Prediction of asphaltene precipitation in crude oil

Prediction of asphaltene precipitation in crude oil

Journal of Petroleum Science and Engineering 68 (2009) 218–222 Contents lists available at ScienceDirect Journal of Petroleum Science and Engineerin...

1010KB Sizes 0 Downloads 19 Views

Journal of Petroleum Science and Engineering 68 (2009) 218–222

Contents lists available at ScienceDirect

Journal of Petroleum Science and Engineering j o u r n a l h o m e p a g e : w w w. e l s ev i e r. c o m / l o c a t e / p e t r o l

Prediction of asphaltene precipitation in crude oil G. Zahedi a,⁎, A.R. Fazlali b, S.M. Hosseini a, G.R. Pazuki c, L. Sheikhattar a a b c

Simulation and Artificial Intelligence Research Center, Department of Chemical Engineering, Faculty of Engineering, Razi University, Kermanshah, Iran Department of Chemical Engineering, Faculty of Engineering, Arak University, Arak, Iran Department of Chemical Engineering, Faculty of Engineering, Malek Ashtar University of Technology, Tehran, Iran

a r t i c l e

i n f o

Article history: Received 29 December 2007 Accepted 29 June 2009 Keywords: asphaltene artificial neural networks Flory–Huggins model crude oil

a b s t r a c t Asphaltene are problematic substances for heavy-oil upgrading processes. Deposition of complex and heavy organic compounds, which exist in petroleum crude oil, can cause a lot of problems. In this work an Artificial Neural Networks (ANN) approach for estimation of asphaltene precipitation has been proposed. Among this training the back-propagation learning algorithm with different training methods were used. The most suitable algorithm with appropriate number of neurons in the hidden layer which provides the minimum error is found to be the Levenberg–Marquardt (LM) algorithm. ANN's results showed the best estimation performance for the prediction of the asphaltene precipitation. The required data were collected and after pre-treating was used for training of ANN. The performance of the best obtained network was checked by its generalization ability in predicting 1/3 of the unseen data. Excellent predictions with maximum Mean Square Error (MSE) of 0.2787 were observed. The results show ANN capability to predict the measured data. ANN model performance is also compared with the Flory–Huggins and the modified Flory–Huggins thermo dynamical models. The comparison confirms the superiority of the ANN model. © 2009 Elsevier B.V. All rights reserved.

1. Introduction Complex and heavy organic compounds, which are called asphaltene, cause many problems in crude oil production, and residual processing. Deposition of asphaltene, in crude oil, can make a number of problems. Solid precipitation around the separators, pumps, tanks and other equipment especially the well pipe in many conditions may threaten the economic oil recovery or considerably increase production costs. Different side effects for asphaltene deposition have been known such as: (CO2, rich gas), ph shift, mixing of crude oil streams, incompatible organic chemicals, stimulation, pressure drop, streaming potential charge and bare metal surfaces (Hirschberg et al., 1984; Pazuki and Nikookar, 2006; Fazlali et al., 2006, 2007). In order to predict the asphaltene precipitation an exact model is necessary to predict the amount and the conditions of precipitation (Hirschberg et al., 1984; Fazlali, 1999; Fazlali et al., 2006; Pazuki and Nikookar, 2006; Fazlali et al., 2007; Söze et al., 2004a,b). The acceptation of a method approach, obtained exclusively from the experimental data, can provide other practical methods for modeling. These models provide a dynamic relationship between input and output variables and bypass underlying complexity inside the system. Statistical models based on regression analysis are an example of such black box modeling. Most of these common approaches rely on linear system identification models. The major ⁎ Corresponding author. Tel/Fax: +98 831 4274542. E-mail address: [email protected] (G. Zahedi). 0920-4105/$ – see front matter © 2009 Elsevier B.V. All rights reserved. doi:10.1016/j.petrol.2009.06.023

processes found in chemical engineering are unfortunately nonlinear processes, and previously mentioned approaches fail to respond correctly because of process nonlinearity. Recently, another promising alternative modeling technique, ANN, has undergone numerous applications in chemical engineering (Zahedi et al., 2005a,b; Zahedi et al., 2006; Vallés, 2006). It should be mentioned that neural networks had been widely applied on different technology disciplines with successful results. The ability to learn the behavior of the data generated by a system gives neural networks its versatility (Vallés, 2006). In the remaining part of current study after brief description of ANN, experimental routines will be presented. In another part attempts to build the best ANN predictor will be described. Finally the results of ANN will be compared with two common thermodynamical models. 2. Artificial neural network (ANN) In order to find a relationship between the input and output data driven from accelerated experimentations, a powerful method than traditional modeling is necessary. ANN is an especially efficient algorithm to approximate any function with finite number of discontinuities by learning the relationships between the input and the output vectors (Vallés, 2006; Hagan et al., 1996). Thus, ANN techniques are especially useful for modeling highly nonlinear systems very well. Artificial neural networks are a biological inspiration based on various characteristics of the brain functionality. Artificial neurons are simple computational devices that are highly

