Abnormal working condition prediction method

Abstract

The invention belongs to the field of abnormal working condition prediction and early warning, and discloses an abnormal working condition prediction method, which comprises the following steps: firstly, according to a process flow diagram and process parameters, establishing a process flow steady-state model, adding a controller, establishing a dynamic mechanism model, simulating a fractionating tower tray overtemperature working condition, and obtaining historical data and dynamic simulation data of a fractionating device production process; secondly, screening characteristic variables through a Pearson correlation coefficient, and reducing the data dimension; and utilizing a GA-BP neural network to deeply learn a potential relationship between variables so as to predict a change trend of process key parameters. And finally, comparing the predicted value of the process parameter with a preset alarm limit value, and selectively executing an alarm operation corresponding to the chemical device with the process parameter.

Description

Technical Field

[0001] The present invention belongs to the technical field of abnormal operating condition prediction and early warning, and specifically relates to an abnormal operating condition prediction method. Background Art

[0002] Reducing the occurrence of abnormal conditions is a prerequisite for the safe and stable operation of the kerosene fractionation process. Therefore, it is crucial to carry out the prediction and early warning of abnormal conditions. At present, the devices of chemical enterprises are basically equipped with distributed control systems (DCS) and safety instrument systems (SIS), which realize the real-time monitoring of device parameters, that is, the safe operation of the equipment, but they cannot predict parameters and realize advance warning; when abnormal conditions occur, the on-site personnel do not have enough operating time, so it is necessary to realize advance prediction to ensure the safety and stable operation of the device. At present, abnormal condition prediction technology can be divided into three categories: analytical model-based methods, data-driven methods, and knowledge-based methods. The analytical model-based method requires the establishment of strict and accurate mathematical models, such as material balance and energy balance; however, due to the strong coupling of the kerosene distillation process, it is difficult to establish a strict analytical model; the data-driven method starts from a large amount of data and uses various analytical methods to explore the changing rules hidden in the data; however, if only the data-driven method is used to predict the abnormal operating conditions of the kerosene distillation unit, many parameters need to be predicted, and the data obtained under the abnormal operating conditions of the distillation tower overheating is relatively small. The types and number of models involved in the algorithm are large, and the generalization is not strong. The knowledge-based method can analyze and reason about abnormal situations through a large amount of expert knowledge and experience, and quickly and accurately obtain correct conclusions; however, it is usually difficult to comprehensively collect and extract massive expert knowledge and operating experience using only knowledge-based methods. Summary of the invention

[0003] In order to solve the above technical problems, the present invention provides a method for predicting abnormal operating conditions, which builds a dynamic mechanism model, constructs an overtemperature scenario of a distillation tower tray, and obtains relevant data to solve the problem of insufficient data under actual abnormal production conditions; and uses a GA-BP neural network to predict key process parameters, and selectively executes alarm operations corresponding to chemical equipment with the process parameters.

[0004] The abnormal operating condition prediction and early warning method described in the present invention comprises the following steps:

[0005] Step 1: Collect historical data during the production process of the fractionation tower, establish a steady-state model of the fractionation system through Aspen software, and convert the steady-state model into a dynamic mechanism model through Aspen Dynamics to obtain dynamic simulation data; construct an overtemperature scenario for the fractionation tower, and realize offline collection of continuous operation data based on the dynamic mechanism model to make up for the abnormal operating condition data in the sample set, and obtain a fractionation tower operation data set consisting of historical data and dynamic simulation data;

[0006] Step 2: pre-process the operation data in the obtained distillation tower operation data set, divide it into a training set and a test set, and perform feature selection on the obtained operation data through expert experience and variable correlation analysis, and select highly correlated data as feature variables;

[0007] Step 3: Optimize the BP neural network using a genetic algorithm, construct a GA-BP neural network prediction model and train it to obtain a trained prediction model;

[0008] Step 4: Use the selected characteristic variables as input parameters of the distillation tower abnormal operating condition prediction model, predict the abnormal operating condition of the distillation tower, compare the predicted value with the preset alarm limit, and selectively perform corresponding alarm operations.

[0009] Furthermore, the data in the distillation tower operation data set are preprocessed, specifically: the outliers are processed using the Laida criterion and the outliers are removed; and the data are normalized.

[0010] Furthermore, the Pearson correlation coefficient is used to calculate the correlation between the preprocessed running data. The formula is:

[0011]

[0012] in, and are the average values ​​of variables x and y respectively, and n is the number of samples; when r(x, y) is positive, it means that there is a positive correlation between variables x and y; when r(x, y) is negative, it means that there is a negative correlation between variables x and y.

