Cyber-physical fusion based industrial thermal network control method and system

Abstract

This invention discloses a control method and system for industrial thermal networks based on cyber-physical fusion, relating to the field of industrial process control technology. The method performs unified time-base processing on multi-source operating data, outputting physically estimated states and virtual sensor estimated states from physical twins and data-driven predictions, respectively. The residuals are normalized based on the historical error variance of each estimate, and fused to obtain a reliable state vector and regional confidence level. Using this as input, a rolling time-domain collaborative optimization problem involving dimensionless diagonal weighted matrix and confidence level modulation of the control increment term is solved to obtain heterogeneous target control quantities. An event-driven mechanism is used at the edge side to determine whether to issue new control quantities. This invention aims to improve the temperature control stability and execution robustness of existing industrial thermal transmission and distribution networks without relying on large-scale additions of terminal hardware.

Description

Technical Field

[0001] This invention relates to the field of industrial process control technology, and in particular to an industrial thermal network control method and system based on cyber-physical fusion. Background Technology

[0002] Industrial thermal networks, widely used in petrochemical, regional energy transmission and distribution, large-scale public utilities, and multi-temperature zone production systems, are multi-stage heat energy transmission and distribution systems. They typically consist of heat source units, primary trunk pipelines, heat exchange nodes, and secondary terminal load units, forming a multi-stage series-parallel structure. They rely on distributed control systems or monitoring and data acquisition systems to collect and monitor operating parameters such as heat source outlet parameters, key node supply and return temperatures, flow rates, and differential pressures. In recent years, with the development of the Industrial Internet of Things (IIoT), edge computing, and high-speed data communication technologies, more and more process industry enterprises are overlaying so-called "intelligent management and control" platforms on top of their traditional monitoring systems. These platforms aggregate data from heat sources, heat exchange stations, terminal load groups, and even some dispersed measuring points onto a unified data base, enabling energy efficiency statistics, operational reports, and remote monitoring.

[0003] At the control strategy level, many current industrial thermal systems commonly use feedforward compensation curves or empirical correction curves based on ambient temperature to set the temperature on the heat source supply side, while circulating power equipment often employs constant head control or fixed frequency operation. The essence of this type of strategy is to use environmental meteorological parameters and time patterns as feedforward inputs, issuing process setpoints according to pre-tuned static curves, with each control loop relatively independent and having a fixed control cycle. In another type of improvement scheme, some systems introduce generalized predictive control or improved proportional-integral-derivative control at the level of a single heat exchange node, performing fine-tuning of the local secondary side output temperature; other schemes use real-time digital twin technology for daily or shift-level operation planning, and then use the planning results to guide scheduling. These schemes, to some extent, improve the temperature control stability of a single node or the energy consumption distribution under specific operating conditions.

[0004] However, industrial thermal networks generally have long-distance transmission time delays, significant load thermal inertia, and fluid dynamic and thermodynamic coupling characteristics formed between heat exchange nodes through shared transmission trunk lines. There are strong interactions across levels and time scales between heat sources, trunk networks, heat exchange nodes, and terminal loads. Feedforward compensation strategies that only issue setpoints based on ambient temperature parameters lack comprehensive processing of transmission time delay, load thermal inertia, and closed-loop feedback of terminal temperature. This makes it difficult to avoid large fluctuations in controlled temperature, oversupply of energy, and persistently high return temperature under transitional conditions such as drastic changes in ambient temperature and peak-valley switching of production load. Predictive control strategies that optimize a single heat exchange node are limited to the local loop and fail to incorporate shared transmission constraints between adjacent heat exchange nodes, heat and mass transfer time delays over long-distance pipelines, and thermal inertia of the terminal load side into the global decision-making process. This makes it difficult to suppress control mismatch caused by cross-site coupling and time delay effects from a network-wide perspective. Digital twin solutions, represented by daily planning, focus more on offline energy scheduling and operation planning, and are insufficient in responding to disturbances, anomalies, and actuator responses at the real-time level.

[0005] Crucially, the aforementioned solutions generally implicitly assume an idealized premise on the control input side: that all measurement data entering the optimizer or controller is accurate and reliable. However, the actual data quality in industrial thermal environments does not support this assumption: temperature, flow, and differential pressure signals are susceptible to noise interference, zero-point drift, dead zones, abrupt changes, and data gaps. Some terminal temperature and heat data exhibit significant transmission lag, and data from different sources are difficult to align precisely in terms of timestamp annotation. Furthermore, systematic deviations introduced by improper control parameter tuning or manual intervention can also enter the control loop as seemingly normal data. Whether it's ambient temperature feedforward compensation, improved proportional-integral-derivative (PI-DE) control, or local predictive control, once the input data itself contains unreliable components, the output control quantity will inevitably deviate from the reasonable operating range, manifesting on the physical side as frequent valve actuation, ineffective speed regulation of power equipment, or continuous rise in return flow temperature. When communication links are interrupted, latency jitter occurs, or data packets are lost, some systems directly revert to local fixed frequency or offline curve operation, resulting in a significant break in control capability.

[0006] It is evident that while existing industrial thermal control solutions have made some progress in data aggregation, local optimization, and offline scheduling, they have not yet simultaneously addressed the two core, coupled issues of "global optimization decision-making under conditions of long time delays, strong coupling, and multi-node collaboration" and "the real-time impact of field data reliability on the control closed loop" within the same control structure. As a result, controller adjustment commands are highly dependent on raw inputs without reliability assessment, network-wide collaborative optimization struggles to effectively connect with safe execution at the edge, and robust degradation transition mechanisms are lacking in situations of communication or sensor anomalies. How to simultaneously solve these two types of problems within a unified real-time closed-loop framework without excessive reliance on large-scale additions of field detection hardware and full reuse of existing monitoring and data acquisition systems is a pressing technical challenge in this field. Summary of the Invention

[0007] The purpose of this invention is to provide an industrial thermal network control method and system based on cyber-physical fusion, in order to solve the problem that existing industrial thermal transmission and distribution control schemes are unable to simultaneously handle the real-time impact of network-wide collaborative decision-making under cross-station coupling and the reliability of field data on the control loop in a unified closed loop, which leads to problems such as controlled temperature fluctuations, oversupply, high reflux, and control capability gaps under abnormal operating conditions.

[0008] To achieve the above objectives, the present invention provides the following technical solution.

