Artificial neural network and response surface methodology for modeling oil content in produced water from an Iraqi oil field

Produced water (PW) is salty water that has been trapped in geological formations and brought to the surface during the development and production of oil and gas. PW usually has a high concentration of pollutants, such as metals and hydrocarbons, thus it needs to be properly managed and treated before being disposed of. Thus, the majority of studies have focused on treating oily water to meet environmental regulations as well as to reuse and recycle produced water. The employment of more advanced treatment techniques is required by some major environmental rules as well as financial limitations (Wei et al. 2019). High concentrations of oil and grease (O&G), total dissolved solids (TDS), suspended solids (SS), dissolved and volatile organic compounds, dispersed oil, radionuclides, heavy metals, bacteria, and dissolved gases are among the contaminants of concern found in the produced water (Konvensional 2017). The process of treating produced water from an oilfield typically consists of multiple stages: pretreatment, which eliminates larger particles and gas along with the bulk oil; main treatment, which eliminates small droplets and particles; polishing treatment, which eliminates ultra-small particles; and optional tertiary treatment, which eliminates dissolved matters and gases (Konvensional 2017). Moreover, BTEX chemicals (benzene, toluene, ethylbenzene, and xylene) and other harmful and toxic air pollutants are mostly produced by petroleum refineries. Additionally, they are a significant source of the following criterion air pollutants: sulfur dioxide (SO₂), carbon monoxide (CO), hydrogen sulfide (H₂S), nitrogen oxides (NO_x), and particulate matter (PM). Less hazardous hydrocarbons are also released by refineries, including light volatile fuels and oils and natural gas (methane) (Jabbar et al. 2019; Alardhi et al. 2023a).

Since oily wastewater contains hazardous substances that inhibit the growth of organisms, treating it is essential (Baig et al. 2022). Many treatments, such as chemical destabilization (coagulation, flocculation) (Zhao et al. 2021), membrane filtering techniques (Barambu et al. 2021; Samuel et al. 2022), adsorption in nanoparticles (Alardhi et al. 2023b), biological methods (Jabbar et al. 2023), and electrochemical processes (electrocoagulation) (Merma et al. 2020), can be used to treat these wastewaters. A total of three parts compose the conventional oil–water separation process: the oil content control and monitoring unit, the filtration unit, and the separating unit. The treatment units include the separator and filter units, which employ a range of designs and theories. Typically, the treatment process begins with the use of gravity and centrifugal separators. Other separating methods, referred to as polishing treatment, are then applied. The flotation, coagulation and flocculation, filtration, biological treatment, absorption, and adsorption processes are examples of the polishing unit (E.P.A.W. DC 2011). Standard procedures such as gravity and centrifugation are applied to oily wastewater that is divided into two phases; in the meantime, the emulsified oily bilge water needs to be separated using chemical or biological de-emulsification (Karakulski et al. 1995). Usually, the first step in treating oily water on board is to use gravity to separate the heavier and lighter portions according to density differences. This technique uses oleophilic polymer coalescing materials in the form of parallel plates or loose-packed media to draw oil droplets to the plate (E.P.A.W. DC 2011). Coalescing plate separators made of oleophilic polymers include nylon, fiberglass, and polyethylene (Amran & Mustapha 2021).

The oil droplets must first drift freely and disperse, separate from the coalescing material, and rise to the surface of the tank before they can no longer stick to the plate or media. After that, when the sensors identify the presence of oil, the oil–water separation process is automatically started, and the collected oil is taken out of a waste oil tank. However, this approach is only effective when the phases of water and oil are separated (Koss 1996). In other words, the gravitational technique is incorrect in some instances where the bilge water is emulsified oil produced by chemical emulsifiers (cleaning and agents solvents) as well as mechanical systems like the ship's motion and transfer system pump (E.P.A.W. DC 2011).

Wastewater treatment is currently optimized and modeled using both artificial neural network (ANN) techniques and response surface methodology (RSM). The response surface technique is used to determine the relationship between experimental factors and responses (Zin et al. 2020).

