Steam pipeline pressure control method

Abstract

The invention discloses a steam pipeline pressure control method, which comprises the following steps of: acquiring operation information of each key position of a pipe network in real time by constructing a data acquisition system and deploying an execution unit; a multivariable analysis method is utilized, topological structure characteristics of a steam pipeline are combined, a load change model of a steam consumption unit is constructed, the pressure change process caused by operation of different devices is recognized in advance, and a long-term pressure control strategy is made; meanwhile, an accumulative deviation curve is constructed based on a prediction and actual measurement residual evolution sequence, the potential fault risk of the system is identified in an auxiliary manner, and structured alarm output is formed, so that the safety guarantee capability of pipe network operation is enhanced. The steam pipeline pressure control method is suitable for the application scene that the overall steam use requirement is stable, but the local high-frequency disturbance behavior exists, the tiny pressure fluctuation trend can be accurately predicted under the high-disturbance working conditions such as frequent starting and stopping, and it is guaranteed that the steam pressure is stable.

Description

Technical Field

[0001] This invention belongs to the field of pressure control technology, specifically, it relates to a method for automatically controlling the pressure in a steam pipeline. Background Technology

[0002] Steam pipeline pressure control refers to a technical means of collecting real-time pressure data within the steam pipeline using sensors and controlling the regulating device based on the collected data to maintain the pressure within the pipeline within a set range. This method effectively avoids excessively high or low steam pressure, ensuring that steam can be safely and stably delivered to various steam-consuming points. It can be widely applied in various industrial scenarios requiring precise pressure control.

[0003] Chinese invention patent CN112132428A discloses a predictive control decision-making method for steam distribution stations in a steam heating network based on big data. Specifically, it includes: Step S1, reading daily operating data of the steam heating network and distribution stations from a database and performing data preprocessing; Step S2, selecting an identification object, determining the control characteristic parameter with the highest correlation to the identification object in the desuperheating and pressure reducing equipment of the distribution station under different operating conditions, identifying it as the basic characteristic, and determining the required response characteristics and other characteristics of the identification object; Step S3, dividing the sliding time window length into two equal segments, calculating the moving arithmetic mean difference of the basic characteristics within the two time windows, and obtaining the time series of the moving arithmetic mean difference; Step S4, setting a stable range for the response characteristics, setting the time point when the value of the response characteristics exceeds the stable range as the operating condition change point, setting a threshold for basic characteristic changes, and setting the time point when the moving arithmetic mean difference of the basic characteristics exceeds the threshold after the operating condition change point as the control start point; Step S5, ... Step S6: The first time window starting point where the value of the response feature re-enters the target stable range and the moving arithmetic mean difference within the continuous time window is less than the stable threshold is set as the control completion point. The time difference between the control completion point and the control start point is taken as the control lag time of the steam distribution station. Step S7: Based on the control lag time of the steam distribution station, a basic feature sequence of a specified length starting from the control start point is generated. Machine learning is used to identify the functional relationship between the control lag time of the steam distribution station and the basic feature sequence, other features, and weather factors, resulting in a calculation model for the control lag time of the steam distribution station under different weather conditions, different cooling water flow changes, and different operating conditions. Step S8: The gradient and upper and lower limits of the single change of the basic feature are set. Based on the basic feature value of the required control start point, a basic feature sequence is generated. Based on the functional relationship obtained in Step S6, the objective function is solved to obtain the parameter vector in the basic feature sequence, so that the steam distribution station staff can control the basic features according to the required lag time.

[0004] The previous invention used the moving average difference method to model the operating characteristics of steam distribution stations, generating predictive control strategies to alleviate pressure fluctuations caused by control delays in traditional methods. However, in the steelmaking industry, the number of steam-consuming devices in production processes is large, including many with small single-use volumes and high frequency of use, such as high-frequency gas injection and temperature control equipment. Therefore, in addition to some devices continuously consuming large amounts of steam, the steam system frequently experiences multiple start-ups and shutdowns, exhibiting dynamic characteristics of high-frequency disturbances, making it difficult to maintain stable steam pressure within the target range. Existing technologies mostly employ static rules or simple statistical modeling methods, which can only supply steam based on approximate steam consumption patterns, leading to frequent fluctuations in steam supply. Furthermore, high-frequency disturbances also break the applicability of conventional strategy models, resulting in increased fluctuations in prediction results and decreased reliability of control response.

[0005] Therefore, there is an urgent need for a steam pipeline pressure control method that integrates multi-source data perception, spatiotemporal feature modeling, and rolling optimization control mechanism to improve the stability and intelligence level of steam supply. Summary of the Invention

[0006] To address the problems of large pressure fluctuations, delayed response, and unstable local steam supply caused by frequent start-ups and shutdowns of steam-using equipment and high disturbance intensity in steam pipelines, this invention provides a steam pipeline pressure control method. This method overcomes the limitations of traditional control strategies, such as large prediction errors and low efficiency in dynamic environments, and improves the stability and intelligence level of system operation.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0008] A method for controlling steam pipeline pressure, comprising:

[0009] A data acquisition system and execution unit are deployed in the steam pipeline to collect pipeline data and operation records of the execution unit in real time, forming real-time operation data.

[0010] A global stress evolution trend reference curve is constructed based on real-time operational data and stored historical operational data;

[0011] A multivariate modeling method is adopted, and a graph convolutional neural network is introduced to perform structural modeling of the topology of the steam pipeline. This model is then combined with a global pressure evolution trend reference curve to construct a load change model.

[0012] Based on the output results and predicted trends of the load change model, a matching dynamic control strategy is formulated to control the actions of the execution unit, and the load change model is corrected based on the feedback of real-time operating data.

[0013] Deep learning is performed using convolutional neural networks to optimize the policy library;

[0014] Real-time monitoring of the discrepancies between model data and actual data helps identify abnormal conditions in steam pipelines.