[0009] A cyber-physical network control method based on cyber-physical fusion includes the following steps: S1, collecting multi-source operating data of the thermal network, and performing protocol adaptation, clock alignment, deduplication, and quality marking on the multi-source operating data to obtain data frames with a unified time base; S2, constructing and maintaining a thermal semantic graph, wherein the thermal semantic graph uses heat source nodes, main pipe nodes, heat exchange nodes, secondary network partition nodes, load group nodes, and observation point nodes as node sets, and uses physical connectivity edges, control influence edges, and information dependency edges as edge sets. Each node carries temperature, flow rate, pressure difference, controlled temperature, heat, confidence level, and time delay attributes, and each edge carries pipe... The attributes include: S1) Segment length, equivalent thermal resistance, equivalent water resistance, design flow direction, real-time flow velocity, and heat loss coefficient; S2) Based on the aforementioned thermal semantic diagram, construct an equivalent time delay and heat loss model for each pipe segment to obtain the physical estimated state, and output the virtual sensing estimated state of key nodes and the future load prediction sequence based on historical operating data, outdoor weather, time calendar characteristics, and some controlled temperature inputs; S3) For each key variable, calculate the physical residual between its original measured value and the physical estimated state, and the virtual residual between its original measured value and the virtual sensing estimated state. Normalize the corresponding residuals using the historical error variance of the physical and virtual estimates respectively to obtain the measurement reliability. The system calculates the confidence level, physical confidence level, and virtual confidence level, and then weights and fuses the original measured values, physical estimated states, and virtual sensor estimated states to obtain the confidence level. The average confidence level of several variables within a preset control area is then used to obtain the regional confidence level. S5. Using the confidence level vector, future load forecast, comfort target, and return-side temperature constraint as inputs, the system solves the rolling time-domain collaborative optimization problem to obtain heterogeneous target control quantities including heat source supply-side temperature, heat source output allocation, circulating pump frequency, secondary-side supply-side temperature target of heat exchange nodes, secondary-side differential pressure target, valve opening offset, and load group heating offset. The objective of the collaborative optimization problem is... The function includes an energy consumption term, a comfort loss term, a return-side temperature penalty term, and a control increment penalty term. The control increment penalty term normalizes the difference between heterogeneous control components to a dimensionless scale using a positive definite diagonal weighted matrix, and modulates the penalty term using a monotonically decreasing function of the region confidence. S6: At the edge side, it is determined whether to update the control quantity according to the trigger indication. The trigger indication is true when the controlled temperature deviation, return-side temperature deviation, region confidence, or dimensionless control increment exceeds the corresponding threshold. If the trigger condition is not met, the control quantity of the previous cycle is maintained. If the trigger condition is met, the control quantity is updated with a bounded step size using a saturation function with upper and lower bounds.S7. When the communication interruption between the central platform and the edge gateway lasts for more than a set threshold, or the confidence level of the key area is lower than the safety threshold, switch to a conservative safety curve mode based on ambient temperature. The ambient temperature is obtained through a multi-level degradation method: firstly, it is taken from the local physical ambient temperature sensor on the edge side; secondly, it is taken from the corresponding time value of the meteorological forecast curve periodically sent and cached locally on the edge side during normal communication; and thirdly, it is taken from the last successfully obtained ambient temperature before the network outage. In this mode, the maximum adjustment step size is tightened and the direct adjustment weight of low-confidence variables is frozen. S8. In each execution cycle, record the confidence status, target control quantity, constraint activity status, and the actual state difference before and after the instruction is issued. Rollingly update the equivalent heat loss parameters of the pipe section, the equivalent thermal inertia time constant of the building group, the error statistics of physical estimation and virtual estimation, the area weight, and the trigger threshold.

[0010] As a preferred technical solution of the present invention, the weighted fusion in step S4 calculates the key variables according to the following formula. Trusted status: ; in , , Key variables The original measured values, physical estimates, and virtual sensing estimates, To prevent small positive numbers from the denominator from approaching zero; to measure reliability. Physical credibility With virtual credibility The following formulas are given respectively: ; ; ; in , These are physical residuals and virtual residuals, respectively. , Physical estimates and virtual estimates for variables, respectively. Historical error standard deviation As an availability indicator, The freshness factor is characterized by exponential decay. , The prior weights are used to normalize the variance of the historical error of each residual by the corresponding estimation channel, thus conforming to the statistical logic of multi-source information fusion.

[0011] As a preferred technical solution of the present invention, in step S3, each pipe segment Its transmission delay and outlet temperature Calculate using the following formula: ; ; in , , , , , , These are, respectively, pipe segment length, real-time flow velocity, equivalent heat dissipation parameters, mass flow rate, specific heat capacity of the heat medium, ambient temperature, and pipe segment inlet temperature; when the transmission delay... When the sampling period is not an integer multiple of the unified time base, By sampling period Decompose into integer parts With the decimal part The historical values ​​of the inlet temperature were obtained by linear interpolation of adjacent discrete sampling points. This avoids out-of-bounds exceptions caused by directly indexing discrete arrays with floating-point latency.

[0012] As a preferred embodiment of the present invention, the objective function of the rolling time-domain collaborative optimization problem in step S5 is: ; in This marks the starting point of this rolling optimization. To predict the time domain, To sum the discrete-time indices for traversal, Let the decision variables be column vectors. It is a positive definite diagonal weighted matrix, and its diagonal elements are taken as the reciprocal of the square of the typical engineering variation range of the corresponding control components; , , , , These are the scalar weights for energy consumption, power consumption, comfort loss, high backflow penalty, and control increment penalty, respectively. For the confidence level of the region The monotonically decreasing function; the collaborative optimization problem is solved in parallel across multiple regions using a boundary variable coordination mechanism based on thermal semantic graphs. The regions are divided based on the similarity of the hydraulic and thermal coupling between heat exchange nodes, the time delay of pipe sections, and the thermal inertia time constant of buildings.

[0013] As a preferred embodiment of the present invention, the trigger indication in step S6 is given by the following formula: ; in This is an indicator function; the value within the parentheses is true when the logical expression is true. Otherwise, the value is ; The difference between the controlled temperature and the target temperature of a representative load group. This is the deviation of the reflux side temperature from its design baseline. For the diagonal weighted matrix Isomorphic positive definite diagonal matrices; when At that time, the control quantity is generated with a bounded step size according to the following formula: ; in It is a component-wise saturated cutoff function with upper and lower bounds. This is for adjusting the coefficient.

