Predictive Study of Flow-Accelerated Corrosion Characteristic Parameters Based on the Neural Network

Energy is a significant material foundation of the economy, and the green and safe development of the oil refining industry guarantees national and societal stability. Crude oil is a nonrenewable resource, which has gradually been depleted after more than half a century of exploitation. Because China possesses more coal than oil, large petrochemical companies have been motivated to expand and transform their capacities.

In the process of equipment localization and large-scale, many core-refining equipment systems (such as electric desalination systems, hydrogenation reaction effluent systems, and atmospheric and vacuum tower systems) have experienced varying degrees of corrosion. As a result, equipment tube wall thicknesses of a heat exchanger or air cooler are reduced, causing perforation, leakage, and accidents, seriously affecting safe production and reducing the economic benefits of the enterprise.

In addition to hydrocarbon compounds, petroleum contains impurity elements such as sulfur (S), nitrogen (N), and chlorine (Cl). Although the content of these impurities is low, they can cause severe corrosion problems for equipment during processing. Furthermore, large amounts of water and Cl-containing chemical additives are applied during crude oil extraction, greatly increasing the content of corrosive impurities and salts, such as S, Cl, and N.¹  These impurities undergo hydrolysis and reaction during crude oil transportation and processing to produce corrosive products such as HCl, H₂S, and NH₃.²  HCl, H₂S, and NH₃ flow, transfer heat, and change phases during production and transportation in different equipment systems and form multicomponent corrosive fluids with separated liquid water and oil.

Common types of corrosion in petroleum processing include chemical corrosion and electrochemical corrosion.^3-6  The crude oil must be desalted and dehydrated before processing. However, it is challenging to remove inorganic salts completely, especially chlorides. Organic Cl cannot be removed by electrical desalination, which increases the risks of safe operation during subsequent processing and in transportation equipment. In the initial distillation stage, the water in the crude oil vaporizes to precipitate out Cl-containing inorganic salts, which react with water vapor to generate HCl, resulting in the corrosion of the condensation system at the top of the primary distillation tower.

In the hydrogenation unit, the temperature of the fluid medium in the tube gradually decreases as the air flows through the air cooler, and ammonium chloride and ammonium sulfide can readily crystallize and deposit. If the amount of water injection is insufficient, ammonium salts can deposit on the walls of the air cooler and block the tubes.⁷  The water injected for scouring the crystallization causes the ammonium chloride to quickly absorb moisture, making the concentration of the ammonium chloride solution in the tube too high to cause significant erosion under scale. If the flow rate of water injection is too high and the flow velocity is too fast, it can result in severe erosion and corrosion.^8-9 

Ammonium salt crystal particles are affected by ambient temperature and humidity. Under certain conditions, high-concentration ammonium chloride solution forms and can be highly corrosive once a small amount of liquid water appears.¹⁰  When there is no liquid water in the environment, the ammonium salt can absorb water vapor from the surrounding air if the relative humidity exceeds 10%, causing corrosion under scale.¹¹ 

To predict the water vapor absorption rate in the deliquescent process, Li, et al.,¹²  used a heat transport model to determine the moisture absorbing ability of deliquescent crystallization in a controlled humidity environment. The model was used to find the mass transfer resistance during water vapor absorption and the uniformity of the sample. They also found that the porosity was not very large, and the particle size had no significant effect on the kinetics of the solid to absorb water vapor during the deliquescent process.

Hu, et al.,¹³  analyzed the volatilization and deliquescent processes of NH₄Cl, NH₄SO₄, and NH₄NO₃ and studied the effects of particle size and humidity on the crystallization characteristics. Especially on the inner wall of the tube, the process medium and tube wall had a complex electrochemical reaction, which seriously damages the corrosion product protective film and the matrix. Gong, et al.,¹⁴  found that the corrosion form changes from general to pit corrosion as the dry/wet ratio (D/W) increases. With decreasing pH, the protection of the corrosion product layer was reduced, causing pitting corrosion and promoting anodic dissolution.

The composition of crude oil is complex. In addition, during the cooling process, the type and content of compositions, and physical parameters are constantly changing in the tube of the heat exchange equipment. Because sensors cannot directly monitor some flow-accelerated corrosion characteristic parameters, the international standard from the American Petroleum Institute¹⁵  provides calculation methods for them. However, basic data must be obtained by process simulation software, resulting in slow calculation speeds. At present, there is no systematic method to predict flow-accelerated corrosion parameters. The corrosion resistance of equipment is mainly enhanced by material upgrades. The existing control methods have limitations and poor results. Therefore, it is necessary to study the prediction method of multifluid flow-accelerated corrosion characteristic parameters.

