A steam pipe network heat preservation adjustment optimization method based on deep learning

Abstract

The application discloses a kind of steam pipe network heat preservation regulation optimization methods based on deep learning, comprising the following steps: steam pipe network operation time series data and environmental time series data are collected, and aligned data sequence is generated;Identify control action change and working condition mutation, and the interaction intensity sequence is calculated using Hawkes point process;Based on the aligned data sequence, the pipe section is identified on line, and the modulated interaction intensity sequence is obtained;HP-DGNN prediction model is constructed, and the rolling prediction distribution in future time domain is output, risk assessment is executed, and the risk upper bound is output;Robust security constraint optimization problem is constructed and solved, and the candidate control sequence is obtained;Execute pre-execution safety check and minimum change correction to candidate control sequence, generate the control sequence of issue;The control sequence of issue is arranged in time sequence, and the steam pipe network heat preservation regulation optimization strategy result is obtained.The application improves the prediction accuracy and safety robustness of steam pipe network heat preservation regulation optimization strategy generation.

Description

Technical Field

[0001] This invention relates to the field of industrial steam pipeline control technology, and in particular to a deep learning-based method for optimizing the insulation and regulation of steam pipelines. Background Technology

[0002] Steam pipeline networks, commonly used in industrial parks, combined heat and power plants, and large-scale public works projects for heating and transporting process media, typically consist of heat sources, main pipes, branch pipes, valves, and terminal steam-using equipment. They are characterized by large pipe spans, frequent fluctuations in operating conditions, and significant influence from the external environment. In existing technologies, to reduce heat loss during transport and ensure the quality of steam supply at the terminals, adjustment methods based on empirical rules or mechanistic models are commonly used: one approach relies on manually setting valve openings, desuperheating and depressurization settings, and bypass strategies, combined with routine inspections or simple threshold alarms for adjustment; another approach is based on thermal-hydraulic calculations or simplified heat transfer models for prediction and optimization, combined with PID or MPC control methods to adjust operating parameters.

[0003] Existing control strategies based on empirical rules or threshold triggers often struggle to structurally model changes in control actions and abrupt changes in operating conditions. They fail to depict the sustained impact of current events on future states and the role of pipeline topology coupling in disturbance propagation at different operating stages. This results in lag or overly conservative responses from strategies to events such as sudden loads, valve switching, and bypass opening / closing. While optimization methods based on mechanistic models can describe the thermal-hydraulic relationship to some extent, these models require extensive parameter calibration and are difficult to update online as insulation aging, scaling, or local resistance changes occur, leading to a gradual accumulation of prediction errors. When external environmental changes are significant or pipeline operating modes switch frequently, model mismatch is further exacerbated, causing optimization results to exhibit end-point temperature and pressure fluctuations, increased condensation risk, or increased energy consumption in actual implementation.

[0004] Therefore, how to provide a deep learning-based method for optimizing the insulation and regulation of steam pipeline networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a deep learning-based optimization method for steam pipeline insulation regulation. This invention couples event-driven interaction strength with online pipe segment identification, combines HP-DGNN rolling prediction distribution and risk upper bound to construct robust safety constraint optimization and with minimum modification correction before execution. This enables adaptive regulation to sudden changes in operating conditions and drift in insulation state. While ensuring the quality of steam supply at the end and reducing condensation risk, it also reduces ineffective regulation and energy consumption, and improves the executability and stability of the strategy.

[0006] A steam pipeline insulation regulation optimization method based on deep learning according to an embodiment of the present invention includes the following steps:

[0007] Collect steam pipeline network operation time series data and environmental time series data, and perform time alignment to generate aligned data sequences;

[0008] Based on the alignment data sequence, control action changes and sudden changes in operating conditions are identified, an event-driven interaction sequence is generated, and the Hawkes point process is used to calculate the interaction intensity sequence under the combined action of the event-driven interaction sequence and the historical sequence of the pipeline structure.

[0009] The pipe segment is identified online based on the aligned data sequence, the heat transfer and resistance related parameters of the pipe segment are updated, and the interaction intensity sequence is modulated using the updated pipe segment parameters to obtain the modulated interaction intensity sequence.

[0010] An HP-DGNN prediction model is constructed using aligned data sequences and modulated interaction intensity sequences as inputs. The model is trained and its parameters are fixed. Simultaneously, rolling inference is performed to output the rolling prediction distribution in the future time domain. Risk assessment is performed on safety constraint-related indicators, and the risk upper bound is output.

[0011] A robust safety constraint optimization problem is constructed and solved based on the rolling prediction distribution and risk upper bound to obtain candidate control sequences;

[0012] Perform pre-execution security checks and minimum modification corrections on candidate control sequences to generate control sequences that meet security constraints;

[0013] The issued control sequences are arranged in chronological order to obtain the results of the steam pipeline insulation regulation optimization strategy.

[0014] Optionally, the generation of the aligned data sequence specifically includes:

[0015] The steam pipeline network is continuously sampled according to the preset sampling period. Operational data and environmental data are obtained separately to form the original operational data sequence and the original environmental data sequence. Each sampling record is written with a sampling time identifier and a source identifier. Sampling records for which no valid observations were obtained are marked as missing.

[0016] Time alignment processing is performed on the original running data sequence and the original environmental data sequence using a unified time base. Each sampled record is mapped to the corresponding time base point. When different records exist at the same time point, the valid record is selected according to the preset rules. When a record is missing, the missing mark is maintained and the missing position is recorded.

[0017] The data that has been aligned to the time is processed for consistency. Missing time points are filled in according to preset rules and the filling marks are retained. The running data and environmental data are combined to generate an aligned data sequence in chronological order.