[0014] An industrial thermal network control system based on cyber-physical fusion includes a multi-source data access unit, a spatiotemporal alignment and data governance unit, a topology semantic graph construction unit, a physical mechanism twin unit, a data-driven prediction unit, a confidence assessment and virtual sensing unit, a partitioned collaborative optimization unit, an event-driven edge execution unit, a safety degradation control unit, and a reinjection self-learning and evidence recording unit. The multi-source data access unit acquires real-time data from heat sources, primary pipe networks, heat exchange nodes, heat meters, controlled temperature sampling points, outdoor meteorological services, and the business operation and maintenance management system. The spatiotemporal alignment and data governance unit performs protocol adaptation, clock alignment, deduplication, and quality marking on the real-time data, outputting data frames with a unified time base. The topology semantic graph construction unit abstracts the thermal network into an attributed thermal semantic graph. The physical mechanism twin unit constructs equivalent physical models of pipe sections, heat exchange nodes, and load groups based on the thermal semantic graph. The system generates and outputs physical estimated states; the data-driven prediction unit outputs virtual sensing estimated states of key nodes and future heat load prediction sequences; the confidence assessment and virtual sensing unit normalizes the residuals according to the historical error variance of each estimation channel, and fuses and outputs a reliable state vector and regional confidence; the partitioned collaborative optimization unit solves the rolling time-domain collaborative optimization problem containing dimensionless diagonal weighted matrix and confidence modulation control increment terms, and outputs heterogeneous target control quantities; the event-driven edge execution unit determines whether to update and distribute the heterogeneous target control quantities with bounded steps according to the trigger indication; the safety degradation control unit switches to a conservative safety curve mode based on multi-level degradation environmental temperature input when communication interruption exceeds a set threshold or the confidence of key areas is lower than the safety threshold; the reinjection self-learning and evidence recording unit records the data throughout the process and updates the model parameters and trigger thresholds on a rolling basis.

[0015] As a preferred embodiment of the present invention, the multi-source data access unit sets up a local physical environment temperature sensor as a redundant acquisition channel on the edge gateway side, and periodically caches the outdoor weather forecast curves for the next few hours issued by the central platform on the edge side, so that it can still provide usable environmental temperature input to the safety degradation control unit during the period of disconnection from the central platform; the data collected by the multi-source data access unit includes at least the temperature of the heat source supply side, the supply side temperature and return side temperature of the trunk key node, the flow rate and pressure difference of the primary side and the secondary side, the frequency of the circulating pump, the opening degree of the electric valve, the heat meter reading, the controlled temperature of the sampling room or load group, the outdoor dry bulb temperature and wind speed, time calendar characteristics, and repair and complaint information.

[0016] As a preferred technical solution of the present invention, the physical mechanism twin unit characterizes the heat transfer and heat dissipation process of the heat medium for each pipe segment using an equivalent time delay and heat loss model, describes the heat transfer relationship between the primary and secondary sides using a heat transfer efficiency model for heat exchange nodes, and characterizes the thermal inertia of the building envelope for the load group using an equivalent first-order inertial element. The response of indoor air temperature to the secondary supply side temperature is characterized as a calculable hysteresis-inertial process. When the pipe segment transmission delay is not an integer multiple of the unified time base sampling period, the physical mechanism twin unit uses a linear interpolation method between adjacent discrete sampling points to obtain the historical value of the inlet temperature.

[0017] As a preferred embodiment of the present invention, when the safety degradation control unit switches to the safety curve mode, it generates a conservative curve of the heat source supply side temperature from the ambient temperature using the following formula: ; in The current available ambient temperature, , Based on the basic curve parameters, To address the bias in the current control area, the safety degradation control unit simultaneously tightens the maximum adjustment step size, freezes the direct adjustment weights for low-confidence variables, and only allows conservative actions to be performed at the levels of heat source temperature and primary circulation pump frequency, while maintaining the secondary side pressure difference and valve opening at the previous steady-state value. When the communication link is restored and the confidence level of the critical area rises above the safety threshold, it re-enters the trusted state generation stage, first reconstructs the trusted state vector, and then makes a decision.

[0018] As a preferred technical solution of the present invention, the objects periodically updated by the reinjection self-learning and evidence recording unit include at least the equivalent heat loss parameters of the pipe section, the equivalent thermal inertia time constant of the building group, the virtual sensing error statistics and physical estimation error statistics, the area weight, the trigger threshold, and the comfort preference parameters; the data recorded by the reinjection self-learning and evidence recording unit includes at least the trusted state snapshot, the optimization target value, the constraint activity status, the state difference before and after the instruction is issued, the fault label, the manual intervention record, and the energy efficiency comparison baseline. The recorded data is used for online model identification, operation auditing, and post-event traceability.

[0019] By implementing the above technical solution, this invention introduces a reliable state generation link on the control input side that combines physical mechanism twinning, data-driven prediction, and confidence assessment. The original measurement, physical estimation, and virtual estimation are weighted and fused according to their respective historical error variance normalized confidence levels. This ensures that the data entering the control loop is no longer the unassessed original data. Measurement points with occasional drift, jamming, or large time delays can be automatically downweighted without direct rejection, thus enabling the output of robust control quantities even under sensor malfunctions and inconsistent data conditions.

[0020] By using a thermal semantic graph as a unified structural carrier, this invention incorporates physical constraints such as pipe segment transmission delay, pipe segment heat loss, heat exchange efficiency, and building thermal inertia, along with heterogeneous decision variables such as heat source, circulating pump, secondary side supply side temperature of heat exchange nodes, secondary side pressure difference, valve opening bias, and load group heating bias, into the same rolling time-domain collaborative optimization problem. Furthermore, it uses a diagonal weighted matrix to render the components of different dimensions dimensionless, avoiding optimization failure due to differences in magnitude. Compared to control methods targeting a single station or single loop, this collaborative optimization can balance heat consumption, pump consumption, controlled temperature comfort loss, high return side temperature penalty, and control increment modulated by regional confidence within the same objective function, thereby suppressing oversupply and high return phenomena caused by cross-station coupling and long-distance time delays.

[0021] By introducing an event-driven execution mechanism at the edge, this invention issues new control quantities only when the controlled temperature deviation, return-side temperature deviation, regional confidence level, or control increment exceeds the corresponding threshold; otherwise, it maintains the control quantity of the previous cycle. When communication interruption exceeds the set duration or the confidence level of the critical area falls below the safety threshold, the safety degradation control loop switches to a conservative curve based on ambient temperature. Multi-level degradation inputs are constructed using local physical ambient temperature sensors, locally cached meteorological forecast curves, and the last successfully acquired ambient temperature. This avoids system instability caused by input loss during network outages, ensuring that the frequency of command issuance and the number of actuator actions are no longer passively equivalent to the number of samples. This reduces ineffective actions of electric valves and circulating pumps and maintains basic heating quality and execution safety under abnormal communication or data conditions.