Yang, et al.,¹⁶  proposed a data-driven corrosion diagnosis scheme and the random forest (RF) algorithm based on particle swarm optimization (PSO) was used to classify and predict the corrosion mechanism, type, and degree with improved accuracy. Mohamed, et al.,¹⁷  proposed a practical implementation of robust ensemble learning models to accurately predict the internal corrosion rate in oil and gas pipelines. Peng, et al.,¹⁸  introduced a hybrid intelligent algorithm to predict the corrosion rate in multiphase flow pipelines for oil and gas transportation. However, the corrosion environment in the above three references was different from this paper.

The artificial neural network (ANN) model is a fast computational tool for treating scientific data having various properties. It has shown great promise in physics and chemistry research in the last few years,^19-20  especially in data prediction. Rocabruno-Valdés, et al.,²¹  developed ANN models for predicting the corrosion rates of metals in different biodiesels. The model had good predictive performance, but the corrosive environment was only biodiesel without fluidity. The current work used this model to study the prediction of corrosion characterization parameters of carbon steel piping in an air cooler under high temperature and high pressure in the multiphase flow environment.

During the cooling process, crystallized components (HCl, NH₃, and H₂S) may generate NH₄Cl and NH₄HS salts, which were deposited on the tube walls, causing tube blockages. Therefore, the water injection points are set before the hot high gas enters the heat exchanger and the air cooler.

The raw material of the device is a mixture of vacuum wax oil, coking wax oil, tank area wax oil, and a small amount of catalytic diesel oil. Coking wax oil comes from the delayed coking unit and contains additional impurities such as S and N. It is the main source of impurities in the mixed feedstock oil.

The physical parameters of raw oil after mixing were shown in Table 1. In actual processing, the N content of the mixed raw oil reaches 5,000 mg/kg (design specification is ≤3,500 mg/kg), the Cl content reaches 1.5 mg/kg (design specification is less than or equal to 1 mg/kg), and the S content is within the design specification range (2.5 wt%).

Through the analysis of the mixed raw oil, it is known that the N and Cl content of the wax oil hydrogenation unit under typical working conditions exceeds the standard, and the S content is very likely to exceed the design specification. Therefore, it is preliminarily judged that during the hydrogenation reaction effluent cooling process, readily crystallized components such as HCl, NH₃, and H₂S may undergo ammonium salt crystallization, deposition, corrosion, and erosion in the tubes. These flow-accelerated corrosion characteristics must be represented by relevant parameters, such as crystallization temperature and flow velocity, obtained by modeling, analysis, and calculation.

Aspen Plus^† (AspenTech Inc., USA) software simulated the physical parameters to explore the flow-accelerated corrosion characteristics of a hydroprocessing reactor effluent system. The simulations were performed according to the calibration report and the on-site material collection and analysis data. The Peng-Robinson cubic equation of state was utilized for three-phase flash evaporation separation. This equation can accurately predict critical and supercritical states when used in phase equilibrium calculations and has good applicability to the gas-liquid equilibrium of nonideal systems.

In Figure 2, the functions of the two-phase separation tank, heat exchanger, air cooler, and three-phase separation tank were respectively, equivalent to the hot high-pressure separator D103, heat exchanger E103, air cooler A101, and cold high-pressure separator D105 in Figure 1. The physical properties of the three product oils (diesel oil, hydrogenated wax oil, and naphtha) were given in Table 2. The volumetric flow rates of low gas and circulating hydrogen were 2,361.3 ncmh and 160,315 ncmh, respectively. The assumptions used to develop the simulation are listed in Table 3.

Along with the continuous changes in temperature, pressure, flow velocity, and other state parameters with the flow of the hydrogenation reaction effluent, the various parameters of the oil-gas-water mixture in the tube affect the timing of the ammonium salt crystallization and multiphase flow erosion. The flow-accelerated corrosion characteristics of the hydrogenation cold-exchange equipment depend on various actual production process factors that cause fluctuations in the working conditions. These factors include the processing volume m_pc, feed oil S content m_s, Cl content m_Cl, and N content m_N changes, the change in the water injection flow rate m_w1 before E103, and the water injection flow rate m_w2 before A101.

In this paper, the ranges of m_pc, m_s, m_Cl, m_N, m_w1, and m_w2 were specified, and several groups of working conditions were randomly generated within each variable range for simulation. Based on the change in feedstock oil varieties over multiple operation cycles of the device, m_pc was 0 to 310 ton/h, m_s was 0 to 3 wt%, m_Cl was 0 to 12 ppm, and m_N was 0 to 7,000 mg/kg. The production load and feedstock oil’s S, N, and Cl content usually fluctuate around typical working conditions, and extreme working conditions rarely occur, conforming to the Gaussian distribution law. Therefore, it was necessary to randomly generate several groups of working conditions for m_pc, m_s, m_Cl, and m_N according to the probabilities. The specific implementation was as follows:

1. Construct the Gaussian function according to the device’s actual operating data:

   

   where μ is the expected value, taken as the parameter values under typical working conditions, and σ is the variance, adjusted according to the actual operating data.