[0015] In some embodiments of this application, the global pressure evolution trend reference curve can be constructed in the following manner:

[0016] Receive real-time operation data and extract all stored historical operation data from the database, including pipeline data collected by the data acquisition system in the past and operation records of the execution unit;

[0017] A time series analysis framework is adopted, and the historical and real-time running data are segmented using the sliding window method;

[0018] Extract the original pressure sequence of the main pipeline and its subordinate distribution nodes in the steam pipeline, that is, the pressure data in the steam pipeline sorted by time, and combine it with the operation records in the corresponding time period to obtain the pressure response curve.

[0019] The pressure response curves for each time period are smoothed to form a reference curve for the global pressure evolution trend.

[0020] In some embodiments of this application, during the construction of the load change model, a reasonable pressure range and fluctuation interval can be extracted based on the generated pressure evolution trend reference curve, and upper and lower boundary constraints for training the load change model can be constructed to provide convergence guidance during model fitting and control deduction.

[0021] In some embodiments of this application, the load variation model can be constructed using the following process:

[0022] Based on a long short-term memory network, the input is a multi-source feature vector based on historical pressure data, operation records and time metadata of the main pipeline and its subordinate distribution nodes in the steam pipeline at different operating stages.

[0023] By using a sliding window mechanism to fragment historical operational data, the model can gradually learn the operational characteristics of different time periods while covering the entire time period.

[0024] The input multi-source feature vector is encoded in time using a long short-term memory network to dynamically capture the dependency relationship between the main pipeline and the distribution node, and extract the potential pattern in the state transmission process. Finally, the predicted steam pressure values ​​of each key position at the next moment are output.

[0025] The loss function is optimized by mean square error to quantify the error between the predicted value and the actual pressure value.

[0026] The backpropagation algorithm is used to continuously correct the parameters in the long short-term memory network;

[0027] A graph convolutional neural network is introduced to model the topology of the steam pipeline and extract graph embedding vectors.

[0028] All graph embedding vectors are organized in chronological order to construct a temporal-spatial joint feature sequence, which is then input into a long short-term memory network to obtain the final load change model.

[0029] In some embodiments of this application, the process of using the graph convolutional neural network to perform structural modeling of the topology of steam pipelines can be designed with the following steps:

[0030] Construct a directed graph network based on the network topology of the steam pipeline to form a graph structure;

[0031] The graph structure information is standardized and encoded to construct a normalized adjacency matrix;

[0032] The spatial topological features of each node are encoded using a graph convolutional neural network. The single-layer graph convolutional propagation function is defined as follows:

[0033]

[0034] Among them, H (l+1) H is the input feature matrix of the (l+1)th layer. (l) Let l be the input feature matrix of the l-th layer. W is the normalized adjacency matrix. (l) Let σ be the learning weight matrix of the l-th layer, and σ be the non-linear activation function ReLU.

[0035] Perform multi-layer graph convolution operations to obtain high-dimensional representation vectors for each node, which constitute the graph embedding vectors of the nodes and are used to construct the load variation model.

[0036] In some embodiments of this application, the process of formulating the dynamic control strategy may include three parts: a model training phase, a model convergence phase, and a model execution phase; wherein:

[0037] During the model training phase, real-time operating data is used as the main input to dynamically analyze the operating status of each node in the steam pipeline. When the pressure value deviates, the control strategy mechanism is triggered to extract a suitable control strategy from the strategy library, schedule the execution unit, and execute the control strategy.

[0038] During the model convergence phase, the model prediction results are used as the main input to formulate a control strategy. When the mean square error of the model is stably lower than the set threshold in multiple consecutive time windows, and the predicted curves of each node tend to coincide with the measured curves, the model is considered to have converged. The predicted curves here are the reference curves for the global pressure evolution trend.

[0039] During the model operation phase, based on the model's prediction of steam demand for future periods, a cross-period pressure control plan is developed, and the target pressure range, allowable fluctuation range, and response threshold for each node of the steam pipeline are set in advance.

[0040] In some embodiments of this application, the dynamic control strategy can be formulated in the following manner:

[0041] A rolling optimization mechanism is introduced to dynamically adjust the control strategy based on real-time operational data and periodic prediction results. This includes:

[0042] Based on the current operating status and model prediction results, formulate a control strategy that covers a window of future time.

[0043] During the execution of the control strategy, real-time operational data is continuously collected and compared with the prediction results of the original model.

[0044] If a deviation is found between the real-time operating data and the prediction results, or if the operating conditions change, the effectiveness of the control strategy will be evaluated, and the content of the control strategy will be readjusted accordingly to generate an updated strategy plan.

[0045] The above process is repeated continuously in each rolling cycle, ensuring that the control strategy always keeps pace with the latest operational situation.

[0046] In some embodiments of this application, the load variation model can be modified in the following ways:

[0047] Compare and analyze the real-time operating data with the prediction results of the load change model;

[0048] Based on the deviations in the analysis results, the load change model is modified and dynamically compensated.

[0049] The length of the time series input window was readjusted, the weight ratio of each feature channel was adjusted, and the adjacency structure of the graph convolutional layer was locally corrected.

[0050] In some embodiments of this application, the policy library can be optimized in the following ways:

[0051] Using the system state before the control strategy is executed, the control actions taken, and the pressure change process after control as training samples, a strategy optimization model based on a convolutional neural network is constructed.

[0052] Align pressure data, control commands, and status indicators by time and unfold them into a two-dimensional tensor by time step;

[0053] By using multi-layer one-dimensional convolution kernels to slide extract two-dimensional tensors, the dynamic pressure response patterns caused by different control actions are identified, and the coupling relationship between their spatial location and temporal change characteristics is automatically captured.

[0054] By stacking convolutional and pooling layers, the original signal is abstracted layer by layer to extract high-level features with discriminative power. The comprehensive score of policy execution effect is combined in the fully connected layer to construct a loss function with the goal of minimizing the control error.

[0055] Backpropagation is used during training to update weights and optimize the network structure.

[0056] After training historical operating data with a convolutional neural network, the nonlinear mapping relationship between the control strategy and the pressure response is extracted, and the control strategy with the best performance under specific operating conditions is summarized. The control strategies that already exist in the current strategy library are then reconstructed and optimized.

