynamic conditions, ANNs can offer a notable level of utility, learn and adapt to new data and evolving conditions, and still provide a more dynamic and responsive approach [10,15], whereas conventional methods are often unable to adapt to rapidly changing operational scenarios. Despite their potential, research exploring the application of ANNs for monitoring the clogging duration cycle in roughing filtration remains relatively sparse. Moreover, conventional techniques often rely on methods of manual inspections or simple threshold-based methods. In addition, they can be unable to detect early indicators of clogging and may be to a larger extent less effective in managing complex and large amounts of data [10].
The primary objective of this study is to predict filter clogging and effluent turbidity in HRF equipment. The ANN was configured with two outputs, the clogging duration and effluent turbidity, which were predicted concurrently. Effluent turbidity was mod-eled to enhance the network’s learning process and improve the accuracy of clogging prediction. Different types of water quality parameters were used for the development of ANN-based prediction of clogging duration and effluent turbidity in HRF equipment (influent turbidity, filtration rate, pH, conductivity, clogging duration, and effluent turbidity), and a certain set of network parameters (learning rate, momentum, and the number of neurons in the hidden layer) are carefully considered for use. This proposed method of predication and monitoring of clogging duration is of paramount importance in HRFs and provides invaluable practical contributions and insights in terms of supporting efficient maintenance planning and performance-based decision-making in HRFs.

2. Materials and Methods

2.1. Study Area and Water Quality Data

This work extends previous studies conducted by the authors, in terms of the application of HRF equipment to assess its efficiency in treating graywater in the study area [17,28,29]. In this study, a new set of data has been acquired experimentally in the HRF equipment, as will be described in the following sections of the methodology, to investigate the method used to investigate the clogging duration and the effluent turbidity. The study area, which is the Umhlabeni informal settlement, is situated within the Umlazi township, which is approximately 24 km southwest of Durban, KwaZulu-Natal, South Africa (Figure 2). This densely populated region grapples with a whole range of substantial environmental issues, notably the lack of adequate water and sanitation infrastructure and the inefficiency of wastewater management practices.
This area generates substantial volumes of domestic wastewater containing different types of physico-chemical pollutants of graywater primarily from a variety of social and graywater-producing activities of the inhabitants. Notably, there are no existing standard practices of graywater use and reuse in the study area. The impact of domestic wastewater significantly influences both the ecological and social aspects of the local population. The selection of the study area was strategic due to its distinctive characteristics and its potential representative nature of informal settlements with similar types of dynamics. This is therefore intended to ensure that the outcomes of the study are robust, generalizable, and applicable across diverse contexts.

2.2. Constructed HRF Equipment

An HRF with multiple layers of gravel was constructed and adapted in the Chemical Engineering Laboratory at Mangosuthu University of Technology for the systematic collection of experimental data. Table 1 and Figure 3 present details of design parameters of the HRF equipment used in this study. This filter equipment adhered to the methods of [2] in terms of its design concepts and operation principles. This type of filter consists of three horizontally arranged compartments, each separated by constructed perforated partitions. Multiple layers of gravel media of varying effective sizes, coarse, medium, and fine, are carefully selected and subsequently charged into each compartment of the filter. Influent graywater transits into the feeding point in the primary compartment of the filter, where the first layer of coarse gravel of 14 mm size facilitates filtration and the removal of larger particulate matter contained in water. As the water advances to the second compartment with the second layer of medium gravel of 10 mm in size, the process of filtration further occurs and enhances the removal of pollutants while acting as the secondary stage of the filtration process. The bulk of water then progresses to the third segment of the filter, where the last layer of fine gravel of 8 mm effectively removes the residual amounts of pollutants in the filter [8,17,30].
In terms of dimensions, the filter had a total length of 3 m, and its operational mechanism was in terms of [2,7] design principles. This filter closely adhered to the concept of the [2] 1/3–2/3 filter theory, which was of importance to ensure the effectiveness of the filtration process. Following established guidelines, different layers of gravel media of specific range or sizes are typically recommended. That is a coarse layer of gravel in the range of 12–18 mm in the first compartment of the filter, a medium layer of gravel fractions in the range of 8–12 mm for the secondary compartment of the filter, and the fine layer of gravel in a range of 4–8 mm in the last compartment of the filter [2]. In this study, the availability of these guidelines is fundamentally of interest for the efficient operation of HRF equipment (Figure 3). A series of trials were conducted on a filter, and it previously demonstrated a remarkable degree of performance in treating large volumes of graywater in the work of [17].

2.3. Experimental Data and Sampling

The set of data from the HRF equipment was leveraged to train and optimize this supervised type of ANN, which subsequently enabled simultaneous prediction of filter clogging duration and effluent turbidity. Systematic sampling of raw and treated graywater was conducted, and analyses of graywater quality were conducted while adhering to rigorous methodological standards of water analyses. The sampling intervals were systematically structured and distributed over an interval of six consecutive days of each week for a period of just over 30 months to ensure comprehensive temporal coverage in the sampling of data. To account for diurnal variations and stochastic fluctuations in household water quality metrics, samples were carefully prepared, including sterilization and calibration of equipment used, and then systematically collected at predetermined intervals corresponding to presumed morning to midday and afternoon periods. The filter received quantities of graywater from a capacity of 0.5 m3 prefabricated rotary-extruded polyethylene-type reservoir. The required amount of influent raw graywater was facilitated by a small size-range mechanically fitted centrifugal pump.
The entire set of experiments was conducted by systematic sampling of both raw and treated graywater using the HRF system’s batch operating mode. All collected samples were recorded as average values. This was of higher interest for comprehensive data collection and capturing of variations in water quality due to daily household activities. In addition, this provides a comprehensive overview of the dimensions of water quality characteristics following the performance of the filter. The entire set of experiments also encompassed detailed steps of analytical analyses and comprehensive recording of various performance metrics at specifically selected filtration rates of 0.3 m/h, 0.6 m/h, and 0.9 m/h, which represent the flow velocities through the filter media. Rigorous standards of water quality analyses were used in the experiments, and this ensured that the findings of this study were reliable and applicable to real-world filtration scenarios.