2. Generate n values of x within a given variable range, and find the corresponding y.

3. Calculate the probability P(x_i) corresponding to each x_i value. The equation used is as follows:

   

4. Take the values x_i according to the probabilities P(x_i) as a working condition parameter set.

5. Repeat step 4 to generate a total of 10,000 sets of samples, with one sample set representing one working condition.

According to the operation experience, the water injection rate before E103 was 0 to 15,000 kg/h, the water injection rate before A101 was 0 to 40,000 kg/h, m_w1 and m_w2 were generated randomly with equal probability. The crystallization temperature of ammonium salt, the concentration of NH₄HS, the content of liquid water, and the flow velocity were important parameters that characterize the risk of flow-accelerated corrosion. The changes in these parameters are of great significance for judging the corrosion risk of the heat exchanger.

After calculating the flow corrosion characteristic parameters for many samples, the parameter rule of hydroprocessing reactor effluent systems for all working conditions was obtained. An equipment risk evaluation index system suitable for the work in this paper was constructed, and the risk grades for different parameters were obtained, as shown in Table 4.

Ammonium salt readily crystallizes in heat exchanger E103, so when the crystallization temperature is between 245°C and 120°C it is considered at high risk. Crystallization does not occur in the air cooler. If crystallization occurs, high-liquid water content can dissolve ammonium salt. It is considered a medium risk when the crystallization temperature is between 108°C and 50°C. When the crystallization temperature is between 120°C and 108°C, the piping between the heat exchanger outlet and the air cooler inlet has a large wall thickness and does not leak during operation, so when the temperature section crystallizes, it is regarded as low risk.

For liquid water content, referring to “API RP 932-B-2019,” lower than 25% is considered high risk, and higher than 25% is considered high risk. As for the flow velocity, considering that too high flow velocity will cause erosion and too low flow velocity will cause ammonium salt deposition, the flow rate at 3.1 m/s to 6.1 m/s is considered low risk, 6.1 m/s to 9.1 m/s is considered as medium risk, and <3.1 m/s or >9.1 m/s is considered as high risk. For ammonia hydride concentration, it is considered low risk at <4%, medium risk at 4% to 8%, and high risk at above 8%. Because there is no absolute measurement standard for hydrogen sulfide partial pressure, it is related to the operating pressure of the device, so there is no risk grade division.

A neural network^26-28  was used to construct a predictive model of the characteristic parameters of flow-accelerated corrosion to provide a reference basis for corrosion evaluation of air coolers and heat exchangers. Every neuron in a neural network has tens of thousands of synapses connected to other neurons.

The error back-propagation algorithm^30-32  is one of the most widely used and best performing neural network learning algorithms, with the learning process summarized as follows.

The implementation process of the back-propagation algorithm is as follows: inputting samples, signal forwarding, calculation of output error, error back propagation, parameter adjustment, and ending iteration until certain conditions are met.

This section constructs a back-propagation neural network (BPNN) flow-accelerated corrosion characteristic parameter prediction model to determine the crystallization temperature of ammonium chloride, the free water content, flow velocity, the concentration of ammonium sulfide, and the partial pressure of hydrogen sulfide.

The simulation environment in detail is as follows: CPU-Inter®^† Core™^† i5-8400 CPU @ 2.80 GHz to 2.81 GHz; RAM-8.00 GB; Matlab^† 2019a.

Before BPNN training, 70% of the samples were randomly selected as the training set, 15% as the test set, and 15% as the verification set. The data normalization method was adopted to realize the conversion from real to coded variables. That is, the original data are linearly transformed so that the result value is mapped to [0,1]. The training parameters of the neural network model were set as follows: the learning rate was 0.01, the epoch was 10,000, and the training goal was 0.00001. Levenberg-Marquardt algorithm was used to train the neural network. Before ANN training, 70% of samples were randomly selected as the training set, 15% samples as the test set, and 15% samples as the verification set.

Each BPNN model was repeatedly trained 50 times, and the average mean squared error (MSE) was 0.000281, 0.000322, 0.00565, 0.00392, and 0.00398, respectively. The number of iterations for each model to reach the optimal MSE were 83, 57, 18, 35, and 10, respectively.

The correlation analysis indicated that the crystallization temperature of NH₄Cl salt is highly correlated with the Cl content and S content in the feedstock oil. Therefore, the Cl and S contents were the input variables, and the crystallization temperature was the output variable. Neural networks with different hidden nodes were evaluated to select a reasonable number of hidden layer nodes. When the number of hidden layer nodes was 13, the root-mean-square of the neural network could not be minimized. Therefore, a “2-13-1” BPNN model structure was adopted.