[0057] In some embodiments of this application, the abnormal condition of the steam pipeline can be determined in the following ways:

[0058] The predicted pressure values ​​at key locations in the steam pipeline are dynamically compared with the collected real-time data. When there is a persistent deviation that cannot be explained by conventional load disturbances, it is determined that there is an anomaly in the area with the persistent deviation.

[0059] The pressure time series at each key location in the steam pipeline is tracked in real time. The global pressure trend reference curve is used as the benchmark background. The deviation accumulation technology is used to construct a cumulative deviation curve that reflects the direction and persistence of the deviation.

[0060] The relative pressure change relationship between nodes is mapped onto the topology of the steam pipeline to analyze the propagation path and expansion rate of the deviation between adjacent nodes.

[0061] By combining the cumulative deviation curve and the propagation path between nodes, representative spatial propagation patterns and persistent fluctuation characteristics are extracted to help determine whether the current deviation has systematic abnormal characteristics.

[0062] Generate structured early warning information or fault analysis reports to push to maintenance personnel for fault investigation.

[0063] Compared with the prior art, the advantages and positive effects of the present invention are mainly reflected in:

[0064] 1. This invention identifies the trend of micro-disturbance changes caused by frequent start-stop by constructing a spatiotemporally coupled load change model, thereby enhancing the adaptability to micro-amplitude high-frequency disturbances.

[0065] 2. By establishing a real-time analysis and control strategy linkage mechanism, this invention can respond immediately at the initial stage of pressure deviation, avoiding problems such as insufficient local steam supply and overpressure, and improving the response speed and accuracy of pressure regulation.

[0066] 3. This invention is based on the prediction and measured residual evolution sequence to assist in identifying potential fault risks of the system and generate alarm output, thereby enhancing the system's safety assurance capability.

[0067] Other features and advantages of the present invention will become clearer after reading the detailed embodiments of the invention in conjunction with the accompanying drawings. Attached Figure Description

[0068] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without any creative effort.

[0069] Figure 1 This is a flowchart illustrating an embodiment of the steam pipeline pressure control method proposed in this invention. Detailed Implementation

[0070] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0071] In the description of this invention, the consecutive numbers of the method steps are for ease of review and understanding. Considering the overall technical solution of this invention and the logical relationship between each step, adjusting the implementation order of the steps will not affect the technical effect achieved by the technical solution of this invention.

[0072] Combination Figure 1 As shown, the steam pipeline pressure control method of this embodiment mainly includes the following steps:

[0073] S100. Construct a data acquisition system and deploy execution units.

[0074] Specifically, the process includes the following:

[0075] S110. Deploy several data acquisition devices at key locations in the steam pipeline to collect data from inside the pipeline.

[0076] In this embodiment, each data acquisition unit has several sensors, which may include pressure sensors, temperature sensors, steam flow sensors, etc., to collect data such as pressure, temperature, and flow rate in the steam pipeline and transmit the collected data to the data processor.

[0077] S120. Preprocess the collected data.

[0078] The data processor preprocesses the data acquired by the sensor. The preprocessing process may specifically include:

[0079] S121, Data format standardization processing;

[0080] After receiving data from various sensors, the data processor performs format standardization processing on the collected data, unifying the data uploaded by different types of sensors into a structured format to ensure data format consistency.

[0081] S122, Unit Conversion;

[0082] Data collected by the same type of sensors and processed through format standardization are converted into a unified unit of measurement to facilitate subsequent analysis.

[0083] S123, Timestamp Alignment;

[0084] Data uploaded at different times is synchronized according to a unified time benchmark to ensure the comparability of multi-source data.

[0085] S124, filtering, noise reduction, and completion processing;

[0086] The collected data undergoes preliminary filtering and noise reduction to remove obviously abnormal or excessively erroneous data, and missing data is supplemented by interpolation.

[0087] S125. For each valid data record, attach its corresponding sensor number, device location identifier and acquisition time tag to generate pipeline data, and upload it to the cloud platform via the network.

[0088] S130. Deploy several execution units at key locations in the steam pipeline to achieve dynamic control of the pressure of each level of the steam pipeline network.

[0089] Specifically, a control system can be configured to connect to a steam generator within the pipeline to control steam generation. Key locations such as the steam generator outlet, the main pipeline inlet, and the inlets or intermediate sections of distribution nodes (branch pipelines at all levels) are designated as critical locations. Actuators such as control strategy valves, pressure relief valves, or bypass valves with control capabilities are installed at these locations. Each actuator is equipped with an electric drive module and a communication interface to receive and execute control strategy commands. After receiving the control strategy commands through its connected communication interface, the electric drive module dynamically adjusts the valve opening according to the commands, thereby achieving active control of the steam flow and pressure in the target pipeline section. Simultaneously, the electric drive module records its own operation data and uploads it to a cloud platform via a network.

[0090] Here, the control system and actuators that control the output parameters of the steam generator constitute the execution unit, which receives control strategy commands.

[0091] S200, Construct a reference curve for the global pressure evolution trend.

[0092] The cloud platform continuously receives real-time operational data from the steam pipeline network (including pre-processed pipeline data and operation records of execution units), and extracts all stored historical pipeline data and execution unit operation records from the database to construct a global pressure evolution trend reference curve. Specific steps include:

[0093] S210. A time series analysis framework is adopted, and the data is segmented using a sliding window method. The operational data is segmented using a fixed-length sliding window, with a certain step size, sliding across the historical pressure time series to divide the continuous operational data into multiple local time periods. Within each time window, the original pressure sequence of the main pipeline and its subordinate distribution nodes is extracted, and combined with the control operation records of the corresponding time period, a pressure response curve is obtained. This curve reflects the pressure change trajectory over time under specific equipment start-up and shutdown, load changes, and other operating conditions. The pressure response curves within each time period are smoothed to form a set of representative reference curves that reflect the global pressure evolution trend, used to describe common pressure evolution paths of the system under different operating rhythms.

[0094] S220. Based on the generated reference curve, extract the reasonable pressure range and fluctuation interval under each typical state, and then construct the upper and lower boundary constraints for predictive model training, providing convergence guidance in the process of model fitting and regulation deduction.