[0018] Optionally, the calculation process of the interaction intensity sequence under the combined effect of the event-driven interaction sequence and the network structure historical sequence specifically includes:

[0019] The aligned data sequence is scanned point by point according to the preset time window. At each sampling moment, the difference between the current sampled value and the previous sampled value is calculated, and the difference is compared with the preset threshold. When the difference meets the threshold trigger condition, the current sampling moment is registered as the event occurrence moment. At the same time, the event type identifier is registered and the event amplitude parameter is calculated to form an event record.

[0020] Sort all event records by the time of occurrence, and write the sorted event records into the event-driven interaction sequence in sequence;

[0021] Based on the alignment data sequence and the topology relationship of the steam pipeline network, a historical sequence of pipeline network structure is generated. The coupling state of each pipe segment and node at each sampling time is calculated with a preset time window. The coupling state of each sampling time is written into the historical sequence of pipeline network structure in chronological order to obtain a historical structural input consistent with the time reference of the alignment data sequence.

[0022] The Hawkes point process intensity calculation is performed using the event-driven interaction sequence and the pipeline structure history sequence as input. At each sampling time, a baseline intensity value is first determined. At the same time, for each event that occurred before the current sampling time, a corresponding weight coefficient is selected according to the event type. The weight coefficient is multiplied by the event amplitude parameter to obtain the event contribution value. The event contribution value is then subjected to a preset attenuation kernel function according to the time interval from the event occurrence time to the current sampling time to obtain the attenuated event contribution value. At the same time, the structural history input is weighted and summed according to the structural item weight coefficient at the sampling time to obtain the structural contribution value. The sum of the baseline intensity value and the attenuated event contribution value is combined with the structural contribution value to obtain the interaction intensity value at the current sampling time.

[0023] Arrange the interaction intensity values ​​at all sampling times in chronological order to generate an interaction intensity sequence and output it.

[0024] Optionally, the process of obtaining the modulated interaction intensity sequence specifically includes:

[0025] Using the aligned data sequence as input, the data is aggregated according to the topology of the steam pipe network, an online identification status is established for each pipe segment, the initial estimated values ​​of the heat transfer parameters and the initial estimated values ​​of the resistance parameters of the current pipe segment are written, and the parameter update cycle, sliding window length and forgetting coefficient are set.

[0026] At each parameter update time, the sliding window data ending at the current time is extracted from the aligned data sequence, the identification input data and identification output data are set, and the corresponding identification samples are generated.

[0027] For each parameter update time, a recursive update operation is performed to update the estimated values ​​of heat transfer parameters and drag parameters. Based on the updated estimated values ​​of heat transfer parameters and drag parameters, the interaction intensity sequence is modulated. The recursive update operation includes generating a prediction output based on the parameter estimates at the previous update time and the identified input data at the current update time, calculating the residual between the prediction output and the identified output data, calculating the update gain based on the forgetting coefficient and the covariance state at the previous update time, updating the parameter estimates using the update gain and residual, and updating the covariance state synchronously. The modulation includes calculating the changes in heat transfer parameters and drag parameters, weighting the changes with modulation coefficients and exponentially decaying them to obtain a modulation factor, and then multiplying the intensity values ​​of the interaction intensity sequence by the modulation factor according to the update time to obtain the modulated intensity values, and arranging them in chronological order to generate the modulated interaction intensity sequence.

[0028] Optionally, the output process of the rolling prediction distribution and the risk upper bound specifically includes:

[0029] The aligned data sequence is segmented in chronological order. For each segment, modulated interaction intensity sequence fragments within the same time range are simultaneously extracted. Each pair of aligned data fragments and modulation intensity fragments are combined as training input samples and combined with the corresponding aligned data fragments within the future time range as training label samples, and the training sample set is output.

[0030] The HP-DGNN prediction model is constructed based on the training sample set and the input and output interfaces are configured. The predicted target in the future time domain is used as the model output.

[0031] The HP-DGNN prediction model is trained and updated. The training sample set is sampled in batches. Each batch of training input samples is input into the HP-DGNN prediction model to obtain the predicted distribution output in the future time domain. The error between the predicted distribution output and the corresponding training label samples is calculated to generate the training loss. Backpropagation is performed based on the training loss to obtain the gradient of the model parameters. The model parameters are updated according to the preset learning rate and optimizer rules until the preset termination condition is met. The model parameters after training are solidified into solidified model parameters.

[0032] Rolling inference is performed based on fixed model parameters. The historical input segment is extracted from the aligned data sequence with the current time as the end, and the modulated interaction intensity sequence segment within the same time range is extracted simultaneously. The event history and pipeline structure history at each time within the window are updated with intensity to obtain the dynamic interaction intensity trajectory. The dynamic interaction intensity trajectory is used as a graph propagation gate signal to drive HP-DGNN to recursively update the node representation and pipeline segment representation. After completing the recursive update within the window, the rolling prediction distribution in the future time domain is output. Then, the current time is moved forward by the rolling step size and the extraction, intensity update, gate propagation and recursive update operations are repeated until the current time reaches the preset inference termination time or the end of the aligned data sequence, and the rolling prediction distribution is output.

[0033] Extract the distribution of prediction indicators related to safety constraints from the rolling prediction distribution, and calculate the upper quantile threshold for each prediction indicator distribution at a preset confidence level as the upper risk bound of the current indicator at the corresponding prediction time.

[0034] Optionally, obtaining the candidate control sequence specifically includes:

[0035] Set the optimization time domain and control step size corresponding to the rolling prediction distribution, set the control quantities at each control time in the optimization time domain as decision variables to be solved, and form control sequence variables in time order;

[0036] Based on the rolling prediction distribution, a sequence of prediction quantities is generated in the optimization time domain to calculate the objective function, and the sequence of prediction quantities is summarized according to the control time to form the objective function;

[0037] A robust safety constraint set is constructed based on the risk upper bound sequence, and the robust safety constraint set and the objective function are combined to form a robust safety constraint optimization problem;

[0038] The robust safety constraint optimization problem is solved by using a preset initial control sequence as the initial solution. The target value of the initial solution under the objective function is calculated, and the robust safety constraint set is checked for feasibility. In each iteration, the corresponding prediction sequence and target value are calculated based on the current control sequence variables. The update direction and update step size of the control sequence variables relative to the previous iteration are calculated. The updated control sequence variables are substituted into the robust safety constraint set for constraint checking. When a constraint is violated, the update step size is reduced or the control variable that violates the constraint is backed up until the constraint is satisfied. The iteration is terminated when the change in the target value is less than a preset threshold or the iteration reaches a preset upper limit. A candidate control sequence that satisfies the robust safety constraint set and makes the objective function optimal is obtained.