[0022] Since the above-mentioned links are interconnected through a trusted state bus and a thermal semantic graph, the implementation of this invention does not rely on a large-scale addition of terminal hardware. It can be deployed on the basis of existing monitoring and data acquisition systems, programmable controllers or remote terminal units of thermal power stations, and some representative controlled temperature measurement points. It has good engineering applicability to existing industrial thermal transmission and distribution networks with heterogeneous equipment and limited transformation resources. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the system architecture according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the method flow according to an embodiment of the present invention.

[0024] Figure labeling: 100, Multi-source data access unit; 200, Spatiotemporal alignment and data governance unit; 300, Topological semantic graph construction unit; 400, Physical mechanism twin unit; 500, Data-driven prediction unit; 600, Confidence assessment and virtual sensing unit; 700, Partition collaborative optimization unit; 800, Event-driven edge execution unit; 900, Security degradation control unit; 1000, Recharge self-learning and evidence recording unit. Detailed Implementation

[0025] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Those skilled in the art should understand that the following description is merely illustrative and not intended to limit the scope of protection of the present invention; any equivalent substitutions, partial modifications, or parameter adjustments based on the present invention without departing from the technical concept of the present invention should be considered to fall within the scope of the present invention.

[0026] like Figure 1 As shown, this embodiment provides an industrial thermal network control system based on cyber-physical fusion. It sequentially constructs a data access and governance layer, a trusted state generation layer, a collaborative decision-making layer, and an edge execution and degradation layer above the physical object layer, and achieves overall closed-loop control through a self-learning and evidence recording unit 1000. The physical object layer includes at least a heat source, a primary pipeline network, several heat exchange nodes, a load group downstream of the secondary pipeline network, and circulating water pumps, electric regulating valves, temperature sensors, pressure sensors, flow meters, heat meters, and controlled temperature measurement points distributed throughout these locations. This physical object layer is not limited to specific manufacturers or models, nor does it require the addition of large-scale terminal hardware; it can be deployed in a software-defined manner on top of existing industrial thermal distribution infrastructure and monitoring and data acquisition systems.

[0027] The cyber-physical network control system based on cyber-physical fusion described in this embodiment can be deployed in city-level, park-level, campus-level, or district-level heating scenarios, and is compatible with single or multiple heat sources. It is applicable to new construction projects, expansion projects, and existing renovation projects. Regarding communication, preferably, data transmission between devices and edge gateways, and between edge gateways and the central platform, uses an industrial interoperability protocol with semantic object interaction capabilities. The trusted state bus and event distribution link use a lightweight publish-subscribe messaging protocol. When field conditions are limited or some older devices only support serial protocols, a protocol adaptation submodule can be added to the edge gateway to standardize it into a unified external interface, thereby maintaining the consistency of the upper-layer logic.

[0028] The multi-source data access unit 100 is responsible for acquiring real-time data from heat source outlets, key trunk line nodes, heat exchange nodes, heat meters, controlled temperature sampling points, outdoor meteorological services, and the business operation and maintenance management system. Specifically, the data collected by the multi-source data access unit 100 includes at least the supply-side temperature of the heat source, the supply-side and return-side temperatures of key trunk line nodes, the flow rate and pressure difference between the primary and secondary sides, the frequency of circulating pumps, the opening degree of electric valves, heat meter readings, the controlled temperature of the sampled room or load group, outdoor dry-bulb temperature and wind speed, time calendar characteristics, and repair and complaint information related to user experience. The multi-source data access unit 100 does not require full deployment to all end users, but adopts a deployment strategy that prioritizes key nodes, supplements controlled temperature sampling with virtual sensing, thereby significantly reducing the transformation cost. To ensure input integrity during communication interruption, the multi-source data access unit 100 sets up a local physical ambient temperature sensor as a redundant acquisition channel on the edge gateway side, and periodically caches the outdoor weather forecast curves for the next few hours issued by the central platform on the edge side, so as to ensure that usable ambient temperature input can still be provided to downstream units during the period of disconnection from the central platform.

[0029] The spatiotemporal alignment and data governance unit 200 undertakes the unified time base processing task for cross-source data. Raw second-level or ten-second-level samples are aligned into one-minute-level state frames within the spatiotemporal alignment and data governance unit 200. For upper-layer optimization scenarios, these can be further aggregated into five-minute or fifteen-minute-level frames. For missing samples, the spatiotemporal alignment and data governance unit 200 does not directly perform simple interpolation replacement, but instead adds a quality flag bit to the sample. For variables with significant transmission delays, such as end-point temperature, the spatiotemporal alignment and data governance unit 200 simultaneously records their freshness coefficient. For suspicious samples such as those with jumps, jams, or drifts, the spatiotemporal alignment and data governance unit 200 only performs initial screening; the final adoption is determined by subsequent confidence assessment and the virtual sensing unit 600, thus avoiding irreversible hard rejection at the front end of the control loop.

[0030] The topological semantic graph building unit 300 abstracts the entire thermal network into a thermal semantic graph with attributes. This graph serves as a unified structural carrier for subsequent physical modeling, load forecasting, confidence assessment, and collaborative optimization. The node set... It includes at least heat source nodes, main pipe nodes, heat exchange nodes, secondary network partition nodes, load group nodes, and observation point nodes; edge set It includes at least physically connected edges, control-influenced edges, and information-dependent edges. Each node carries attributes such as temperature, flow rate, pressure difference, controlled temperature, heat, confidence level, and time delay; each edge carries attributes such as pipe segment length, equivalent thermal resistance, equivalent water resistance, design flow direction, real-time flow velocity, and heat loss coefficient. Thermal semantic graph. The system is automatically generated from equipment ledgers, pipeline design documents, and historical operating data at the beginning of system power-on, and is dynamically corrected as the self-learning and evidence-keeping unit 1000 is updated.

