Artificial Neural Network (ANN) for Prediction of Viscosity Reduction of Heavy Crude Oil using Different Organic Solvents

The increase globally fossil fuel consumption as it represents the main source of energy around the world, and the sources of heavy oil more than light, different techniques were used to reduce the viscosity and increase mobility of heavy crude oil. this study focusing on the experimental tests and modeling with Back Feed Forward Artificial Neural Network (BFF-ANN) of the dilution technique to reduce a heavy oil viscosity that was collected from the southIraq oil fields using organic solvents, organic diluents with different weight percentage (5, 10 and 20 wt.% ) of (nheptane, toluene, and a mixture of different ratio toluene / n-Heptane) at constant temperature. Experimentally the higher viscosity reduction was about from 135.6 to 26.33 cP when the mixture of toluene/heptane (75/25 vol. %) was added. The input parameters for the model were solvent type, wt. % of solvent, RPM and shear rate, the results have been demonstrated that the proposed model has superior performance, where the obtained value of R was greater than 0.99 which confirms a good agreement between the correlation and experimental data, the predicate for reduced viscosity and DVR was with accuracy 98.7%, on the other hand, the μ and DVR% factors were closer to unity for the ANN model.


INTRODUCTION
The oil is the main nerve of the energy in the world, oil prices are changed depending on supply and demand, the heavy crude oil (HO) price is half the price of light oil attributed to contains high quantities of sulfur and heavy metals like nickel and vanadium, which is difficulties through production, transportation in the pipeline, (Hasan, et al., 2010). One of the major difficulties with the production of heavy oils is transporting through the pipeline has attributed to high viscosity, particularly when there is no pre-viscosity reduction to ensure oil flow through pipelines, (Alomair and Almusallam, 2013; Ali, et al., 2019a). Crude oil has various compounds, among them, asphaltenes are responsible for high viscosity and are the most polar active fraction of oil that is insoluble in saturated components like n-C5 and n-C7 while in aromatics are soluble such as toluene and benzene, (Aristizábal, et al., 2017;Ali, et al.,2019b). Heteroatoms (O, S, and N) are located inside the asphalt molecule give high polarity which leads to the formation of large aggregates and causes a significant increase in viscosity, (Franco, et al., 2015). Viscosity is an important characteristic in determining the quality of crude oil, so the production of oil from the reservoir or transporting in pipelines has become one of the most important challenges today due economically and environmentally expensive, (Taborda, et al., 2017). Various techniques were used to reduce the viscosity of heavy oil such as heating, emulsion, electric or magnetic field, nanotechnology, and dilution with organic solvents like kerosene, light naphtha and gas condensing, (Al-Adwani and Al-Mulla, 2019;Al-Hashmi, et al., 2017), each of these techniques has benefits and drawbacks. Dilution is effective technique for transport HO in pipelines, it has drawn the attention of most researchers to reduce the viscosity of HO and facilitate flow oil in pipelines due to reduce the cost of pumping and avoid pressure drop ,(Mohammadi, et al., 2019). Another benefit including maintaining the original character of hydrocarbon in contrast to the emulsification method, in addition, it can be applied at any sites regardless of weather conditions, while heating is not effective in cold places, (Luo, et al., 2007), reduce the viscosity of crude oil needs to select the best diluents to get an acceptable viscosity, to ensure the smooth flow of oil transported as considered an influencing factor economically. Recently, some researchers have used intelligent programming to find appropriate relationships between input and output data, a fast and accurate prediction model was designed to predicate the reduction in viscosity of heavy crude oil with various organic solvents to reduce the time in practical experiments, (Daryasafar and Shahbazi, 2017). Advanced technology using intelligent tools has been widely applied for finding the non-linear relationships between inputs and outputs values, like artificial neural networks and adaptive neurofuzzy interference systems, the fuzzy systems were used to model scientific constraints, while Artificial Neural Networks (ANNs) were used as an efficient tool of modeling, predict, solved the problems and identifies errors, the ANN a computing model is widespread and highly flexible and can be used in all application of science and engineering, (Tatar, et  , predicted the density of Athabasca bitumentetradecane mix, at different conditions (temperature, pressure, and weight percentage of diluents) using a radial basis function neural network (RBF-NN) technique, they conclude the proposed model is a suitable model for density forecasting of bitumentetradecane mix. (Eghtedaei, et al.,2017), presented accuracy calculating for viscosity reduction by proposed a radial artificial neural network function(RBF-ANN) for relationship between heavy oil viscosity and the Athabasca bitumen mix, as a function of the temperature, pressure and weight% of tetradecane when comparing the obtained results with previous studies. However, the literature focuses on a prediction of the effect of n-heptane, toluene, and mixture from different volume percentages of toluene/n-heptane as dilutes solvents on the viscosity reduction of heavy oil using intelligent model Back Feed Forward Artificial Neural Network (BFF-ANN) looks to be rare. As a consequence, in this study, the results obtained from practical laboratory experiments were used as input data to build a model of an artificial neural network, then proposed model was used to examine a section of laboratory results to know its accuracy for future approval in obtaining direct results without referring to the laboratory, the input parameters to ANN are solvent types, solvent addition, RPM, and shear rate, while the output parameters are viscosity and DVR.

