Green ammonia process synthesis section abnormal condition early warning method
Through BiLSTM-BiGRU hybrid neural network and multi-view angle aggregation adaptive encoder, a key variable library and early warning threshold library of the chlorammonia process are built, and early warning and diagnosis of abnormal conditions in the chlorammonia process is realized, and the equipment operation is not stable and safety risks issues improve the safety and stability of the process.
Patent Information
- Application Number
- CN202411188868.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-28
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2044-08-28
AI Technical Summary
In the urinary ammonia process, the equipment operation is unstable due to the volatility of renewable energy and frequent load regulation, which increases safety risks, and it is difficult for the existing technology to effectively warn and manage abnormal conditions.
Using BiLSTM-BiGRU hybrid neural network model combined with multi-view angle aggregation adaptive encoder (MVAAE), a key variable library and early warning threshold library for the chlorammonia process are constructed. Through real-time data prediction and adaptive sliding window adjustment, the abnormal situation is achieved in advance Early warning and diagnosis.
The safety and stability of the chlorammonia process are improved, and through accurate early warning and diagnosis, the occurrence of equipment failures and safety accidents is reduced, and the controllability of the production process is enhanced.
Smart Images

Figure CN119087930A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of green hydrogen to ammonia production, and in particular to a method for early warning of abnormal conditions in a synthesis section of a green ammonia process. Background Art
[0002] As a clean and efficient energy carrier, green ammonia plays an important role in the modern chemical industry. Compared with the traditional Haber-Bosch ammonia synthesis process, the green ammonia process has lower carbon emissions due to the different original source of hydrogen, and mainly uses renewable energy. It has received more attention in the global pursuit of carbon neutrality goals. With the increasing attention paid to environmental protection and the popularization of the concept of sustainable development, the research and application of green ammonia technology has become an important development direction of the chemical industry. However, in the green ammonia production process, due to the volatility and intermittent nature of renewable energy such as wind energy and solar energy, it is difficult to match with the traditional continuous ammonia synthesis process. Therefore, in recent years, scholars have proposed a flexible ammonia synthesis technology for application in the green ammonia process. Flexible ammonia synthesis technology can support the regulation of production load, match the fluctuating wind and solar resources, realize "production follows source movement", and effectively improve the energy utilization efficiency and adaptability of the ammonia synthesis process. However, the synthetic ammonia reaction itself is already quite risky, and the frequent load regulation in the green ammonia process brings many challenges to the safety of the production process: frequent load adjustment may cause changes in operating parameters such as temperature, pressure and flow, which increases the instability of equipment operation, thereby causing problems such as loss of control and increased equipment wear; gas tanks, compressors and other equipment may face additional stress and wear under frequent load adjustment, resulting in safety problems such as equipment failure or leakage. The above problems pose a greater challenge to the production safety of the green ammonia process. Since the green ammonia process needs to react in a synthesis tower under high temperature and high pressure conditions, and ammonia is flammable and explosive and toxic to the human body, once a safety accident such as explosion or leakage occurs, the consequences are disastrous. Therefore, the requirements for abnormal condition management of the green ammonia process cannot be limited to real-time monitoring and diagnosis, but also need to be warned in advance before the abnormal situation occurs, giving on-site technicians sufficient operating time to avoid abnormalities and failures as much as possible, and ensure the safe and stable production process. Summary of the invention
[0003] The present invention provides a method for early warning of abnormal conditions in a synthesis section of a green ammonia process to solve at least one of the above technical problems.
[0004] To solve the above problems, as one aspect of the present invention, a method for early warning of abnormal conditions in a synthesis section of a green ammonia process is provided, comprising:
[0005] Step 1: Collect historical data of the green ammonia process with time series, classify it according to different working conditions, and build a green ammonia process database under different working conditions;
[0006] Step 2: Build a key variable library for the synthesis process, and use the variables in it as candidate output variables for the BiLSTM-BiGRU prediction model;
[0007] Step 3: For different output variables, a potential related variable library is constructed based on the process mechanism, and the related variable library is feature extracted through the multi-view aggregation adaptive encoder (MVAAE) to determine the model input variables;
[0008] Step 4: Construct a BiLSTM-BiGRU hybrid neural network model for trend prediction of key parameters of the green ammonia process, and train it based on the different operating condition data sets obtained by the above classification;
[0009] Step 5: Construct a warning threshold library, and determine the warning thresholds of the prediction variables under different working conditions based on the aforementioned different working condition data sets;
[0010] Step 6: Collect real-time data of the green ammonia process and import it into the model. After the data is preprocessed, the model trained with different data sets is called to predict the key parameter values and change rates of the green ammonia process;
[0011] Step 7: According to the different output variables of the model, select the corresponding threshold input from the warning threshold library constructed in step 5, and implement the early warning of the process through the adaptive sliding window adjustment method based on the feedback mechanism to determine whether there is any abnormal situation. If it is determined that there is no abnormality in the system, return to step 6 to continuously warn the process; if it is determined that an abnormal situation has occurred, an alarm is issued to remind the on-site operator of the upcoming abnormal situation, and the initial cause of the abnormality is diagnosed and processed.