[0013] Furthermore, the network connection weights and thresholds of the GA-BP neural network model are represented by real number coding, and the fitness function uses the inverse of the root mean square error as an evaluation of whether each individual is close to the optimal solution; the GA-BP neural network is trained using a training set, and the GA-BP neural network is tested using a test set to obtain an optimized GA-BP neural network.

[0014] Furthermore, the fitness function is:

[0015]

[0016] in, is the predicted value of the model in the training set,

[0017] T={t1, t 2,...t n} is the actual value of the model in the training set, and n is the number of samples in the training set.

[0018] The beneficial effects described in the present invention are as follows: the method described in the present invention is based on the problem of insufficient abnormal operating condition data in actual production, and uses software to build a dynamic simulation process, obtains enough abnormal data, performs dynamic simulation, constructs a distillation tower tray overtemperature scenario, and enriches the data set of the neural network; the present invention processes complex and high-dimensional production process big data through correlation analysis and neural networks, extracts the essential characteristics of abnormal operating conditions from a large amount of data, and accurately predicts the key process variables related to the abnormal operating conditions; highly correlated variables are selected as feature variables to predict the key process variables related to the abnormal operating conditions, and can predict abnormalities in advance. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of the method of the present invention;

[0020] Figure 2 This is a fractionation process flow chart of a petrochemical enterprise implementing the present invention;

[0021] Figure 3 This is a steady-state flow chart of the fractionation process implemented in the present invention;

[0022] Figure 4 It is a dynamic flow chart of the fractionation process implemented in the present invention;

[0023] Figure 5 It is an image of the change of characteristic variables over time when the 15-layer tray of the fractionation tower is overheated according to the present invention;

[0024] Figure 6 is a heat map of the correlation of PCCs implemented in the present invention;

[0025] Figure 7 It is a structural diagram of a GA-BP neural network implemented in the present invention;

[0026] Figure 8 is an error curve diagram during the training process of the present invention;

[0027] Fig. 9 It is a prediction effect diagram of the GA-BP neural network implemented in the present invention;

[0028] Fig.10 It is a comparison chart of the predicted values ​​implemented by the present invention and the measured values ​​of the DCS system. DETAILED DESCRIPTION

[0029] In order to make the contents of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments in conjunction with the accompanying drawings.

[0030] like Figure 1 As shown, the abnormal operating condition prediction and early warning method described in the present invention includes the following steps:

[0031] Step 1: Collect historical data during the production process of the fractionation tower, establish a steady-state model of the fractionation system through Aspen software, and convert the steady-state model into a dynamic mechanism model through Aspen Dynamics to obtain dynamic simulation data; construct an overtemperature scenario for the fractionation tower, and realize offline collection of continuous operation data based on the dynamic mechanism model to make up for the abnormal operating condition data in the sample set, and obtain a fractionation tower operation data set consisting of historical data and dynamic simulation data;

[0032] Step 2: pre-process the operation data in the obtained distillation tower operation data set, divide it into a training set and a test set, and perform feature selection on the obtained operation data through expert experience and variable correlation analysis, and select highly correlated data as feature variables;

[0033] Step 3: Optimize the BP neural network using a genetic algorithm, construct a GA-BP neural network prediction model and train it to obtain a trained prediction model;

[0034] GA-BP neural network is that in the learning process of BP neural network, weights and thresholds are described as chromosomes, and appropriate fitness functions are selected, and then GA iterations are performed until convergence in a certain sense. When optimizing, individuals must first be encoded; global search effectively prevents the algorithm from falling into local optimality; the optimization process purposefully guides the entire search process to move towards the optimal value in the search space; it can handle complex problems well.

[0035] Step 4: Use the selected characteristic variables as input parameters of the distillation tower abnormal operating condition prediction model, predict the abnormal operating condition of the distillation tower, compare the predicted value with the preset alarm limit, and selectively perform corresponding alarm operations.

[0036] The method of the present invention is described in conjunction with a fractionation process flow chart of a petrochemical enterprise. The fractionation process flow chart is as follows: Figure 2 As shown, Figure 2 Among them, C-101 is a stripping tower, C-102 is a stabilizing tower, C-103 is a fractionating tower, E-136 is an air cooler at the top of the light oil fractionating tower, V-120 is a receiver at the top of the light oil fractionating tower, and V-109 is a receiver at the top of the stabilizing tower.