When the wastewater treatment procedure is carried out under substandard conditions, more materials may be used, more time may be spent, and the results may be less favourable. As a result, it is crucial to model and optimize operational parameters on processes. ANN and RSM techniques are commonly used in the optimization and modelling approach. RSM and ANN models were improved for these issues because of their accuracy in making predictions about complex nonlinear systems.

ANNs, a type of machine learning technique that is highly predictive, have been used in the petroleum and energy sectors (Sheikhoushaghi et al. 2022). The feed-forward back-propagation algorithm-trained multilayer perceptron (MLP) architecture is the most widely used artificial neural network (ANN) model. The MLP design consists of at least three layers of processing units coupled by weighted connections. The input vector is found in the first layer, while the output vector is found in the last layer. Neural networks are represented by the intermediate layers, which are also called hidden layers, which modify the input data by creating several weighted connections (Elmabrouk et al. 2014). One of the many reasons ANNs are preferable to mathematical models is their ability to recognize various types of interdependence in biological systems without requiring a previous understanding of the kinetic and metabolic processes of the systems (Heydari et al. 2021). The use of ANNs and other models to enhance water treatment has been the subject of numerous studies, as Table 1 illustrates.

In order to improve the production of oil content removal efficiency, this work attempts to employ RSM and ANNs to simulate the oily water treatment process. RSM-based central composite design with four parameters developed in (DESIG EXPERT software version 7) and an ANN based on the feed-forward neural network (FFNN) principle is created using MATLAB software. Two back-propagation algorithms were utilized to train the feed-forward network utilizing an array of training data [96 data sets] collected over a month from a real operating plant. It is suggested that plant engineers can use this ANN model as a platform to estimate how plants would behave when handling oily water, giving them important information for implementing appropriate control strategies and unsuitable operating conditions.

Previous research studied the oily water treatment system using mathematical models and relied on laboratory tests. This study, however, is based on operational conditions and realistic data from an Iraqi oil field. It presents the impact of temperature, pressure, flow rate, and oil content, all of which were modeled and optimized through the use of two different approaches, ANN and RSM on the overall performance of oil content reduction. Next, the best model is chosen by comparing the anticipated outcomes of these models with the observed values.

The aim of this study is to predict the oil content of produced water from an Iraqi oil field depending on the dataset that was obtained from an operational plant in Iraq in June 2023. An artificial neural network ANN technique was employed in this study to improve the central composite design model for oil content in wastewater. The dataset included a variety of parameters, such as the temperature, pressure, unit flow rate, and oily water concentration (ppm) of the feed water. Despite the significant differences in parameters, the ANN model was remarkably accurate at simulating the process under investigation.

There are four initial sections to this study. An outline of the main issue the study addresses and a list of the most recent solutions are provided in the first section, the introduction. The second section explains the data collection procedure and gives a thorough explanation of the RSM and ANN algorithms used. Results and explanations for both models are presented in the third section, along with a comparison of them. The main conclusions and impacts are finally summed up in the conclusion.

There are 10 kinds of chemicals required to inject in the oily water treatment system. The details of chemical injection are given in Table 2.

The oily water treatment system mainly includes the coalescing oil separators, parallel plate skimmer vessel, gas stripping column, filters, the back-flushing wastewater tank and pumps, etc. Two trains with a capacity of 10,000 m³/day for each one will include the equipment mentioned above. The oily water treatment plant will receive the produced water from de-salters and oil storage tanks (water draw-off, oily water from closed drain, and new CPF). The water quality could be complex and could be dissolved hydrogen sulphide. In order to maximize the performance of the dissolved hydrogen sulphide removal, gas stripping columns are designed. Oily water comes from inlet water pumps will go to the coalescing oil separators and parallel plate skimmer vessel to remove the oil and SS, to the gas stripping column to remove H₂S, to the filters to remove oil and SS, to meet the standard specification for treated water, and then it will be sent to the water injection station.

2 × 350 m³/h coalescing oil separators (for one train) will receive water from the 2 × 3,000 m³ control-storage tanks by 3 × 250 m³/h inlet water pumps (for one train) and reduce the oil concentration to below 80 mg/L under normal operating conditions, rendering it suitable for further treatment in the downstream. It will also remove a proportion of suspended solids, and reduce it to below 80 mg/L under normal operating conditions. This outlet water quality should be suitable for further treatment in the downstream parallel plate skimmer vessel.