G. Zahedi et al. / Journal of Petroleum Science and Engineering 68 (2009) 218–222

interconnected and the connections between neurons determine the transfer function of the network. An artificial neural network determines an empirical relationship between the inputs and the outputs of a given system. Where the inputs of the system are the independent variables and the outputs are the dependent variables. Therefore, it is important for the user to have a good understanding of the science behind the underlying system to provide the appropriate input and, consequently, to support the identified relationship. A network is composed of units or nodes, which represent the neuron body. The units are interconnected by links that act like axons and dendrites of their biological counterparts. A typical interconnected neural network is shown in Fig. 1. (Zahedi et al., 2005a,b, 2006; Vallés, 2006). In the above figure an input layer, a central or hidden layer and an output layer can be seen. In a network each connecting line has an associated weight. Two important abilities of the neural network (NN) are supplying fast answers to a problem and the capability of generalizing answers, providing acceptable results for unknown samples. In this way, they should learn about the problem under study and this learning is commonly named training process. Training usually begins with random value for the weight of NN. Then, NNs are supplied with a set of samples belonging to a problem domain to modify the values of their weights. There are various learning algorithms to train neural networks. Some of the famous algorithms are: Perceptron, Hebbian, Widrow–Hoff, and Back propagation. The latter is a suitable algorithm for this research because it has the capability of training multilayer neural networks for function approximation. The key idea of the variable learning rate is to modify the weights and biases update by changing the momentum and the learning rate, based on the behavior of the squared errors. One of the well-known topologies of neural networks for learning is the MultiLayer Perceptron (MLP), which is used for the classification and the estimation problems (Zahedi et al., 2005a,b; 2006). An MLP is an NN with three layers, an input layer, a hidden layer and an output layer. The input layer represents the incoming pattern and the output layer is the output of the network. Each layer consists of a series of nodes, interconnected with weights. During the learning cycle, the MLP is presented with an input pattern on the input nodes and a target pattern on the output layer. The weights are then updated so that the network gives the desired output. Each node contains an activation function, which is a function that decides whether the neuron should fire depending on its inputs. After training (when the network is put to use), the values of the weights and the activation functions decide which nodes fire. These activation functions come in many different forms, the classics being threshold, sigmoid Gaussian and etc. (Lang, 2006). For more details of the various activation functions see Bulsari (Bulsari, 1995).

Fig. 1. A typical MLP neural network.


3. Experimental data 3.1. Materials and methods To build an ANN for predicting the asphaltene precipitation, the experimentation in the lab was done (Fig. 2). In selecting the data for modeling, and to ensure that they represent normal operating ranges, off data were deleted from the data list. The data which are not in the normal trend of the process and are not rational are considered as off data and are deleted. The data sets were collected from the six samples of crude oils. Two data sets of collection were made ready from Buenrostro-Gonzalez et al. (2004) Rassmdana et al. (1996) which they reported the experimental data of the asphaltene precipitation and the other data sets (4 sets) were prepared from the different reservoirs of Iranian crude oil from the EOR research center of the NIOC laboratory (Fazlali, 1999). At this experimental section for the recognition of components, the separation of crude oil into two portions (light and heavy) was done by a simple distillation at first step. Then the light section was analyzed by gas chromatography for concentration analysis and determining the average molecular weight. Also the average molecular weight of heavy part was determined by the freezing point method. All of these equipments were prepared based on the IP-86 standard. At the second step, calculating the amount of asphaltene deposition was performed based on the IP standard. In fact, the amount of asphaltene precipitation in oil samples have been calculated by adding n-C5, n-C6, n-C7, n-C9 and n-C12 precipitants in different proportions of solvent volume ratio. The specified amount of precipitant as n-C5, n-C6, n-C7, n-C9 or n-C12 was used for an obvious crude oil. The vessels of samples were shacked during these experiments by a shaker. After 2 h, by using a paper filter (WattmanNo. 42) the formed precipitate was separated from the mixture and the weighted. This residue is named asphalt. After that, the asphalt was boiled for 30 min and filtered again. The precipitation is solved in toluene and filtered in another time, the remaining is pure asphaltene solution that must be dried in the oven. All of the experiments were performed based on the IP-143 standard method. As shown in Figs. 5 and 6, the mass fraction of asphaltene with the solvent ratio was studied. At a high solvent ratio the high asphaltene's mass fraction is due to the differential solubility parameter which decreases for the asphaltene and the oil solutions. In fact the differential solubility parameter acts as a driving force in the asphaltene precipitation formation. The variation of the solubility parameter is due to change in the nature of the liquid phase molecules, their interaction, the variation of molecular weight and the amount of polar compound in the liquid phase.