[0012] Preferably, the green ammonia process database under different working conditions described in step 1 is specifically classified as follows:
[0013] All data are divided into two categories: steady state and load switching state. The following 9 categories are specifically classified in the steady state category: production load 30%, 40%, 50%, 60%, 70%, 80%, 90%, 100%, 110%;
[0014] The load switching category is divided into load increase adjustment group and load decrease adjustment group. The load increase adjustment group is divided into the following 9 categories: production load 30%-40%, 40%-50%, 50%-60%, 60%-70%, 70%-80%, 80%-90%, 90%-100%, 100%-110%;
[0015] The load reduction adjustment groups are divided into: production load 40%-30%, 50%-40%, 60%-50%, 70%-60%, 80%-70%, 90%-80%, 100%-90%, 110%-100%.
[0016] Preferably, the green ammonia stability index is introduced to determine whether the process is in a steady state. For a single variable, the stability index is calculated as follows:
[0017]
[0018] Among them, S A is the stability index of variable A, β A is the real-time rate of change of variable A, ΔV A is the range of A within 10 minutes, V A,max -V A,min Indicates the range of variable A within the entire determination time range;
[0019] The stability index was calculated every 10 seconds. When the S of all key variables was less than 0.1 within 10 minutes, it was considered a steady-state process.
[0020] Preferably, the key variable library of the synthesis section described in step 2 specifically includes the following variables: temperature of the first bed of the synthesis tower, temperature of the second bed of the synthesis tower, temperature of the third bed of the synthesis tower, pressure of the first bed of the synthesis tower, pressure of the second bed of the synthesis tower, pressure of the third bed of the synthesis tower, and ammonia content of the gas discharged from the synthesis tower.
[0021] Preferably, the potential related variable library described in step 3 is specifically: a potential related variable library of synthesis tower temperature, a potential related variable library of synthesis tower pressure, and a potential related variable library of ammonia content in synthesis tower outlet gas; the variables contained in each library are:
[0022] The potential variable library of synthesis tower temperature includes 21 variables, including flash tower top outlet flow, flash tower top outlet temperature, flash tower bottom outlet pressure, etc.
[0023] The potential variable library of synthesis tower pressure includes 28 variables such as synthesis tower inlet pressure, synthesis tower outlet pressure, synthesis tower inlet flow, synthesis tower outlet flow, etc.
[0024] The library of potential related variables for the ammonia content in the gas outlet of the synthesis tower includes 22 variables, such as the ammonia content at the inlet of the synthesis tower, the pressure at the top outlet of the gas-liquid separation tower, the flow rate at the top outlet of the gas-liquid separation tower, the temperature at the top outlet of the gas-liquid separation tower, the ammonia content at the top outlet of the gas-liquid separation tower, and the pressure at the bottom outlet of the gas-liquid separation tower.
[0025] Preferably, the multi-view aggregation adaptive encoder (MVAAE) in step 3 includes the following modules:
[0026] A multi-view encoder module for extracting features from multiple data sources in the ammonia synthesis section;
[0027] A graph neural network (GNN) fusion module, used to fuse the features extracted by the multi-view encoder module to capture the relationship between different data sources;
[0028] Attention mechanism module, which is used to dynamically adjust the weight of different data sources in feature fusion according to their importance;
[0029] The adaptive dimension reduction module is used to reduce the dimension of the fused high-dimensional features while retaining key information.
[0030] Preferably, the multi-view encoder module includes multiple encoders, each encoder is used to extract features from different data sources; the data sources include: reaction temperature sensor data, reaction pressure sensor data, hydrogen-nitrogen ratio data, catalyst activity monitoring data, reaction product concentration data; the encoder types include: multi-layer perceptron (MLP), convolutional neural network (CNN), principal component analysis (PCA); wherein, MLP is used to process reaction temperature sensor data and reaction pressure sensor data, CNN is used to process hydrogen-nitrogen ratio data and reaction product concentration data, and PCA is used to process catalyst activity monitoring data.
[0031] Preferably, the MVAAE model training module is optimized by minimizing the following loss function:
[0032] L=L recon +ωL reg
[0033] Among them, L represents the total loss function, L recon represents the reconstruction loss function, L reg represents the regularized loss function, ω is the regularization coefficient; where L recon The calculation formula is as follows
[0034]
[0035] x represents the original input data, is the data reconstructed from the low-dimensional feature representation zlow by the decoder part of the adaptive dimensionality reduction module; L reg The calculation formula is as follows
[0036]
[0037] α i represents the weight assigned to the i-th feature in the attention mechanism, and N is the number of features.