[0039] Optionally, the generation of the control sequence that satisfies the security constraints specifically includes:

[0040] The candidate control sequence is expanded into a set of candidate control quantities according to the control time order, and the preset benchmark control quantity is read as the reference quantity for rate of change verification.

[0041] Perform safety verification on candidate control quantities at each control moment. Candidate control quantities that pass the verification are registered as control quantities that can be issued, and candidate control quantities that fail the verification are registered as control quantities that need to be corrected.

[0042] The control quantity to be corrected is modified with minimal changes. The control quantity to be corrected is subjected to boundary projection and backscaling in time sequence. The corrected control quantity is then combined and output as a control sequence to be sent down according to the control time.

[0043] Optionally, obtaining the results of the steam pipeline insulation and regulation optimization strategy specifically includes:

[0044] Read the execution time and control quantity of each control record in the issued control sequence, and sort them according to the execution time to generate a time-ordered control record set;

[0045] Merge and organize the time-ordered control record set, merge records with unchanged control quantities at consecutive times into the same strategy entry and write the start and end times, and form new strategy entries for control quantities that have changed.

[0046] The strategy items are summarized and output in chronological order to obtain the optimization strategy results for steam pipeline insulation regulation.

[0047] The beneficial effects of this invention are:

[0048] This invention aligns runtime timing data with environmental timing data to form an aligned data sequence that can be directly used for subsequent modeling and optimization. This allows the state changes of the steam pipeline network at different measuring points and sampling sources to be uniformly expressed on the same time base, thereby providing a stable data foundation for the joint analysis of operating condition fluctuations and external environmental disturbances.

[0049] After identifying changes in control actions and sudden changes in operating conditions, this invention introduces a Hawkes point process to explicitly quantify the interaction intensity sequence under the combined action of event-driven interaction sequences and historical sequences of pipeline structures. This allows the continuous impact of historical events on future states and the propagation of disturbances caused by topological coupling to be dynamically characterized, overcoming the shortcomings of existing technologies that rely solely on thresholds or static models, which are unable to reflect the intensity of event propagation and the effect of structural coupling.

[0050] The present invention further identifies pipe sections online and modulates the interaction intensity sequence with updated heat transfer and resistance parameters, so that model drift caused by insulation degradation, scaling or local resistance changes can be absorbed in time and applied to the interaction intensity, thereby improving the adaptability of subsequent predictions to actual working conditions.

[0051] This invention constructs an HP-DGNN prediction model to output a rolling prediction distribution and performs risk assessment to obtain an upper risk bound. Based on this, a robust safety constraint optimization is established, and a control sequence is generated by combining pre-execution safety verification and minimum modification correction. This makes the optimization strategy more executable and stable while meeting safety constraints. It can effectively reduce the fluctuation of terminal steam supply quality and condensation risk under sudden changes in operating conditions, and reduce ineffective regulation and energy consumption caused by conservative margin or model mismatch. Attached Figure Description

[0052] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0053] Figure 1 This is a flowchart of a deep learning-based steam pipeline insulation regulation optimization method proposed in this invention;

[0054] Figure 2 This is a schematic diagram of the core algorithm structure of a deep learning-based steam pipeline insulation regulation optimization method proposed in this invention. Detailed Implementation

[0055] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0056] refer to Figure 1-2 A deep learning-based method for optimizing the insulation and regulation of steam pipeline networks includes the following steps:

[0057] Collect steam pipeline network operation time series data and environmental time series data, and perform time alignment to generate aligned data sequences;

[0058] Based on the alignment data sequence, control action changes and sudden changes in operating conditions are identified, an event-driven interaction sequence is generated, and the Hawkes point process is used to calculate the interaction intensity sequence under the combined action of the event-driven interaction sequence and the historical sequence of the pipeline structure.

[0059] The pipe segment is identified online based on the aligned data sequence, the heat transfer and resistance related parameters of the pipe segment are updated, and the interaction intensity sequence is modulated using the updated pipe segment parameters to obtain the modulated interaction intensity sequence.

[0060] An HP-DGNN prediction model is constructed using aligned data sequences and modulated interaction intensity sequences as inputs. The model is trained and its parameters are fixed. Simultaneously, rolling inference is performed to output the rolling prediction distribution in the future time domain. Risk assessment is performed on safety constraint-related indicators, and the risk upper bound is output.

[0061] A robust safety constraint optimization problem is constructed and solved based on the rolling prediction distribution and risk upper bound to obtain candidate control sequences;

[0062] Perform pre-execution security checks and minimum modification corrections on candidate control sequences to generate control sequences that meet security constraints;

[0063] The issued control sequences are arranged in chronological order to obtain the results of the steam pipeline insulation regulation optimization strategy.

[0064] In this embodiment, the generation of the aligned data sequence specifically includes:

[0065] The steam pipeline network is continuously sampled according to a preset sampling period to obtain operational data and environmental data, forming a raw operational data sequence and a raw environmental data sequence. Each sampling record is marked with a sampling time identifier and a source identifier. Sampling records that do not obtain valid observation values ​​are marked as missing. The operational time series data is used to characterize the changes in the operating status of each pipe section and node of the steam pipeline network over time, including temperature, pressure, flow rate, valve status, desuperheating and depressurization settings, bypass status, and condensate status. The environmental time series data is used to characterize the changes in the external environmental conditions of the steam pipeline network over time, including ambient temperature, wind speed, humidity, and rainfall or snowfall.