2.4. Water Measuring Instruments

Advanced types of water quality measuring instruments are used for precise monitoring of water quality parameters. These instruments were important for monitoring several key water quality parameters such as conductivity, pH, turbidity, and filtration rate. The amounts of conductivity were measured by a calibrated Orion Star conductivity meter, while recordings of pH and turbidity were measured by a calibrated pH meter and an Orbeco Hellige turbidity meter. In addition, a flowmeter device was used for flow regulation at the inlet section of the filter, ensuring a consistent and steady supply of graywater into the HRF equipment under varying head conditions. All measuring instruments were regularly calibrated to ensure good degrees of accuracy and reliability in the measurement of the entire recorded dataset of this study.

2.5. ANN Software and Structure

One of the objectives of the study was to perform the modeling of the filter equipment. The entire process of the training phase of the network was performed on MATLAB version R2015a, including multi-steps of network calibration. Both outputs of the network were modeled simultaneously using MATLAB’s neural network toolbox. This type of software is a highly capable and effective software that can be useful for the assessment of various training aspects of ANN. It offers a vast array of capabilities with robust functionalities that meet performance measures and requirements for the evaluation of ANN models. Therefore, it was possible to efficiently perform the design and training of the network and analyses of the ANN structures. The optimal architecture of the ANN was systematically evaluated using various types of performance metrics to ensure robust statistical validation and comparison of the developed ANN structures.

2.6. Data Normalization

The steps of the network involved in training require the careful normalization of the input-output datasets. A few moderate steps of this process transform the entire set of captured data to fit a standard range of 0 to 1. The normalization of the datasets speeds up the process of convergence and reduces several training issues related to local minima by using the minimum and maximum values from the training dataset. The normalization of the data is crucial for ensuring the stability and efficiency of the ANN training process, which then allows for a faster process of convergence and better performance of the network [31,32].

2.7. ANN Performance

With constructed ANN structures employed, different types of performance indicators such as MSE and R-coefficients were used to assess the training performance of the training, testing, and validation datasets. The main objective of network development focused on the minimization of values of MSE and the maximization of the R-coefficients of the ANN. Consequently, the optimal ANN structure was identified as the configuration that effectively achieved the minimization of the MSE values and the maximization of the R-coefficients, thereby attaining the highest level of predictive accuracy. The selection of performance metrics was carefully made to provide a comprehensive evaluation of the ANN’s predictive capabilities while ensuring good accuracy and reliability of the developed model.

2.8. ANN Training and Performance

The entire dataset for the network was generated experimentally using the HRF equipment described in this study. The data consisted of a set of three different types of categories: training set, test set, and validation dataset. Out of the total data of 637 datasets, a proportion of the dataset equal to 70% was allocated to the training set, while the remaining proportion of 30% of the original set was divided equally between the testing and validation sets. When training the ANNs, all developed structures of network designs used a single hidden layer and followed the method of a trial-and-error approach. Various training parameters of the network were monitored for network development, including learning algorithm, learning functions, and activation functions. An iterative approach was used for the development of optimal topologies while varying the number of neurons in the hidden layer between a minimum and maximum of 4 and 10. The optimal network used different types of learning algorithms and functions based on the iterative method of model development.
Roughing filters are widely recognized for their robust performance in water treatment processes, particularly for their outstanding effectiveness in significantly reducing various types of water contaminants, including turbidity, which is commonly used as an indicator for the assessment of water quality and can obtain high removal of turbidity including other pollutants [33,34,35,36]. Effluent turbidity measurements were carried out in a filter and were recorded for use as a lower-to-upper limit indicator of clogging duration cycle. Recorded values of the effluent turbidity were utilized as a lower-to-upper limit indicator of the clogging duration variable. All of the water quality data metrics of the filter performance were recorded throughout filter operations to as far as the effluent turbidity consistently exceeded the permissible threshold or was close at the point of approaching the upper limit due to the gradually increasing accumulation of scale from diverse water contaminants. The ANN was configured with two outputs, the clogging duration and effluent turbidity, which were predicted concurrently. Effluent turbidity was modeled to enhance the network’s learning process and improve the accuracy of clogging prediction. At each filter run, the iterative training process utilized a range of input parameters, including influent turbidity, filtration rate, pH, conductivity, and effluent turbidity. Throughout the operational period, all ANN-related parameters and outputs were systematically recorded. Both influent and effluent turbidity levels were continuously monitored and recorded, along with the corresponding clogging durations and turbidity values. A turbidity removal efficiency of 70% was established as the minimum acceptable threshold for evaluating filter performance [3,15,37]. During training, network parameters such as learning rate, momentum, and the configuration of neurons in the hidden layer were also recorded.
The investigation of various training techniques and ANN attributes resulted in the configuration of several different ANN topologies based on two types of training algorithms: LM and SCG training algorithms (Table 2). These types of learning algorithms effectively learn and generalize instances of non-linear problems and offer a strong type of features such as the fastest backpropagation, all in accordance with the principles and theoretical frameworks underpinning ANN design. As shown in Table 2, a variety of ANN training parameters included the LM and SCG training algorithms, GDM adaptation learning functions, GD learning functions, transfer functions, and either Tansig or Logsig in the hidden layer with purelin utilized in the output layer.

