CN103914622A - Quick chemical leakage predicating and warning emergency response decision-making method - Google Patents

Abstract

The invention discloses a chemical leakage rapid prediction, early warning and emergency response decision-making method, which combines the diffusion model simulation with the neural network and the gas sensor system, and is applied to the rapid early warning and auxiliary decision-making of harmful gas leakage in industrial parks, including: park risk factor identification, It is used to identify various possible leakage accidents; numerical simulation, to simulate all possible accidents to obtain the influence range of harmful gases; data screening, to extract and reorganize the effective part of the numerical simulation results according to the actual sensor layout; neural network Training, using the filtered data to train a specific neural network model to obtain model parameters for specific industrial parks and surrounding conditions, and use redundant data parameter verification; sensor system and neural network model integration, combining the model with sensor DCS stand up.

Description

A decision-making method for rapid prediction, early warning and emergency response of chemical leakage

technical field

The invention belongs to the technical field of safety early warning in industrial production, and in particular relates to a decision-making method for rapid chemical leakage prediction, early warning and emergency response.

Background technique

Many process industries will use or produce some poisonous and harmful gases (such as chlorine gas, phosgene, etc.) that are harmful to human body during the production process. It can cause serious harm to humans within a certain range. When a toxic and harmful gas leakage accident occurs, the leaked substance and the approximate location of the leak can be relatively easily determined, but it is difficult to obtain the leakage rate or leakage rate of the harmful gas on site. It is an important part of the accident emergency response process to predict the diffusion trend and influence range of toxic gas with limited information in a limited time. At present, traditional numerical diffusion prediction models, including Gaussian plume models, computational fluid dynamics models, and some relatively mature diffusion simulation software require users to provide detailed leakage rates of leakage sources, and it takes a certain amount of time to calculate and obtain results. Therefore, the leakage rate or leakage of the leakage source and the long calculation time limit the application of these traditional gas diffusion models in the accident emergency response and auxiliary decision-making process, and they are more used in the investigation and analysis after the accident. Due to historical reasons, a considerable number of industrial parks in my country where toxic and harmful gas leakage accidents may occur have residential areas within 3-5km. When serious toxic and harmful gas leakage accidents occur, these toxic and harmful gases are likely to spread to the industrial park boundary Outside the area, it poses a threat to residential areas. When there is no method to quickly and effectively predict the spread of toxic and harmful gases, decision makers often consider the coverage of harmful gases in the worst case, and the worst case prediction results often mean that government departments need to evacuate tens of thousands or even hundreds of thousands Man, this is very unrealistic. Therefore, a method to quickly and effectively predict the diffusion range of harmful gases is of great significance for accident emergency response and auxiliary decision-making.

Factories dealing with toxic and harmful gases will set up one or more harmful gas sensors near the toxic and harmful gas storage tanks to monitor whether there is any gas leakage. These gas sensors are all connected with the control center to provide toxic gas leakage alarm information. However, at present, many industrial parks do not give full play to the role of gas sensors. When an accident occurs, emergency response personnel can only obtain the type and instantaneous concentration of the leaked gas through the sensor, but cannot quickly predict the distribution of the leaked gas through the information, and then take measures. Effective control measures and a reasonable evacuation plan.

At present, the accident consequence analysis methods commonly used in the world mainly include the use of different types of numerical models (Gaussian plume model, computational fluid dynamics model (CFD) and commercial models PHAST, FLACS, etc. to reproduce the accidents that have occurred, as shown in The scope of influence of the leaked gas under the conditions of the accident leak at that time (including the death zone, serious injury zone and impact zone), and the influence of atmospheric diffusion conditions on the diffusion range and concentration distribution of the leaked gas. But the disadvantage of using numerical model analysis is that the leak must be known The source strength (leakage rate) and leakage form (explosion leakage/aperture leakage, etc.) of the source are simulated in combination with meteorological parameters and mass transfer and diffusion equations. The model is complex and the calculation takes a long time, so it cannot be used for real-time or rapid prediction under accident conditions.

Contents of the invention

In order to overcome the above-mentioned shortcomings of the prior art, the object of the present invention is to provide a decision-making method for rapid prediction, early warning, and emergency response of chemical leakage, which can conduct risk analysis on industrial parks, simulate leakage scenarios, and appropriately supplement and optimize the existing toxic and harmful gases in the park. The sensor system combines the toxic gas alarm system in the park with gas diffusion prediction and analysis to provide technical support for accident emergency response decision-making.