Fig. 2. Schematic of laboratory as phaltene precipitation.


G. Zahedi et al. / Journal of Petroleum Science and Engineering 68 (2009) 218–222

4. Modeling

Table 1 Comparison of performance of optimum MLP with various algorithms for training and testing.

4.1. Neural network applied to asphaltene precipitation predicting


Inputs of a network should be selected carefully if the best results are expected to be achieved. The input variables should reflect the underlying physics of the process to be analyzed. In the asphaltene precipitation type of the solvent, the ratio of the solvent to the oil, the pressure, the temperature and the composition of oil have strong effect on the formation of the asphaltene precipitations. Therefore the inputs for the network's model are the ratio of the solvent to the oil, the pressure, the temperature and the composition of oil; the output is the amount of the asphaltene precipitation. GC analysis and the total asphaltene content are considered as the compositional analysis of the oil which is used as the input for the ANN model. Also the experimental measurements were conducted at ambient temperature and atmospheric pressure. The back-propagation learning with one hidden layer network has been used in this work. Scaled Conjugate Gradient (SCG), Levenberg–Marquardt (LM), Gradient Descent with Momentum (GDM) and adaptive learning rate Back propagation (GDX) has been implemented for simulation. Inputs and outputs are normalized between the ranges of 0–1. Logistic Sigmoid and purelin transfer functions have been used in building ANNs. Each ANN has been trained with 2/3 of the data set and 1/3 of the data have been used for testing the predictions of NN. Some statistical methods were used for comparison. The criterion for comparison in this work was mean-squared-error between the net output and the training data. MSE is defined as: 2 P exp x −xsim MSE = ð1 = nÞ Where xexp is the target value, xsim is the output value and n is the number of the experimental data. 4.2. Thermodynamic models applied to asphaltene precipitation predicting At the next steps the prediction of the asphaltene precipitation is performed using the Flory–Huggins and the modified Flory–Huggins models (Fazlali et al., 2007; Pazuki and Nikookar, 2006). 4.2.1. Flory–Huggins model In this case a Vapor–Liquid–Liquid Equilibria (VLLE) model is assumed, to predict the phases which can split, by using separate VLE and LLE calculations. The model uses the concept of material balance by coupling it with a thermodynamic model. The SRK equation of state is used for the estimation of the component properties for the lighter fraction and equilibrium calculation. Asphaltene is assumed a polymeric substance; the polymeric solution theory was applied for the prediction of the asphaltene precipitation in crude oil. The simple model which has been used for the study of the polymeric solution behavior is Flory–Huggins (Fazlali et al., 2007; Pazuki and Nikookar, 2006). Based on the Flory–Huggins theory, the chemical potential of the polymeric compound is predicted as below:   0 μP − μP V V 2 2 = LnφP + 1 + P φS + P ðδP −δS Þ φS ð1Þ RT VS RT The solubility parameter in Eq. (1) is calculated as:  0:5 ΔU δi = v i





0.000609 0.003190 0.085960 0.172901

0.010600 0.012108 0.172574 0.278702

For this calculation, as an important assumption, the asphaltene precipitation phase is assumed a pure liquid pseudo-component that has no effect on the liquid–vapor equilibrium. Also, crude oil is considered as a binary homogeneous mixture of asphaltene and solvent (Pazuki and Nikookar, 2006). 4.2.2. Modified Flory–Huggins model (Fazlali et al., 2007; Pazuki and Nikookar, 2006) The Flory–Huggins model didn't have any suitable prediction of the asphaltene precipitation, in our previous study (Fazlali et al., 2006; Pazuki and Nikookar, 2006; Fazlali et al., 2007). Therefore, the modified Flory–Huggins model is implemented as below (Fazlali et al., 2007; Pazuki and Nikookar, 2006): 0