EXPERIMENTAL 2.1 Materials and Experimental Methods
The crude oil from south of Iraq (Amara oil field) was used in the present study; the density, API, and viscosity of this oil are 0.979 g/cm 3 , 16, and 135.6 cP respectively at 298.15 K. The solvents toluene and n-Heptane (purity = 99%), the diluents solvents n-Heptane and toluene were used as received, and the prepared mixture from different volume percentages of toluene / n-Heptane (25/75, 50/50 and 75/25) were also used as diluents in this study. The mixtures of heavy crude oil with different weight concentration of solvents (5, 10 and 20 wt. %) were done in a closed cylindershaped beaker size 125 ml, with working volume 100 ml under continues mixing using a magnetic stirrer for an hour to approve a homogenous blend, (Mortazavi-Manesh and Shaw, 2016). The viscosities of samples were measured at temperature 298.15 K and shear rate range from 2 to 42 s -1 using Brookfield viscometer model DV-11.

Experimental Data Analysis using ANN
Artificial neural networks are a system inspired by the biological nervous system, in which problems or information are treated as inputs, the brain processes this information, and the problem is solved. The basis of the functioning of the biological network depends on the modifications that occur in the neural connections and thus reaching the goal, this idea was adopted in the work of artificial neural networks, the artificial neural network solves the complex problems that the brain sometimes fails to recognize, the artificial neural network consists of three layers, the input layer for input data, the output layer that represents the output data and a hidden layer through which the network is learned. In each layer there is a group of neurons that are associated with two factors, w (weights) and b (biases), whose value is random at the beginning and their values are continuously updated in each attempt until the desired result is achieved which represents the goal. Determining the number of neurons in each layer and choosing the appropriate transfer function is very important in determining the speed and accuracy of the obtained results, (Usman and Ademola, 2013). The feed-forward back propagation net is using a back propagation training algorithm. In this type of network, the results are calculated, the error rate is calculated and compared with the permitted error rate, and this process is repeated a number of times until the goal is reached.

= ∑ + =1
(2) In this study, the ANN model was accomplished using MATLAB software to analyze and predict viscosity reduction of heavy crude oil using different solvent from the experimental data, the data used in this study consists of (177) sets of (Solvent type, wt. % of Solvent, RPM, and Shear rate) as inputs, and viscosity and DVR as outputs as shown in Appendix (A). The training set is used for neural network training, depending on the number of attempts (epochs) through which the error rate is improved. The validation set is used to calculate network generalization. The testing group is used to measure the strength of network improvement in predicting results for previously untrained values. Thus give an independent measure of network performance during and after training. In the current work, the data were divided into three sets, namely, training set (80%), validation set (10%) and testing set (10%).

Experimental Results
Reductions of heavy crude oil viscosity using different lighter hydrocarbons were considered the first and a preferred methodology, (Martínez, et al., 2011). It was noted from the obtained results as shown in Figs.2 and 3, the little viscosity decline for pure crude oil has occurred from 135.6 to 111.5 cP during the shearing rate, whereas, the addition of different weight fractions of solvent decrease the viscosity of heavy oil, the higher reduction appeared at 20 wt. % of solvent. Moreover, the viscosity reduction with additions of the solvent mixture from toluene and n-heptane were greater than toluene and n-heptane solvent alone and the higher viscosity reduction was about from 135.6 to 26.33 cP when the mixture of toluene/heptane (75/25 vol. %) was added. These results as an aspect of the aromatic distinctive of the toluene which consents to an interfere in asphaltene aggregation attitude by aggravating asphaltene self-congregation, (Tao, 2006), self-aggregation as a result of heteroatoms contained in asphaltene structures that firstly lead to colloidal aggregates then permit the growth of aggregates and subsequently increasing the viscosity of crude oil,   The results for degree of viscosity reduction (DVR %) as a function of shear for the crude oil blended with different weight fractions of solvents were calculated by Eq. (4), (Quan, et al., 2019), as shown in a Figs.4 to 8.
(4) and are the viscosity of crude oil before and the after dilution.
The DVR% was enhanced with increasing the solvent fractions, the results display the maximum change in viscosity were happens at 20 wt. % for all tests, as was noted the uppermost values were achieved with additions of the solvent mixture from toluene and n-heptane, moreover it increased with increasing the toluene percentage, and the higher value was about (80.58) when the mixture of toluene/heptane (75/25 vol. %) was added, these results recognized to the structure formation of crude oil within solvent under a shear rate, (Ghannam, et al.,2012).

ANN Model Results
ANN's proposed model code was written using Matlab2018a software. The ANN topology was proposed (5: 25: 2). The input layer consists of five neurons, their names are Solvent1, Solvent2, Solvent add (wt.%), RPM, Shear Rate (1/s), 25 hidden neurons, two-output neurons are viscosity and DVR%. Fig. 9 shows the relationship between the different layers of the developed ANN. Figure 9. Design of artificial neural network.
As it was known, the ideal MSE is 0, and the R is 1. So we need to minimize the MSE and maximize R as we could, the Mean Square Error (MSE) is given in Eq. (5), while the R is calculated by Eq. (6), (Fahriye, et al., 2017). These equations are used to indicate the success of the algorithms.
In this study, throughout the training, the best model that has MSE and coefficient of correlation equals 1.0445 and 0.99949 respectively. When it was testing, provide MSE and R about 1.6697e-1 and 0.9997 respectively. Figs. 10 to 14 show the relevant results during the repetition for training and data testing, the solid 45 ° reveals that there is an ideal overlap between the predicted and experimental data. These

Figure14.
Relationship between experimental, predicted viscosities and R values for the testing data.