[0066] Time alignment processing is performed on the original running data sequence and the original environmental data sequence using a unified time base. Each sampled record is mapped to the corresponding time base point. When different records exist at the same time point, the valid record is selected according to the preset rules. When a record is missing, the missing mark is maintained and the missing position is recorded.

[0067] The data that has been aligned to the time is processed for consistency. Missing time points are filled in according to preset rules and the filling marks are retained. The running data and environmental data are combined to generate an aligned data sequence in chronological order.

[0068] In this embodiment, the calculation process of the interaction intensity sequence under the combined effect of the event-driven interaction sequence and the network structure historical sequence specifically includes:

[0069] The aligned data sequence is scanned point by point according to the preset time window. At each sampling moment, the difference between the current sampled value and the previous sampled value is calculated, and the difference is compared with the preset threshold. When the difference meets the threshold trigger condition, the current sampling moment is registered as the event occurrence moment. At the same time, the event type identifier is registered and the event amplitude parameter is calculated to form an event record.

[0070] All event records are sorted by the time of event occurrence, and the sorted event records are written into the event-driven interaction sequence in sequence. Each record in the event-driven interaction sequence includes the time of event occurrence, event type identifier, and event magnitude parameter.

[0071] Based on the alignment data sequence and the topological relationship of the steam pipeline network, a historical sequence of the pipeline network structure is generated. The coupling state of each pipe segment and node at each sampling time is calculated with a preset time window, and the coupling state of each sampling time is written into the historical sequence of the pipeline network structure in chronological order to obtain a structural history input consistent with the time reference of the alignment data sequence. The topological relationship of the steam pipeline network refers to the structural information used to describe the connection method and interrelationship between each pipe segment and node in the steam pipeline network, which is used to characterize the steam transport path and coupling relationship in the pipeline network.

[0072] The Hawkes point process intensity calculation is performed using the event-driven interaction sequence and the pipeline structure history sequence as input. At each sampling time, a baseline intensity value is first determined. At the same time, for each event that occurred before the current sampling time, a corresponding weight coefficient is selected according to the event type. The weight coefficient is multiplied by the event amplitude parameter to obtain the event contribution value. The event contribution value is then subjected to a preset attenuation kernel function according to the time interval from the event occurrence time to the current sampling time to obtain the attenuated event contribution value. At the same time, the structural history input is weighted and summed according to the structural item weight coefficient at the sampling time to obtain the structural contribution value. The sum of the baseline intensity value and the attenuated event contribution value is combined with the structural contribution value to obtain the interaction intensity value at the current sampling time.

[0073] Arrange the interaction intensity values ​​at all sampling times in chronological order to generate an interaction intensity sequence and output it.

[0074] This invention transforms changes in control actions and abrupt changes in operating conditions into quantifiable event records and forms an event-driven interaction sequence. Simultaneously, it combines steam pipeline network topology coupling to construct a structural history input and utilizes Hawkes point processes to generate an interaction intensity sequence under time decay and structural weighting mechanisms. This unifies the continuous excitation effect of historical events on the current state and the topological propagation effect, significantly improving the temporal traceability and computability of disturbance propagation intensity. It reduces response lag and misjudgment caused by relying solely on thresholds or static models, providing a stable and constrained interaction-driven foundation for subsequent intensity modulation, HP-DGNN gated propagation, and robust safety optimization to reduce the risk of exceeding limits and condensation.

[0075] In this embodiment, the process of obtaining the modulated interaction intensity sequence specifically includes:

[0076] Using the aligned data sequence as input, the data is aggregated according to the topology of the steam pipe network, an online identification status is established for each pipe segment, the initial estimated values ​​of the heat transfer parameters and the initial estimated values ​​of the resistance parameters of the current pipe segment are written, and the parameter update cycle, sliding window length and forgetting coefficient are set.

[0077] At each parameter update time, a sliding window of data ending at the current time is extracted from the aligned data sequence. Identification input data and identification output data are set. The identification input data is the temperature, pressure, flow rate of the current pipe segment within the window and the difference between adjacent times. The identification output data is the heat loss related observation and pressure drop related observation of the current pipe segment within the window. Corresponding identification samples are generated.

[0078] For each parameter update time, a recursive update operation is performed to update the estimated values ​​of heat transfer parameters and drag parameters. Based on the updated estimated values ​​of heat transfer parameters and drag parameters, the interaction intensity sequence is modulated. The recursive update operation includes generating a prediction output based on the parameter estimates at the previous update time and the identified input data at the current update time, calculating the residual between the prediction output and the identified output data, calculating the update gain based on the forgetting coefficient and the covariance state at the previous update time, updating the parameter estimates using the update gain and residual, and updating the covariance state synchronously. The modulation includes calculating the changes in heat transfer parameters and drag parameters, weighting the changes with modulation coefficients and exponentially decaying them to obtain a modulation factor, and then multiplying the intensity values ​​of the interaction intensity sequence by the modulation factor according to the update time to obtain the modulated intensity values, and arranging them in chronological order to generate the modulated interaction intensity sequence.

[0079] This invention performs online pipe segment identification based on aligned data sequences and steam pipeline topology aggregation. It recursively updates heat transfer and resistance parameters and generates modulation factors by weighting and exponentially decaying the parameter changes. These factors are then applied to the interaction intensity sequence, allowing the interaction intensity to adaptively adjust with insulation degradation, scaling, and operating condition fluctuations. This reduces model mismatch and error accumulation, improves the accuracy of event propagation characterization and the stability of prediction input, reduces overly conservative strategies and frequent actuator actions caused by parameter drift, and lowers the risk of increased energy consumption and terminal temperature and pressure fluctuations. This provides a more reliable dynamic interaction drive and safety margin foundation for subsequent HP-DGNN prediction and robust safety optimization.