[0038] Preferably, in the early warning threshold library described in step 5, the index of each threshold includes: variable name, load condition, threshold type, threshold; wherein, the types included in the load condition are consistent with the condition types in the green ammonia process data set under the aforementioned different conditions; and the threshold types include: a fixed threshold for the variable value, an adaptive dynamic threshold for the variable value, a threshold for the rate of change of the variable value, and a residual threshold for the difference between the actual value of the variable and the prediction; wherein, the adaptive dynamic threshold for the variable value is obtained based on the Bayesian method, and the variable change rate threshold and the residual threshold are obtained based on the ROC method.
[0039] Preferably, when collecting real-time data of the green ammonia process in step 6, it is required to display the current load status on the central control display screen, display the current load in steady state, and display the load switching task target being executed during load switching.
[0040] Preferably, the corresponding threshold input model is selected from the warning threshold library constructed in step 5 for early warning as described in step 7, and the input threshold is selected as follows: when the output variable is the synthesis tower temperature, the input threshold is the adaptive dynamic threshold of the variable value under the corresponding working condition, the threshold of the rate of change of the variable value, and the residual threshold; when the output variable is the synthesis tower pressure, the input threshold is the fixed threshold of the variable value under the corresponding working condition, the threshold of the rate of change of the variable value, and the residual threshold; when the output variable is the ammonia content of the synthesis tower outlet gas, the input threshold is the fixed threshold of the variable value under the corresponding working condition.
[0041] Preferably, when the process load adjustment task collected in step 6 cannot find a corresponding threshold match in the green ammonia process database and the threshold library, a GNN-based operating condition interval adaptive determination method is adopted. By constructing a graph structure with process load and related process parameters as nodes, the association and similarity between the nodes are learned in real time, so as to dynamically determine the operating condition interval in which the current load is located, and adjust the corresponding parameter prediction model and early warning strategy; for example: when the load adjustment task of the real-time process read in step 6 is the production load from 30% to 70%, and the process data and threshold value directly corresponding to the production load 30% to 70% operating condition cannot be found in the green ammonia process database and the threshold library, the early warning task in step 7 can be split into four sections of production load 30%-40%, 40%-50%, 50%-60%, and 60%-70% existing in the library, and the production load adjustment task is determined in which interval based on GNN, and early warning is performed in sections. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] Figure 1 Schematic diagram of the overall operation process of the present invention
[0043] Figure 2 This is a diagram showing the prediction and warning results of the present invention when used for temperature warning of the synthesis tower DETAILED DESCRIPTION
[0044] The embodiments of the present invention are described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.
[0045] (1) Step 1: Collect historical data of the green ammonia process with time series, classify it according to different working conditions, and build a green ammonia process database under different working conditions;
[0046] The classification is as follows: all data are divided into two categories: steady state and load switching state. The steady state category is specifically classified into the following 9 categories:
[0047] Production load 30%, 40%, 50%, 60%, 70%, 80%, 90%, 100%, 110%;
[0048] The load switching category is divided into load increase adjustment group and load decrease adjustment group. The load increase adjustment group is divided into the following 9 categories: production load 30%-40%, 40%-50%, 50%-60%, 60%-70%, 70%-80%, 80%-90%, 90%-100%, 100%-110%;
[0049] The load reduction adjustment groups are divided into: production load 40%-30%, 50%-40%, 60%-50%, 70%-60%, 80%-70%, 90%-80%, 100%-90%, 110%-100%.
[0050] In the above classification, the green ammonia stability index is introduced to determine whether the process is in a steady state. For a single variable, the stability index is calculated as follows:
[0051]
[0052] Among them, S A is the stability index of variable A, β A is the real-time rate of change of variable A, ΔV A is the range of A within 10 minutes, V A,max -V A,min Indicates the range of variable A within the entire determination time range;
[0053] The stability index was calculated every 10 seconds. When the S of all key variables was less than 0.1 within 10 minutes, it was considered a steady-state process.