[0031] Physical mechanism twin unit 400 based on thermal semantic graph For each pipe section An equivalent time delay and heat loss model is constructed to characterize the heat transfer and heat dissipation process of the heat medium in the pipe network online. Preferably, for a length of... Real-time flow rate is pipe section Its transmission delay With outlet temperature The following formulas are given respectively: ; ; in, For pipe section The equivalent heat dissipation parameters, For quality flow, Specific heat capacity of the heat medium The ambient temperature of this pipe section. Let be the inlet temperature of the pipe section. Considering that all variables in this embodiment are discretized using a unified time base, the continuous time delay calculated from the above formula... It is usually not an integer multiple of the discrete sampling period; the physical mechanism twin units number 400 pairs. The value is obtained by linear interpolation between adjacent discrete sampling points: By sampling period Decompose into integer parts With the decimal part Then there is This avoids the out-of-bounds anomalies caused by directly indexing discrete arrays with floating-point delays. The twin model constructed by the physical mechanism twin unit 400 does not pursue the level of detail required for offline simulation, but rather focuses on the computability and real-time performance needed for online control, retaining only dynamic terms that have a significant impact on adjustment decisions. For heat exchange nodes, the physical mechanism twin unit 400 further describes the heat exchange relationship between the primary and secondary sides using a heat exchange efficiency model; for load clusters, the physical mechanism twin unit 400 characterizes the thermal inertia of the building envelope using an equivalent first-order inertial element, representing the response of indoor air temperature to the secondary supply temperature as a computable hysteresis-inertial process.

[0032] The data-driven prediction unit 500 performs two core tasks. The first is heat load prediction: using historical heat data, outdoor weather, holidays, time characteristics, and some controlled temperatures as inputs, it preferably uses a graph spatiotemporal network or a dynamic sparse graph model to output a predicted heat load sequence for each heat exchange node and load group for the next few hours. ,in For the prediction time domain, it represents the time from the current moment. The first type is the total distance predicted forward, with frames at the one-minute or five-minute level as the basic unit. The second type is virtual sensing: when a critical node lacks a physical sensor, or its measurement is marked as suspicious at the spatiotemporal alignment and data governance unit 200, the data-driven prediction unit 500 outputs a virtual estimate of the node's temperature, flow rate, or pressure difference based on its upstream and downstream related variables, twin estimation results, and historical samples to supplement the system's observability. This virtual sensing capability allows for the acquisition of state information that meets control requirements in existing pipeline networks where hardware modification resources are limited, thereby avoiding reliance solely on large-scale new sensor deployments.

[0033] The confidence assessment and virtual sensing unit 600 is one of the most important software logic layers in this embodiment. (For the thermal semantic graph...) Each key variable in Record its current moment The original measurement components are The physical twin estimate is The virtual sensing estimate is Unless otherwise specified, if no time index is used in the following text, it will be assumed to refer to the current time of the determination. The confidence assessment and virtual sensing unit 600 first calculate the physical residuals and virtual residuals: ; Further introduce freshness factor To reflect the time validity of this variable, the exponential decay form is preferred: ; in This represents the time lag of the current sample relative to the reference time. The attenuation constant can be tuned according to the variable type. For example, different constants can be used for the heat source supply side temperature, the secondary supply side temperature, and the controlled temperature. Based on this, the confidence assessment and the virtual sensing unit 600 respectively construct the measurement confidence level. Physical credibility With virtual credibility The residuals of each path are normalized by the historical error variance of their corresponding estimated channels to conform to the statistical logic of multi-source information fusion. ; ; ; in This variable serves as an availability indicator, reflecting whether the sensor is online and whether its engineering range complies with regulations. and Physical estimates and virtual estimates for variables, respectively. The historical error standard deviation is updated on a rolling basis by the 1000 self-learning and evidence tracking units. , These are the corresponding prior weights. Finally, the confidence level assessment and the virtual sensing unit 600 are weighted and fused to obtain the confidence state of the variable: ; in To prevent small positive numbers where the denominator approaches zero, a control region consisting of several variables is used. The confidence assessment and virtual sensing unit 600 further define the area confidence level: ; Therefore, each variable not only participates in subsequent control in the form of a credible state, but its confidence level is also directly passed to the collaborative decision-making layer as a first-class control variable, so that the controller can distinguish whether the current error is caused by the actual load disturbance or by the abnormality of the measurement point or the link.

[0034] The partitioned collaborative optimization unit 700 is based on the thermal semantic graph. The entire network is divided into zones, not based on administrative divisions, but on the similarity of hydraulic and thermal coupling between heat exchange nodes, pipe segment time delay magnitudes, and building thermal inertia time constants. Preferably, the zoned collaborative optimization unit 700 employs a rolling time-domain optimization strategy. Upper-level optimization has a 15-minute cycle, with a prediction time domain of two to eight hours; the middle-level rolling update cycle is one to five minutes; and lower-level commands are triggered as needed by the event-driven edge execution unit 800. Decision variables are determined at each decision moment. Organized as column vectors Its components include at least the temperature on the heat source supply side. Heat source output distribution, primary circulation pump frequency Temperature targets on the secondary supply side of each heat exchange node Secondary side pressure difference target Key valve opening offset and load group heating offset Considering that the above components are heterogeneous in physical dimensions and numerical magnitudes (temperature in degrees Celsius, frequency in Hertz, pressure difference in kilopascals, and valve opening in normalized percentage), the partitioned collaborative optimization unit 700 introduces a positive definite diagonal weighting matrix when constructing the control increment penalty term. Its diagonal element The diagonal element corresponding to the typical engineering variation range of the corresponding component is taken as the reciprocal of the square of the temperature rise corresponding to the single allowable adjustment range. For example, the diagonal element corresponding to the temperature on the heat source supply side is taken as the reciprocal of the square of the temperature rise corresponding to the single allowable adjustment range. The diagonal elements corresponding to the circulation pump frequency, secondary side pressure difference, and valve opening are respectively tuned according to the reciprocal of the square of their respective allowable single-step variation range, so that each component is on a comparable dimensionless scale after weighting. The optimization objective function is constructed as follows: ; in This is the starting time of this rolling optimization, which is consistent with the current decision time when the optimization solution was initiated; To predict the time domain, the prediction step size should be consistent with that used by the data-driven prediction unit 500; For summation and traversal of discrete-time indices; , , , , These are the scalar weights for energy consumption, power consumption, comfort, high backflow, and control increment, used to balance the relative importance of energy consumption indicators, comfort indicators, and performance stability indicators. Their values ​​are determined by on-site commissioning. Characterizes the heat supply or equivalent heat consumption in the prediction time domain. Characterizing the power consumption of the circulating pump, This is the comfort loss term, typically expressed as the square of the difference between the controlled temperature and the target temperature. This is a penalty term for high reflux side temperature. It is a monotonically decreasing function of the region confidence level; the lower the confidence level, the more likely it is to be true. The larger the value, the more amplified the control increment term becomes, causing the optimizer to tend to maintain small adjustments or even remain inactive when data reliability is low. Constraints include at least the time delay and heat loss constraints for each pipe segment, upper and lower limits of heat source capacity, the relationship between pump frequency and head, the physical range of valve opening, the safety boundaries of temperature and pressure difference on the secondary side supply side, the comfort range of the controlled temperature, and the upper limit of the return side temperature. The optimization solution can employ a strategy combining rolling time-domain linearization and quadratic programming, or a predictive control solver based on data-driven approximate dynamics. In the case of multiple zones, the zone-based collaborative optimization unit 700 uses a thermal semantic graph-based approach. The boundary variable coordination mechanism enables each region to solve in parallel while satisfying the constraints of the entire network.