In order to achieve the above object, the technical scheme adopted in the present invention is:

A decision-making method for rapid prediction, early warning, and emergency response of chemical leakage, including five stages of risk factor identification, scenario numerical simulation, simulation result screening, neural network training, and integration of training results and sensor systems, in which:

Identification of risk factors includes: identification of risk factors in chemical industrial parks, quantification of each risk factor, and rational selection and combination of various possible leakage scenarios based on the identified risk factors and their value ranges;

The scenario numerical simulation stage includes: simulating all possible leakage scenarios formed in the park risk factor identification step to obtain the influence range of leaked gas under different leakage scenarios;

Simulation result screening includes:

The key parameters in a layout plan including 4 gas sensors are optimized, and the leaked gas concentration distribution obtained by simulation is used for simple optimization to obtain the most suitable sensor layout;

According to the determined sensor layout scheme, extract the virtual detection data within the measurable wind direction range of the sensor layout and the matching leakage gas diffusion data of environmentally sensitive points, and prepare training data and verification data according to the training logic of the neural network;

Neural network training steps include:

Establish a forward neural network for function fitting, use the prepared training data as the input and output of neural network training, and calculate network parameters;

Input the test data input part into the neural network, and compare the result part with the prediction result of the neural network to evaluate the prediction accuracy;

Steps to integrate neural network with sensors and park control system:

Combine the simulation analysis and training results of each risk source with the geographical location information of risk sources and environmental sensitive points in the form of database and call program, and provide a data interface with the sensor system to realize rapid prediction from accident occurrence-sensor alarm-model - Aided decision-making workflow.

The park risk identification part consists of the following steps:

Step 1. Use risk assessment to identify risk factors and accident scenario factors and determine the necessary elements of various possible leakage scenarios, including risk elements and accident scenario elements. These elements include but are not limited to the type of risk substances (toxic/flammable gas or volatile liquid), storage/amount, storage location (spatial coordinates), storage form (pressure, temperature, equipment), wind speed, humidity, temperature, atmosphere Stability, the location of sensitive points in the surrounding environment (spatial coordinates), etc.

Step 2. Determine the value range of each risk factor and accident situation factor (for continuous variables), and divide the numerical sequence with a certain step size within the value range of each factor (the step size must not exceed 10% of the value range of the factor). %), and finally combine the values of different risk factors/accident scenario factors into a large number of possible leakage scenarios. If there are n risk factors in total, and each risk factor has N _i values (N _i ≥ 1, integer), then there are a leak scenario.

The numerical models used in the numerical simulation stage of the scenario include: Gaussian and Gaussian-like diffusion models (Gaussian Plume Model), computational fluid dynamics (CFD) models, and integrated models such as PHAST, FLACS, SLAB, ALOHA, and HGSYSTEM. Finally, the simulation results of J leakage scenarios are obtained.

The simulation result screening step needs to optimize the layout of a group of 4 sensors, and the 4 sensors adopt a regular polygonal structure, the sequence numbers are 1-4, and the distances between adjacent numbered sensors are equal. The optimization steps are as follows:

Step 1: Determine the value ranges of the basic layout parameters L ₁ , d and α according to the location of the leakage source, the layout of the park, and the detection limit of the sensor. Among them, _L1 is the distance between No. 1 sensor and the leakage source, d is the distance between sensors, and α is 1/2 of the angle between 1-2 and 1-4 sensors.

Step 2: Determine the prevailing wind direction according to the wind frequency information, the sensors are arranged symmetrically in the downwind direction of the prevailing wind, and the angle between the actual wind direction and the prevailing wind direction is β. And take a larger value β ₀ as the optimization interval, discretize it into the actual wind direction sequence (β ₁ ,β ₂ ,…,β _M ), the sequence length is M. (Generally take β ₀ =50°, and discretize the interval [-50°, 50°] into a wind direction sequence with an interval of 1°)

Step 3. Discretize the layout parameters L1, d, and α into parameter sequences within the value range, with lengths P ₁ , P ₂ , and P ₃ , consisting of layout scheme.