[0054] (2) Step 2: Construct a key variable library for the synthesis process and use the variables therein as candidate output variables for the BiLSTM-BiGRU prediction model;
[0055] The key variable library of the synthesis stage described in step 2 specifically includes the following variables:
[0056] Temperature of the first bed of the synthesis tower, the temperature of the second bed of the synthesis tower, the temperature of the third bed of the synthesis tower, the pressure of the first bed of the synthesis tower, the pressure of the second bed of the synthesis tower, the pressure of the third bed of the synthesis tower, and the ammonia content of the gas discharged from the synthesis tower
[0057] (3) Step 3: For different output variables, a potential related variable library is constructed based on the process mechanism, and the model input variables are determined by feature extraction of the related variable library;
[0058] The library of potentially relevant variables described in step 3 is specifically:
[0059] The potential related variable library of synthesis tower temperature, the potential related variable library of synthesis tower pressure, and the potential related variable library of ammonia content in the gas discharged from the synthesis tower; the variables contained in each library are:
[0060] The library of potential related variables for synthesis tower temperature includes the following 21 variables: synthesis tower inlet temperature, synthesis tower outlet temperature, synthesis tower inlet flow, synthesis tower outlet flow, hydrogen raw material flow, nitrogen raw material flow, gas-liquid separation tower top outlet pressure, gas-liquid separation tower top outlet flow, gas-liquid separation tower top outlet temperature, gas-liquid separation tower bottom outlet pressure, gas-liquid separation tower bottom outlet temperature, gas-liquid separation tower bottom outlet flow, flash tower top outlet pressure, flash tower top outlet flow, flash tower top outlet temperature, flash tower bottom outlet pressure, flash tower bottom outlet flow, flash tower bottom outlet flow, boiler steam temperature, boiler steam flow, and feed four-stage compressor outlet temperature.
[0061] The library of potential related variables for synthesis tower pressure includes the following 28 variables: synthesis tower inlet pressure, synthesis tower outlet pressure, synthesis tower inlet flow, synthesis tower outlet flow, feed first stage compressor outlet temperature, feed first stage compressor outlet pressure, feed second stage compressor outlet temperature, feed second stage compressor outlet temperature, feed third stage compressor outlet pressure, feed third stage compressor outlet temperature, feed third stage compressor outlet pressure, feed third stage compressor outlet temperature, feed fourth stage compressor outlet pressure, gas-liquid separation tower top outlet flow, gas-liquid separation tower top outlet temperature, gas-liquid separation tower bottom outlet pressure, gas-liquid separation tower bottom outlet temperature, gas-liquid separation tower bottom outlet flow, flash tower top outlet pressure, flash tower top outlet flow, flash tower top outlet temperature, flash tower bottom outlet pressure, flash tower bottom outlet flow, flash tower bottom outlet flow, boiler steam flow, cooling water temperature.
[0062] The library of potential related variables for ammonia content in the gas outflowing from the synthesis tower includes the following 22 variables: synthesis tower inlet temperature, synthesis tower inlet pressure, synthesis tower inlet ammonia content, gas-liquid separation tower top outlet pressure, gas-liquid separation tower top outlet flow rate, gas-liquid separation tower top outlet temperature, gas-liquid separation tower top outlet ammonia content, gas-liquid separation tower bottom outlet pressure, gas-liquid separation tower bottom outlet temperature, gas-liquid separation tower bottom outlet flow rate, gas-liquid separation tower bottom outlet ammonia content, flash tower top outlet pressure, flash tower top outlet flow rate, flash tower top outlet temperature, flash tower top outlet ammonia content, flash tower bottom outlet pressure, flash tower bottom outlet flow rate, flash tower bottom outlet flow rate, flash tower bottom outlet ammonia content, ammonia product purity, hydrogen feed flow rate, nitrogen feed flow rate.
[0063] The multi-view aggregation adaptive encoder (MVAAE) described in step 3 includes the following modules:
[0064] A multi-view encoder module for extracting features from multiple data sources in the ammonia synthesis section;
[0065] A graph neural network (GNN) fusion module, used to fuse the features extracted by the multi-view encoder module to capture the relationship between different data sources;
[0066] Attention mechanism module, which is used to dynamically adjust the weight of different data sources in feature fusion according to their importance;
[0067] The adaptive dimension reduction module is used to reduce the dimension of the fused high-dimensional features while retaining key information.
[0068] The multi-view encoder module includes multiple encoders, each encoder is used to extract features from different data sources; the data sources include: reaction temperature sensor data, reaction pressure sensor data, hydrogen-nitrogen ratio data, catalyst activity monitoring data, reaction product concentration data; the encoder types include: multi-layer perceptron (MLP), convolutional neural network (CNN), principal component analysis (PCA); among them, MLP is used to process reaction temperature sensor data and reaction pressure sensor data, CNN is used to process hydrogen-nitrogen ratio data and reaction product concentration data, and PCA is used to process catalyst activity monitoring data.
[0069] The GNN module fuses the features of different data sources to capture the complex relationships between them; i As nodes in the graph, the edges between nodes represent the relationship between data sources. GNN is used to update the feature representation of each node to capture global process information:
[0070]
[0071] The attention mechanism module dynamically adjusts the weights of each data source in the final feature fusion according to the importance of each data source and calculates the attention score e for each feature. i , and normalized to weight α through the softmax function i :
[0072] e i =FFN(z i )
[0073]
[0074] The final aggregate feature is expressed as:
[0075]
[0076] The adaptive dimensionality reduction module can reduce the dimensionality of the fused high-dimensional features through a linear layer while retaining key information for subsequent analysis:
[0077] z low =g θ (z fusion )
[0078] In addition, the MVAAE model training module is optimized by minimizing the following loss function:
[0079] L=L recon +ωL reg
[0080] Among them, L represents the total loss function, L recon represents the reconstruction loss function, L reg represents the regularized loss function, ω is the regularization coefficient; where L recon The calculation formula is as follows
[0081]
[0082] x represents the original input data, is the data reconstructed from the low-dimensional feature representation zlow by the decoder part of the adaptive dimensionality reduction module; L reg The calculation formula is as follows
[0083]
[0084] α i represents the weight assigned to the i-th feature in the attention mechanism, and N is the number of features.