[0080] In this embodiment, the output process of the rolling prediction distribution and the risk upper bound specifically includes:

[0081] The aligned data sequence is segmented in chronological order. For each segment, modulated interaction intensity sequence fragments within the same time range are simultaneously extracted. Each pair of aligned data fragments and modulation intensity fragments are combined as training input samples and combined with the corresponding aligned data fragments within the future time range as training label samples, and the training sample set is output.

[0082] The HP-DGNN prediction model is constructed based on the training sample set and the input and output interface configuration is completed. The predicted target in the future time domain is used as the model output. The predicted target is supervised by the training label sample during the training phase.

[0083] The HP-DGNN prediction model is trained and updated. The training sample set is sampled in batches. Each batch of training input samples is input into the HP-DGNN prediction model to obtain the predicted distribution output in the future time domain. The error between the predicted distribution output and the corresponding training label samples is calculated to generate the training loss. Backpropagation is performed based on the training loss to obtain the model parameter gradient. The model parameters are updated according to the preset learning rate and optimizer rules until the preset termination condition is met. The model parameters after training are solidified into solidified model parameters. The preset termination condition is that the training loss decreases by less than a preset threshold within a preset number of consecutive rounds or the number of training iterations reaches a preset upper limit.

[0084] Rolling inference is performed based on fixed model parameters. The historical input segment is extracted from the aligned data sequence with the current time as the end, and the modulated interaction intensity sequence segment within the same time range is extracted simultaneously. The event history and pipeline structure history at each time within the window are updated with intensity to obtain the dynamic interaction intensity trajectory. The dynamic interaction intensity trajectory is used as a graph propagation gating signal to drive HP-DGNN to recursively update the node representation and pipeline segment representation. This makes the information propagation range, propagation weight and propagation direction within each rolling step change with the dynamic interaction intensity trajectory. After completing the recursive update within the window, the rolling prediction distribution in the future time domain is output. Then, the current time is moved forward by the rolling step size and the extraction, intensity update, gating propagation and recursive update operations are repeated until the current time reaches the preset inference termination time or the end of the aligned data sequence, and the rolling prediction distribution is output.

[0085] Extract the distribution of prediction indicators related to safety constraints from the rolling prediction distribution, and calculate the upper quantile threshold for each prediction indicator distribution at a preset confidence level as the upper risk bound of the current indicator at the corresponding prediction time.

[0086] This invention couples the aligned data sequence with the modulated interaction intensity sequence to construct and solidify the HP-DGNN prediction model. In the rolling inference, the dynamic interaction intensity trajectory is used as the graph propagation gating signal to adaptively adjust the information propagation range, propagation weight, and propagation direction. This outputs the future time-domain rolling prediction distribution and generates risk upper bounds for safety constraint-related indicators under a preset confidence level. This enables a quantifiable characterization of event-driven disturbances and topological coupling propagation and a constrained expression of uncertainty, thereby improving prediction accuracy, stability, and generalization ability. It also provides directly callable risk boundaries for subsequent robust safety optimization to reduce the risk of exceeding limits and condensation.

[0087] In this embodiment, obtaining the candidate control sequence specifically includes:

[0088] Set the optimization time domain and control step size corresponding to the rolling prediction distribution, set the control quantities at each control time in the optimization time domain as decision variables to be solved, and form control sequence variables in time order;

[0089] Based on the rolling prediction distribution, a sequence of prediction quantities is generated in the optimization time domain to calculate the objective function. The sequence of prediction quantities includes prediction quantities related to energy consumption, heat loss, terminal steam supply quality deviation and actuator action cost. The sequence of prediction quantities is then summarized according to the control time to form the objective function.

[0090] A robust safety constraint set is constructed based on the risk upper bound sequence. The robust safety constraint set includes safety constraints obtained by replacing or tightening the constraint boundaries corresponding to the terminal temperature lower limit, terminal pressure lower limit, and condensation risk upper limit with the risk upper bound sequence; actuator constraints that include valve opening, desuperheating and depressurization settings, bypass status and their rate of change limits; and operation constraints that include constraint boundaries related to pipeline operation stability. The robust safety constraint set and the objective function together constitute a robust safety constraint optimization problem.

[0091] The robust safety constraint optimization problem is solved by using a preset initial control sequence as the initial solution. The target value of the initial solution under the objective function is calculated, and the robust safety constraint set is checked for feasibility. In each iteration, the corresponding prediction sequence and target value are calculated based on the current control sequence variables. The update direction and update step size of the control sequence variables relative to the previous iteration are calculated. The updated control sequence variables are substituted into the robust safety constraint set for constraint checking. When a constraint is violated, the update step size is reduced or the control variable that violates the constraint is backed up until the constraint is satisfied. The iteration is terminated when the change in the target value is less than a preset threshold or the iteration reaches a preset upper limit. A candidate control sequence that satisfies the robust safety constraint set and makes the objective function optimal is obtained.

[0092] This invention constructs robust safety constraint optimization based on rolling prediction distribution and risk upper bound. By dynamically replacing or tightening the terminal temperature lower limit, terminal pressure lower limit, and condensation risk upper limit according to the risk upper bound, and linking valve opening, desuperheating and depressurization settings, bypass status and their rate of change limits, it comprehensively minimizes energy consumption, heat loss, terminal steam supply quality deviation and actuator action cost while ensuring operational feasibility and stability. It significantly reduces the probability of exceeding limits and over-adjustment under sudden operating conditions and enhances the adaptability to insulation degradation and resistance drift, thereby improving the safety, economy and executability of the control strategy.

[0093] In this embodiment, the generation of the control sequence that satisfies the security constraints specifically includes:

[0094] The candidate control sequence is expanded into a set of candidate control quantities according to the control time order, and the preset benchmark control quantity is read as the reference quantity for rate of change verification.