Step 4: Use the data screening algorithm for the kth layout scheme (k=1,2,...,K), and filter out the layout scheme in The number of effective layouts when the wind direction is β _m (m=1,2,…,M)

Step 5. Calculate the scenario applicability rate of different wind directions β _m (m=1,2,...,M) for the kth layout scheme (k=1,2,...,K) records meet The maximum wind direction range W _k of m=1,2,...,M, the layout satisfying {max(W _k ),k=1,2,...,K} is taken as the optimal layout.

The criterion for judging the effective layout in the simulation result screening step is: for any sensor layout, for each leakage scenario in a specific wind direction, if 3 or more of the 4 sensors are covered by the plume, the layout is said to be in the air. Effective layout for this wind direction and leak scenario.

In the screening step of the simulation results, the scenario applicability rate is (m=1,2,...,Mk=1,2,...,K).

The simulation result data screening part includes the following screening algorithms:

Step 1. Check whether the sensor layout parameters L1, d, α, the actual wind direction β, the position of the sensitive source in the downwind direction (L2, Y2) and the serial number J of the leakage scenario meet the requirements

Step 2. Interpolate the gas diffusion distance, concentration distribution, diffusion time and plume width of the simulation results of the jth leakage scenario (j=1,2,...,J), and determine the starting point of screening

Step 3. Calculate parameters γ, θ ₂ , θ ₄ , W ₁ , and W ₂ , where γ is the angle between sensors 2 and 4 with the leakage source as the vertex and the dominant wind direction, θ ₂ and θ ₄ are sensors 2 and 4 Taking the leakage source as the angle between the vertex and the actual wind direction, W ₁ is the vertical distance from the environmental sensitive point to the plume centerline, W ₂ is the half-width of the plume at the environmentally sensitive point

Step 4. Check whether the sensor layout is effective in the leakage scenario and the actual wind direction, if yes, continue; if not, skip and record the leakage scenario, and continue to step 2

Step 5: Extract the pollution source concentration and diffusion time of the three effective sensor locations.

Step 6: Check whether the leaked gas can diffuse to the environmental sensitive point in the downwind direction, that is, calculate the location parameter Q _j of the environmentally sensitive point in the jth leakage scenario. If yes, proceed to step seven, if not, record the serial number and proceed to step two.

Step 7: Extract the concentration and diffusion time of the harmful gas at the environmental sensitive point in the downwind direction.

Step 8: Arranging and saving all extracted sensor position concentration and time values and downwind sensitive point concentration and time values.

The method for judging whether the leaked gas can diffuse to the environmentally sensitive point in the downwind direction is: calculate the position parameter Q _j =W ₁ /W ₂ of the environmentally sensitive point in the gas leakage plume, if Q _j ≤ 1, the environment is sensitive Points are affected by the diffusion of leaked gas, and vice versa.

The implementation process of the neural network training part is as follows:

Step 1. Determine the network input, using parameters available on site as input, including storage pressure (p), wind speed (v), wind direction (β), temperature (T), humidity (h), atmospheric stability (S), The effective concentration measurement values (c ₁ , c ₂ , c ₃ ) and alarm time (t ₁ , t ₂ , t ₃ ) of the three gas sensors in the downwind direction of the risk source, and the positions of the downwind sensitive points (L ₂ , Y ₂ )

Step 2: Determine the network output, the concentration (c) and time (t) of the harmful gas arriving at the environmental sensitive point in the downwind direction are taken as the output.

Step 3: Determine the internal structure of the network, using a feed-forward neural network (BP network), including a nonlinear hidden layer (sigmoid transfer function) and a linear output layer (linear transfer function), and the network does not include a feedback loop.

Step 4. According to the input and output requirements, use the data screening algorithm to prepare the input and output data matrix and redundant verification data. (The columns of the matrix are the input and output elements, and the input and output data of different scenarios)

Step 5. Use the MATLAB neural network toolbox to train the neural network using input and output data according to a single unified training method to obtain network parameters.

Step 6. Use the redundant input data as the input of the neural network, compare the difference between the network output and the redundant output, and evaluate the prediction accuracy.

The neural network and sensor system integration part includes the following steps:

Step 1. Establish a regional map of a single risk source and environmentally sensitive points in the leeward direction of the dominant wind.

Step 2: Carry out risk analysis on a single risk source, identify possible leakage scenarios and simulate the leakage scenarios.