[0095] The typical state here can be determined based on the specific application scenario. For example, for a steel plant, the steam-using processes in the converter steel refining, hot rolling preheating, billet insulation, equipment cleaning, and pipeline purging sections can be taken as typical states.

[0096] Meanwhile, the reference curve can be used to assess whether the implementation of the strategy may lead to excessive volatility or violate the evolutionary trend of the system.

[0097] In addition, during operation, the prediction results can be dynamically compared and analyzed with the reference curve in real time to help identify systemic risks such as abnormal deviations and trend drift, thereby improving the robustness and safety of the model.

[0098] S300: Construct a load change model for critical nodes.

[0099] Based on the time-series evolution background reference provided by the global pressure trend reference curve, a load change model oriented towards key nodes is constructed, which specifically includes the following steps:

[0100] S310. Using a multivariate modeling method, the extracted features of the main pipeline and distribution nodes are integrated to construct a load change model.

[0101] Specifically, a Long Short-Term Memory (LSTM) network with temporal memory capabilities can be used as the core framework to capture the temporal dependence and spatial correlation between pressure changes and equipment operating modes, extract common patterns of systemic steam consumption behavior under different operating states, and focus on identifying micro-amplitude, high-frequency fluctuation patterns caused by frequent start-stop cycles and low-intensity disturbances. The specific process is as follows:

[0102] S311. Based on the Long Short-Term Memory (LSTM) network structure, the input is a multi-source feature vector based on historical pressure data, operation records and time metadata of the main pipeline and multiple distribution nodes at different operating stages.

[0103] The multi-source feature vector is specifically as follows: pressure data is extracted in the form of a time series and formed into a fixed-length real vector in continuous time steps; the operation record is structurally transformed according to its data attributes, the binary state signal is represented as a binary vector, the multi-state signal is generated into a dense vector through one-hot encoding, and the continuous control quantity is directly retained as a real-valued feature; the time data is represented as a numerical vector with time position significance after periodic mapping.

[0104] S312. By using a sliding window mechanism to fragment historical operating data, the model can gradually learn the operating characteristics of different time periods while covering the entire time period, and identify the typical pressure response patterns of key locations under specific steam usage combinations under continuous time series input.

[0105] S313. Long Short-Term Memory (LSTM) networks encode the input multi-source feature vectors over time, dynamically capture the dependencies between the main pipeline and multiple distribution nodes, extract potential patterns during state propagation, and finally output the predicted values ​​of steam pressure at each key location at the next moment.

[0106] S314. The loss function is optimized using the mean square error (MSE) to quantify the error between the predicted and actual pressure values. The expression for the mean square error (MSE) is:

[0107]

[0108] Where n is the total number of samples; y i Let i be the actual pressure value of the i-th sample. Let be the predicted stress value for the i-th sample.

[0109] S315. Using the backpropagation algorithm, continuously adjust the parameters in the Long Short-Term Memory (LSTM) network.

[0110] In the forward propagation phase, the model's prediction of the input features is calculated, and a loss function is constructed based on the error between the actual and predicted pressure values. Subsequently, in the backpropagation phase, the error signal is propagated forward layer by layer starting from the output layer. The chain rule is used to calculate the partial derivatives of each network parameter with respect to the loss, obtain the gradient information of each weight parameter, and use the adaptive moment estimator optimizer Adam to iteratively update the network weights. This causes the loss function value to gradually decrease during training iterations, thereby guiding the model to continuously correct the prediction error of the pressure trend.

[0111] S320. Use graph convolutional neural networks to perform structural modeling of the topology of steam pipelines.

[0112] To more comprehensively reflect the structural relationships and interaction characteristics between the main pipeline and multiple distribution nodes (multi-level branch pipelines), this embodiment introduces a Graph Convolutional Neural Network (GCN) to structurally model the topology of the steam pipeline. The specific process is as follows:

[0113] S321. Construct a directed graph network based on the network topology of the steam pipeline to form a graph structure.

[0114] A directed graph network can be represented as G = (V, E); where V represents the set of nodes, and each key location can be considered as a node. Each node records the pipeline data and operation records of that node; E is the set of edges, representing the physical connection relationship between nodes and the direction of steam flow.

[0115] By using graph structures, the spatial connections between nodes in a steam pipeline network can be revealed, thereby capturing the pressure transmission paths and spatial coupling characteristics between different pipe segments.

[0116] S322. Use a graph convolutional neural network (GCN) to encode and model the graph structure, and extract the graph embedding vector.

[0117] The process of standardizing and encoding graph structure information is as follows:

[0118] First, define the adjacency matrix A. Then, by adding self-loops to the original network structure and performing normalization, we obtain the normalized adjacency matrix. The specific formula for normalization is:

[0119]

[0120] in, Let A be the normalized adjacency matrix, D be the degree matrix, and I be the identity matrix.

[0121] An adjacency matrix is ​​a storage structure for graphs that represents the adjacency relationships between nodes in a graph. It describes the connection state or weight between nodes using a two-dimensional array and is widely used in graph theory and data structures.

[0122] After constructing the normalized adjacency matrix, a graph convolutional neural network is used to encode the spatial topological features of each node. The single-layer graph convolutional propagation function is defined as follows:

[0123]

[0124] Among them, H (l+1) H is the input feature matrix of the (l+1)th layer. (l) Let l be the input feature matrix of the l-th layer. W is the normalized adjacency matrix. (l) Let σ be the learning weight matrix of the l-th layer, and σ be the non-linear activation function ReLU.

[0125] By performing multi-layer graph convolution operations, a high-dimensional representation vector for each node can be obtained, denoted as H. (l) The high-dimensional representation vector H (l) The graph embedding vector of a node is formed by integrating the node's own attribute information and the adjacency structure features.

[0126] S330. Organize all graph embedding vectors in chronological order to construct a temporal-spatial joint feature sequence, which is then input into the LSTM network to obtain the final load change model.

[0127] Introducing spatial features encoded by graph structure during the temporal modeling process can enhance the spatial dependencies brought about by the topological structure.