[0085] (4) Step 4: Construct a BiLSTM-BiGRU hybrid neural network model for trend prediction of key parameters of the green ammonia process, and train it based on the different operating condition data sets obtained by the above classification;
[0086] The BiLSTM-BiGRU hybrid model described in step 4 has the following specific structure and implementation: the variable data after feature extraction is input into the model, and the input data is in the form of
[0087] X={x1,x2,…,x T}
[0088] x i is the input data vector at time step t; through the parallel BiLSTM and BiGRU layers, the time series data is predicted and the output is obtained:
[0089]
[0090] The output is fed into a DENSE layer to generate the weight parameter α t , the weight parameter is updated by the back propagation algorithm with the goal of minimizing RMSE. The calculation formula of the weight function is:
[0091] α t =sigmoid(W α ·x t +b α )
[0092] Among them, W α is the weight matrix, b α is the bias vector; after obtaining the weight parameter α t After that, the weight distribution results are input into the Multiply layer to weight the respective output results. The calculation formula is as follows:
[0093] h Weighted,BiLSTM,t =α t ·h BiLSTM,t
[0094] h Weighted,BiGRU,t =(1-α t )·h BiGRU,t
[0095] Then the Add layer is merged into the final output result in the BiLSTM-BiGRU parallel structure:
[0096] h Weighted,t =h Weighedt,BiGRU,t +h Weighedt,BiLSTM,t
[0097] H weighted,t =[h Weighted,1 ,h Weighted,2 ,...,h Weighted,T-1 ,h Weighted,T ]
[0098] Finally, through the fully connected layer, the final output value of the model is obtained:
[0099] Y=[y1,y2,...,y T-1 ,y T ]
[0100] in, Represents the output result at time step t.
[0101] (5) Step 5: Construct a warning threshold library, and determine the warning thresholds of the prediction variables under different working conditions based on the aforementioned different working condition data sets;
[0102] In the early warning threshold library described in step 5, the index of each threshold is: variable name, load condition, threshold type, threshold; among them, the types included in the load condition are consistent with the condition types in the green ammonia process data set under the aforementioned different conditions; and the threshold types include: fixed threshold of variable value, adaptive dynamic threshold of variable value, threshold of variable value change rate and residual threshold of the difference between the actual value and the prediction of the variable; among them, the adaptive dynamic threshold of the variable value is obtained based on the Bayesian method, and the variable change rate threshold and residual threshold are obtained based on the ROC method.
[0103] (6) Step 6: Collect real-time data of the green ammonia process and import it into the model. After the data is preprocessed, the model trained with different data sets is called to predict the key parameter values and change rates of the green ammonia process;
[0104] The real-time data of the green ammonia process is collected as described in step 6, and the current load status is required to be displayed on the central control display screen. The current load is displayed in steady state, and the load switching task target being executed is displayed during load switching.
[0105] (7) Step 7: According to the different output variables of the model, select the corresponding threshold input from the warning threshold library constructed in step 5, and implement the early warning of the process through the adaptive sliding window adjustment method based on the feedback mechanism to determine whether there is any abnormal situation. If it is determined that there is no abnormality in the system, return to step 6 and continue to warn the process; if it is determined that an abnormal situation has occurred, an alarm is issued to remind the on-site operator of the upcoming abnormal situation, and the initial cause of the abnormality is diagnosed and processed.
[0106] The corresponding threshold input model is selected from the warning threshold library constructed in step 5 for early warning as described in step 7, and the input threshold is selected as follows: when the output variable is the synthesis tower temperature, the input threshold is the adaptive dynamic threshold of the variable value under the corresponding working condition, the threshold of the rate of change of the variable value, and the residual threshold; when the output variable is the synthesis tower pressure, the input threshold is the fixed threshold of the variable value under the corresponding working condition, the threshold of the rate of change of the variable value, and the residual threshold; when the output variable is the ammonia content of the synthesis tower outlet gas, the input threshold is the fixed threshold of the variable value under the corresponding working condition.