3. Results

3.1. Experimental Results

Table 3 provides a comprehensive summary of the descriptive statistics for water quality parameters obtained for the entire set of ANN training data. It illustrates several variations observed in these parameters, particularly following the pre-and post-application of graywater treatment. The data include the range of unstandardized lower and upper bounds for both input and output parameters. The methods of data collection were by systematic measurement and recording of water quality parameters. The values of filtration rate intervals typically ranged from 0.3 to 0.9 m/h. Regarding trends of water quality, high levels of turbidity removal efficiency and low coefficients of particulate matter concentrations were evident. The monitoring of turbidity profiles at the discharge point of the filter provided significant insights into its efficiency and overall performance.

Source Graywater Quality Indicators

The graywater was characterized for water quality conditions prior to ANN application. A variety of key parameters including measured parameters were influent turbidity, pH, conductivity, filtration rate, clogging duration, and the effluent turbidity from the filter. Specifically, the influent turbidity of greywater ranged from 120 to 319 NTU, with a mean value of 217 NTU. The effluent turbidity ranged from 6 to 95 NTU, with a mean of 32 NTU. The pH varied between 6.8 and 11.4, while conductivity ranged from 301 to 1094 µS/cm. The filtration rate was observed between 0.3 to 0.9 m/h, and the clogging duration spanned 10 to 25 days. These values reflect the variability typically found in domestic graywater from washing activities and those obtained following the performance of HRF equipment. A detailed statistical summary of these parameters is presented in Table 3.

3.2. ANN Training Results

A comprehensive summary of the findings in terms of the performance of ANN with two of the network LM and SCG algorithms was recorded in this study. A systematic and thorough training of the ANN showed that both algorithms produced satisfactory results as evidenced by values of correlation coefficients across all recorded values of data subsets and notably low values of MSE coefficients. Moreover, the ANN of the LM algorithm demonstrated a superior performance for the prediction of clogging duration and effluent turbidity in HRF equipment. The performance of the LM algorithm offered a more precisely accurate degree of precision and a higher degree of accuracy as indicated by its number of performance indices recorded and by recorded performance characteristics. The training results of the ANN with LM algorithm are summarized in Table S1 of the Supplementary Materials. By employing a supervised type of ANN and optimizing its number of learning and training functions as shown in Table S1, the number and types of ANN structures—the optimal number of neurons in the hidden layer, training and learning functions, and algorithms, along with gradient descent with momentum backpropagation, it was therefore possible to effectively assess the efficacy of the ANN results through the application of the trial-and-error method [38].
The structure of the optimal network essentially comprised four input neurons, a single hidden layer containing eight neurons, and two output neurons in the output layer. This type of configuration consistently demonstrated superior characteristics in terms of learning efficiency and a good power of predictive accuracy as evidenced by network performance indices recorded. This configuration of optimal-performing ANN yielded a minimum value of MSE of less than 0.001 and recorded the fastest number of iterations. The ANN with the application of the LM algorithm achieved a good degree of performance with %AAD of 9.5 for the prediction of clogging duration. In Table S1, the R-coefficients specifically recorded for all values were rounded to three decimal places for clarity and ease of reporting. These coefficients were all above 0.999 for the training, testing, validation, and overall datasets, respectively. The network used learngdm, the tansig function for the hidden layer, and the purelin function in the output layer of this network.
Table S2 in the Supplementary Materials comprehensively presents the performance of the ANN using the SCG training algorithm across different types of network structures. Using the SCG algorithm, after several steps of training, the optimal configuration of four neurons in the input layer, a single hidden layer with six neurons, and two neurons in the output layer was obtained at the completion of training. The minimum value of MSE achieved with the corresponding type of training algorithm showed a lower degree of improvement in terms of performance compared to the LM algorithm. A remarkably higher degree of predictive attributes in terms of accuracy for R-coefficients was obtained to be above 0.999 for the training, testing, validation, and overall dataset. This type of configuration consistently demonstrated a remarkable learning efficiency and predictive accuracy. The network used learning functions of learngdm and tansig for the hidden layer and the purelin function for the output layer of the network.

3.3. ANN Validation

Figure 4 shows a plot of MSE coefficients recorded from two different types of training algorithms. The plots illustrate the MSE coefficient as a function of the number of iterations or epochs. The recorded number of training, testing, and validation errors showed consistently decreasing trends with the number of iterations. Model validation and overfitting prevention were addressed using MATLAB’s nntraintool, which employs early stopping based on validation performance. This demonstrates the efficiency of ANN in terms of its robust learning process. The performance of the network recorded in terms of MSE was evident with a value of less than 0.001 being evident for up to 1000 epochs. The SCG algorithm also achieved an MSE performance of less than 0.001 with specified epochs of training. Of note, the findings of test and validation curves for both types of algorithms exhibited similarly decreasing trends with no significantly noticeable indicators of overfitting in the structures of developed ANNs. A consistent trend of performance in ANN is of interest, as it clearly relates to the performance characteristics of the model and its ability to effectively learn from and adapt to the given type of dataset [39].
The training state of ANN with the LM algorithm in terms of validation checks, gradient, and momentum update parameters was also investigated. Of note, different types of training parameters served as key indicators of the ANN’s degree of overall performance including the effectiveness of the training cycle. The iterative steps of detailed learning were evident for validation of the network’s training parameters. The network was able to reach a point of convergence after the minimization of gradient and momentum update parameters of the network. The network values of the training coefficient recorded were as follows: 0.99998 for the training, 0.99995 for the test set, 0.99993 for the validation set, and 0.99997 for the overall dataset after a total number of 1000 epochs. For clarity and consistency in reporting, these coefficients were rounded to three decimal places and are presented in Table S1. Most of the data points aligned closely with a 45-degree plot, and this indicated a near-perfect indication of a close to perfect fit between values of network outputs and target outputs. Positive trends were also observed in instances of the validation and testing phase with no significant indicators of overfitting patterns in the training dataset (Figure 5).