[0035] The event-driven edge execution unit 800 is deployed on the field edge gateway or hot-swap node controller side, and performs target control on the partition collaborative optimization unit 700. Instead of indiscriminately periodically issuing commands, an event-triggered mechanism is introduced to suppress invalid actions. Specifically, the event-driven edge execution unit 800 calculates a trigger indication within each decision cycle: ; in This is an indicator function; its value is true when the logical expression within the parentheses is true. Otherwise, the value is ; The deviation between the controlled temperature and the target temperature of a representative load group; This refers to the deviation of the reflux side temperature from its design baseline. To be consistent with the aforementioned weighting matrix An isomorphic positive definite diagonal matrix is ​​used to quantify the differences between heterogeneous control components into a unified dimensionless scale, avoiding the judgment distortion caused by differences in dimensional magnitude; , , , These are the corresponding judgment thresholds. When At the same time, the event-driven edge execution unit 800 maintains the control quantity of the previous cycle. This remains unchanged, thus significantly reducing the frequency of command issuance and minimizing ineffective actions of electric valves and variable frequency pumps; when At that time, the event-driven edge execution unit 800 updates the control variables with bounded steps: ; in It is a component-wise saturated cutoff function with upper and lower bounds. The adjustment coefficient can be determined jointly by comfort level, urgency level, and the maximum allowable rate of change of the actuator. This approach transforms the control frequency from a passive, sampling-cycle equivalent into an optimizable object, allowing both command density and actuator lifespan to benefit simultaneously.

[0036] The security degradation control unit 900 handles three types of boundary scenarios: communication anomalies, data anomalies, and low confidence levels. This includes scenarios where the communication interruption between the central platform and the edge gateway lasts for a certain duration. Exceeding the set threshold or confidence level of key areas Below the safety threshold At this time, the safety degradation control unit 900 immediately takes over the output of the event-driven edge execution unit 800, switching the system to safety degradation mode. In this mode, the temperature on the heat source supply side is conservatively calculated from the ambient temperature using the following formula: ; in The current available ambient temperature, , Based on the basic curve parameters, In order to target the current area The bias is determined offline based on the historical operational deviations of the area, building usage types, and complaint sensitivity. Considering the synchronous failure of the meteorological data channel provided by the central platform under communication interruption conditions, the safety degradation control unit 900 pairs... The value is determined using a multi-level degradation strategy: It is preferentially taken from the local physical ambient temperature sensor deployed on the edge gateway side by the multi-source data access unit 100; if the local physical ambient temperature sensor is also unavailable, the value corresponding to the outdoor weather forecast curve for the next few hours, periodically sent and cached in the local storage of the edge gateway during normal communication, is used; if both of the above are unavailable, the last successfully acquired ambient temperature before the network outage is used as a fallback input, and the maximum adjustment step size is tightened simultaneously, waiting for data recovery. It is important to emphasize that the safety curve used by the safety degradation control unit 900 is not the normal master control quantity, but a fallback strategy for low-confidence scenarios; in this mode, the safety degradation control unit 900 simultaneously tightens the maximum adjustment step size and freezes the direct adjustment weights for low-confidence variables, allowing only conservative actions at the heat source temperature and primary pump frequency levels, while maintaining the previous steady-state value for secondary side differential pressure and valve opening, avoiding drastic adjustments due to erroneous data or link jitter. When the communication link is restored and the confidence level of the critical area recovers... After the above, the system does not directly continue the optimization results before the switch, but re-enters the trusted state generation stage, first reconstructs the trusted state vector and then makes a decision, thereby suppressing the shock adjustment that may occur at the moment of disconnection recovery.

[0037] The reinjection self-learning and evidence recording unit 1000 is responsible for recording the entire process of the system and updating parameters on a rolling basis. Specifically, the objects periodically updated by the reinjection self-learning and evidence recording unit 1000 include at least: the equivalent heat loss parameters of the pipe section. Equivalent thermal inertia time constant of building complex, virtual sensing error statistics Error statistics of physical estimation Regional weight to Trigger threshold and In addition to comfort preference parameters, the recorded data includes at least: trusted state snapshots, optimization target values, constraint activity, state differences before and after instruction issuance, fault labels, human intervention records, and energy efficiency comparison baselines. The recorded data from the self-learning and evidence-based recording unit 1000 can be used simultaneously for online model identification, operational auditing, and post-event traceability.

[0038] like Figure 2 As shown, the intelligent heating control method provided in this embodiment can be described as several steps executed sequentially. In step S1, the multi-source data access unit 100 and the spatiotemporal alignment and data governance unit 200 determine the current time. Real-time raw measurement vector Protocol adaptation, clock alignment, deduplication, and quality marking are performed to obtain data frames with a unified time base. ,in By weight The key variables are represented in a specific format. In step S2, the topology semantic graph construction unit 300 constructs or updates the thermal semantic graph based on the equipment ledger, pipeline topology, and real-time connectivity status. In step S3, the physical mechanism twin unit 400 estimates the temperature propagation and heat loss of each pipe section based on the current flow rate, pressure difference, and environmental boundary, and outputs the physical estimation state. Simultaneously, the data-driven prediction unit 500 outputs virtual sensing results. With future load forecasts In step S4, the confidence level is calculated for each key variable by the confidence assessment and virtual sensing unit 600. And fuse them to obtain a reliable state vector. With regional confidence In step S5, the partitioned collaborative optimization unit 700 solves the above-mentioned rolling time-domain optimization problem using the reliable state vector, future load, comfort target, backflow constraint, and energy consumption weight as inputs, to obtain the upper-level control target. In step S6, the event-driven edge execution unit 800 determines whether to issue a new control quantity based on the trigger indication: when it is determined that no trigger is needed, the status is maintained. Unchanged; when a trigger is detected, it is generated with a bounded step size. In step S7, if a communication disconnection or a drop in regional confidence level is detected... Then the safety degradation control unit 900 switches to safety curve mode output. and order In step S8, the self-learning and evidence recording unit 1000 records the execution results of this cycle and their actual impact on the state, and updates the twin parameters, virtual sensing error model, slice weights and trigger thresholds accordingly, before entering the next execution cycle.