[0128] Specifically, each graph embedding vector not only contains attribute information such as the node's own pressure state and control behavior, but also integrates its connection relationships with neighboring nodes and historical interaction features. During temporal modeling, the LSTM network learns the spatial dependency information carried in the graph embedding vector to identify the pressure coupling effect between nodes caused by physical connections. Thus, while capturing the temporal evolution trend, it can more accurately model the spatial dynamics of multi-node collaborative changes in the steam pipeline network. When predicting steam pressure, the model can not only identify the evolution trend in the time dimension, but also perceive the collaborative changes between different nodes caused by topological coupling. This allows for a more accurate simulation and prediction of the dynamic pressure distribution of the entire steam pipeline network. The resulting load change model can comprehensively characterize the spatiotemporal coupling law of steam use and pressure change in the steam pipeline network, reflect the pressure evolution trend under the overall operation arrangement, and accurately capture the local dynamic features of each allocation node.

[0129] S400. Analyze the obtained load change model, and formulate a matching dynamic control strategy based on the output results and predicted trends of the load change model.

[0130] This process can specifically include: the model training phase, the model convergence phase, and the model execution phase.

[0131] S410, Model Training Phase

[0132] During the model training phase, multi-source data such as real-time collected pressure, flow rate, temperature, and operation records are used as the main input to dynamically analyze the operating status of each node in the current steam pipeline network and evaluate the pressure distribution of each pipe segment in real time. When the pressure value deviates from the set target range, or when abnormal trends such as sudden pressure rise, violent fluctuations, or slow response occur, it will be determined that there is a potential instability risk in that area, and the control strategy mechanism will be triggered immediately. A suitable control strategy will be extracted from the strategy library, and the execution unit will be scheduled to execute the targeted rapid control strategy.

[0133] Specific control strategies include: on the one hand, directly regulating the power output of the steam generator and adjusting the combustion intensity or heating rate according to real-time pressure feedback to control changes in source pressure; on the other hand, sending control commands to the pipeline execution unit to execute a rapid control strategy, effectively buffering local pressure fluctuations by controlling the rate of change of valve opening, and preventing potential faults such as gas blockage, short-term overpressure, or sudden pressure drop.

[0134] Through the above mechanism, abnormal pressure states can be quickly identified and intervened in in a timely manner, thereby improving the overall stability and security of the system.

[0135] Simultaneously, based primarily on real-time operational data analysis, the system also references the predictive results of load change models to assist in formulating relatively long-term control strategies. By identifying trends and analyzing periodic fluctuations in historical operational data, combined with model predictions of pressure distribution within short-term time windows, target pressure ranges and control strategy plans for each pipeline segment can be pre-planned, and continuously fine-tuned based on real-time data during operation. This strategy ensures rapid short-term response while possessing a certain degree of forward-looking control capability.

[0136] S420, Model Convergence Phase

[0137] During the model convergence phase, when the mean squared error (MSE) of the model is consistently below a set threshold over multiple consecutive time windows, and the predicted curves at each node tend to coincide with the measured curves, it indicates that the model has strong generalization ability and environmental adaptability, and model convergence is considered complete. At this point, the model will enter an operational phase centered on prediction and based on real-time feedback for correction, and a longer-term, more stable control strategy will be developed.

[0138] The threshold setting is primarily determined based on the actual pressure prediction accuracy requirements of the steam pipeline system. This threshold is not a fixed value but is dynamically set considering multiple factors. The core principle is to ensure that the model's predictive ability is within the system's acceptable error range and possesses sufficient steady-state reliability. Specifically, the threshold setting typically considers the following three factors:

[0139] First, the control method itself has requirements for pressure control accuracy, namely, the allowable pressure fluctuation range of each node under normal operating conditions, which is used to define the tolerable upper limit of prediction deviation.

[0140] Second, the error convergence characteristics during model training, that is, the fluctuation range exhibited when the MSE gradually decreases and tends to stabilize as training iterations proceed, serves as an important reference for judging model stability.

[0141] Third, operational safety and fault tolerance redundancy in policy execution. That is, when model prediction results are used for regulatory decisions, they must have a certain level of confidence and error boundaries to prevent over-regulation or delayed response caused by prediction deviations during policy execution.

[0142] The specific threshold should be set according to the actual working conditions.

[0143] S430, Model Execution Phase

[0144] During the model operation phase, based on the model's prediction of steam demand for future periods, a cross-period pressure control plan is developed, pre-setting the target pressure range, allowable fluctuation range, and response threshold for each steam node. Specifically, this includes: pre-setting the operating schedule and output rhythm of the steam generator to match the basic steam supply capacity required for future load changes; and simultaneously, through various execution units, pre-configuring the control strategies of each control node in the pipeline, enabling key pipeline sections to gradually adapt to steam pressure changes within the expected time window, thereby effectively buffering the impact of pressure fluctuations.

[0145] When formulating long-term control strategies, a certain control strategy margin is preset to meet the pressure fluctuation boundary provided by the global pressure trend reference curve, ensuring that the strategy response is executed within a controllable range, preventing drastic intervention from causing system instability, and coping with the operating deviations caused by model prediction errors, load disturbances or boundary condition fluctuations, thereby maintaining the stability and flexibility of the control system.

[0146] S440. Introducing a rolling optimization mechanism, the control strategy is dynamically adjusted based on real-time feedback and periodic prediction results to avoid new response lags or local imbalances caused by rigid execution of the control strategy.

[0147] The specific steps include:

[0148] S441. Based on the current operating status and model prediction results, formulate a control strategy that covers a window of time in the future;

[0149] S442. During the execution of the control strategy, real-time operating data is continuously collected and compared with the original predicted values ​​for analysis.

[0150] S443. If a deviation is found between the real-time operating data and the predicted value, or if the operating conditions change, the effect of the control strategy should be evaluated immediately, and the content of the control strategy should be readjusted accordingly to generate an updated strategy plan.

[0151] The above process is repeated continuously in each rolling cycle, ensuring that the control strategy always keeps pace with the latest operational situation.