[0037] In step 1, the process flow is built based on the existing equipment model in Aspen software to obtain the TBP (temperature, boiling point, yield) data, density, and light-end components of the feed reaction oil and gas. The virtual component input of the feed oil sample generated by the petroleum fraction curve is used for simulation. The main composition and property parameters of the oil product are shown in Table 1. The steady-state process of the fractionation system is as follows Figure 3 As shown, the main operating parameters of the fractionation system are shown in Table 2. The high-temperature oil and gas mixture produced by the hydrogenation reaction system is separated into top oil and gas and bottom liquid by the fractionation of the fractionation tower. The top oil and gas are ≤C11 light fractions (only a small amount of C10 and C11) and gases such as H2S. They are cooled to 50°C by an air cooler and enter the receiver. The liquid at the bottom of the receiver is pumped out by a pump (hot reflux), part of which is used as the top reflux of the stripping, and part of which is sent to the light oil fractionation tower as light oil for further treatment. The bottom liquid is the hydrorefined kerosene with ≤C9 components removed. The top material is heated by the heater and sent to the light oil fractionation tower to separate the No. 2 solvent oil. The bottom oil left is sent to the stabilization tower for further separation. The bottom oil at the bottom of the tower is No. 1 solvent oil.

[0038] The whole process of the fractionation system is simulated by the method of the present invention, and the comparison between the main parameter values ​​and the actual values ​​of the stripping tower C-101, the light oil fractionation tower C-103, and the stabilization tower C-102 is shown in Table 3. As can be seen from Table 3, most of the simulated calculated values ​​of the main operating parameters show a small error with the actual operating parameter values. Since the loss in chemical production is generally around 5.00%, the error result verifies the effectiveness of the mechanism model.

[0039] Table 1 Main components and property parameters of oil products

[0040]

[0041] Table 2 Main operating parameters of the distillation system

[0042]

[0043] Table 3 Comparison of main simulation parameters and actual values ​​of fractionation system

[0044]

[0045] Use Aspen Dynamics to convert the steady-state model into a dynamic mechanism model and obtain dynamic simulation data. Figure 4As shown, PID controllers are added to the stripping tower C-101, the light oil fractionation tower C-103, the stabilization tower C-102 and the three flash tanks. When disturbed, the process parameters (such as pressure, liquid level, etc.) and product composition of the equipment are quickly adjusted to restore them to the preset set values. In this embodiment, Aspen Dynamic has preset the pressure controller, and the controllers such as the sensitive plate temperature control loop and the tower bottom liquid level control are further configured according to the actual situation.

[0046] In order to solve the problem of insufficient data under abnormal production conditions, the present invention constructs a scenario of overheating of the tray C-103 of the fractionating tower, and realizes offline collection of continuous operation data. Figure 5 shown.

[0047] The present invention mainly obtains relevant historical data and dynamic simulation data of the kerosene fractionation tower. Through on-site investigation and discussion with on-site technicians, the following parameters are preliminarily determined as acquisition parameters: E-113 outlet temperature, E-138 hot oil flow, C-103 reflux flow, C-103 bottom pressure, C-103 top pressure, C-103 bottom temperature, C-103 feed temperature, C-103 bottom discharge, C-103 top temperature and other parameters are shown in Table 4. Since in the enterprise DCS, there are fewer abnormal conditions of overheating of the 15-layer tray of the fractionation tower C-103, and the data generated is insufficient, for safety reasons, the DCS system operating parameters cannot be changed in actual production to construct abnormal scenarios; therefore, the constructed fractionation tower dynamic model is run, and a set of data is obtained at intervals of 90 seconds. The output of 500 sets of parameter prediction simulation data when the 15-layer tray of the tower C-103 is overheated is shown in Table 5 below, and the abnormal data set is supplemented, thereby effectively reducing the imbalance between normal and abnormal data in the data set.

[0048] Table 4 Historical data collection table of kerosene distillation tower

[0049]

[0050] Table 5 Parameter collection table under overtemperature of 15 trays in fractionation tower C-103

[0051]

[0052] To ensure the integrity and accuracy of the acquired data, data preprocessing is required. In the actual production of the fractionation unit, the generation of abnormal values ​​may be caused by sudden fluctuations in process variables or recording errors by recorders. Therefore, abnormal values ​​need to be eliminated.

[0053] The present invention processes outliers according to the Laida criterion; using x1, x2, ... x i ,...,xn Represented as n samples, is the average value;

[0054] Calculate the deviation: i=1,2...n;

[0055] The standard deviation is calculated using the Bessel formula:

[0056]

[0057] If a sample data x i The deviation v i , 1≤i≤n satisfies:

[0058] |v i |>3σ,

[0059] Then we think x i The abnormal data will be removed.

[0060] In order to eliminate the impact of dimensions between features and improve the model's data processing efficiency, the acquired data is normalized.

[0061] like Figure 2 As shown in Figure 2, due to the complex process flow and numerous process variables in the fractionation process, too many redundant variables seriously affect the prediction efficiency of the GA-BP neural network and have a certain impact on the prediction accuracy. PCCs are used to select feature variables for process variables, reduce data dimensions, improve the prediction accuracy of the neural network, and shorten the calculation time of the prediction model.

[0062] PCCs: (Pearson product-moment correlation coefficient, PPMCC or PCCs for short) is used to measure the correlation (linear correlation) between two variables X and Y, and its value is between -1 and 1. The Pearson correlation coefficient formula is:

[0063]

[0064] in, and are the average values ​​of variables x and y, respectively. n represents the number of samples, that is, the total number of data samples collected. When the value of r(x,y) is positive, it means that there is a positive correlation between variables x and y; when the value of r(x,y) is negative, it means that there is a negative correlation between variables x and y.

[0065] The relevant criteria for PCCs are shown in Table 6, and variables with high correlation are selected as characteristic variables.

[0066] Table 6 Comparison table of absolute value of correlation coefficient and degree of correlation

[0067]

[0068] like Figure 6 As shown, the hot oil flow rate of the distillation tower reboiler E-138 and the bottom temperature of the distillation tower C-103 are strongly correlated with the temperature of the 15-layer tray of the distillation tower C-103; the bottom material temperature of the distillation tower C-101, the top temperature of the distillation tower C-103, the bottom discharge amount of the distillation tower C-103 and the feed temperature of the distillation tower C-103 are moderately correlated with the temperature of the 15-layer tray of the distillation tower C-103; the reflux flow rate of the distillation tower C-103 and the top pressure of the distillation tower C-103 are weakly correlated with the temperature of the 15-layer tray of the distillation tower C-103, and the correlation degree of the remaining variables is relatively low.

[0069] According to the experience of on-site technicians and the results of correlation analysis, the present invention screens out the hot oil flow rate of the distillation tower reboiler E-138, the bottom temperature of the distillation tower C-103, the material temperature at the bottom of the distillation tower C-101, the top temperature of the distillation tower C-103, the bottom discharge amount of the distillation tower C-103, the feed temperature of the distillation tower C-103, the reflux flow rate of the distillation tower C-103 and the top pressure of the distillation tower C-103 as characteristic variables, and the 15-layer tower plate temperature of the distillation tower C-103 as the target variable, and obtains these variables as input and output parameters of the neural network model.

[0070] In step 3, based on the kerosene fractionation unit operation data and real-time data collected by Aspen software, a prediction model was constructed with the eight variables selected in step 2 as input variables and the fractionation tower tray temperature as output variable. By using the trial and error method, different numbers of hidden layer neurons were selected for simulation training, and the number of hidden layer neurons was finally determined to be 10. Figure 7 The schematic diagram of the GA-BP neural network constructed based on the device, the input parameters of the device are located in the input layer of the neural network, the parameters to be predicted are located in the output layer of the neural network, and the corresponding rules implied by the two are automatically calculated through multiple nodes of the hidden layer. The layers are fully connected through transfer functions, and the transfer function in the present invention uses a logarithmic S-type function.

[0071] Figure 8 is the error performance of the GA-BP neural network operation results. The errors of the training samples, validation samples, and test samples gradually decrease with the increase of the number of iterations. In the early stage of network training, the error decreases rapidly, and as the number of iterations increases, the decrease gradually slows down. This shows that the network training has not fallen into a local minimum. The mean square error of the training set is 0.08, and the mean square error of the test set is 0.09.

[0072] Save the trained GA-BP neural network and do not change the parameters. Read in the 73 sets of data collected after the training sample and observe the relative error between the true value and the predicted value. Fig. 9 As shown in the figure, there is a comparison between 73 groups of predicted values ​​and measured values. The relative errors between the true values ​​and the predicted values ​​in the first 30 groups are small. After the 50th group of data, the prediction accuracy of the model decreases, and the relative errors between the true values ​​and the predicted values ​​become larger, but the relative errors are all less than 1%. Generally speaking, the predicted values ​​can follow the measured values ​​well and can achieve a good prediction purpose.

[0073] As a process indicator of the kerosene fractionation tower production process, overheating of the fractionation tower will reduce the fractionation effect and easily lead to unqualified product quality. When the overheating is serious, it may even cause an explosion. Eight process parameters related to the fractionation tower tray are selected as the input of the prediction model. The temperature of the fractionation tower tray varies from 164 to 182 °C when the temperature is stable. For the accuracy of the model, 2456 groups of samples of 8 parameters when an abnormal condition occurs are selected. Data are collected every 90 seconds, of which 1670 groups are training samples, 716 groups are test samples, and 70 groups are verification samples. Some sample data are shown in Table 7.

[0074] Table 7 Sample data of 9 parameters under overtemperature of distillation tower

[0075]

[0076] The present invention uses GA-BP neural network to predict the temperature of 15 trays of kerosene fractionation tower in the early stage, and the steps are as follows:

[0077] (1) The 2386 data sets and 8 input variables of the sample set were normalized and preprocessed. The normalized data are shown in Table 8;

[0078] (2) Select the parameter settings of the genetic algorithm and establish the GA-BP neural network;

[0079] (3) The GA-BP neural network model predicts the data for the next 1800 seconds. The prediction results are as follows: Fig.10 As shown in the figure, the alarm time of the prediction model is 9:37:00, and the actual alarm data exceeds the alarm limit at 9:37:30. The proposed method alarms 30s earlier than DCS. The prediction accuracy of the GA-BP neural network model decreases after 9:37:45. The GA-BP neural network is used to more effectively and accurately predict the temperature changes in the next 1440s.

[0080] Table 8 Partial sample data after normalization of 8 input variables under the condition of overtemperature of distillation tower

[0081]

[0082] In a chemical process, when dangerous parameters (such as pressure, temperature, etc.) change rapidly or fluctuate severely, the alarm method used by the DCS can only alert the operator when the parameters exceed the preset upper and lower limits, which may cause the operator to be unable to make adjustments and judge the response to abnormal situations in time. The method described in the present invention can predict the abnormality in advance and accurately predict the temperature change within a certain period of time in the future, which is of great significance for predicting the changes of important parameters in the chemical process.

[0083] The above description is only a preferred embodiment of the present invention and is not intended to be a further limitation of the present invention. All equivalent changes made using the contents of the present specification and drawings are within the protection scope of the present invention.

Claims

1. A method for predicting abnormal operating conditions, characterized in that: The following steps are involved: Step 1: Collect historical data during the production process of the fractionation tower, establish a steady-state model of the fractionation system through Aspen software, and convert the steady-state model into a dynamic mechanism model through Aspen Dynamics to obtain dynamic simulation data; construct an overtemperature scenario for the fractionation tower, and realize offline collection of continuous operation data based on the dynamic mechanism model to make up for the abnormal operating condition data in the sample set, and obtain a fractionation tower operation data set consisting of historical data and dynamic simulation data; Step 2: pre-process the operation data in the obtained distillation tower operation data set, divide it into a training set and a test set, and perform feature selection on the obtained operation data through expert experience and variable correlation analysis, and select highly correlated data as feature variables; Step 3: Optimize the BP neural network using a genetic algorithm, construct a GA-BP neural network prediction model and train it to obtain a trained prediction model; Step 4: Use the selected characteristic variables as input parameters of the distillation tower abnormal operating condition prediction model, predict the abnormal operating condition of the distillation tower, compare the predicted value with the preset alarm limit, and selectively perform corresponding alarm operations.

2. The abnormal operating condition prediction method according to claim 1, characterized in that: The data in the distillation tower operation data set are preprocessed, specifically: outliers are processed and removed using the Laida criterion; and the data are normalized.

3. The abnormal operating condition prediction method according to claim 2, characterized in that: For the preprocessed running data, the Pearson correlation coefficient is used to calculate the correlation between the data. The formula is: in, and are the average values ​​of variables x and y respectively, and n is the number of samples; when r(x, y) is positive, it means that there is a positive correlation between variables x and y; when r(x, y) is negative, it means that there is a negative correlation between variables x and y.

4. The abnormal operating condition prediction method according to claim 1, characterized in that: The network connection weights and thresholds of the GA-BP neural network model are represented by real number coding, and the fitness function uses the inverse of the root mean square error as an evaluation of whether each individual is close to the optimal solution; the GA-BP neural network is trained using a training set, and the GA-BP neural network is tested using a test set to obtain an optimized GA-BP neural network.

5. The abnormal operating condition prediction method according to claim 4, characterized in that: The fitness function is: in, is the predicted value of the model in the training set, T={t1,t2,...t n } is the actual value of the model in the training set, and n is the number of samples in the training set.

Images