[0107] When the process load adjustment task collected in step 6 cannot find the corresponding threshold match in the green ammonia process database and threshold library, the GNN-based operating condition interval adaptive determination method is adopted. By constructing a graph structure with process load and related process parameters as nodes, the association and similarity between the nodes are learned in real time, so as to dynamically determine the operating condition interval of the current load and adjust the corresponding parameter prediction model and early warning strategy; for example: the load adjustment task of the real-time process read in step 6 is the production load from 30% to 70%, and the process data and threshold value of the production load 30% to 70% working condition directly corresponding to the green ammonia process database and threshold library cannot be found. The early warning task in step 7 can be divided into four sections of production load 30%-40%, 40%-50%, 50%-60%, and 60%-70% in the library. The interval of the production load adjustment task is determined based on GNN, and early warning is carried out in sections.
[0108] The specific implementation method of the adaptive sliding window adjustment method with feedback mechanism described in step 6 described in step 7 is as follows:
[0109] ①Using the true positive rate (TPR) and false positive rate (FPR) of the warning results as feedback indicators, dynamically adjust the size of the sliding window:
[0110]
[0111] Among them, FP is the number of false alarms, FN is the number of missed alarms, TP is the number of normal samples without alarms, FN is the number of abnormal sample alarms, FP+TN represents the total number of normal samples, and TP+FN represents the total number of fault samples;
[0112] ② Set the initial window size W0 and the minimum window size W of the sliding window method min , the maximum window size W max ,window
[0113] The adjustment step size P, the window alarm step size F, and the FPR and TPR indicator thresholds T F , T T ;
[0114] ③ Input the corresponding alarm threshold;
[0115] ④ If there are F consecutive points in the window that exceed any of the thresholds described in right 6, it is determined that an abnormal situation has occurred;
[0116] ⑤ Combine the warning results with the actual abnormal situation to calculate FPR and TPR;
[0117] ⑥ If the calculated FPR and TPR satisfy TPR <T T 、FPR>T F If any one or more of the following occurs, it means that
[0118] The front window is too small and needs to be increased, but it cannot exceed the maximum window size:
[0119] W t+1 =min(W t +P,W max )
[0120] If none of them are satisfied, it means that the current window setting may be too large, and the window size can be reduced to improve computing efficiency, but
[0121] The maximum window size cannot be exceeded:
[0122] W t+1 =max(W t -P,W min )
[0123] The abnormal diagnosis and treatment guidance described in step 7 is achieved through the combination of SDG-HAZOP analysis, and the specific implementation method is as follows:
[0124] ① Conduct HAZOP hazard and operability analysis on the green ammonia process, build a case library of abnormal situations, and classify the risks of abnormal situations in the case library according to the possible harmful consequences, and formulate disposal requirements for abnormal situations of different risk levels;
[0125] ② All measurable variables in the green ammonia process are abstracted as nodes, and the correlation between two nodes is abstracted as edges. On this basis, the SDG symbol directed graph G = (V, E) of the measurable variables in the green ammonia process is constructed to reflect the correlation between variables, where V is a node set and E is a directed edge set;
[0126] ③ Assuming that the process variable of the early warning is temperature T, when the system is judged to be abnormal, all nodes that have a causal relationship with T are found through the symbolic directed graph, that is, all nodes that satisfy (v i , v j ) and v j =T nodes, and recursively find all the predecessor nodes of these variables to build the causal chain I;
[0127] ④Starting from the abnormal variable T, trace back each predecessor node recursively until a variable without a predecessor node is found, which is the initial cause of the abnormality. Specifically:
[0128] I: put all nodes in I into the queue to be processed;
[0129] II: Take a variable from the queue and find all predecessor nodes. If the predecessor node also has an exception, add it to the queue to be processed;
[0130] III: Traverse the entire queue. When the queue is empty, the variable without a predecessor node among all the nodes in the queue is the initial cause of the exception.
[0131] ⑤ After finding the initial cause, search for the corresponding case and its risk classification result in the case library, and handle the abnormal situation according to the disposal requirements corresponding to the risk classification result proposed in (1).
[0132] The beneficial effects of the present invention are:
[0133] (1) The present invention determines the status of green ammonia historical data based on the stability index method, clearly and accurately classifies a large amount of process data, and constructs a green ammonia historical database under different working conditions, providing solid data support for early warning of the green ammonia process.
[0134] (2) The present invention establishes a library of potential related variables corresponding to different key variables, removes a large number of irrelevant variables based on the process mechanism, and effectively improves the efficiency of feature variable extraction.
[0135] (3) The present invention is based on the multi-view aggregation adaptive encoder (MVAAE) method to implement feature engineering extraction from the potential correlated variable library, accurately extracting the strongly correlated variables corresponding to the key variables, eliminating the interference of redundant variables, and effectively improving the prediction effect of the prediction model.
[0136] (4) The present invention constructs a diversified threshold library of variables under different working conditions and different threshold types, and selects different types of thresholds for different variables. On the basis of ensuring the generalization ability of the model, it realizes the targeted ability for each early warning task.
[0137] (5) The present invention proposes a method for decomposing warning tasks, which is used to issue warnings for load adjustment conditions that have no matching conditions in the case library, effectively improving the robustness and versatility of the warning method.