[0039] To facilitate a detailed explanation of the core components of the system described in this embodiment, an implementation example for a district heating scenario is provided below. Assume a city area has an industrial thermal distribution building area of ​​approximately 1.2 million square meters, consisting of a peak-shaving heat source, four relay heat exchange nodes, and several building units. The original control strategy involved each heat exchange node querying the ambient temperature curve based on the ambient temperature, with circulating pumps operating at industrial frequency or using simple constant differential pressure control. This resulted in consistently high return-side temperatures, and some load groups experiencing controlled temperature fluctuations exceeding two degrees Celsius during cold waves. After deploying this system, the multi-source data access unit 100 collects data on the heat source supply side temperature, the primary and secondary side temperatures and differential pressures of the four heat exchange nodes, circulating pump frequency, electric valve opening, heat meter readings, controlled temperatures of approximately 10% of the representative load groups, and real-time data from nearby weather stations. Redundant data is then collected at the edge gateway via local physical ambient temperature sensors. The spatiotemporal alignment and data governance unit 200 aligns this data to a one-minute data frame. The topology semantic graph construction unit 300 divides the entire network into four sub-regions centered around four heat exchange nodes, each sub-region further containing several load group nodes. The physical mechanism twin unit 400 models the time delay and heat loss of several long-distance pipe segments of the primary trunk line, and uses linear interpolation of adjacent sampling points to obtain historical inlet temperature values ​​for non-integer multiple time delays; it uses a heat transfer efficiency model for heat exchange nodes, and models the thermal inertia of each load group using an equivalent first-order inertial model. The data-driven prediction unit 500 trains the load prediction model based on historical heating season operation data, and constructs a virtual sensing model for the secondary side return side temperature that is missing in some terminals. The confidence assessment and virtual sensing unit 600 simultaneously tracks the measurement, physical estimation, and virtual estimation results for each key variable, and automatically downweights sensors with occasional jumps or freezes instead of directly removing them. The partitioned collaborative optimization unit 700 uses a 15-minute upper-level optimization cycle to jointly solve for the heat source supply-side temperature, circulation pump frequency, secondary-side supply-side temperature targets and pressure difference targets of each heat exchange node, and controls the incremental penalty term using a diagonal weighted matrix. The system achieves dimensionless scaling for each component. The event-driven edge execution unit 800 determines whether to issue a new instruction every minute, maintaining the control quantity from the previous cycle for operating conditions that have not reached the trigger threshold. When the safety degradation control unit 900 detects that the communication interruption between the central platform and the edge gateway exceeds a set threshold, it immediately switches to a conservative curve based on ambient temperature on the edge side and freezes the opening of the electric valve. The ambient temperature is preferentially taken from the local physical ambient temperature sensor, with the cached prediction curve and the last successfully acquired value used as backup input. The reinjection self-learning and evidence retention unit 1000 continuously updates the building thermal inertia parameters, physical estimation error statistics, and virtual sensing error statistics every heating season weekend.

[0040] See Figure 1 As can be seen, this embodiment does not require a full-scale hardware overhaul of existing heating stations and load groups. In other words, the minimum feasible implementation version only requires adding a unified data access program, a trusted state engine, a zoned collaborative optimizer, and an event-driven edge agent to the existing heat source control system, heat exchange node programmable controllers or remote terminal units, and some representative controlled temperature measurement points for pilot deployment in a single area. This implementation method is particularly suitable for large-scale, heterogeneous equipment, and existing industrial thermal transmission and distribution networks with limited retrofit resources.

[0041] It should be understood that the above division of the units in the system is only a logical division, and in actual deployment, they can be merged or split according to computing resources and network conditions. For example, the physical mechanism twin unit 400 and the confidence assessment and virtual sensing unit 600 can be deployed together on the central platform or regional computing nodes; the event-driven edge execution unit 800 and the security degradation control unit 900 can be deployed together on the edge gateway, and retain the necessary security curve parameters and the most recent legal control quantity after the edge gateway loses power or restarts, so as to ensure the local executability of degradation capabilities. The bearer protocol adopted by the multi-source data access unit 100 can also be replaced with other industrial communication protocols that support the publish-subscribe mode according to the field conditions; as long as its semantic layer can carry the unified data frame structure described in this embodiment, it will not affect the implementation of the present invention.

[0042] Furthermore, this embodiment does not strictly limit the interfaces and data structures between the units. Trusted State Vector A thermal semantic graph-based approach can be adopted. The attribute-based structure can be represented using a flattened tensor structure, and cross-unit alignment can be achieved through a unified timestamp and node number. The weights in the objective function are optimized. to Diagonal weighted matrix and confidence modulation function The specific form can be adjusted according to the load characteristics of the actual thermal network, user sensitivity, and actuator response characteristics. Trigger threshold , , , With security threshold , All parameters can be configured in segments according to different stages of the operating cycle, ambient temperature ranges, and user types; the specific values ​​of these parameters do not constitute a limitation on the scope of protection of this invention.

[0043] In summary, this embodiment reconstructs the previously loosely structured monitoring, simulation, prediction, optimization, alarm, and edge execution into a closed-loop system with clear causal relationships through a unified information and control flow formed between the multi-source data access unit 100, the spatiotemporal alignment and data governance unit 200, the topological semantic graph construction unit 300, the physical mechanism twin unit 400, the data-driven prediction unit 500, the confidence assessment and virtual sensing unit 600, the partitioned collaborative optimization unit 700, the event-driven edge execution unit 800, the security degradation control unit 900, and the feedback self-learning and evidence recording unit 1000. Its technical effect does not originate from any single prediction model or single controller, but from the synergistic relationship between the trusted state as the control input, the confidence level as the control variable, the event triggering as the execution mechanism, and the security degradation as the fallback strategy. Based on the technical concept provided by this invention, those skilled in the art can make various equivalent substitutions and parameter adjustments without departing from the essence of this invention; all such modifications should be considered as one of the embodiments of this invention.