[0152] At the same time, it retains the ability to quickly analyze real-time operating data, ensuring that when prediction errors, sudden disturbances, or nonlinear changes exceed the model's perception range, it can promptly capture the actual performance of pressure changes in the steam pipeline and make temporary adjustment plans accordingly, so as to ensure sufficient pressure control flexibility and fault response capability.

[0153] S500 performs model corrections based on real-time operational data, evaluates control strategies, and uses deep learning to optimize the strategy library.

[0154] During operation, real-time operating data of the steam pipeline is continuously acquired and compared with the prediction results of the load change model. When a deviation is found between the predicted value and the actual measured data, the model parameters are corrected and dynamically compensated accordingly. This includes readjusting the length of the time series input window, adjusting the weight ratio of each feature channel, and locally correcting the adjacency structure of the graph convolutional layer, thereby improving the model's prediction accuracy and robustness under the current operating conditions.

[0155] Building upon this, we further employ deep learning methods to model and learn the nonlinear correlation between policy execution behavior and system response results, which is used to continuously optimize the control policies in the policy library.

[0156] The specific method is as follows: using the state-action-feedback triplet for each control operation—that is, the system state before the control strategy is executed, the control action adopted, and the pressure change process after control—as training samples, a policy optimization model based on a convolutional neural network (CNN) is constructed to extract features and summarize patterns from the system response patterns of various typical control operations under different operating states. Specific steps include:

[0157] S510. Align the pressure data, control commands, and status indicators by time, and expand them into a two-dimensional tensor by time step. Rows represent different time points, and columns represent various feature variables, representing the state evolution trajectory and control response process within different time windows.

[0158] S520 uses multi-layer one-dimensional convolution kernels to slide extract two-dimensional tensors, identifies the dynamic pressure response patterns caused by different regulatory behaviors, and automatically captures the coupling relationship between their spatial location and temporal change characteristics.

[0159] S530: By stacking convolutional and pooling layers, the original signal is abstracted layer by layer to extract high-level features with discriminative power. These features are then combined with a comprehensive score of policy execution performance, such as target pressure achievement, control policy timeliness, and energy consumption cost, in the fully connected layer as a supervision signal to construct a loss function that minimizes the control error. Backpropagation is used during training to update weights and optimize the network structure.

[0160] S540. After training historical operating data through deep learning using a convolutional neural network (CNN), the nonlinear mapping relationship between control strategy and pressure response is extracted, and the regulation feature pattern with optimal performance under specific working conditions is summarized.

[0161] S550. Reconstruct and optimize the existing control strategies in the current strategy library, correct strategies with slow response, frequent control strategies, or poor effects, and replace them with new parameter configurations or execution logic that match the efficient strategies, so as to realize the dynamic updating of strategy content.

[0162] S560: In actual operation, the system combines the current operating conditions and the output results of the CNN model to recommend the optimal adjustment scheme and make real-time corrections to the control strategy being executed, thereby improving the pertinence and stability of the control actions.

[0163] S600 monitors the difference between real-time model data and actual data to determine abnormal conditions in steam pipelines.

[0164] Specifically, it can include the following steps:

[0165] S610. After the model is running stably, the predicted pressure values ​​at each key location are continuously compared with the real-time operational data collected in real time.

[0166] When a significant deviation occurs between the actual pressure data and the model prediction results in a certain area, that is, within a continuous monitoring period, the difference between the actual pressure value and the model prediction value exceeds the reasonable operating range, and the difference is far from the normal fluctuation range in terms of numerical amplitude, showing an abnormal state of abnormally low or high pressure value, reaching an extreme level that cannot effectively maintain the steam supply function, and has a continuous characteristic that cannot be explained by conventional load disturbances, it is determined that there is an abnormal operating state or potential fault risk in the area, and an alarm process is automatically triggered.

[0167] S620. When determining that there is an abnormal situation in a certain area, the abnormal situation is comprehensively analyzed by combining the direction of data deviation, evolution trend and spatial propagation characteristics in the steam pipeline network, and abnormal pressure behavior patterns are identified.

[0168] Specific methods include:

[0169] S621. Real-time tracking of the pressure time series at each key location, subtracting the real-time monitored pressure value from the pressure value predicted by the model at each time point to obtain a set of prediction error values ​​that change over time.

[0170] S622. Using the global pressure trend reference curve as a baseline background helps identify systemic or continuous operational deviation patterns, improving the accuracy and timeliness of anomaly identification.

[0171] S623. Using the deviation accumulation technique, the residual values ​​are continuously accumulated and analyzed to construct a cumulative deviation curve that reflects the direction and persistence of the deviation. This curve is used to identify whether the pressure deviation exhibits significant unidirectional accumulation characteristics, thereby helping to determine whether the abnormal deviation is systematic and persistent.

[0172] S624. Map the relative pressure change relationship between nodes onto the pipeline topology and analyze the propagation path and expansion rate of the deviation between adjacent nodes.

[0173] S625. By comparing the deviation evolution pattern during normal operation with the spatial distribution characteristics of the current deviation in the steam pipeline network, determine whether the deviation exhibits abnormal clustering, nonlinear expansion trend, or directional enhancement along a specific path.

[0174] S626. By combining the cumulative change curve of the deviation and the propagation path between nodes, extract representative spatial propagation patterns and continuous fluctuation characteristics to help determine whether the current deviation has systematic abnormal characteristics.

[0175] When the analysis results show a significant deviation from the system's usual operating characteristics, and the trend does not return to the expected track with the control behavior, it is determined that the anomaly may originate from the system's inherent disturbance or potential failure mechanism.

[0176] S630. After completing the anomaly identification, generate structured early warning information or fault analysis report.

[0177] The fault analysis report may include the location information of the abnormal node, the degree of difference between the predicted and measured data, the characteristics of the anomaly's persistence, the potential system risk level, and the recommended response measures.

[0178] S640 transmits early warning information or fault analysis reports to maintenance personnel through the system interface.

[0179] Industrial applicability

[0180] Taking the application of the steam pipeline pressure control method of this embodiment to the centralized steam supply network of a large steel plant as an example, the industrial applicability of the control method of this embodiment is illustrated.