4. Discussion

4.1. ANN Performance Analysis

Although HRFs are widely used in many decentralized water treatment systems, most studies focus on effluent water quality treatment and its effectiveness [9,10,15,17,27,28]. These studies typically rely on empirical models, which are often unable to capture the component of nonlinear dynamics of graywater filtration. This study utilized ANN for the prediction of clogging duration and effluent turbidity in HRF equipment. The ANN was configured with two outputs, the clogging duration and effluent turbidity, which were predicted concurrently (Table 2 and Table 3). Effluent turbidity was modeled to enhance the network’s learning process and improve the accuracy of clogging prediction. The application of ANN in this study incorporated filtration rate, pH, turbidity, and electrical conductivity as input variables. The model architecture (4-8-2) consists of four input neurons, eight hidden neurons, and two output neurons representing clogging duration and effluent turbidity. With LM algorithm, the ANN also achieved a low value of MSE of 0.001 and R-coefficients exceeding 0.999 across training, validation, testing, and the entire dataset.
To assess the performance capability of this ANN, various performance metrics for ANN were recorded. A variety of the main parameters of the ANN and measured parameters were influent turbidity, pH, conductivity, filtration rate, clogging duration, and the effluent turbidity from the HRF equipment (Table 3). These values reflect the variability typically found in domestic graywater from washing activities and those obtained following the performance of HRF equipment, also indicating the extent of variation in water quality throughout the monitoring period of the study. The %AAD of 9.5 was recorded specifically in relation to the prediction of clogging duration. As a widely accepted indicator of filter performance, turbidity is commonly monitored to guide maintenance scheduling and interventions [40,41].
Turbidity was included as a complementary output to strengthen the model’s operational relevance and broaden its interpretive capacity in assessing filter performance. Turbidity is widely recognized as a practical and reliable indicator of filtration efficiency, particularly in the case of decentralized graywater treatment systems where continuous monitoring of complex parameters is often not feasible. The study of [42] demonstrated the application of machine learning for the derivation of sea water turbidity from Sentinel-2 satellite imagery and highlighted its robustness in handling large volumes of remote sensing data for environmental monitoring. Turbidity also serves as one of the early warning signals for clogging and is commonly used in the field to guide maintenance decisions [8]. In this study, a turbidity removal threshold of 70% was used to define the end of a filter run. This threshold was selected based on its alignment with operational practices in decentralized graywater reuse systems, where treated water is typically used for non-potable purposes such as toilet flushing, garden irrigation, or surface cleaning. Ref. [18] recommends that effluent turbidity for graywater reuse in agriculture and other non-potable applications should ideally remain below a value of 10 NTU to ensure safe and effective use.
Similarly, Ref. [43] and other national guidelines reference turbidity thresholds in the range of 5 to 10 NTU as benchmarks for acceptable performance in non-potable reuse contexts. Several studies on HRFs and slow sand filters also use turbidity removal efficiency or effluent turbidity thresholds within this range as indicators of filter performance. A study of [44] demonstrated the progress in terms of integration of machine learning models with RGB sensors for the quantification and classification of water turbidity. The study offered a compelling demonstration of the potential of ANN-based methods in the application of real-time environmental monitoring in marine areas. The significance of this work lies in the advancement of intelligent systems for the management of water resources as well as the value of responsiveness and accuracy in modern water quality monitoring frameworks. These benchmarks hold particular significance in decentralized systems, where ease of operation and dependable performance are essential. In this work, the ANN developed supports proactive maintenance and operational planning in HRFs, including data-driven scheduling of backwashing based on predicted clogging trends. In addition, the integration of turbidity as one of the network outputs aligns with the findings of [45], who emphasized the importance of integrating multiple water quality indicators to improve the accuracy and operational value of graywater treatment models.

4.2. ANN Applications and Implications for HRF Operation

This network architecture, with its two outputs, enabled the model to capture both the temporal and quality-related aspects of HRF equipment performance. The variability in the input parameters demonstrated the complexity of graywater treatment and hence a need for a robust predictive ANN. The network, trained using the LM algorithm, achieved high correlation coefficients across all data subsets. The results of this study clearly showed that the application of the LM algorithm led to a better predictive performance than that of the SCG algorithm. This is largely due to the strength of the LM algorithm in handling nonlinear regression problems with a high degree of precision, particularly when applied to structured datasets of the size and nature obtained in this study. The LM, with its advantage of gradient descent and the application of Gauss-Newton methods, allowed for faster convergence and more accurate minimization of error, resulting in lower values of MSEs compared to that of the SCG algorithm (Figure 4 and Figure 5).
In addition, the network trained using the LM algorithm consistently produced higher values of the correlation coefficient across all subsets of data, which showed a strong relationship between the predicted outputs of the model and the actual observed values of clogging duration. The ability of the ANN to capture the complex interactions of the input parameters with the clogging behavior of the HRF highlights the effectiveness of the ANN and LM algorithm. Therefore, the application of the LM algorithm proved to be more suitable and offered a sensible level of predictive reliability that is essential for the practical implementation of ANN-based monitoring systems in water treatment operations. Previous studies have reported the effectiveness of HRFs in reducing turbidity and other different types of pollutants in graywater [9,10,15,17,27,28]; however, there has been limited application of ANNs to study clogging duration in HRF systems. For instance, Ref. [28] reported average removal efficiencies of 90% for turbidity and 70% for chemical oxygen demand in a pilot-scale HRF system in Durban, South Africa. Similarly, an earlier effort by [9] proposed a useful physical model for estimating clogging duration in HRFs.
The ANN developed in this study demonstrated a highly robust level of performance in line with studies of ANN applications in the field of water treatment systems. For instance, the work of [46] reported R-values exceeding 0.99 and a MSE of 5.962 for the prediction of biological oxygen demand (BOD), total nitrogen (TN), and total suspended solids (TSS) in the context of wastewater treatment plants. Similarly, the study of [47] involved the development of an ANN model for the estimation of BOD in wastewater using a combination of water quality parameters, achieving near-ideal levels of performance with low values of error metrics, thereby supporting the reliability of ANN-based approaches in the domain of wastewater treatment applications. A broader systematic review of ANN applications by [48], which involved the analysis of 44 studies, confirmed that feedforward ANN models trained using backpropagation are among the most commonly employed tools for the prediction of effluent quality in full-scale systems of wastewater treatment. However, many of these models are designed for the prediction of a single output. In addition, the work of [49] noted the limited application of ANN in the context of water treatment plant systems, particularly for the purposes of predictive maintenance or operational planning. Therefore, this study not only aligns with the broader body of ANN applications in the field of water treatment but also advances the state of the art by applying ANN to the prediction of clogging duration in HRFs, which is a domain traditionally modeled using empirical or theoretical approaches.