The model was evaluated using the coefficient of determination R to represent the proportion of the output explained by the neural network model. The closer R is to 1, the greater the explanatory power of the neural network model

The correlation analysis indicated that there was a strong correlation between the liquid water content at the outlet of E103 and the amount of water injected before E103. Therefore, the amount of water injected before E103 is used as the input of the neural network, and the amount of liquid water at the outlet of E103 is the output. When predicting the E103 outlet liquid water content from the amount of water injected before E103, the root-mean-square of the neural network could not be minimized when the number of hidden layer nodes was 10. Therefore, a “1-10-1” BPNN model structure was adopted.

The correlation analysis indicated that the inlet and outlet flow velocities of A101 have a strong correlation with the processing volume of feedstock oil. Therefore, feedstock oil processing volume was used as the input volume of the neural network, and the inlet and outlet flow velocities of the A101 were the output volume. Evaluate the structure of different numbers of hidden layer nodes. When the number of hidden layer nodes was 16, the root-mean-square of the neural network could not be minimized when predicting the A101 inlet and outlet flow velocities from the raw oil processing volume. Therefore, a “1-16-2” BPNN model structure was adopted.

The R of the model is lower than 0.9, which has an influence on the subsequent prediction accuracy. The main reason is that the flowmeter is installed in the inlet and outlet main piping of the air cooler, which is relatively close to the diverging point of the piping, where the fluid is easy to produce strong turbulence during the diverging, resulting in the numerical instability detected by the flowmeter. Therefore, as the output variable of the neural network, the large fluctuation of data is easy to cause the unsatisfactory training results for the model.

The correlation analysis indicated that the concentration of NH₄HS has a strong correlation with the water injection flow rate in front of the air cooler and a weak correlation with the water injection flow rate in front of the heat exchanger. Therefore, the water injection flow rates were input variables, and NH₄HS concentration was output variable. The network of nodes in different hidden layer was evaluated. The root-mean-square of the neural network could not be minimized when the number of hidden layer nodes is 18 when predicting the NH₄HS concentration from the water injection flow rate before A101 and the water injection flow rate before E103. Therefore, a “2-18-1” BPNN model structure was adopted.

The correlation analysis indicated that the partial pressure of H₂S has a strong correlation with the S content in the feedstock oil. Therefore, the S content in the feedstock oil was input variable, while the partial pressure of H₂S was the output. The network of nodes in different hidden layer was evaluated. During the prediction of the hydrogen sulfide partial pressure from the S content, the root-mean-square of the neural network could not be minimized when the number of hidden layer nodes was 7. Therefore, a “1-7-1” BPNN model structure was adopted.

An additional 2,000 sets of data were selected to further test the overall model performance.

Although the fitting degree of the neural network model for velocity prediction in A101 Inlet and Outlet Flow Rate Prediction section was not very good, it was found through data testing that its predictive accuracy could meet the requirements.

From the testing results just described, the BPNN models were found to have some errors. Possible reasons for these errors are as follows: (i) the amount of sample data was not large enough; (ii) there were individual abnormal data in the sample; and (iii) the test data sets and the training data sets were collected at different times, resulting in possible operating condition fluctuations. Moreover, a better predictive model than BPNN could be used to improve the results.

The behavior of the six flow-accelerated corrosion characteristic parameters predicted in this paper basically conformed to Gaussian distributions for the actual working conditions. Most of the sample data were concentrated near a particular working condition, and some data samples deviated substantially because of random working condition fluctuations during long-term device operation.

• This study proposed a Gaussian distribution-based expansion method for the hydrogenation reaction effluent workload parameter population. A BPNN prediction model was constructed, validated, and tested, and each model had good predictive performance, for example, with a faster prediction speed (millisecond level) than Aspen Plus software (minute level).

• The output variables were the NH₄HS crystallization temperature, E103 outlet liquid water content, A101 inlet flow velocity, A101 outlet flow velocity, and NH₄HS concentration, with relative prediction errors of about 0.2%, 1%, 1%, 3%, and 5%, respectively. Based on the BPNN prediction results and the corrosion risk evaluation method proposed in this paper, engineers can monitor real-time equipment corrosion status under different working conditions and execute timely preventive measures.

• Although many flow-accelerated corrosion parameters have been predicted in this paper, there is still room for accuracy improvement. Because of the flow complexity, the diversity of material composition, and the fluctuations in working conditions, more process parameters related to flow-accelerated corrosion parameters can be mined as neural network inputs to improve the model’s accuracy. In addition, the neural network structure can be modified to improve the model’s predictive accuracy.

The authors are grateful to the National Natural Science Foundation of China (No. 51876194, No. 52176048, and No. U1909216) and the General Research Project of Zhejiang Provincial Department of Education (Y201942785) for supporting this study.