Step 3: Perform sensor layout optimization in the data screening step to obtain the maximum applicable angle of the sensor layout in the leeward direction of the prevailing wind.

Step 4: Perform data screening on all environmental sensitive points within the maximum use angle range of the dominant wind downwind of the risk source, and obtain a set of input and output matrices for neural network training.

Step 5: Use the input and output matrix of the neural network obtained through data screening for training and evaluation of prediction accuracy.

Step 6: Making the trained neural network parameters into an application program that can be called automatically, so that the DCS data of the sensor can be collected through the data interface and calculated using the application program.

Step 7. When a real accident occurs, the sensor alarms—starts the prefabricated application—predicts the impact of leakage diffusion on environmentally sensitive points (including whether it affects the environment sensitive point or the time and concentration of the leaked gas reaching the environment sensitive point)—judgments by decision makers Whether evacuation is required.

Compared with the prior art, the present invention proposes a simple optimization method for the sensor layout of factories dealing with toxic and harmful gases, and introduces a rapid prediction method using early risk identification, numerical simulation and neural network training, so that when an accident occurs, it can Timely and effectively predict the diffusion range of toxic and harmful gases, and provide decision support for the evacuation and protection of people in residential areas around the park. On the basis of the existing toxic and harmful gas sensors in the park, according to the production process and the surrounding environment of the park, the risk identification in the production process of the park, the simulation of leakage scenarios, the improvement of the sensor layout and the training of the neural network have been completed to overcome the toxic gas leakage accident. The insufficiency of the alarm information to support the emergency response to the accident ensures the applicability and accuracy of the method.

Description of drawings

Fig. 1 is a flowchart of a method of an embodiment of the present invention.

Fig. 2 is a schematic diagram of an application scenario of a method according to an embodiment of the present invention.

Fig. 3 is a sensor layout optimization curve obtained by a method of an embodiment of the present invention.

Fig. 4 is the basic algorithm used for data screening by an embodiment method of the present invention

FIG. 5 is a structural diagram of a neural network used in a method according to an embodiment of the present invention.

Fig. 6 is a verification of redundant data of the neural network applied by the method of an embodiment of the present invention.

Fig. 7 is a judgment of whether the sensitive area is affected by the neural network applied in the method of an embodiment of the present invention.

Detailed ways

The implementation of the present invention will be described in detail below in conjunction with the drawings and examples.

Fig. 1 is the flowchart of the method according to one embodiment of the present invention, the method according to one embodiment of the present invention comprises:

-Risk factor identification steps

Use risk assessment to identify risk factors and accident scenario factors and determine the necessary elements of various possible leakage scenarios, including risk factors and accident scenario elements. These elements include but are not limited to the type of risk substances (toxic/flammable gas or volatile liquid), storage/amount, storage location (spatial coordinates), storage form (pressure, temperature, equipment), wind speed, humidity, temperature, atmosphere Stability, the location of sensitive points in the surrounding environment (spatial coordinates), etc. Then determine the value ranges of each risk factor and accident situation factor (for continuous variables), and divide the numerical sequence with a certain step size within the value range of each factor (the step size must not exceed 10% of the value range of the factor) , and finally combine the values of different risk factors/accident scenario factors into a large number of possible leakage scenarios. If there are n risk factors in total, and each risk factor has N _i values (N _i ≥ 1, integer), then there are a leak scenario.

—Simulation steps of leakage scenario

In this step, the traditional gas diffusion model is used to numerically simulate the various leakage scenarios identified in the previous step to obtain the diffusion data and distribution of toxic and harmful gases under different conditions. Available models include Gaussian and Gaussian-like diffusion models (Gaussian Plume Model), computational fluid dynamics (CFD) models, and PHAST, FLACS, SLAB, ALOHA, HGSYSTEM, etc.

— Valid data screening steps

Figure 2 shows the location relationship diagram of the risk source-plume-sensor-downwind sensitive point in the effective data screening step. The simulation result screening step needs to optimize the layout of a group of 4 sensors, and the 4 sensors adopt a regular polygonal structure, numbered sequentially from 1 to 4, and the distance between adjacent numbered sensors is equal. The optimization steps are as follows:

Step 1: Determine the value ranges of the basic layout parameters L ₁ , d and α according to the location of the leakage source, the layout of the park, and the detection limit of the sensor. Among them, _L1 is the distance between No. 1 sensor and the leakage source, d is the distance between sensors, and α is 1/2 of the angle between 1-2 and 1-4 sensors.