4.3. Implications for HRF Operation and Maintenance

The conventional methods of tracking and monitoring maintenance in HRFs are less proactive, and despite being widely adopted and well-established, these methods are often limited, as they may be unable to detect early indicators of clogging and may be to a larger extent less effective in managing complex and large amounts of data [10]. In many real-world scenarios, the scheduling of backwashing is often based on the use of fixed intervals or the judgment of operators, both of which can lead to inefficiencies in the use of water, labor, and system downtime [1,2]. The introduction of an ANN tool capable of providing data-driven predictions of the onset of clogging allows for the development of a more responsive and efficient maintenance approach. The prediction of clogging duration enables the planning of backwashing activities based on the actual condition of the filtration system rather than assumptions, thereby preventing the risk of performance degradation. This approach supports the optimization of the use of operational resources, the reduction of unnecessary consumption of water, and the extension of the lifespan of the filter media. By delivering precise, data-driven predictions of clogging onset, the proposed ANN offers a systematic and objective basis for optimizing backwashing operations and enhancing maintenance efficiency and ensuring more reliable performance of the filter equipment. Moreover, the effectiveness of artificial intelligence-based methods in the enhancement of water treatment management has been demonstrated by [50], who emphasized the role of intelligent models in the management of water treatment systems.

4.4. Limitations and Future Perspectives

This study utilized collected data from the study area, which was transported to the laboratory for treatment in an HRF system. While the use of such a set-up is valuable for controlled experimentation, future work should take into account the complexity of full-scale and real-world systems. In practical applications, the composition of graywater can vary significantly due to the diversity of household activities, the influence of seasonal changes, and the presence of different types of influent contaminants. Furthermore, the operation of full-scale systems is subject to fluctuations of flow rates, variations of maintenance practices, and changes in environmental conditions such as temperature and humidity, which are some of the factors that are not always captured in a controlled laboratory environment. As a result of these differences, the predictive accuracy and generalizability of the model may be limited when applied to larger, decentralized systems without further calibration. Additionally, the availability of complete and high-quality input data may not always be obtainable in field conditions. These limitations highlight the importance of future studies focused on the validation of the model using pilot- or full-scale systems under dynamic and realistic conditions.

5. Conclusions

This study utilized ANN for the prediction of clogging duration and effluent turbidity in a HRF equipment. The ANN was configured with two outputs, the clogging duration and effluent turbidity, which were predicted concurrently. Effluent turbidity was modeled to enhance the network’s learning process and improve the accuracy of clogging prediction. The optimized structure of the ANN, configured in a 4-8-2 architecture and trained with the application of the LM algorithm, achieved reasonable predictive performance, with an MSE of less than 0.001 and correlation coefficients exceeding 0.999 across all subsets of data. This ANN outperformed the SCG algorithm and achieved a %AAD of 9.5 for the prediction of clogging duration. This optimal structure of the ANN proved to be a robust tool for the tracking of clogging duration in HRF equipment. This method supports proactive maintenance and operational planning in HRFs, including data-driven scheduling of backwashing based on predicted clogging patterns. The findings are most applicable during the operational stage of HRF systems, where early detection of clogging can prevent performance decline, reduce head loss, and extend filter lifespan. These results indicate that the ANN offers a reliable and economically efficient alternative to conventional manual monitoring methods.
Its advantages also include high predictive accuracy, reduced need for frequent physical inspections, and potential cost savings in maintenance planning. The composition of graywater can vary significantly due to the diversity of household activities, the influence of seasonal changes, and the presence of different types of influent contaminants. In light of this, the validation of the optimal ANN under real-world conditions may add a much bigger role to the study. For instance, the performance of the ANN should be evaluated in full-scale HRF equipment and where the variability of graywater in terms of composition, the influence of seasonal changes, and the fluctuations of flow rates and environmental conditions such as temperature and humidity remain the key parameters. This study demonstrated the effectiveness of ANN as an accurate, cost-effective, and practical option for predicting clogging duration and effluent turbidity in HRF systems, while supporting smarter maintenance and operational decisions.

Supplementary Materials

The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/w17152319/s1, Table S1. Summary of training results of various ANN structures with LM algorithm; all R-values are rounded to three decimal places for uniformity and readability. Table S2. Summary of training results for SCG algorithm; all R-values are rounded to three decimal places for uniformity and readability.

Author Contributions

Conceptualization, S.M.; methodology, S.M.; software, S.M.; validation, S.M. and B.F.B.; formal analysis, S.M. and B.F.B.; investigation, S.M.; resources, B.F.B.; data curation, S.M.; writing—original draft preparation, S.M.; writing—review and editing, S.M., B.F.B. and S.R.; visualization, S.M., B.F.B. and S.R.; supervision, B.F.B. and S.R. All authors have read and agreed to the published version of the manuscript.

Funding

This research received no funding.