[0095] Safety verification is performed on the candidate control quantities at each control moment. The safety verification includes value range verification, relative reference quantity change amplitude verification, and corresponding moment safety constraint verification. Candidate control quantities that pass the verification are registered as control quantities that can be issued, and candidate control quantities that fail the verification are registered as control quantities that need to be corrected.

[0096] The control variable to be modified is modified with minimal changes. Boundary projection and backscaling are performed on the control variable to be modified in time sequence to ensure that it simultaneously meets the value range, rate of change limit and safety constraint. The modified control variable is then combined and output as a control sequence to be sent down according to the control time.

[0097] In this embodiment, obtaining the results of the steam pipeline insulation and regulation optimization strategy specifically includes:

[0098] Read the execution time and control quantity of each control record in the issued control sequence, and sort them according to the execution time to generate a time-ordered control record set;

[0099] Merge and organize the time-ordered control record set, merge records with unchanged control quantities at consecutive times into the same strategy entry and write the start and end times, and form new strategy entries for control quantities that have changed.

[0100] The strategy items are summarized and output in chronological order to obtain the optimization strategy results for steam pipeline insulation regulation.

[0101] Example 1:

[0102] To verify the feasibility of this invention in practice, it was applied to a typical scenario of thermal regulation and operational optimization of a park-level steam pipeline network. This scenario includes a main steam supply pipe, several trunk pipes, and multiple branch line end users. The pipeline network has a mixed tree and ring topology, and the end users include intermittent and continuous steam loads. There are mixed sections of overhead and underground laying along the pipeline, and the external environment experiences diurnal temperature differences, rainfall, wind speed fluctuations, and other operating conditions within the same operating cycle. The core problem that has long existed in this scenario is that when events such as sudden load changes, valve switching, and bypass opening and closing occur, the disturbance propagation path in the pipeline topology is complex, and the end temperature and pressure are prone to short-term exceedances. At the same time, the aging of the insulation layer, local scaling, and resistance drift cause the heat transfer and pressure drop characteristics to change over time. Traditional mechanism models and rule strategies based on fixed parameters are difficult to adapt in a timely manner, which commonly manifests as delayed control actions, excessive conservative margins leading to increased energy consumption, and increased condensation risk under low temperature, high humidity, and strong wind conditions that are difficult to predict in advance.

[0103] In this scenario, the present invention accesses existing measurement points and control system data in the pipeline network, collects operational and environmental data, and unifies them to the same time reference, forming an aligned data sequence that can be directly used for modeling. The data sampling covers temperature, pressure, flow rate, valve opening, desuperheating and depressurization settings, bypass status, and drainage-related status of the steam supply main pipe and key branches, as well as ambient temperature, wind speed, humidity, rainfall, and snowfall. To ensure that "abrupt changes" in the scenario are truly utilized by the model, the present invention identifies rapid changes in valve opening, sudden changes in desuperheating and depressurization settings, bypass status switching, and step changes in terminal flow as control action changes and abrupt changes in operating conditions. The events are written into the event-driven interaction sequence with their occurrence time, type, and amplitude. At the same time, the historical sequence of the structure is calculated by combining the pipeline network topology relationship, so that the propagation differences of the same event in different topological locations and under different coupling states are explicitly presented. Subsequently, the Hawkes point process is used to calculate the intensity of the combined effect of event history and structural history, generating an interaction intensity sequence. This allows us to express which events occurred in the past, in which directions these events might propagate topologically, and how their propagation intensity decays over time, as a continuous and computable intensity trajectory. To address insulation degradation and drag drift, this invention performs online identification at the pipe segment scale, continuously updating equivalent heat transfer and drag-related parameters. These parameter changes drive the modulation of the interaction intensity sequence, thus incorporating the effect of "changes in insulation / drag state amplifying or weakening event propagation intensity" into the key input for subsequent predictions.

[0104] In the prediction and decision-making stages, this invention constructs an HP-DGNN prediction model using aligned data sequences and modulated interaction intensity sequences. During training, a rolling window is used to construct samples and solidify model parameters. During inference, the dynamic interaction intensity trajectory is used as a graph propagation gating signal to drive the recursive updating of node representations and pipe segment representations, so that the information propagation range, propagation weight, and propagation direction change in real time with the interaction intensity, outputting a rolling prediction distribution in the future time domain. Simultaneously, risk upper bounds are calculated for prediction indicators related to safety constraints, and these risk upper bounds are directly used as boundary inputs for robust safety constraint optimization to obtain candidate control sequences. Before being issued, the candidate control sequences undergo pre-execution safety verification and minimum modification correction to ensure that they meet valve opening and rate of change limits, de-temperature and de-pressure setting boundaries, bypass switching constraints, and safety boundaries related to terminal temperature and pressure and condensation risk. Finally, an executable control sequence is formed and organized into a strategy result arranged in chronological order for output to the control system.

[0105] To objectively demonstrate the beneficial effects, two comparative schemes were selected under the same pipeline network and data window: one was a common field rule and fixed margin strategy, mainly based on threshold triggering and manual experience tuning; the other was a combined strategy of "time series deep model prediction and conventional MPC," which does not introduce event-structure strength and online identification modulation, nor output risk upper bounds but adopts a fixed safety margin. The comparative evaluation used indicators such as prediction error of terminal temperature and pressure, number of constraint overruns, and accuracy of condensation risk early warning. The statistical window covered a 12-week operating cycle. Specific comparative data are shown in Table 1.