2 × 350 m³/h capacity parallel plate skimmer vessel (for one train) will receive water from the coalescing oil separators and reduce the oil concentration to below 30 mg/L under normal operating conditions, rendering it suitable for further treatment in the downstream. It will also remove a proportion of suspended solids, and reduce it to below 20 mg/L under normal operating conditions. This outlet water quality should be suitable for further treatment in the downstream gas stripping column.

2 × 250 m³/h capacity gas stripping column (for one train) will receive water from the parallel plate skimmer vessel and reduce the dissolved hydrogen sulphide (H₂S) to 30 mg/L under normal operating conditions, rendering it suitable for further treatment in the downstream nutshell filters.

The treated water from the gas stripping column will be pumped into the downstream system. In order to maximize the performance of the oily water treatment system, the oil droplets should be as large as possible. This means that the pumps must be designed to minimize the shearing of the droplets.

3 × 180 m³/h capacity units will be installed for the base case flow of 10,000 m³/day, allowing one vessel to be withdrawn from service for backwashing, and maintenance. AII will be online under normal operation. Treated water from the water injection station will be used to backwash the filters.

6 × 90 m³/h capacity units will be installed for the base case flow of 10,000 m³/day, allowing one vessel to be withdrawn from service for backwashing, and maintenance. Before the water injection station is built, treated water will be sent to the existing evaporation water pool.

RSM is considered an effective tool to optimize effective parameters in wastewater refinery oil. RSM based on central composite design gives an accurate set of experiments for investigation and data obtained for 30 runs were generalized and analyzed by Design-Expert software version 7.

The best prediction accuracy with the appropriate correlation coefficient (R²) of the ANN model with the mean squared error (MSE) would be employed. The oil content of produced water is predicted using an ANN model that was developed. The predicted data from the ANN model will be compared with the industry data (experimental data). Two training techniques in the MATLAB R2021a program will be used to train the ANN. Several performance metrics will be taken into consideration to select the optimal training algorithm. The oil field in south Iraq will supply the training data set for the ANN model.

To develop the characterization model for the oily water treatment facility, we used the database of an Iraqi oil field. Four essential process parameters for this investigation were obtained for a month from a local field in Iraq: the oil content (ppm), temperature (°C), pressure (Mpa), and flow rate (m³/day) of the input oily water, furthermore operational parameters ranges as follow: 840.02–991.9 ppm, 46.4–68.2 °C, 18.1–20.55 Mpa, and 27,000–29,390 m³/day.

This system's raw dataset is made up of 96 data points that were sampled in June 2023. The oil content of treated water is taken into consideration as an output parameter. Table 3 illustrates the units of measurement together with the maximum and minimum values for each parameter. Oily water up to 20,000 (m³/day) can be processed by the plant.

The upper and lower levels of each effective parameter were pre-coded according to initial data from refinery oil as follows: oil content (840–991 ppm), temperature (46.4–68.2 °C), and pressure (18–21 Mpa) and flow rate (27,000–29,390 m³/day).

The experimental design was carried out with six replications of the center point and one replication of the factorial and axial points. Each factor was changed in five levels. The data were constructed and randomized using the Design-Expert software 7, as explained in Table 3.

Samples of wastewater were taken to evaluate the oil content of each sample, and given data were then analyzed via the Design-Expert software 7. Experiment matrix and regression data analysis were designed by Design-Expert software 7. The model was statistically analyzed with the analysis of variance (ANOVA) (Zin et al. 2020).

The ANN model simulates the functions of the human brain using mathematical calculations (Alardhi et al. 2024). The architecture format of the ANN model is unique and draws inspiration from the structure of the biological nervous system. The complex and nonlinear organization of neurons in ANN models is similar to that of the human brain (Malekian & Chitsaz 2021). Three groups include the ANN applications that address real-world issues: (i) pattern categorization, (ii) prediction, and (iii) control and optimization (Malekian & Chitsaz 2021). An ANN typically consists of one input layer, one or more hidden layers, and one output layer. Several nodes, or neurons in biological terms, are present in each layer of the neural network architecture and link to the layer after it. Throughout the training process, the weights assigned to the nodes can be changed. To adjust the strength of the signal from the linked neuron, the weights are increased or reduced (Fiyadh et al. 2023).