[0181] A steel plant has multiple production sections involving continuous or intermittent steam consumption processes such as converter steel refining, hot rolling preheating, billet insulation, equipment cleaning, and pipeline purging. The plant has high requirements for the stability of steam pressure and often experiences problems such as concentrated steam consumption during peak periods and instantaneous pressure fluctuations caused by frequent equipment start-ups and shutdowns.

[0182] To achieve precise steam supply and automated pressure control strategies, the steam pipeline pressure control method of this embodiment is applied to the steam supply system of this steel plant.

[0183] First, configure a control system to connect to the steam generator, and install pressure sensors and controllable actuators on the main pipeline and the main branch pipes of the steelmaking section. Then, upload real-time operating data to the cloud platform via the network.

[0184] The cloud platform continuously receives historical and real-time operational data from the pipeline network. Combined with pressure information and control records, it constructs a global pressure trend reference curve covering the entire operating cycle, extracts the pressure response curves of the steel plant under high load, low load, start-stop switching and other conditions, and determines the acceptable pressure fluctuation boundary accordingly.

[0185] Based on this reference curve, a load change model is constructed to analyze and learn data such as steam consumption rhythm, pressure changes and control response at key nodes of the steam pipeline, and extract the regular characteristics of different operating stages.

[0186] When it is detected that a certain area is about to enter a period of load fluctuation or that the pressure is deviating from the trend, the control system and / or actuators are driven to take action in advance through control strategies, such as adjusting valve opening, limiting steam flow, and adjusting steam generator output parameters, so as to keep the actual operating data within the trend boundary.

[0187] By comparing the predicted results with the actual measurement data in real time through the cloud platform, if a persistent deviation is found, the model adjustment mechanism is triggered to dynamically correct the relevant parameters and improve the matching accuracy of subsequent prediction and control.

[0188] In the initial stages of deployment, the steam supply process was dynamically controlled primarily by relying on real-time steam consumption data from the steel plant, and each execution unit was controlled based on on-site pressure feedback. Due to the frequent start-ups and shutdowns of metallurgical equipment and the overlap of steam consumption at multiple points, the steam supply was somewhat unstable.

[0189] As the model gradually converges, it gradually identifies the steam consumption patterns of the steel plant. For example, it detects that after each round of molten steel processing in the converter, frequent and small-volume purging and injection operations are required. Therefore, steam resources are allocated in advance, the output of the steam generator is appropriately increased, and the pressure of the main pipe and key branch pipes is increased in advance to cope with the upcoming instantaneous high load.

[0190] Subsequently, based on the operating rhythm of steam consumption for staggered heat treatment and cooling in different production lines, the steam output is gradually reduced, and the control strategy is switched accordingly. For example, if a heating section is identified as operating every 30 minutes, with each steam consumption lasting approximately 8 minutes and accompanied by a pressure drop of 0.1–0.15 MPa, the pressure can be pre-pressurized and stabilized before its start-up to prevent sudden drops from affecting steam supply to other branches. The model covers typical operating modes of a steel plant, achieving forward-looking prediction and dynamic control of steam consumption behavior across the entire plant, ensuring stable pipeline pressure under complex load conditions.

[0191] Meanwhile, the system compares the predicted results with the actual measurement data in real time. When a systematic deviation trend is detected, it automatically fine-tunes the parameters and corrects the strategy to ensure that subsequent prediction and control strategies are more in line with the current operating conditions.

[0192] During continuous operation, control response data under multiple typical operating stages are accumulated, and verified effective control strategies are incorporated into the strategy library to optimize the control strategies in the library. This strategy library contains steam supply control parameters and execution commands for common load modes in steel plants, such as converter switching, hot rolling switching, and intensive injection operation cycles. When a similar operating mode is detected to reappear, existing solutions in the strategy library can be preferentially invoked to achieve rapid control response.

[0193] Furthermore, the system continuously compares actual pressure data with predicted values ​​to dynamically calculate deviation sequences. If a significant deviation persists in a certain area and cannot be explained by normal load fluctuations, it is identified as a potential abnormal behavior and an automatic warning is issued. By combining the direction and speed of deviation propagation in the pipeline network, the cloud platform can further pinpoint possible fault locations, such as pressure imbalances caused by valve blockage, branch pipe water accumulation, or abnormal steam return from equipment. This assists maintenance personnel in quickly completing diagnosis and repair, significantly improving the safety and response efficiency of the steam supply system.

[0194] Of course, the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions claimed by the present invention.

Claims

1. A method for controlling steam pipeline pressure, characterized in that, include: A data acquisition system and execution unit are deployed in the steam pipeline to collect pipeline data and operation records of the execution unit in real time, forming real-time operation data. A global stress evolution trend reference curve is constructed based on real-time operational data and stored historical operational data; A multivariate modeling method is adopted, and a graph convolutional neural network is introduced to perform structural modeling of the topology of the steam pipeline. This model is then combined with a global pressure evolution trend reference curve to construct a load change model. Based on the output results and predicted trends of the load change model, a matching dynamic control strategy is formulated to control the actions of the execution unit, and the load change model is corrected based on the feedback of real-time operating data. Deep learning is performed using convolutional neural networks to optimize the policy library; Real-time monitoring of the discrepancies between model data and actual data helps identify abnormal conditions in steam pipelines.

2. The steam pipeline pressure control method according to claim 1, characterized in that, The process of constructing the global pressure evolution trend reference curve includes: Receive real-time operational data and extract all stored historical pipeline data and execution unit operation records from the database; A time series analysis framework is adopted, and the data is segmented using the sliding window method; The original pressure sequence of the main pipeline and its subordinate distribution nodes in the steam pipeline is extracted, and the pressure response curve is obtained by combining it with the operation record of the corresponding time period. The pressure response curves for each time period are smoothed to form a reference curve for the global pressure evolution trend.

3. The steam pipeline pressure control method according to claim 2, characterized in that, In the process of constructing the load change model, based on the generated pressure evolution trend reference curve, a reasonable pressure range and fluctuation interval are extracted, and upper and lower boundary constraints for training the load change model are constructed.