Claims
1. A method for early warning of abnormal conditions in a green ammonia process synthesis section, characterized by: The following steps are involved: Step 1: Collect historical data of the green ammonia process with time series, classify it according to different working conditions, and build a green ammonia process database under different working conditions; Step 2: Build a key variable library for the synthesis process, and use the variables in it as candidate output variables for the BiLSTM-BiGRU prediction model; Step 3: For different output variables, a potential related variable library is constructed based on the process mechanism, and the related variable library is feature extracted through the multi-view aggregation adaptive encoder (MVAAE) to determine the model input variables; Step 4: Construct a BiLSTM-BiGRU hybrid neural network model for trend prediction of key parameters of the green ammonia process, and train it based on the different operating condition data sets obtained by the above classification; Step 5: Build a warning threshold library, and determine the warning thresholds of the prediction variables under different working conditions based on the aforementioned different working condition data sets; Step 6: Collect real-time data of the green ammonia process and import it into the model. After the data is preprocessed, the model trained with different data sets is called to predict the key parameter values and change rates of the green ammonia process; Step 7: According to the different output variables of the model, select the corresponding threshold input from the warning threshold library constructed in step 5, and implement the early warning of the process through the adaptive sliding window adjustment method based on the feedback mechanism to determine whether there is any abnormal situation; if it is determined that there is no abnormality in the system, return to step 6 to continuously warn the process; If an abnormal situation is determined to have occurred, an alarm is issued to alert on-site operators of the impending abnormal situation, and the initial cause of the abnormal situation is diagnosed and processed.
2. The method for early warning of abnormal conditions in the synthesis section of a green ammonia process according to claim 1, characterized in that: The green ammonia process database under different working conditions described in step 1 is specifically classified as follows: All data are divided into two categories: steady state and load switching state. The following 9 categories are specifically classified in the steady state category: production load 30%, 40%, 50%, 60%, 70%, 80%, 90%, 100%, 110%; The load switching category is divided into load increase adjustment group and load decrease adjustment group. The load increase adjustment group is divided into the following 9 categories: production load 30%-40%, 40%-50%, 50%-60%, 60%-70%, 70%-80%, 80%-90%, 90%-100%, 100%-110%; The load reduction adjustment groups are divided into: production load 40%-30%, 50%-40%, 60%-50%, 70%-60%, 80%-70%, 90%-80%, 100%-90%, 110%-100%.
3. The method for early warning of abnormal conditions in the synthesis section of the green ammonia process according to claim 1, characterized in that The green ammonia stability index is introduced to determine whether the process is in a steady state. For a single variable, the stability index is calculated as follows: Among them, S A is the stability index of variable A, β A is the real-time rate of change of variable A, ΔV A is the range of A within 10 minutes, V A,max -V A,min Indicates the range of variable A within the entire determination time range; The stability index was calculated every 10 seconds. When the S of all key variables was less than 0.1 within 10 minutes, it was considered a steady-state process.
4. The method for early warning of abnormal conditions in the synthesis section of a green ammonia process according to claim 1, characterized in that: The key variable library of the synthesis section described in step 2 specifically includes the following variables: temperature of the first bed of the synthesis tower, temperature of the second bed of the synthesis tower, temperature of the third bed of the synthesis tower, pressure of the first bed of the synthesis tower, pressure of the second bed of the synthesis tower, pressure of the third bed of the synthesis tower, and ammonia content of the gas discharged from the synthesis tower.
5. The method for early warning of abnormal conditions in the synthesis section of a green ammonia process according to claim 1, characterized in that: The potential related variable library described in step 3 is specifically: a potential related variable library of synthesis tower temperature, a potential related variable library of synthesis tower pressure, and a potential related variable library of ammonia content in the gas discharged from the synthesis tower; the variables contained in each library are: The potential variable library of synthesis tower temperature includes 21 variables, including flash tower top outlet flow, flash tower top outlet temperature, flash tower bottom outlet pressure, etc. The potential variable library of synthesis tower pressure includes 28 variables such as synthesis tower inlet pressure, synthesis tower outlet pressure, synthesis tower inlet flow, synthesis tower outlet flow, etc. The library of potential related variables for the ammonia content in the gas outlet of the synthesis tower includes 22 variables, such as the ammonia content at the inlet of the synthesis tower, the pressure at the top outlet of the gas-liquid separation tower, the flow rate at the top outlet of the gas-liquid separation tower, the temperature at the top outlet of the gas-liquid separation tower, the ammonia content at the top outlet of the gas-liquid separation tower, and the pressure at the bottom outlet of the gas-liquid separation tower.