One of the most important performance factors in determining if the ANN model can be considered usable is its MSE number, which gives information about the model's accuracy. It displays the variance between the experimental data and the expected values. The fit's accuracy rises with a decreasing MSE value. To assess and indicate how closely the data fit the regression line, the value of R² is utilized. There is less of a variance between the fitted values and the observed data when the R-value is nearer 1. The range of the R² value is 0 to 1. An acceptable R² value is defined as more than 0.8.

An RSM depending on CCD was developed to evaluate the feasible combinations of four studied parameters (oil content ppm, temperature C, pressure Mpa and flow rate m³/day). The lower and upper bound of parameters are identified initially conditions from oil refinery. Optimization of selected effective parameters is considered crucial parameters for reducing oil content in wastewater. RSM is a strategy of experimentation in which designed parameters are changed together instead of a single parameter at a time. The design based on CCD is satisfactory to apply in this set of experiments because each parameter varies in five levels desired to study a wide range of values for each parameter as shown in Table 7.

The value of alpha was selected based on the hypothesis of rotatability to ensure uniform variance at points that are equal distance from the origin (center point) and hence provide equal precision of response evaluation in any design direction (Montgomery 2017). The center point was run with six replications to get a reasonable estimation of an experimental error (pure error). The center point values are utilized to reveal curvature in the response (they play a part in estimating the coefficient of quadratic terms). The axial point is also used to calculate the coefficient of quadratic terms, while factorial points are used mainly to estimate the coefficients of linear terms and two-way interaction (Montgomery 2017).

The ANOVA for CCD is shown in Table 4. The goodness of the linear model was revealed to be highly significant, with P > F equal to 0.0001. Based on the ANOVA Table 4, all terms showed an insignificant effect on the response except flowrate shows highly significant, where the p value is below 0.05 (p < 0.05). D (flow rate) revealed the most significant factor affecting oil content with the highest p value of 0.0001.

Model quality was determined via correlation coefficient (R²), equal to 0.747, R adjusted was 0.706, and R predicted 0.643. It is considered a good result because the difference between R adjusted and R predicted is less than 0.2, which increases the model adequacy. Furthermore, the adequate precision utilized to identify signal-to-noise is thought to be recommendable more than 4. Here the value 17.1 showed that the empirical model is of adequate signal and can be used to navigate the design space (Breig & Luti 2021).

From the determined regression coefficient in Table 5, the resulting matrix from CCD was fitted and can be described in first-order equation (correlation) (Equation (4)) where A is the oil content, B is the temperature, C is the pressure, and D is the flowrate). Correlation is explained by the ‘correlation coefficient, a dimensionless signifier, always between −1 and 1. However, such a simple evaluation carries uncertainty of the correlation coefficient being influenced by a distance from the center point, which can have an enormous impact. Furthermore, the relationship may not be linear but parabolic; therefore, values are interpreted with care (Breig & Luti 2021).

Based on the optimization result, the optimum parameters that minimize oil content are an initial oil content of 991 ppm, a temperature of 46.4 °C, and pressure 21 Mpa, and a flow rate of 27,000 m³/day. A validation experiment was performed using optimized parameters to verify the predicted result. The experiment revealed an average oil content of 8.5 ± 0.7 ppm, near the expected (assumed) value of 8.55 ppm (Fard et al. 2021; Thangavelu & Zou 2022).

The experimental final oil content values obtained by the different CCD combinations during the 30 runs were described in Table 7, together with the anticipated CCD values and their residuals. Eight axial points, 16 factorial points, and 6 replications of the center points are included in the experiment matrix that was created using CCD in order to determine the pure error. All RSM inputs and outputs were used in the training, testing, and validation of the ANN system. Every experiment was conducted twice, and samples were gathered over the course of the same hour.