4. The steam pipeline pressure control method according to claim 1, characterized in that, The process of constructing the load change model includes: Based on a long short-term memory network, the input is a multi-source feature vector of historical pressure data, operation records and time metadata of the main pipeline and its subordinate distribution nodes in the steam pipeline at different operating stages. By using a sliding window mechanism to fragment historical operational data, the model can gradually learn the operational characteristics of different time periods while covering the entire time period. The input multi-source feature vector is encoded in time using a long short-term memory network to dynamically capture the dependency relationship between the main pipeline and the distribution node, and extract the potential patterns in the state transmission process. Finally, the predicted steam pressure values ​​of each key position at the next moment are output. The loss function is optimized by mean square error to quantify the error between the predicted value and the actual pressure value. The backpropagation algorithm is used to continuously correct the parameters in the long short-term memory network; A graph convolutional neural network is introduced to model the topology of the steam pipeline and extract graph embedding vectors. All graph embedding vectors are organized in chronological order to construct a temporal-spatial joint feature sequence, which is then input into a long short-term memory network to obtain the final load change model.

5. The steam pipeline pressure control method according to claim 4, characterized in that, The process of using the graph convolutional neural network to perform structural modeling of the topology of steam pipelines includes: Construct a directed graph network based on the network topology of the steam pipeline to form a graph structure; The graph structure information is standardized and encoded to construct a normalized adjacency matrix; The spatial topological features of each node are encoded using a graph convolutional neural network. The single-layer graph convolutional propagation function is defined as follows: Among them, H (l+1) H is the input feature matrix of the (l+1)th layer. (l) Let l be the input feature matrix of the l-th layer. W is the normalized adjacency matrix. (l) Let σ be the learning weight matrix of the l-th layer, and σ be the non-linear activation function ReLU. Perform multi-layer graph convolution operations to obtain the high-dimensional representation vector of each node, which constitutes the graph embedding vector of the node.

6. The steam pipeline pressure control method according to claim 1, characterized in that, The process of formulating the dynamic control strategy includes a model training phase, a model convergence phase, and a model execution phase, wherein: During the model training phase, real-time operating data is used as the main input to dynamically analyze the operating status of each node in the steam pipeline. When the pressure value deviates, the control strategy mechanism is triggered to extract a suitable control strategy from the strategy library, schedule the execution unit, and execute the control strategy. During the model convergence phase, the model prediction results are used as the main input to formulate a control strategy. When the mean square error of the model is stably lower than the set threshold in multiple consecutive time windows, and the predicted curves of each node tend to coincide with the measured curves, the model is considered to have converged. During the model operation phase, based on the model's prediction of steam demand for future periods, a cross-period pressure control plan is developed, and the target pressure range, allowable fluctuation range, and response threshold for each node of the steam pipeline are set in advance.

7. The steam pipeline pressure control method according to claim 6, characterized in that, The process of formulating the dynamic control strategy also includes: A rolling optimization mechanism is introduced to dynamically adjust the control strategy based on real-time operational data and periodic prediction results. This includes: Based on the current operating status and model prediction results, formulate a control strategy that covers a window of future time. During the execution of the control strategy, real-time operational data is continuously collected and compared with the prediction results of the original model. If a deviation is found between the real-time operating data and the prediction results, or if the operating conditions change, the effectiveness of the control strategy will be evaluated, and the content of the control strategy will be readjusted accordingly to generate an updated strategy plan. The above process is repeated continuously in each rolling cycle, ensuring that the control strategy always keeps pace with the latest operational situation.

8. The steam pipeline pressure control method according to claim 4 or 5, characterized in that, The process of correcting the load variation model includes: Compare and analyze the real-time operating data with the prediction results of the load change model; Based on the deviations in the analysis results, the load change model is modified and dynamically compensated. The length of the time series input window was readjusted, the weight ratio of each feature channel was adjusted, and the adjacency structure of the graph convolutional layer was locally corrected.

9. The steam pipeline pressure control method according to claim 1, characterized in that, The process of optimizing the strategy library includes: Using the system state before the control strategy is executed, the control actions taken, and the pressure change process after control as training samples, a strategy optimization model based on a convolutional neural network is constructed. Align pressure data, control commands, and status indicators by time and unfold them into a two-dimensional tensor by time step; By using multi-layer one-dimensional convolution kernels to slide extract two-dimensional tensors, the dynamic pressure response patterns caused by different control actions are identified, and the coupling relationship between their spatial location and temporal change characteristics is automatically captured. By stacking convolutional and pooling layers, the original signal is abstracted layer by layer to extract high-level features with discriminative power. The comprehensive score of policy execution effect is combined in the fully connected layer to construct a loss function with the goal of minimizing the control error. Backpropagation is used during training to update weights and optimize the network structure. After training historical operating data with a convolutional neural network, the nonlinear mapping relationship between the control strategy and the pressure response is extracted, and the control strategy with the best performance under specific operating conditions is summarized. The control strategies that already exist in the current strategy library are then reconstructed and optimized.

10. The steam pipeline pressure control method according to claim 1, characterized in that, The process for determining abnormal conditions in the steam pipeline includes: The predicted pressure values ​​at key locations in the steam pipeline are dynamically compared with the collected real-time data. When there is a persistent deviation that cannot be explained by conventional load disturbances, it is determined that there is an anomaly in the area with the persistent deviation. The pressure time series at each key location in the steam pipeline is tracked in real time. The global pressure trend reference curve is used as the benchmark background. The deviation accumulation technology is used to construct a cumulative deviation curve that reflects the direction and persistence of the deviation. The relative pressure change relationship between nodes is mapped onto the topology of the steam pipeline to analyze the propagation path and expansion rate of the deviation between adjacent nodes. By combining the cumulative deviation curve and the propagation path between nodes, representative spatial propagation patterns and persistent fluctuation characteristics are extracted to help determine whether the current deviation has systematic abnormal characteristics. Generate structured early warning information or fault analysis reports.

Images