Data Availability Statement

All relevant data are included in the manuscript, and any additional datasets are available on request from the first author of the manuscript.

Acknowledgments

Our utmost gratitude is extended to the Mangosuthu University of Technology and chemical engineering department for supporting this research work that was conducted in the in the chemical engineering department’s laboratory.

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Boller, M. Filter mechanisms in roughing filters. J. Water Supply Res. Technol. AQUA 1993, 42, 174–185. [Google Scholar]
  2. Wegelin, M. Surface Water Treatment by Roughing Filters: A Design, Construction and Operation Manual; SANDEC Report No. 2/96; Swiss Centre for Development Cooperation in Technology and Management (SKAT): St. Gallen, Switzerland, 1996. [Google Scholar]
  3. Nkwonta, O.; Ochieng, G. Roughing filter for water pre-treatment technology in developing countries: A review. Int. J. Phys. Sci. 2009, 4, 455–463. [Google Scholar]
  4. El-Taweel, G.E.; Ali, G.H. Evaluation of roughing and slow sand filters for water treatment. Water Air Soil Pollut. 2000, 120, 21–28. [Google Scholar] [CrossRef]
  5. Clarke, B.A.; Lloyd, B.J.; Crompton, J.L.; Major, I.P. Cleaning of up-flow gravel pre-filters in multi-stage filtration water treatment plants. In Advances in Slow Sand and Alternative Biological Filtration; Graham, N., Collins, R., Eds.; Wiley & Sons: Chichester, UK, 1996; pp. 312–326. [Google Scholar]
  6. Collins, M.R.; Westersund, C.M.; Cole, J.O.; Roccaro, J.V. Evaluation of Roughing Filtration Design Variables; American Water Works Association Research Foundation and American Water Works Association: Denver, CO, USA, 1994. [Google Scholar]
  7. Galvis, G.; Latorre, J.; Visscher, J.T. Multi-Stage Filtration: An Innovative Water Treatment Technology; IRC International Water and Sanitation Centre: The Hague, The Netherlands, 1998. [Google Scholar]
  8. Wegelin, M.; Schertenleib, R.; Boller, M. The decade of roughing filters-Development of a rural water-treatment process for developing countries. J. Water SRT-Aqua 1991, 40, 304–316. [Google Scholar]
  9. Sittivate, D. How to estimate and design the filter run duration of a horizontal-flow roughing filter. Thammasat Int. J. Sci. Technol. 2000, 5, 2. [Google Scholar]
  10. Patil, V.B.; Kulkarni, G.S.; Kore, V.S. Performance of Horizontal Roughing Filters for Wastewater: A Review. Int. Res. J. Environ. Sci. 2012, 1, 53–55. [Google Scholar]
  11. Visscher, J.T.; Paramasivam, R.; Raman, A.; Heijnen, H.A. Slow Sand Filtration for Community Water Supply: Planning, Design, Construction, Operation and Maintenance; Technical Paper No. 24; IRC: The Hague, The Netherlands, 1987; pp. 20–32. [Google Scholar]
  12. Sorrentino, G.; Chellappah, K.; Biscontin, G. Understanding clogging mechanisms in filter media: An integration of laboratory findings and theoretical perspectives. Sep. Purif. Technol. 2025, 359, 130602. [Google Scholar] [CrossRef]
  13. Baveye, P.; Vandevivere, P.; Hoyle, B.L.; DeLeo, P.C.; Jacobson, A.R. Environmental impact and mechanisms of the biological clogging of saturated soils and aquifer materials. Crit. Rev. Environ. Sci. Technol. 1998, 28, 123–191. [Google Scholar] [CrossRef]
  14. Environmental Protection Agency (EPA). Water Treatment Manuals: Filtration; Johnstown Castle: Wexford, Ireland, 1995.
  15. Mukhopadhay, B.; Mujumder, M.; Barman, R.N.; Roy, P.K.; Mazumdar, A. Verification of filter efficiency of horizontal roughing filter by Wegelin’s design criteria and artificial neural network. Drink. Water Eng. Sci. Discuss. 2009, 2, 21–27. [Google Scholar] [CrossRef]
  16. Awasthi, A.; Gandhi, K.; Rayalu, S. Greywater treatment technologies: A comprehensive review. Int. J. Environ. Sci. Technol. 2024, 21, 1053–1082. [Google Scholar] [CrossRef]
  17. Mtsweni, S. Performance of a Horizontal Roughing Filtration System for the Pretreatment of Greywater. Master’s Thesis, Durban University of Technology, Durban, South Africa, 2016. [Google Scholar]
  18. World Health Organization (WHO). Guidelines for the Safe Use of Wastewater, Excreta and Greywater: Volume 4—Excreta and Greywater Use in Agriculture; WHO: Geneva, Switzerland, 2006.
  19. Dastanaie, J.; Nabi Bidhendi, G.R.; Nasrabadi, T.; Habibi, R.; Hoveidi, H. Use of horizontal flow roughing filtration in drinking water treatment. Int. J. Environ. Sci. Technol. 2007, 4, 379–382. [Google Scholar] [CrossRef]
  20. Fereidonian Dashti, A.; Abdul Aziz, H.; Ibrahim, A.H.; Zahed, M.A. Suspended Solid Removal of Palm Oil Mill Effluent Using Horizontal Roughing Filter and Calcinated Limestone. Water Air Soil Pollut. 2020, 231, 393. [Google Scholar] [CrossRef]
  21. Liu, Z.; Muhammad, T.; Puig-Bargués, J.; Han, S.; Ma, Y.; Li, Y. Horizontal roughing filter for reducing emitter composite clogging in drip irrigation systems using high sediment water. Agric. Water Manag. 2021, 258, 107215. [Google Scholar] [CrossRef]
  22. Mara, D. Domestic Wastewater Treatment in Developing Countries; Earthscan: London, UK, 2003. [Google Scholar]