Step 2: Determine the prevailing wind direction according to the wind frequency information, the sensors are arranged symmetrically in the downwind direction of the prevailing wind, and the angle between the actual wind direction and the prevailing wind direction is β. And take a larger value β ₀ as the optimization interval, discretize it into the actual wind direction sequence (β ₁ ,β ₂ ,…,β _M ), the sequence length is M. (Generally take β ₀ =50°, and discretize the interval [-50°, 50°] into a wind direction sequence with an interval of 1°)

Step 3. Discretize the layout parameters L1, d, and α into parameter sequences within the value range, with lengths P ₁ , P ₂ , and P ₃ , consisting of layout scheme.

Step 4: Use the data screening algorithm for the kth layout scheme (k=1,2,...,K), and filter out the layout scheme in The number of effective layouts when the wind direction is β _m (m=1,2,…,M) The effective layout is defined as: for any sensor layout, for each leakage scenario in a specific wind direction, three or more of the four sensors are covered by the plume (that is, the plume concentration can be detected).

Step 5. Calculate the scenario applicability rate of different wind directions β _m (m=1,2,...,M) for the kth layout scheme (k=1,2,...,K) (m=1,2,…,M k=1,2,…,K), records satisfy The maximum wind direction range W _k of m=1,2,...,M, the layout satisfying {max(W _k ),k=1,2,...,K} is taken as the optimal layout.

Figure 3 is the layout of 27 layout schemes consisting of the values of L ₁ =[40,60,80]m, d=[20,30,40]m and α=[30°,45°,60°] Optimization results (Fig. 3d). Under the condition that the scenario usage rate is 95%, the sensor layout scheme L ₁ =40m, d=20m and α=60° is the optimal layout, and the maximum wind direction applicable range of this layout is -20°～20°. After determining the optimal sensor layout, use the screening algorithm shown in Figure 4 to screen the simulation results:

Step 1. Check whether the sensor layout parameters L1, d, α, the actual wind direction β, the position of the sensitive source in the downwind direction (L2, Y2) and the leakage scenario number j meet the requirements

Step 2. Interpolate the gas diffusion distance, concentration distribution, diffusion time and plume width of the simulation results of the jth leakage scenario (j=1,2,...,J), and determine the starting point of screening

Step 3. Calculate parameters γ, θ ₂ , θ ₄ , W ₁ , and W ₂ , where γ is the angle between sensors 2 and 4 with the leakage source as the vertex and the dominant wind direction, θ ₂ and θ ₄ are sensors 2 and 4 Taking the leakage source as the angle between the vertex and the actual wind direction, W ₁ is the vertical distance from the environmental sensitive point to the plume centerline, W ₂ is the half-width of the plume at the environmentally sensitive point

Step 4. Check whether the sensor layout is effective in the leakage scenario and the actual wind direction (the number of effective sensors is greater than 3), if so, continue; if not, skip and record the leakage scenario, and continue to step 2

Step 5: Extract the pollution source concentration and diffusion time of the three effective sensor locations.

Step 6: Check whether the leaked gas can spread to the downwind environmentally sensitive point, that is, calculate the location parameter Q _j =W ₁ /W ₂ of the environmentally sensitive point in the jth leakage scenario. If yes, ie Q _j ≤ 1, go to step 7; if not, ie Q _j > 1, record the sequence number and go to step 2.

Step 7: Extract the concentration and diffusion time of the harmful gas at the environmental sensitive point in the downwind direction.

Step 8: Arranging and saving all extracted sensor position concentration and time values and downwind sensitive point concentration and time values.

—Neural network training steps

The implementation process of the neural network training part is as follows:

Step 1. Determine the network input, using parameters available on site as input, including storage pressure (p), wind speed (v), wind direction (β), temperature (T), humidity (h), atmospheric stability (S), The effective concentration measurement values (c ₁ , c ₂ , c ₃ ) and alarm time (t ₁ , t ₂ , t ₃ ) of the three gas sensors in the downwind direction of the risk source, and the positions of the downwind sensitive points (L ₂ , Y ₂ )

Step 2: Determine the network output, the concentration (c) and time (t) of the harmful gas arriving at the environmental sensitive point in the downwind direction are taken as the output.