μP − μP = LnφP + RT

    V Vp  2 2 δP − δS φS + FPS 1 − P φS + RT VS

where FPS function is described as follow: FPS = 2λPS δP δS


Also λPS is the interaction parameter between the polymer and the solvent and it is a function of the asphaltene, the solvent and the solvent ratio molecular weight, that several functions have been examined for this respect by Pazuki and Nikookar and Fazlali et al. For more information about it see Fazlali et al. (2007) and Pazuki and Nikookar (2006). 5. Results and discussion In this section various ANN architectures with different training algorithms have been examined (Neural Network Toolbox, 2006). Table 1 depicts the best obtained results for each training algorithm. The LM training algorithm was found to have the best performance among these best obtained networks. As the network trained with LM gave much better results for the training sets than the other algorithm, it was used for modeling the asphaltene precipitation. Fig. 3 shows the performance of the LM algorithm with different hidden layer neurons. There is not a general and accurate method for obtaining the optimum number of hidden layers of the neurons and


Where μ, ϕ and V, δ are the chemical potential, the volume fraction, the molar volume and the solubility parameter, and subscriptions P and S are related to the polymer and the solvent phases. Also ΔU is the vaporization internal energy at constant temperature. The values of ΔU and v are calculated by the SRK equation (Riazi and Al-Sahhaf, 1996).


Fig. 3. Effect of number of hidden layer neurons on network estimation.

G. Zahedi et al. / Journal of Petroleum Science and Engineering 68 (2009) 218–222


this quantity is obtained by trial and error. The optimum number of hidden layer neurons was determined to be 15 for this network. A scatter plot of measured experimental data against the ANN model predictions (Fig. 4) represents the comparison between the predicted data by the ANN model and the experimental data which have not been used in the training of the ANN (1/3 remaining data). It is obvious from Fig. 4 that the ANN provides results very close to the laboratory measurements. The predictions which match measured values should fall on the diagonal line. Almost all data fall on this line, which confirms the accuracy of the ANN model. As shown in Figs. 5 and 6, a good agreement between the experimental data and the predicted ANN results can be observed. The Flory–Huggins prediction is far from experimental. Based on our expectation the modified Flory–Huggins method provides better

Fig. 4. Scatter plot of experimental (wt.% aspaltene) vs. predicted values by ANN model for unseen data.

Fig. 5. (a,b,c) A typically comparison between predicted data by ANN, Flory–Huggins, modified Flory–Huggins models and the experimental data in different solvent volume ratio of precipitants.

Fig. 6. (a,b,c) A typical comparison between predicted data by ANN and the experimental data in different solvent volume ratios of precipitants.


G. Zahedi et al. / Journal of Petroleum Science and Engineering 68 (2009) 218–222

Table 2 A typically comparison between the lab values and the predicted values of asphaltene precipitation by ANN, Flory–Huggins and modified Flory–Huggins models. Mass fraction of asphaltene precipitation A crude oil sample Solvent


ANN Model


Modified F.H.M

% Error for ANN model

% Error for F.H model

% Error for modified F.H model

n-C6 n-C7

1.45 1.40

1.443 1.354

0.830 0.745

2.09 1.29

0.482 3.285

42.758 46.785

44.137 7.857

results than the Flory–Huggins method. ANN has an excellent overlap with the laboratory experimental data. At lower solvent ratio n-C7 has a maximum yield (Figs. 5 and 6). By increasing the solvent ratios n-C5 provides better asphaltene precipitation. To check the performance of the ANN model, its estimation ability is compared with the Flory–Huggins and the modified Flory–Huggins models. The results carried out with the Flory– Huggins and the modified Flory–Huggins were compared with the ANN and also the experimental data which were not used in the training of the ANN. Table 2 compares the error of ANN , the Flory–Huggins and the modified Flory–Huggins models. The estimation error of the ANN is unbelievably low. ANN is also able to handle the solvent type. The obtained ANN model can be updated where new data are available. This task is applicable by retraining the ANN using the old ANN weights as initial weights for the new ANN. 6. Conclusions and remarks In this work, the ability of ANN in the modeling and in the prediction of the asphaltene precipitation has been investigated. Specifically, the asphaltene precipitation in the specific crude oil was modeled with MLP neural network architectures. By using this topology a good agreement with the experimental data was obtained. An important feature of the model is it doesn't require any theoretical knowledge or human experience during the training process. So prior knowledge hasn't been used and the model has been trained based on the experimental data only. All unknown relationships have been represented with NN, which can approximate instead of the traditional relationships. Finally the ANN ability has been compared with two thermo dynamical models. ANN was found as the most accurate model. 7. List of symbols xexp xsim