  23. Morel, A.; Diener, S. Greywater Management in Low and Middle-Income Countries: Review of Different Treatment Systems for Households or Neighbourhoods; Sandec Report No. 14/06; Swiss Federal Institute of Aquatic Science and Technology (Eawag): Dübendorf, Switzerland, 2006. [Google Scholar]
  24. Jiménez, B.; Asano, T. Water Reuse: An International Survey of Current Practice, Issues and Needs; IWA Publishing: London, UK, 2008. [Google Scholar]
  25. Nisaa, A.F.; Krauss, M.; Spuhler, D. Adapting Santiago Method to Determine Appropriate and Resource Efficient Sanitation Systems. Water 2021, 13, 1197. [Google Scholar] [CrossRef]
  26. Cescon, A.; Jiang, J.-Q. Filtration Process and Alternative Filter Media Material in Water Treatment. Water 2020, 12, 3377. [Google Scholar] [CrossRef]
  27. Nkwonta, O.I.; Olufayo, O.A.; Ochieng, G.M.; Adeyemo, J.A.; Otieno, F.A.O. Turbidity Removal: Gravel and Char-coal as Roughing Filtration Media. S. Afr. J. Sci. 2010, 106, 1–5. [Google Scholar] [CrossRef]
  28. Bakare, B.F.; Mtsweni, S.; Rathilal, S. Pilot study of a horizontal roughing filtration system treating greywater generated from a peri-urban community in Durban, South Africa. J. Water Reuse Desal. 2019, 9, 123–134. [Google Scholar] [CrossRef]
  29. Mtsweni, S. Performance Optimization Modelling of a Horizontal Roughing Filter for the Treatment of Mixed Greywater. Ph.D. Thesis, Durban University of Technology, Durban, South Africa, 2021. [Google Scholar]
  30. Wegelin, M. Horizontal Flow Roughing Filtration (HRF): Design, Construction and Operation Manual; IRCWD Report No. 06/86; IRCWD: Dübendorf, Switzerland, 1986. [Google Scholar]
  31. Gyparakis, S.; Trichakis, I.; Daras, T.; Diamadopoulos, E. Artificial neural networks (ANNs) and multiple linear regression (MLR) analysis modelling for predicting chemical dosages of a water treatment plant (WTP). Water 2025, 17, 227. [Google Scholar] [CrossRef]
  32. Kim, Y.-S.; Kim, M.K.; Fu, N.; Liu, J.; Wang, J.; Srebric, J. Investigating the impact of data normalization methods on predicting electricity consumption in a building using different artificial neural network models. Sustain. Cities Soc. 2025, 118, 105570. [Google Scholar] [CrossRef]
  33. Mushila, C.; Ochieng, G.; Otieno, F.; Shitote, S.M. Performance of horizontal-flow roughing filter units under varied raw water conditions in Kenya. J. Glob. Resour. 2016, 2, 136–144. [Google Scholar]
  34. Hasnain, G.; Khan, Z. Treatment of high turbidity water through plain and coagulated up-flow roughing filter. Desalin. Water Treat. 2014, 52, 1035–1043. [Google Scholar] [CrossRef]
  35. Karki, B.K.; Amatya, I.M. A comparative study of water turbidity removal efficiency of anthracite and gravel in roughing filter. J. Innov. Eng. Educ. 2020, 3, 42–49. [Google Scholar] [CrossRef]
  36. Cahyana, G.H.; Firdaus, A.R.; Mulyani, T. Effectiveness of combined up-flow roughing filter and up-flow slow sand filter to reduce turbidity in Citarum water. IOP Conf. Ser. Earth Environ. Sci. 2021, 896, 012020. [Google Scholar] [CrossRef]
  37. Mahvi, A.H. Performance of a DHRF system in treatment of highly turbid water. Iran J. Environ. Health Sci. Eng. 2004, 1, 1–4. [Google Scholar]
  38. Huang, X.; Cao, H.; Jia, B. Optimization of Levenberg Marquardt algorithm applied to nonlinear systems. Processes 2023, 11, 1794. [Google Scholar] [CrossRef]
  39. Lee, C. Artificial neural networks and machine learning modeling employee behavior with management towards economic advancement of workers. Sustainability 2024, 16, 9516. [Google Scholar] [CrossRef]
  40. Yadav, R.K.; Sahoo, S.; Yadav, A.K.; Patil, S.A. Green wall system coupled with slow sand filtration for greywater management. NPJ Clean Water 2023, 6, 73. [Google Scholar] [CrossRef]
  41. Goffin, A.; Vasquez-Vergara, L.A.; Guérin-Rechdaoui, S.; Rocher, V.; Varrault, G. Temperature, turbidity, and inner filter effect correction in urban sewage analysis. Environ. Sci. Pollut. Res. 2020, 27, 35712–35723. [Google Scholar] [CrossRef]
  42. Magrì, S.; Ottaviani, E.; Prampolini, E.; Besio, G.; Fabiano, B.; Federici, B. Application of ML techniques to derive sea water turbidity from Sentinel-2 imagery. Remote Sens. Appl. Soc. Environ. 2023, 30, 100951. [Google Scholar]
  43. United States Environmental Protection Agency (EPA). Onsite Non-Potable Water Reuse Resources; United States Environmental Protection Agency (EPA): Washington, DC, USA, 2024.
  44. Parra, L.; Ahmad, A.; Sendra, S.; Lloret, J.; Lorenz, P. Combination of ML and RGB sensors to quantify water turbidity. Chemosensors 2024, 12, 34. [Google Scholar] [CrossRef]
  45. Samayamanthula, D.R.; Sabarathinam, C.; Bhandary, H. Treatment and effective utilization of greywater. Appl. Water Sci. 2019, 9, 90. [Google Scholar] [CrossRef]
  46. Alnajjar, H.Y.H.; Üçüncü, O. ANN-based pollution prediction model in WWTPs. Arab Gulf J. Sci. Res. 2023, 41, 1–15. [Google Scholar]