Step 3: Determine the internal structure of the network, as shown in Figure 5, use the feedforward neural network (BP network), including a nonlinear hidden layer (sigmoid transfer function) and a linear output layer (linear transfer function), the network does not contain feedback loop.

Step 4. According to the input and output requirements, use the data screening algorithm to prepare the input and output data matrix and redundant verification data. (The columns of the matrix are the input and output elements, and the input and output data of different scenarios)

Step 5. Use the MATLAB neural network toolbox to train the neural network using input and output data according to a single unified training method to obtain network parameters.

Step 6. Use the redundant input data as the input of the neural network, compare the difference between the network output and the redundant output, and evaluate the prediction accuracy. Figure 6 shows the neural network prediction situation of a chlorine gas leakage embodiment. The vertical coordinates of all the subgraphs in the figure are under the specific leakage situation, according to the alarm information of the three sensors, weather conditions and the location of the sensitive area in the downwind direction, the neural network is used to analyze the location of the sensitive area in the downwind direction.

(a) Chlorine concentration (ppm)

(b) The time (s) for chlorine gas to reach the sensitive location

(c) The half-width W2 of the plume at the sensitive position (m)

(d) The vertical distance W1 (m) between the sensitive position and the actual wind direction

The abscissas in Figures 6a to 6d are the results of Phast simulation. In theory, the fitting result of the scattered points in the figure should be a straight line y=x. The actual results show that only the neural network prediction results of chlorine concentration are compared with the Phast simulation results. On the small side (slope about 0.93), the chlorine arrival time, W1 and W2 predictions are very accurate.

Using the values of W1 and W2 obtained in Figures 6c and 6d can predict whether the downwind sensitive area is threatened by chlorine gas leakage, as shown in Figure 7, the abscissa is the number of leakage scenarios used:

(a) Relative error between neural network concentration prediction value and Phast simulation value

(b) W1/W2 ratio calculated from W1 and W2 predicted by the neural network in Fig. 6

(c) W1/W2 ratio based on Phast simulation and selected points in the plume

(d) The logical result obtained by subtracting the predicted W1/W2 and theoretical W1/W2 values after rounding

According to (a), it can be found that the maximum error between the predicted value of the neural network and the theoretical value does not exceed 30%. Figures 7b and 7c are very close, and the difference is shown in 7d: Among the 2780 leakage scenarios, there are 2 leakage scenarios Sensitive points in wind direction are theoretically threatened by chlorine gas leakage, but the results of neural network predictions show that they are not threatened, which is called "missing negative", with a false negative rate of 0.072%; In theory, it is not threatened, but the neural network prediction shows that it is threatened, which is called "false positive", and the false positive rate is 0.504%.

Step 7: Making the trained neural network parameters into an application program that can be called automatically, so that the DCS data of the sensor can be collected through the data interface and calculated using the application program.

Step 8. When a real accident occurs, the sensor alarms—starts the prefabricated application—predicts the impact of leakage diffusion on environmentally sensitive points (including whether it affects environmentally sensitive points or the time and concentration of leaked gas reaching environmentally sensitive points)—judgment by decision makers Whether evacuation is required.

Claims (10)