target value output value

N pattern P pressure (MPa) R universal gas constant T temperature (K) V molar volume (m3 kmolK1) U internal energy ϕ volume fraction μ chemical potential δ solubility parameter l liquid o standard state P (subscripts and superscripts) asphaltene S (subscripts and superscripts) solvent References Buenrostro-Gonzalez, E., Lira-Galeana, C., Gil-Villegas, A., Wu, J., 2004. AIChE J. 50, 2552–2570. Bulsari, A.B., 1995. Neural Networks for Chemical Engineers. Elsevier Science Press, Amsterdam. Fazlali, A., 1999. The asphaltene precipitation in crude oil of Iran, PhD Thesis , Amir Kabir University, Iran, Fazlali, A., Orangi, H., Modarress, H., Namazi, M., 2006. Phase behavior of binary mixture of asphaltene + solvent and ternary mixture of asphaltene + solvent + precipitant. Fluid Phase Equilib. 245, 117–124. Fazlali, A., Hosseini, M., Khosrobeigi, E., 2007. A new thermodynamic modified Flory–Huggins model for prediction of asphaltene precipitation in crude oil. International Conference on mining, Materials and Petroleum Engineering, Phuket, Thailand, 10–12 May. Hagan, M.T., Demuth, H.B., Beal, M., 1996. Neural network design. PWS Publishing Company, Boston. Hirschberg, A, de Jong, LKJ, Schipper, BA, Meijer, JG, 1984. Influence of temperature and pressure on asphaltene flocculation. Soc. Pet. Eng. J. 283–293 (June). Lang, R.I.W, 2006. A future for dynamic neural networks. Dept. Cybernetics, University of Reading, UK. Neural Network Toolbox, 2006. For use with MATLAB, user's guide version 4. Pazuki, G., Nikookar, M, 2006. A modified Flory–Huggins model for prediction of asphaltene precipitation in crude oil. Fuel 85, 1083–1086. Rassmdana,, H., Dabir,, H., Nematy,, M., Farhani, M., Sahimi,, M., 1996. Asphalt flocculation and deposition: I, the onset of precipitation. AIChE J. 42, 10. Riazi,, M.R., Al-Sahhaf,, T.A., 1996. Properties of heavy petroleum fractions and crude oils. Fluid Phase Equilib. 117 (1–2), 217–224. Sözen, Adnan, Özalp, Mehmet, Arcakioglu, Erol, 2004a. Formulation based on artificial neural network of thermodynamic properties of ozone friendly refrigerant/absorbent couples. Appl. Therm. Eng. 25. Sözen, Adnan, Özalp, Mehmet, Arcakioglu, Erol, 2004b. Investigation of thermodynamic properties of refrigerant/absorbent couples using artificial neural networks. Chemical Engineering and Processing 43, 1253–1264. Vallés, Harry Rodríguez,“ A neural networks method to predict activity coefficients for binary systems based on molecular functional group contribution” Master thesis, University of Puerto Rico, 2006. Zahedi, G, Jahanmiri, A., Rahimpor, M.R., 2005a. A neural network approach for prediction of the CuO–ZnO–Al2O3 catalyst deactivation. Int. J. Chem. React. Eng. 3 Article A8. Zahedi, G.R, Elkamel, A, Lohi, A, Jahnmiri, A., Rahimpor, M.R., 2005b. Hybrid artificial neural network—first principle model formulation for the unsteady state simulation and analysis of a packed bed reactor for CO2 hydrogenation to methanol. Chem. Eng. J. 115. Zahedi, G, Fgaier, H, Jahanmiri, A, Al-Enezi, G, 2006. Artificial neural network identification and evaluation of hydrotreater plant. Petrol. Sci.Technol. 24, 1447–1456.