The ability of RSM-CCD and the created ANN model to forecast a decrease in oil content was assessed. Table 8 displays the calculated statistical parameters that compare and evaluate the degrees of correctness of the two models. The model's lowest MSE and highest R² scores indicate how accurate it is. Since R² evaluates the connection between the response under study and the operating parameters, higher findings (up to 1) suggest a strong correlation between the two datasets. Regression analysis usually employs MSE to validate experimental results since a lower value indicates that the data are concentrated around the line of best-fit (prediction-errors). In light of the aforementioned data, both models demonstrated high predictive ability. Nonetheless, an analysis of the two models reveals that the ANN model outperforms the RSM-CCD model for R² and another goodness-of-fit metric. Consequently, compared to the CCD model, the ANN model predicts a higher reduction in oil content through the west water process.

The present result is consistent with that put forth by Ezemagu et al. (2021), which found that the ANN model performed better than the RSM.

In a second overall comparison, linear regression analysis was used to evaluate the actual and predicted values of the ANN and CCD models (Figure 6). Once more, Figure 14 of ANN shows considerable fitting with a higher R² value, suggesting that ANN produces better optimization results than CCD. Consequently, when compared to the CCD model, the ANN model's capacity for generalization is significantly higher.

Nonetheless, there are a few advantages to CCD modeling. For example, by emphasizing each factor's contribution to regression models, the CCD's structure can help remove unnecessary components from the model. Additionally, because ANN modeling involves multiple iterative calculations, it requires more computation time than CCD. However, the ANN-generated model showed better prediction accuracy than the CCD model did. This can be explained by its comprehensive ability to replicate the nonlinearity of the system; CCD's restriction to the second-order polynomial requires only one computation step for a response surface model. However, ANN consistently yields superior results than CCD in every aspect.

In order to maximize the performance of the system, environmental engineering factors are crucial for optimization. The majority of wastewater treatment systems involve several variables, making conventional approach optimization rigid, unreliable, and time-consuming. Therefore, it is preferred to use a different approach that will be more efficient and adaptable for parameter optimization of different wastewater treatment procedures. There are many studies that dealt with mathematical models with different methods of water treatment, but regarding water produced from oil fields, we did not find many. Therefore, this study includes some studies that dealt with the subject of water treatment in general with the use of different mathematical models:

A Box–Behnken design was used by Sanchez-Sanchez et al. (2024) for the analysis using a gray wolf optimizer (GWO)-coupled ANN model and response surface methodology (RSM) to analyze the effect of three operating parameters (volumetric exchange ratio [VER], aeration rate [AR], and cycle time [CT]) manipulated during an aerobic granular sludge process (AGS) sequencing batch reactor on modeling the removal of chemical oxygen demand (COD) in mixed wastewater. The most efficient architecture for COD showed the highest efficiency for modeling the AGS. The RSM model and plot results indicate that the CT and AR were the most influential on COD removal efficiency. When compared with models with statistical indices, GWO-ANN demonstrated higher performance compared to RSM.

The effectiveness of the electrocoagulation approach, which uses aluminum electrodes in a batch bi-polar system, was tested by Moneer et al. (2023) in order to ascertain how well oil was removed from oily wastewater. The impacts of four independent factors: oil volume (X₁), temperature (X₂), initial pH (X₃), and treatment duration (X₄) were investigated to look into turbidity recovery and conductivity changes. The Box–Behnken design was used to improve conditions. The findings showed that reducing oil volume and lengthening treatment times greatly improve turbidity removal and conductivity changes. When the treatment method is used under ideal operating conditions, it can remove turbidity and conductivity with removal efficiencies of 97.3 and 73.4%, respectively.

In 2020, Najafzadeh & Oliveto (2022) looked into a novel approach in which machine learning (ML) models are used to generate equations for the estimation of the rate at which currents will cause scouring to spread along pipes. The primary dimensionless parameters from the credible literature were identified as the following: the current angle of attack on the pipeline, the Shields parameter, the approaching flow Froude number, and the ratio of embedment depth to pipeline diameter. Various optimization techniques and setting parameters derived from evolutionary and classification components were used to construct machine learning models. Additionally, the explicit equations produced by machine learning models were utilized to show how the suggested methods agree with experimental findings.