1. A decision-making method for rapid prediction, early warning, and emergency response to chemical leakage, comprising five stages of risk factor identification, scenario numerical simulation, simulation result screening, neural network training, and integration of training results and sensor systems, characterized in that: Identification of risk factors includes: identification of risk factors in chemical industrial parks, quantification of each risk factor, and rational selection and combination of various possible leakage scenarios based on the identified risk factors and their value ranges; The scenario numerical simulation stage includes: simulating all possible leakage scenarios formed in the park risk factor identification step to obtain the influence range of leaked gas under different leakage scenarios; Simulation result screening includes: The key parameters in a layout plan including 4 gas sensors are optimized, and the leaked gas concentration distribution obtained by simulation is used for simple optimization to obtain the most suitable sensor layout; According to the determined sensor layout scheme, extract the virtual detection data within the measurable wind direction range of the sensor layout and the matching leakage gas diffusion data of environmentally sensitive points, and prepare training data and verification data according to the training logic of the neural network; Neural network training steps include: Establish a forward neural network for function fitting, use the prepared training data as the input and output of neural network training, and calculate network parameters; Input the test data input part into the neural network, and compare the result part with the prediction result of the neural network to evaluate the prediction accuracy; Steps to integrate neural network with sensors and park control system: Combine the simulation analysis and training results of each risk source with the geographical location information of risk sources and environmental sensitive points in the form of database and call program, and provide a data interface with the sensor system to realize rapid prediction from accident occurrence-sensor alarm-model - Aided decision-making workflow. 2. The chemical leakage rapid prediction, early warning and emergency response decision-making method according to claim 1, wherein the risk identification part of the park consists of the following steps: Step 1. Use risk assessment to identify risk factors and accident scenario factors and determine the necessary elements of various possible leakage scenarios, including risk elements and accident scenario elements; Step 2. Determine the value range of each risk factor and accident situation factor respectively, and divide the numerical sequence with a certain step size within the value range of each factor, and the step size shall not exceed 10% of the value range of the factor. Finally, the The values of different risk factors/accident scenario factors are combined into a large number of possible leakage scenarios. When there are n risk factors in total, and each risk factor has N _i values (N _i ≥ 1, integer), then there are a leak scenario. 3. The chemical leakage rapid prediction early warning emergency response decision-making method according to claim 1 is characterized in that, the numerical model used in the numerical simulation stage of the scenario comprises: Gaussian and Gaussian-like diffusion model (Gaussian Plume Model), computational fluid Mechanical (CFD) model and integrated model of PHAST, FLACS, SLAB, ALOHA and HGSYSTEM, finally obtained the simulation results of J kinds of leakage scenarios. 4. The decision-making method for chemical leakage rapid prediction, early warning and emergency response according to claim 1, characterized in that, the simulation result screening step optimizes the layout of a group of 4 sensors, and the 4 sensors adopt a regular polygonal structure, and the order The numbers are 1 to 4, and the distance between adjacent numbered sensors is equal. The optimization steps are as follows: Step 1. Determine the value ranges of the basic layout parameters L ₁ , d and α according to the location of the leakage source, the layout of the park, and the detection limit of the sensors, where L ₁ is the distance from the No. 1 sensor to the leakage source, and d is the distance between the sensors. α is 1/2 of the angle between 1-2 and 1-4 sensors; Step 2. Determine the dominant wind direction according to the wind frequency information. The sensors are symmetrically arranged in the downwind direction of the dominant wind direction. The angle between the actual wind direction and the dominant wind direction is β, and a larger value β ₀ is taken as the optimization interval, which is discretized into the actual wind direction Sequence (β ₁ ,β ₂ ,…,β _M ), the sequence length is M; Step 3. Discretize the layout parameters L ₁ , d, and α into parameter sequences within the value range, with lengths P ₁ , P ₂ , and P ₃ , consisting of a layout scheme; Step 4: Use the data screening algorithm for the kth layout scheme (k=1,2,...,K), and filter out the layout scheme in The number of effective layouts when the wind direction is β _m (m=1,2,…,M) Step 5. Calculate the scenario applicability rate of different wind directions β _m (m=1,2,...,M) for the kth layout scheme (k=1,2,...,K) records meet The maximum wind direction range W _k of m=1,2,...,M, the layout satisfying {max(W _k ),k=1,2,...,K} is taken as the optimal layout. 5. The decision-making method for chemical leakage rapid prediction, early warning and emergency response according to claim 4, characterized in that, the criterion for judging the effective layout in the simulation result screening step is: for any sensor layout, for each In the leakage scenario, if three or more of the four sensors are covered by the plume, the layout is called an effective layout under the conditions of the wind direction and leakage scenario. 