  47. Asteris, P.G.; Alexakis, D.E.; Tsoukalas, M.Z.; Gamvroula, D.E.; Guney, D. Machine learning approach for estimating five-day biochemical oxygen demand in wastewater. Water 2023, 15, 103. [Google Scholar] [CrossRef]
  48. Dantas, A.D.C.; De Oliveira, M.C.; De Oliveira, L.F.; De Oliveira, J.L.; De Oliveira, R.A. Artificial neural networks for performance prediction in full-scale wastewater treatment plants: A systematic review. Water Sci. Technol. 2023, 88, 1447–1465. [Google Scholar] [CrossRef]
  49. Mathaba, M.; Banza, J. A Comprehensive Review on Artificial Intelligence in Water Treatment for Optimization. J. Environ. Sci. Health A 2023, 58, 1047–1060. [Google Scholar] [CrossRef] [PubMed]
  50. Alprol, A.E.; Mansour, A.T.; Ibrahim, M.E.E.-D.; Ashour, M. Artificial intelligence technologies revolutionizing wastewater treatment: Current trends and future prospective. Water 2024, 16, 314. [Google Scholar] [CrossRef]
Figure 1. Different types of roughing filters, as classified in [2].
Figure 1. Different types of roughing filters, as classified in [2].
Water 17 02319 g001
Figure 2. Topographic map of the study area in Durban South Africa, Durban KwaZulu Natal, South Africa.
Figure 2. Topographic map of the study area in Durban South Africa, Durban KwaZulu Natal, South Africa.
Water 17 02319 g002
Figure 3. Schematic representation of a multi-compartment HRF system, adapted from earlier work [17].
Figure 3. Schematic representation of a multi-compartment HRF system, adapted from earlier work [17].
Water 17 02319 g003
Figure 4. MSE results of the ANN with the first plot corresponding to the LM algorithm and the second to SCG.
Figure 4. MSE results of the ANN with the first plot corresponding to the LM algorithm and the second to SCG.
Water 17 02319 g004
Figure 5. Correlation coefficients for training, testing, validation, and overall datasets with LM algorithm; values are displayed in full precision to accurately reflect ANN performance.
Figure 5. Correlation coefficients for training, testing, validation, and overall datasets with LM algorithm; values are displayed in full precision to accurately reflect ANN performance.
Water 17 02319 g005
Table 1. Design parameters of the HRF equipment used in this study.
Table 1. Design parameters of the HRF equipment used in this study.
Filter ParametersRecommended Literature Values by [2,7]HRF Design Values
Gravel mediaGravel L1 (mm)18–1214 mm
Gravel L2 (mm)12–810 mm
Gravel L3 (mm)8–48 mm
Gravel type (Granite, Quartz, Local)Quartzite
Filter depth (m) 0.2–1.20.3
Total height (m) ≈1.21
Filtration velocity (m/hr) 0.3–1.50.3–0.9
Hydraulic loading rate (m3/m2/day) 7.2–21.6
Filter length (m)L1 (m) 1.5
L2 (m) 1
L3 (m) 0.5
Filter width (m) 1–2.31
Filter material Steel and PVC
Table 2. Characteristics and training specifications of the ANN.
Table 2. Characteristics and training specifications of the ANN.
ANN CharacteristicsSpecification
1. Type of neural networkFeed Forward Back Propagation
2. Number of neurons in the input layer4
3. Number of neurons in the hidden layer(s)4–10
4. Number of neuron(s) in the output layer2
5. Number of input parameters4
6. Number of output parameter2
7. Size of the ANN data cases637
8. Performance functionMSE, R-coefficients
9. Training algorithmsLM and SCG
10. Learning functionslearngdm, learngd
11. Activation function in the hidden layertansig, logsig
12. Activation function in the output layerlinear
13. Maximum number of epochs1000
14. Minimum MSE Value<0.001
Table 3. The statistical summary of the input and output variable ranges for graywater.
Table 3. The statistical summary of the input and output variable ranges for graywater.
InputsOutputs
Filtration Rate (m/h)pH
(-)
Turbidity
(NTU)
Conductivity
(µS/cm)
Clogging Duration (days)Turbidity (NTU)
Mean-9.12176811832
Min0.36.8120301106
Max0.911.431910942595
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Mtsweni, S.; Bakare, B.F.; Rathilal, S. Artificial Neural Network-Based Prediction of Clogging Duration to Support Backwashing Requirement in a Horizontal Roughing Filter: Enhancing Maintenance Efficiency. Water 2025, 17, 2319. https://doi.org/10.3390/w17152319

AMA Style

Mtsweni S, Bakare BF, Rathilal S. Artificial Neural Network-Based Prediction of Clogging Duration to Support Backwashing Requirement in a Horizontal Roughing Filter: Enhancing Maintenance Efficiency. Water. 2025; 17(15):2319. https://doi.org/10.3390/w17152319

Chicago/Turabian Style

Mtsweni, Sphesihle, Babatunde Femi Bakare, and Sudesh Rathilal. 2025. "Artificial Neural Network-Based Prediction of Clogging Duration to Support Backwashing Requirement in a Horizontal Roughing Filter: Enhancing Maintenance Efficiency" Water 17, no. 15: 2319. https://doi.org/10.3390/w17152319

APA Style

Mtsweni, S., Bakare, B. F., & Rathilal, S. (2025). Artificial Neural Network-Based Prediction of Clogging Duration to Support Backwashing Requirement in a Horizontal Roughing Filter: Enhancing Maintenance Efficiency. Water, 17(15), 2319. https://doi.org/10.3390/w17152319

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop