CN112330082A

CN112330082A - Intelligent quantitative risk assessment method based on dynamic mechanism model

Info

Publication number: CN112330082A
Application number: CN202010811837.XA
Authority: CN
Inventors: 田文德; 刘子健; 崔哲
Original assignee: Qingdao University of Science and Technology
Current assignee: Qingdao University of Science and Technology
Priority date: 2020-08-13
Filing date: 2020-08-13
Publication date: 2021-02-05
Anticipated expiration: 2040-08-13
Also published as: CN112330082B

Abstract

The invention discloses an intelligent quantitative risk assessment method based on a dynamic mechanism model, which comprises the following steps: establishing a steady-state model of the process flow according to the process flow diagram and the process parameters, adding a controller, establishing a dynamic mechanism model, and simulating process leakage to obtain dynamic data sets with different disturbances; selecting a variable for manual intervention, dividing a dynamic data set into a training set and a testing set, and adjusting network hyper-parameters by adopting an orthogonal experiment method; predicting variables which directly affect accident consequences through neural network learning, so that the network achieves the optimal prediction effect on a training set; and simulating accident hazard ranges under different process disturbances by a dynamic risk assessment method. The method realizes the construction of an industrial process dynamic mechanism model, accurately predicts the variable directly influenced by the accident, dynamically simulates the damage range of the accident, and has higher reliability compared with the classical safety assessment method.

Description

Intelligent quantitative risk assessment method based on dynamic mechanism model

Technical Field

The invention relates to the field of chemical process prediction and safety evaluation, in particular to an intelligent quantitative risk evaluation method based on a dynamic mechanism model.

Background

Safety assessment is a prerequisite for the proper operation of chemical processes. Predictive risk analysis (PHA) is a widely used method for security assessment. Although it has the advantages of reducing the cost and reducing the occurrence of accidents, it still has some drawbacks, it depends heavily on the expertise of experts and it is difficult to obtain reliable quantitative results, and it may cause misjudgment if the experience is insufficient. Therefore, the reliable, accurate and dynamic risk assessment method can better ensure the production safety and the stable operation of the device.

Risk assessment methods in chemical processes are divided into qualitative and quantitative methods. Qualitative methods such as HAZOP and SIL analysis can estimate the possible risks of the process and the reasons of the risks through the brainstorms of expert groups, obtain the severity and frequency of the consequences when an accident occurs, and further provide corresponding preventive measures. Compared with a qualitative method, the quantitative risk assessment calculates the damage range caused by the accident based on the process parameters and the external environment parameters. However, it also requires constant loop calculations and cannot update the hazard range based on process upsets.

Disclosure of Invention

Aiming at the problems in the prior art, the invention provides an intelligent quantitative risk assessment method based on a dynamic mechanism model, which is characterized in that a dynamic data set is built through the mechanism model, the interrelation among variables is learned by a deep learning technology, the variables directly influencing the consequences are predicted, and then the risk assessment method is adopted to simulate the accident hazard range.

In order to achieve the purpose, the invention provides the following technical scheme:

an intelligent quantitative risk assessment method based on a dynamic mechanism model comprises the steps of building a process flow mechanism model, predicting variables with direct influence on consequences by an LSTM, and simulating an accident hazard range by a QRA method, and comprises the following steps:

step one, establishing a process flow steady-state model through Aspen according to a process flow diagram and process parameters to ensure that the steady-state model is consistent with industrial actual data;

secondly, adding a controller through Aspen Dynamics, building a dynamic mechanism model, and simulating process leakage to obtain dynamic data sets with different disturbances;

thirdly, selecting a variable for human intervention from the dynamic data set obtained in the second step, dividing the dynamic data set into a training set and a testing set according to a proportion, and adjusting network hyper-parameters by adopting an orthogonal experiment method;

predicting variables which directly affect accident consequences through neural network learning, so that the network achieves the optimal prediction effect on the training set;

and fifthly, inputting network prediction and external environment parameters through a dynamic risk assessment method, and simulating the accident hazard range under different process disturbances.

Preferably, the steady-state model in the step one is consistent with the industrial practice, and the standard is that the relative error between the simulated data and the actual data is less than 3%.

Preferably, in the second step, each column of the dynamic data set represents a variable for human intervention, a row represents time, and reflects the change of the variable along with time, and the last column is a network prediction variable and is also a variable which directly affects an accident.

Preferably, the neural learning tool in step four is an LSTM neural network.

Preferably, in the third step, the dynamic data set except the last column is used as the input of the LSTM network, the dynamic data set is divided into a training set and a test set according to the proportion of 8:2, an LSTM model is built based on an open-source python 3.5 code packet, and the network hyper-parameter is adjusted by adopting an orthogonal experiment method

Preferably, in the fourth step, the characteristics of the human intervention variables are learned through a three-door mechanism of the LSTM network, deep relationships among the variables are memorized, and the variables which directly affect the accident consequence are predicted, so that the network achieves the optimal prediction effect on the training set.

The invention also takes the leakage accident of the synthetic ammonia engineering as an example, and particularly describes the flow of the intelligent quantitative risk assessment method provided by the invention, and provides the application of the method in the safety assessment of the chemical process.

The invention has the following beneficial effects:

(1) the invention provides an intelligent quantitative risk assessment method based on a dynamic mechanism model, which realizes the construction of the industrial process dynamic mechanism model, accurately predicts the variable directly influenced by an accident, dynamically simulates the hazard range of the accident, and has higher reliability compared with the classical safety assessment method.

(2) The method of the invention takes the leakage accident of the synthetic ammonia engineering as an example, effectively solves the problems that the safety assessment of the chemical process is mostly based on the heuristic method and needs a large amount of manual experience, improves the safety of the process, prevents unnecessary casualties and property loss caused by the accident, and also provides reference opinions for safe evacuation and plant area arrangement.

Drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, are included to provide a further understanding of the invention, and are incorporated in and constitute a part of this specification, illustrate exemplary embodiments of the invention and together with the description serve to explain the invention and not to limit the invention.

FIG. 1 is a block diagram of DYN-LSTM-QRA method;

FIG. 2 is a PID control schematic;

FIG. 3 is a diagram illustrating a machine learning sequence;

FIG. 4 is a diagram of a long term memory network;

fig. 5 is a flow chart of the QRA method;

FIG. 6 is a steady state flow diagram of a process for ammonia synthesis;

FIG. 7 is a dynamic flow diagram of a process for ammonia synthesis;

FIG. 8 is a graph of feed and leak as a function of time under a feed disturbance;

FIG. 9 is an image of cooling water flow and leakage with time under a water break;

FIG. 10 is a comparison of the orthogonal experimental parameter adjustment process;

FIG. 11 is a graph comparing predicted and actual results for 5 examples;

FIG. 12-1 is a simulated diagram of a jet fire of example c;

FIG. 12-2 is a simulated diagram of an injection fire of example d;

FIG. 12-3 is a simulated view of an injection fire of example e;

FIG. 13 is a simulated diagram of a flash fire of example c;

FIG. 14 is a graph of explosion overpressure versus distance for example c;

fig. 15 is a diagram of a heat exchanger PFD.

Detailed Description

It is to be understood that the following detailed description is exemplary and is intended to provide further explanation of the invention as claimed. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

It is noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments in accordance with the invention. As used herein, the singular is intended to include the plural unless the context clearly dictates otherwise, and it should be further understood that the terms "comprises" and/or "comprising," when used in this specification, specify the presence of features, steps, operations, devices, components, and/or combinations thereof.

Terms and names interpretation:

the PID control is a general name of proportional-integral-derivative feedback control, and a controller applying a PID control law is called a PID controller. The PID controller compares the set value and the output value to form a control deviation, and after proportional, integral and differential operations are carried out on the control deviation, the control deviation is formed by linear combination. The control principle is shown in fig. 2.

Machine learning, as shown in fig. 3, includes three parts, input, network operation and output, and as an important method of machine learning, RNN and LSTM are each characterized, LSTM is a variant RNN and its essence is to introduce a concept of cell state, which is different from RNN only considering the most recent state, and the cell state of LSTM determines which states should be left and which should be forgotten.

The internal working principle of the LSTM neuron is shown in fig. 4. It consists of layer-by-layer and point-by-point operations that provide information on the state of the LSTM unit and maintain long-term memory through the network and input. In contrast to RNN, LSTM has three gates to protect and control the cell state. The input gate controls how much information can flow into the memory unit. Forgetting controls the proportion of information accumulated in the unit from the last time to the current time. The output gate controls the proportion of information in the cell from the current time to the current hidden state. The formula for the three gates is as follows:

i_t＝σ(W_ih_t-1+W_iX_t+b_i) (1)

f_t＝σ(W_fh_t-1+W_fX_t+b_f) (2)

o_t＝σ(W_oh_t-1+W_oX_t+b_o) (3)

i_tis the input gate of the LSTM cell, f_tIs a forgetting gate of the LSTM cell, o_tIs the output gate of the LSTM cell, σ is the activation function, W₍₎Is the weight matrix of the gate, h_t-1Is the output vector of the LSTM cell, X, computed at time t-1_tIs an input value at time t, b₍₎Is the biasing of the door.

The storage unit accumulates history information, which restricts information in a cell to a previous time by forgetting a gate, and restricts the input of new information by an input gate, as shown in the following formula:

wherein

Is a carrier of the LSTM cells,C_tis the state of the LSTM cells at time t.

By introducing memory states and three gates, the LSTM is able to find memorable history data as well as data that should be forgotten.

Quantitative Risk Assessment (QRA) is a Risk prediction method, and is an Analysis technique for assessing system reliability and availability, and is established on the basis of the theory of probability statistics to simulate accidents, and the flow of the method is shown in fig. 5, and the QRA method has been widely used in recent years.

In this embodiment, leakage accidents in the ammonia synthesis engineering are taken as an example, and specifically, the flow of the intelligent quantitative risk assessment method provided by the present invention is described, it should be noted that, after reading this embodiment, a person skilled in the art may transfer the flow to other chemical processes, and should also be used as an application of the technical solution provided by the present invention.

An intelligent quantitative risk assessment method based on a dynamic mechanism model is shown in a general flow chart of fig. 1, and comprises the following specific steps:

step one, establishing a process flow steady-state model through Aspen according to a process flow diagram and process parameters, and ensuring that the steady-state model is consistent with industrial actual data.

Based on the existing equipment model building process flow in Aspen software, the physical property selection method is referred to in the literature. And determining parameters such as pressure drop and temperature difference of each heat exchanger, pipeline feeding amount, temperature, pressure, mole fraction and the like according to the PFD diagram. As shown in the PFD diagram of the heat exchanger in FIG. 15, the temperatures of the

materials

242 and 241 are respectively 40 and 55 degrees Celsius, the flow rate of 242 is 30388kg/h, the heat load of the heat exchanger 207 is-261.04 kw, and the parameters such as the inlet and outlet temperature, the material flow rate and the temperature of the heat exchanger during the Aspen simulation can be determined through the PFD diagram. And finally selecting a Newton iteration method to finish convergence. The steady state flow chart of the synthesis of ammonia is shown in FIG. 6, and the main variables are shown in Table 1. First, fresh synthesis gas enters the compressor (COM1), while fresh air is pressurized through the compression section. The compressed gas is cooled and enters the circulation section of the compressor together with the circulation gas coming out of the cold exchanger (E104). The mixture is finally pressurized to 15mpa and sent to a compressor. The synthesis gas is preheated by a heat exchanger (E102) and then sent to three synthesis ammonia reactors for synthesis ammonia reaction. Reaction gas generated by the synthetic ammonia reactor enters a high-pressure boiler feed water heat exchanger (E101) for heat recovery and then enters a heat exchanger (E102) for preheating, gas at the outlet of a compressor (COM3) is finally cooled to 273K through a water cooler (E103), a cold exchanger (E104) and an ammonia cooler (E105), and condensed liquid ammonia is separated by two-stage ammonia separators (D1 and D2). The circulating gas passes through the cold exchanger and is then introduced into the circulating section of the compressor to be combined with fresh gas, and the cycle is repeated.

Since it is difficult to simulate three-stage ammonia synthesis reaction by one column, the ammonia synthesis column is replaced by 3 reactors (R1, R2, R3) connected in series in the flow. Comparison of the final simulation results with the plant data is shown in table 2, where the steady state model is consistent with the industry practice, with the criterion that the relative error between the simulated data and the actual data is less than 3%.

Table 1 main variables table for ammonia synthesis process

TABLE 2 simulation and comparison Table for ammonia synthesis process

And secondly, adding a controller through Aspen Dynamics, building a dynamic mechanism model, and simulating process leakage to obtain dynamic data sets with different disturbances.

As shown in fig. 7, 3 reactors and 2 separators are respectively added with a pressure controller (PC1, PC2, PC3) and a liquid level controller (LC1, LC2), the parameters of the dynamic analog controller are shown in table 4, and since the controller adopts the PID control principle which is shown in fig. 2, when the controlled material fluctuates, the proportion control immediately generates a control action to reduce the error; the integration link memorizes errors and is mainly used for eliminating static errors and improving the zero-difference degree of the system; the differentiation link reflects the variation trend (change rate) of the deviation signal, and can introduce an effective early correction signal into the system before the value of the deviation signal becomes too large, so that the action speed of the system is accelerated, the regulation time is shortened, and the stable operation of the flow is ensured when disturbance is added. Adding a splitter to simulate leakage, and introducing a code to automatically open a valve when the pressure alarms; the dynamic data set was established by adding feed controllers to the feed stream (FC1), ammonia chiller (E105-FC3), and water cooler (E103-FC2) to set the disturbances. Each column of the dynamic data set represents a variable which is manually intervened, each row represents the change of the reflected variable along with the time at each moment, and the last column is a network prediction variable and is also a variable which has direct influence on accidents. As shown in table 3:

TABLE 3 dynamic data set (part)

As shown in fig. 8 and 9, (a) and (b) in fig. 8 are graphs showing the time-dependent changes of the feeding amount and the leakage amount when the disturbance of 10% and 15% is added to the feeding flow of the system, respectively; fig. 9 (a) and (b) are images showing changes over time in the flow rate and leakage rate of the cooling water when the cooling water is interrupted in the water cooler (E103) and the ammonia cooler (E105), respectively.

TABLE 4 dynamic analog controller parameter table

And step three, except for the last row of network prediction variable rows, taking the dynamic data set as the input of the LSTM network, dividing the dynamic data set into a training set and a testing set according to the ratio of 8:2, building an LSTM model based on an open source python 3.5 code packet, and adjusting network hyper-parameters by adopting an orthogonal experiment method.

According to the step (two), 3000 continuous-time data are collected from each group of dynamic data sets, wherein 2400 is a training set, 600 is a test set, and 20 groups are provided, and the disturbance of 5 cases is simulated, wherein the 5 cases are shown in table 5. The number of layer nodes, the node size and the activation function of the network are discussed respectively by adopting a three-factor three-level orthogonal experiment, and the experimental scheme is shown in table 6. The comparison between the predicted values and the actual values of the 9 sets of experimental networks is shown in fig. 10, and 9 graphs in fig. 10 are images of the predicted values and the actual values of the networks under 9 sets of different hyper-parameters, respectively. As can be seen from the figure, the leakage amount predicted by the network in the scheme D is the smallest difference from the actual leakage amount, so that the final network adopts 30 node numbers, a sigmoid activation function and 100 node sizes for prediction.

TABLE 5 perturbation example information Table

TABLE 6 orthogonal experimental parameters table

And step four, learning the characteristics of the human intervention variables through a three-door mechanism of the LSTM network, memorizing deep relationships among the variables, and predicting the variables which directly affect the accident consequences, so that the network achieves the optimal prediction effect on the training set.

As shown in fig. 11, 5 leakage cases of 10%, 15%, 20%, and 20% of the feed disturbance were predicted, and 5 examples are shown in table 4, where (a), (b), and (c) are images of predicted values and actual values of the network under 10%, 15%, and 20% of the feed disturbance, (d) are predicted images of the network when the cooling water of the water cooler is interrupted, and (e) are predicted images of the network under interruption of the cooling water of the ammonia cooler. It can be seen that the predicted values and the true values of the network differ very little. MAPE is adopted as an index for measurement, and not only is the error between a predicted value and a real value considered, but also the ratio of the predicted value and the real value is considered. The formula is as follows:

MAPE was calculated for samples between 500 and 600 in fig. 11, and the results are shown in table 7. As can be seen from the table, the minimum prediction error occurs in the case (b), and the MAPE in all cases does not exceed 3%, which proves the reliability of the prediction data source for safety analysis, and also proves that the network can well learn the influence of each artificial disturbance variable, and finally accurately predict the leakage amount when a single disturbance occurs. Since there is a small fluctuation in the leakage amount with time due to the presence of industrial noise, in order to input a reasonable net prediction amount in step (5), an average value is taken here, and the prediction and actual values in 5 cases in a state where the leakage amount is relatively stable are shown in table 8.

TABLE 7 prediction MAPE TABLE

Table 8 prediction and comparison table of leakage amount in disturbance example

And (4) taking the leakage quantity predicted by the network obtained in the step four as a main index, wherein the input parameters are shown in a table 9, and finally, the relation between the overpressure and the distance of the injection fire, the flash fire and the explosion is simulated, as shown in figures 12-1, 12-2, 12-3, 13 and 14.

TABLE 9 QRA hazard simulation parameter table

Fig. 12 shows the range of damage of the jet fire caused by cases (c), (d) and (e), respectively, and it can be seen from fig. 12-1 that the heat flux gradually decreases with increasing distance. An elliptical region with a heat flux of 37.5kW/m2 and a major radius of 2.5m was formed about 5m from the center of the leakage. As the leakage distance increases to 35m, the heat flux decreases to 4kW/m 2. In conjunction with table 10, it is shown that the person feels pain in the blue line and outside the green line, the wood burns, first-order burns in the green line and outside the red line, all the operating equipment in the red line is damaged, and the person dies within one minute. The cases (d) and (e) in fig. 12 also show that the dangerous range of the jet fire is proportional to the leak flow rate. In case (d) the leakage flow is small, the simulated hazard range is less than 30 m.

TABLE 10 information table of the danger of the fire

FIG. 13 shows the hazard range of the flashover caused by case (c), with red line being the lower explosion limit and green line being the upper explosion limit. The results show that the gas concentration in the circle with a radius of 20m exceeds the upper explosion limit. Although the mixture does not explode when it comes into contact with air, a combustion phenomenon occurs. When the distance from the leak point is 20m to 45m, explosion occurs in the presence of an open flame. A gas concentration of 45m below the lower explosive limit does not pose a hazard.

The relation between overpressure and explosion distance for case (c) given in fig. 14 shows that the overpressure reaches 18.8kPa within 8m from the center of the leak, which would lead to death of the person. An overpressure (14kPa) of less than 11m will result in partial wall and roof collapse. Less than 20m (7.5kPa) may cause damage to a part of the equipment structure and deformation of the metal plate.

A report of the synthesis of ammonia by HAZOP and DYN-LSTM-QRA is shown in Table 11. The results of HAZOP analysis of process leakage of synthetic ammonia in a factory are as follows: the yield of ammonia gas is reduced, the leakage loss of materials, fire and explosion occur, and equipment is damaged in a large area. The recommendations given are: during the inspection, the pipeline condition is noticed, a toxic gas detector is installed, the open fire operation in the area is eliminated, and emergency leakage repair measures are made.

TABLE 11 Risk analysis report of DYN-LSTM-QRA and HAZOP

The consequences obtained by the proposed DYN-LSTM-QRA method are: the method has the advantages that people die within 8m, walls and roofs collapse within 11m, a combustion phenomenon occurs within 20m, equipment structures are damaged, metal plates deform, heat flux is within a dangerous range of 35m, and 20m to 45m can explode when exposed fire occurs. The recommendations given are: fire-proof coating is added to important equipment within 40m of a leakage point, an electrostatic discharge device is installed within 50m, a flammable and toxic gas detector is installed at a flange interface, and necessary interlocking and alarm equipment is added.

So far, the intelligent quantitative risk assessment method based on the dynamic mechanism model is successfully finished.

The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention, and various modifications and changes may be made by those skilled in the art. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. An intelligent quantitative risk assessment method based on a dynamic mechanism model is characterized by comprising the following steps:

secondly, adding a controller through Aspen Dynamics, and building a dynamic mechanism model to obtain dynamic data sets with different disturbances;

2. The method according to claim 1, wherein the steady-state model in the first step is consistent with industrial practice with a relative error between simulated data and actual data of less than 3%.

3. The intelligent quantitative risk assessment method based on dynamic mechanism model according to claim 1, wherein in the second step, each column of the dynamic data set represents a variable for human intervention, a row represents time, and represents the change of the variable along with time, and the last column is a network predictive variable and is also a variable having direct influence on accidents.

4. The intelligent quantitative risk assessment method based on dynamic mechanism model according to claim 1, wherein the neural learning tool in step four is an LSTM neural network.

5. The intelligent quantitative risk assessment method based on the dynamic mechanism model as claimed in claim 4, characterized in that in the third step, the dynamic data set except the last column is used as the input of the LSTM network, the dynamic data set is divided into a training set and a testing set according to the ratio of 8:2, the LSTM model is built based on an open-source python 3.5 code packet, and the network hyper-parameters are adjusted by adopting an orthogonal experiment method.

6. The intelligent quantitative risk assessment method based on the dynamic mechanism model according to claim 4, characterized in that in the fourth step, the characteristics of the human intervention variables are learned through the three-door mechanism of the LSTM network, the deep relationship between the variables is memorized, and the variables which directly affect the accident outcome are predicted, so that the network achieves the optimal prediction effect on the training set.