6. The method for early warning of abnormal conditions in the synthesis section of a green ammonia process according to claim 1, characterized in that: The multi-view aggregation adaptive encoder (MVAAE) described in step 3 includes the following modules: A multi-view encoder module for extracting features from multiple data sources in the ammonia synthesis section; A graph neural network (GNN) fusion module, used to fuse the features extracted by the multi-view encoder module to capture the relationship between different data sources; Attention mechanism module, which is used to dynamically adjust the weight of different data sources in feature fusion according to their importance; The adaptive dimension reduction module is used to reduce the dimension of the fused high-dimensional features while retaining key information.
7. The method for early warning of abnormal conditions in the synthesis section of a green ammonia process according to claim 1, characterized in that: The multi-view aggregation adaptive encoder (MVAAE) described in step 3 includes multiple encoders, each encoder is used to extract features from different data sources; the data sources include: reaction temperature sensor data, reaction pressure sensor data, hydrogen-nitrogen ratio data, catalyst activity monitoring data, reaction product concentration data; the encoder types include: multi-layer perceptron (MLP), convolutional neural network (CNN), principal component analysis (PCA); wherein, MLP is used to process reaction temperature sensor data and reaction pressure sensor data, CNN is used to process hydrogen-nitrogen ratio data and reaction product concentration data, and PCA is used to process catalyst activity monitoring data.
8. The method for early warning of abnormal conditions in the synthesis section of a green ammonia process according to claim 1, characterized in that: The multi-view aggregation adaptive encoder (MVAAE) described in step 3 optimizes the training model by minimizing the following loss function: L=L recon +ωL reg Among them, L represents the total loss function, L recon represents the reconstruction loss function, L reg represents the regularized loss function, ω is the regularization coefficient; where L recon The calculation formula is as follows x represents the original input data, is the data reconstructed from the low-dimensional feature representation zlow by the decoder part of the adaptive dimensionality reduction module; L reg The calculation formula is as follows α i represents the weight assigned to the i-th feature in the attention mechanism, and N is the number of features.
9. The method for early warning of abnormal conditions in the synthesis section of a green ammonia process according to claim 1, characterized in that: In the early warning threshold library described in step 5, the index of each threshold is: variable name, load condition, threshold type, threshold; among them, the types included in the load condition are consistent with the condition types in the green ammonia process data set under the aforementioned different conditions; and the threshold types include: fixed threshold of variable value, adaptive dynamic threshold of variable value, threshold of variable value change rate and residual threshold of the difference between the actual value and the prediction of the variable; among them, the adaptive dynamic threshold of the variable value is obtained based on the Bayesian method, and the variable change rate threshold and residual threshold are obtained based on the ROC method.
10. The method for early warning of abnormal conditions in the synthesis section of a green ammonia process according to claim 1, characterized in that: When collecting real-time data of the green ammonia process in step 6, it is required to display the current load status on the central control display screen, display the current load in steady state, and display the load switching task target being executed during load switching.
11. The method for early warning of abnormal conditions in the synthesis section of a green ammonia process according to claim 1, characterized in that: The corresponding threshold input model is selected from the warning threshold library constructed in step 5 for early warning as described in step 7, and the input threshold is selected as follows: when the output variable is the synthesis tower temperature, the input threshold is the adaptive dynamic threshold of the variable value under the corresponding working condition, the threshold of the rate of change of the variable value, and the residual threshold; when the output variable is the synthesis tower pressure, the input threshold is the fixed threshold of the variable value under the corresponding working condition, the threshold of the rate of change of the variable value, and the residual threshold; when the output variable is the ammonia content of the synthesis tower outlet gas, the input threshold is the fixed threshold of the variable value under the corresponding working condition.
12. The method for early warning of abnormal conditions in the synthesis section of a green ammonia process according to claim 1, characterized in that: When the process load adjustment task collected in step 6 cannot find the corresponding threshold match in the green ammonia process database and threshold library, the GNN-based operating range adaptive determination method is adopted. By constructing a graph structure with process load and related process parameters as nodes, the association and similarity between the nodes are learned in real time, so as to dynamically determine the operating range of the current load and adjust the corresponding parameter prediction model and early warning strategy; For example: the load adjustment task of the real-time process read in step 6 is the production load from 30% to 70%. When the process data and threshold value of the production load 30% to 70% condition directly corresponding to the process data cannot be found in the green ammonia process database and threshold library, the warning task in step 7 can be divided into four sections of production load 30%-40%, 40%-50%, 50%-60%, and 60%-70% in the library. Based on GNN, it is determined in which interval the production load adjustment task is located, and warnings are issued in segments.
Citation Information
Patent Citations
Wind turbine generator fault early warning method based on graph neural network
CN114372504A
Gaussian mixture model embedded GRU auto-encoder blast furnace anomaly monitoring method
CN114854920A
Short-term power load prediction method based on SMBO-BiGRU-Attention
CN117875365A
Dangerous chemical safety monitoring method based on big data
CN118506281A
Cited By
Marine distributed green ammonia production scheduling system and method based on intelligent management
CN120598249A