[0106] Table 1. Comparison of the overall performance of different control strategies

[0107] index Comparison with Option A: Rules + Fixed Margin Comparison Option B: Time Series Deep Forecasting + Conventional MPC This invention Mean absolute error of end-point pressure prediction (MPa) 0.028 0.019 0.012 Mean absolute error of terminal temperature prediction (°C) 4.6 3.2 2.1 Number of times the terminal pressure exceeds the lower limit (times / 12 weeks) 37 21 7 Confirmation of condensation risk events (times / 12 weeks) 14 10 3 Recovery time after sudden load change (minutes, mean) 18.4 12.7 6.9 Equivalent heat consumption per unit transport volume (GJ / thousand tons of steam) 78.6 74.1 69.3 Average number of valve operations per day (times / day) 96 121 83 Number of manual interventions (per 12 weeks) 22 15 4

[0108] As shown in Table 1, the present invention outperforms the two comparative schemes in terms of prediction accuracy, safety, and economy. Specifically, the average absolute error of the terminal pressure prediction is reduced from 0.028 MPa (Comparative Scheme A) and 0.019 MPa (Comparative Scheme B) to 0.012 MPa, and the average absolute error of the terminal temperature prediction is reduced from 4.6℃ and 3.2℃ to 2.1℃. Regarding safety indicators, the number of times the terminal pressure exceeds the lower limit is reduced from 37 and 21 times to 7 times, the number of condensation risk confirmation events is reduced from 14 and 10 times to 3 times, and the recovery time after load mutation is shortened from 18.4 minutes and 12.7 minutes to 6.9 minutes. In terms of economic and feasibility indicators, the equivalent heat consumption per unit of transport is reduced from 78.6 and 74.1 to 69.3 GJ / thousand tons of steam, while the average daily valve operation frequency is reduced from 96 and 121 times to 83 times, and the number of manual interventions is reduced from 22 and 15 times to 4 times. This demonstrates that the present invention improves strategy stability and engineering feasibility while reducing energy consumption and ineffective regulation.

[0109] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for optimizing the insulation and regulation of steam pipeline networks based on deep learning, characterized in that, Includes the following steps: Collect steam pipeline network operation time series data and environmental time series data, and perform time alignment to generate aligned data sequences; Based on the alignment data sequence, control action changes and sudden changes in operating conditions are identified, an event-driven interaction sequence is generated, and the Hawkes point process is used to calculate the interaction intensity sequence under the combined action of the event-driven interaction sequence and the historical sequence of the pipeline structure. The pipe segment is identified online based on the aligned data sequence, the heat transfer and resistance related parameters of the pipe segment are updated, and the interaction intensity sequence is modulated using the updated pipe segment parameters to obtain the modulated interaction intensity sequence. An HP-DGNN prediction model is constructed using aligned data sequences and modulated interaction intensity sequences as inputs. The model is trained and its parameters are fixed. Simultaneously, rolling inference is performed to output the rolling prediction distribution in the future time domain. Risk assessment is performed on safety constraint-related indicators, and the risk upper bound is output. A robust safety constraint optimization problem is constructed and solved based on the rolling prediction distribution and risk upper bound to obtain candidate control sequences; Perform pre-execution security checks and minimum modification corrections on candidate control sequences to generate control sequences that meet security constraints; The issued control sequences are arranged in chronological order to obtain the results of the steam pipeline insulation regulation optimization strategy.

2. The steam pipeline insulation and regulation optimization method based on deep learning according to claim 1, characterized in that, The generation of the aligned data sequence specifically includes: The steam pipeline network is continuously sampled according to the preset sampling period. Operational data and environmental data are obtained separately to form the original operational data sequence and the original environmental data sequence. Each sampling record is written with a sampling time identifier and a source identifier. Sampling records for which no valid observations were obtained are marked as missing. Time alignment processing is performed on the original running data sequence and the original environmental data sequence using a unified time base. Each sampled record is mapped to the corresponding time base point. When different records exist at the same time point, the valid record is selected according to the preset rules. When a record is missing, the missing mark is maintained and the missing position is recorded. The data that has been aligned to the time is processed for consistency. Missing time points are filled in according to preset rules and the filling marks are retained. The running data and environmental data are combined to generate an aligned data sequence in chronological order.

3. The steam pipeline insulation and regulation optimization method based on deep learning according to claim 1, characterized in that, The calculation process of the interaction intensity sequence under the combined effect of the event-driven interaction sequence and the historical sequence of the pipeline structure specifically includes: The aligned data sequence is scanned point by point according to the preset time window. At each sampling moment, the difference between the current sampled value and the previous sampled value is calculated, and the difference is compared with the preset threshold. When the difference meets the threshold trigger condition, the current sampling moment is registered as the event occurrence moment. At the same time, the event type identifier is registered and the event amplitude parameter is calculated to form an event record. Sort all event records by the time of occurrence, and write the sorted event records into the event-driven interaction sequence in sequence; Based on the alignment data sequence and the topology relationship of the steam pipeline network, a historical sequence of pipeline network structure is generated. The coupling state of each pipe segment and node at each sampling time is calculated with a preset time window. The coupling state of each sampling time is written into the historical sequence of pipeline network structure in chronological order to obtain a historical structural input consistent with the time reference of the alignment data sequence. The Hawkes point process intensity calculation is performed using the event-driven interaction sequence and the pipeline structure history sequence as input. At each sampling time, a baseline intensity value is first determined. At the same time, for each event that occurred before the current sampling time, a corresponding weight coefficient is selected according to the event type. The weight coefficient is multiplied by the event amplitude parameter to obtain the event contribution value. The event contribution value is then subjected to a preset attenuation kernel function according to the time interval from the event occurrence time to the current sampling time to obtain the attenuated event contribution value. At the same time, the structural history input is weighted and summed according to the structural item weight coefficient at the sampling time to obtain the structural contribution value. The sum of the baseline intensity value and the attenuated event contribution value is combined with the structural contribution value to obtain the interaction intensity value at the current sampling time. Arrange the interaction intensity values ​​at all sampling times in chronological order to generate an interaction intensity sequence and output it.

4. The steam pipeline insulation and regulation optimization method based on deep learning according to claim 1, characterized in that, The process of obtaining the modulated interaction intensity sequence specifically includes: Using the aligned data sequence as input, the data is aggregated according to the topology of the steam pipe network, an online identification status is established for each pipe segment, the initial estimated values ​​of the heat transfer parameters and the initial estimated values ​​of the resistance parameters of the current pipe segment are written, and the parameter update cycle, sliding window length and forgetting coefficient are set. At each parameter update time, the sliding window data ending at the current time is extracted from the aligned data sequence, the identification input data and identification output data are set, and the corresponding identification samples are generated. For each parameter update time, a recursive update operation is performed to update the estimated values ​​of heat transfer parameters and drag parameters. Based on the updated estimated values ​​of heat transfer parameters and drag parameters, the interaction intensity sequence is modulated. The recursive update operation includes generating a prediction output based on the parameter estimates at the previous update time and the identified input data at the current update time, calculating the residual between the prediction output and the identified output data, calculating the update gain based on the forgetting coefficient and the covariance state at the previous update time, updating the parameter estimates using the update gain and residual, and updating the covariance state synchronously. The modulation includes calculating the changes in heat transfer parameters and drag parameters, weighting the changes with modulation coefficients and exponentially decaying them to obtain a modulation factor, and then multiplying the intensity values ​​of the interaction intensity sequence by the modulation factor according to the update time to obtain the modulated intensity values, and arranging them in chronological order to generate the modulated interaction intensity sequence.

5. The steam pipeline insulation regulation optimization method based on deep learning according to claim 1, characterized in that, The specific process for outputting the rolling prediction distribution and risk upper bound includes: The aligned data sequence is segmented in chronological order. For each segment, modulated interaction intensity sequence fragments within the same time range are simultaneously extracted. Each pair of aligned data fragments and modulation intensity fragments are combined as training input samples and combined with the corresponding aligned data fragments within the future time range as training label samples, and the training sample set is output. The HP-DGNN prediction model is constructed based on the training sample set and the input and output interfaces are configured. The predicted target in the future time domain is used as the model output. The HP-DGNN prediction model is trained and updated. The training sample set is sampled in batches. Each batch of training input samples is input into the HP-DGNN prediction model to obtain the predicted distribution output in the future time domain. The error between the predicted distribution output and the corresponding training label samples is calculated to generate the training loss. Backpropagation is performed based on the training loss to obtain the gradient of the model parameters. The model parameters are updated according to the preset learning rate and optimizer rules until the preset termination condition is met. The model parameters after training are solidified into solidified model parameters. Rolling inference is performed based on fixed model parameters. The historical input segment is extracted from the aligned data sequence with the current time as the end, and the modulated interaction intensity sequence segment within the same time range is extracted simultaneously. The event history and pipeline structure history at each time within the window are updated with intensity to obtain the dynamic interaction intensity trajectory. The dynamic interaction intensity trajectory is used as a graph propagation gate signal to drive HP-DGNN to recursively update the node representation and pipeline segment representation. After completing the recursive update within the window, the rolling prediction distribution in the future time domain is output. Then, the current time is moved forward by the rolling step size and the extraction, intensity update, gate propagation and recursive update operations are repeated until the current time reaches the preset inference termination time or the end of the aligned data sequence, and the rolling prediction distribution is output. Extract the distribution of prediction indicators related to safety constraints from the rolling prediction distribution, and calculate the upper quantile threshold for each prediction indicator distribution at a preset confidence level as the upper risk bound of the current indicator at the corresponding prediction time.

6. The steam pipeline insulation and regulation optimization method based on deep learning according to claim 1, characterized in that, The process of obtaining the candidate control sequence specifically includes: Set the optimization time domain and control step size corresponding to the rolling prediction distribution, set the control quantities at each control time in the optimization time domain as decision variables to be solved, and form control sequence variables in time order; Based on the rolling prediction distribution, a sequence of prediction quantities is generated in the optimization time domain to calculate the objective function, and the sequence of prediction quantities is summarized according to the control time to form the objective function; A robust safety constraint set is constructed based on the risk upper bound sequence, and the robust safety constraint set and the objective function are combined to form a robust safety constraint optimization problem; The robust safety constraint optimization problem is solved by using a preset initial control sequence as the initial solution. The target value of the initial solution under the objective function is calculated, and the robust safety constraint set is checked for feasibility. In each iteration, the corresponding prediction sequence and target value are calculated based on the current control sequence variables. The update direction and update step size of the control sequence variables relative to the previous iteration are calculated. The updated control sequence variables are substituted into the robust safety constraint set for constraint checking. When a constraint is violated, the update step size is reduced or the control variable that violates the constraint is backed up until the constraint is satisfied. The iteration is terminated when the change in the target value is less than a preset threshold or the iteration reaches a preset upper limit. A candidate control sequence that satisfies the robust safety constraint set and makes the objective function optimal is obtained.

7. The steam pipeline insulation regulation optimization method based on deep learning according to claim 1, characterized in that, The generation of the control sequence that satisfies the security constraints specifically includes: The candidate control sequence is expanded into a set of candidate control quantities according to the control time order, and the preset benchmark control quantity is read as the reference quantity for rate of change verification. Perform safety verification on candidate control quantities at each control moment. Candidate control quantities that pass the verification are registered as control quantities that can be issued, and candidate control quantities that fail the verification are registered as control quantities that need to be corrected. The control quantity to be corrected is modified with minimal changes. The control quantity to be corrected is subjected to boundary projection and backscaling in time sequence. The corrected control quantity is then combined and output as a control sequence to be sent down according to the control time.

8. The steam pipeline insulation regulation optimization method based on deep learning according to claim 1, characterized in that, The results of the steam pipeline insulation and regulation optimization strategy are specifically obtained as follows: Read the execution time and control quantity of each control record in the issued control sequence, and sort them according to the execution time to generate a time-ordered control record set; Merge and organize the time-ordered control record set, merge records with unchanged control quantities at consecutive times into the same strategy entry and write the start and end times, and form new strategy entries for control quantities that have changed. The strategy items are summarized and output in chronological order to obtain the optimization strategy results for steam pipeline insulation regulation.

Images