Saberi-Movahed et al. (2020) reported the development of a novel GMDH method called the GMDH network by using an extreme learning machine (GMDH-ELM). This method eliminates the need to update the weighting coefficients of the quadratic polynomials used in conventional GMDH during the training stage through the use of other evolutionary algorithms or backpropagation techniques. In actuality, a relationship between the input and output in each neuron of the GMDH model was established using an intermediate parameter. From a different perspective, this study demonstrated for the first time that it is possible to enhance the procedure and obtain comparably higher levels of accuracy by successfully integrating the idea of ELM into the basic structure of GMDH.

In a piping modeling study by Najafzadeh (2019), the conjugate depths of the hydraulic jump in the circular pipes were assessed using three numerical models based on evolutionary computing: gene-expression programming (GEP), model tree (MT), and evolutionary polynomial regression (EPR). To find a functional relationship between the input and output variables, ideas of particular force produced two non-dimensional factors. The results achieved with conventional methods were compared with the performances of the proposed ways. When MT was compared to those other artificial intelligence (AI) models and empirical equations, its performance showed an accurate prediction of conjugate depths (R² = 0.995 and RMSE = 0.023). Instead of using polynomial quadratic neurons in each GMDH neuron, GEP was carried out in each, as demonstrated by another work by the same last author (Najafzadeh & Saberi-Movahed 2019) The size of the sediment, the shape of the pipeline, and the features of the waves upstream of the pipeline are effective parameters on the three-dimensional scour rates. The dimensional analysis technique considers four-dimensionless parameters as input variables. In addition, scour rates in the pipeline's left and right segments as well as vertical scour rates are established as output parameters. Several statistical indices are used to assess the training and testing stage results of the proposed GMDH-GEP models. The GMDH-GEP models' performances are contrasted with those of the GEP, GMDH, ANN, and conventional equation-based regression models.

Rastegar et al. (2011) used RSM to forecast influent COD, up-flow velocity (V_up), and HRT behaviors in the bioreactor. Reduced quadratic and cubic models were shown to be the most effective models for COD removal and biogas generation rate, respectively, according to RSM. Based on two crucial reactions, the ideal region was determined to have an influent COD of 630 mg/L, a V_up of 0.27 m/h, and an HRT of 21.4 h. As a consequence, the COD removal efficiency was 76.3%, and the biogas production rate was 0.25 L/L feed.

Tir & Mostefa (2008) conducted an initial experimental investigation (Tir & Moulai-Mostefa 2008) in order to determine the most accurate operating parameters, which are subsequently utilized to calculate the efficiency of oil removal. Turbidity and COD measurements were used to estimate oil separation in an experimental design that employed the response surface approach. It was possible to identify an ideal area with low COD and turbidity levels. The primary influences of the operational factors were also looked into as part of the improved procedure. The experimental findings demonstrated the high efficiency of electrocoagulation, which achieved 99% turbidity, 90% COD in less than 22 min, and a current density of 25 mA cm(⁻²). The variance coefficient (R²) value was high in the ANOVA results.

It is still challenging for the petroleum industry to reduce the environmental impact of produced water from oil fields and wastewater from petrochemical processes. RSM and ANN were used to simulate the oily water treatment procedure in an Iraqi oil field case study. Actual industrial parameters were entered into RSM/CCD to minimize oil content and used to validate the model, and the findings demonstrated the model's high accuracy. The ANN model was created by implementing the optimal design of the neural network with one hidden layer for the provided architecture, which is the 4–10–1 architecture. The testing results showed an R² of more than 0.999, and the training set as a whole had a low MSE of less than 1.55 × 10⁻⁷, demonstrating the effectiveness of ANN models in predicting the oil content of treated water. The results are highly accurate based on the several indices, as well as the anticipated and actual outcomes, used in this study to test the reliability of the ANN models. R², RE, and MSE are these markers which are equal to 0.994, 2.5, and 1.55, respectively. The created ANN model could be integrated with various optimization techniques to accelerate learning and prediction.

All relevant data are included in the paper or its Supplementary Information.

The authors declare there is no conflict.