6. The chemical leakage rapid prediction, early warning, emergency response decision-making method according to claim 4 or 5, characterized in that, in the step of screening the simulation results, the scenario applicability rate is (m=1, 2, . . . , Mk=1, 2, . . . , K). 7. The chemical leakage rapid prediction, early warning and emergency response decision-making method according to claim 6, wherein the simulation result data screening part includes the following screening algorithms: Step 1. Check whether the sensor layout parameters L1, d, α, the actual wind direction β, the position of the sensitive source in the downwind direction (L2, Y2) and the serial number J of the leakage scenario meet the requirements; Step 2. Interpolate the gas diffusion distance, concentration distribution, diffusion time and plume width of the simulation results of the jth leakage scenario (j=1,2,...,J), and determine the screening starting point; Step 3. Calculate the parameters γ, θ ₂ , θ ₄ , W ₁ , and W ₂ , where γ is the angle between sensor No. 2 or 4 with the leak source as the vertex and the prevailing wind direction, and θ ₂ is the angle between sensor No. 2 and the leak source is the angle between the apex and the actual wind direction, θ ₄ is the angle between the apex of sensor No. 4 and the actual wind direction with the leakage source as the apex, W ₁ is the vertical distance from the environmental sensitive point to the plume centerline, W ₂ is the distance between the environmental sensitive point Plume half width; Step 4. Check whether the sensor layout is effective in the leakage scenario and the actual wind direction, if so, continue; if not, skip and record the leakage scenario, and continue to step 2; Step 5, extract the pollution source concentration and diffusion time of the three effective sensor positions; Step 6. Check whether the leaked gas can spread to the environmental sensitive point in the downwind direction, that is, calculate the position parameter Q _j of the environmental sensitive point in the jth leakage scenario. If yes, continue to step 7. If not, record the serial number and continue to step two; Step 7, extracting the concentration and diffusion time of the harmful gas at the environmental sensitive point in the downwind direction; Step 8: Arranging and saving all extracted sensor position concentration and time values and downwind sensitive point concentration and time values. 8. The decision-making method for rapid prediction, early warning and emergency response of chemical leakage according to claim 1 or 7, characterized in that the method for judging whether the leaked gas can diffuse to environmentally sensitive points in the downwind direction is: calculating the environmental sensitive points at The position parameter Q _j =W ₁ /W ₂ in the gas leakage plume, if Q _j ≤1, the environmental sensitive point is affected by the diffusion of the leakage gas, otherwise it is not affected. 9. The chemical leakage rapid prediction, early warning, and emergency response decision-making method according to claim 1, wherein the neural network training part is implemented as follows: Step 1. Determine the network input, using parameters available on site as input, including storage pressure (p), wind speed (v), wind direction (β), temperature (T), humidity (h), atmospheric stability (S), The effective concentration measurement values (c ₁ , c ₂ , c ₃ ) and alarm time (t ₁ , t ₂ , t ₃ ) of the three gas sensors in the downwind direction of the risk source, and the positions of the downwind sensitive points (L ₂ , Y ₂ ) ; Step 2, determine the network output, the concentration (c) and time (t) of the harmful gas arriving at the environmental sensitive point in the downwind direction are used as the output; Step 3. Determine the internal structure of the network. Use the feedforward neural network, that is, the BP network, which includes a nonlinear hidden layer and a linear output layer. The nonlinear hidden layer uses a sigmoid transfer function, and the linear output layer uses a linear transfer function. The network does not Contains feedback loops; Step 4. According to the input and output requirements, use the data screening algorithm to prepare the input and output data matrix and redundant verification data; Step 5. Use the MATLAB neural network toolbox to train the neural network using input and output data according to a single unified training method to obtain network parameters; Step 6. Use the redundant input data as the input of the neural network, compare the difference between the network output and the redundant output, and evaluate the prediction accuracy. 10. The chemical leakage rapid prediction, early warning and emergency response decision-making method according to claim 1, wherein the neural network and sensor system integration part includes the following steps: Step 1. Establish a regional map of single risk sources and environmentally sensitive points in the leeward direction of the dominant wind; Step 2. Conduct risk analysis on a single risk source, identify possible leakage scenarios and simulate the leakage scenarios; Step 3, perform the sensor layout optimization of the data screening step, and obtain the maximum applicable angle of the sensor layout in the leeward direction of the dominant wind; Step 4. Perform data screening on all environmentally sensitive points within the maximum use angle range of the dominant wind downwind of the risk source, and obtain a set of input and output matrices for neural network training; Step 5. Use the input and output matrix of the neural network obtained by data screening for training and evaluation of prediction accuracy; Step 6. Making the trained neural network parameters into an application program that can be called automatically, so that the DCS data of the sensor can be collected through the data interface and calculated using the application program; Step 7. When a real accident occurs, the sensor reports an alarm—starts a prefabricated application—predicts the impact of leakage and diffusion on environmentally sensitive points—decision makers judge whether evacuation is necessary.