[0044] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. An industrial thermal network control method based on cyber-physical fusion, characterized in that, Includes the following steps: S1. Collect multi-source operating data of the thermal network, and perform protocol adaptation, clock alignment, deduplication and quality marking on the multi-source operating data to obtain a data frame with a unified time base. S2. Construct and maintain a thermal semantic graph. The thermal semantic graph uses heat source nodes, main pipe nodes, heat exchange nodes, secondary network partition nodes, load group nodes, and observation point nodes as node sets, and physical connectivity edges, control influence edges, and information dependence edges as edge sets. Each node carries temperature, flow rate, pressure difference, controlled temperature, heat, confidence level, and time delay attributes. Each edge carries pipe length, equivalent thermal resistance, equivalent water resistance, design flow direction, real-time flow velocity, and heat loss coefficient attributes. S3. Based on the thermal semantic graph, construct an equivalent time delay and heat loss model for each pipe segment to obtain the physical estimated state. Based on historical operating data, outdoor weather, time calendar features, and some controlled temperature inputs, output the virtual sensing estimated state of key nodes and the future load prediction sequence. S4. For each key variable, calculate the physical residual between its original measurement value and the physical estimated state, and the virtual residual between its original measurement value and the virtual sensing estimated state. Normalize the corresponding residuals with the historical error variance of the physical and virtual estimates respectively to obtain the measurement confidence, physical confidence and virtual confidence. Based on this, perform weighted fusion of the original measurement value, physical estimated state and virtual sensing estimated state to obtain the confidence state. Average the confidence of several variables in the preset control area to obtain the regional confidence. S5. Using the reliable state vector, future load forecast, comfort target, and return side temperature constraint as input, solve the rolling time-domain collaborative optimization problem to obtain heterogeneous target control quantities including heat source supply side temperature, heat source output allocation, circulation pump frequency, secondary side supply side temperature target of heat exchange node, secondary side pressure difference target, valve opening bias, and load group heating bias. The objective function of the collaborative optimization problem includes energy consumption term, comfort loss term, return side temperature penalty term, and control increment penalty term. The control increment penalty term normalizes the difference of heterogeneous control components to a dimensionless scale using a positive definite diagonal weighted matrix, and modulates the penalty term by the monotonically decreasing function of the region confidence. S6. On the edge side, determine whether to update the control quantity according to the trigger indication. The trigger indication is true when the controlled temperature deviation, the return flow side temperature deviation, the area confidence level, or the dimensionless control increment exceeds the corresponding threshold. If the trigger condition is not met, maintain the control quantity of the previous cycle. If the trigger condition is met, update the control quantity with bounded step size using a saturation function with upper and lower bounds. S7. When the communication interruption between the central platform and the edge gateway lasts for more than a set threshold, or the confidence level of the key area is lower than the safety threshold, switch to the conservative safety curve mode based on ambient temperature. The ambient temperature is obtained in a multi-level degradation manner: firstly, it is taken from the local physical ambient temperature sensor on the edge side; secondly, it is taken from the value of the meteorological forecast curve that is periodically sent and cached locally on the edge side during normal communication; and thirdly, it is taken from the last successfully obtained ambient temperature before the network outage. In this mode, the maximum adjustment step size is tightened and the direct adjustment weight of low confidence variables is frozen. S8. Record the trusted state, target control quantity, constraint activity, and actual state difference before and after instruction issuance in each execution cycle. Update the equivalent heat loss parameters of pipe sections, equivalent thermal inertia time constant of building groups, error statistics of physical and virtual estimation, area weights, and trigger thresholds on a rolling basis.

2. The industrial thermal network control method based on cyber-physical fusion according to claim 1, characterized in that: The weighted fusion in step S4 is calculated using the following formula for key variables. Trusted status: ; in , , Key variables The original measured values, physical estimates, and virtual sensing estimates, To prevent small positive numbers from the denominator from approaching zero; to measure reliability. Physical credibility With virtual credibility The following formulas are given respectively: ; in , These are physical residuals and virtual residuals, respectively. , Physical estimates and virtual estimates for variables, respectively. Historical error standard deviation As an availability indicator, The freshness factor is characterized by exponential decay. , These are the prior weights.

3. The industrial thermal network control method based on cyber-physical fusion according to claim 1, characterized in that: In step S3, for each pipe segment Its transmission delay and outlet temperature Calculate using the following formula: ; in , , , , , , These are, respectively, pipe segment length, real-time flow velocity, equivalent heat dissipation parameters, mass flow rate, specific heat capacity of the heat medium, ambient temperature, and pipe segment inlet temperature; when the transmission delay... When the sampling period is not an integer multiple of the unified time base, By sampling period Decompose into integer parts With the decimal part The historical values ​​of the inlet temperature were obtained by linear interpolation of adjacent discrete sampling points. .

4. The industrial thermal network control method based on cyber-physical fusion according to claim 1, characterized in that: The objective function of the rolling time-domain co-optimization problem described in step S5 is: ; in This marks the starting point of this rolling optimization. To predict the time domain, To sum the discrete-time indices for traversal, Let the decision variables be column vectors. It is a positive definite diagonal weighted matrix, and its diagonal elements are taken as the reciprocal of the square of the typical engineering variation range of the corresponding control components; , , , , These are the scalar weights for energy consumption, power consumption, comfort loss, high backflow penalty, and control increment penalty, respectively. For the confidence level of the region The monotonically decreasing function; the collaborative optimization problem is solved in parallel across multiple regions using a boundary variable coordination mechanism based on thermal semantic graphs. The regions are divided based on the similarity of the hydraulic and thermal coupling between heat exchange nodes, the time delay of pipe sections, and the thermal inertia time constant of buildings.

5. The industrial thermal network control method based on cyber-physical fusion according to claim 1, characterized in that: The trigger indication mentioned in step S6 is given by the following formula: ； in This is an indicator function; the value within the parentheses is true when the logical expression is true. Otherwise, the value is ; The difference between the controlled temperature and the target temperature of a representative load group. This is the deviation of the reflux side temperature from its design baseline. For the diagonal weighted matrix Isomorphic positive definite diagonal matrices , , , These are the corresponding judgment thresholds; when At that time, the control quantity is generated with a bounded step size according to the following formula: ; in It is a component-wise saturated cutoff function with upper and lower bounds. This is for adjusting the coefficient.

6. An industrial thermal network control system based on cyber-physical fusion, characterized in that, It includes a multi-source data access unit, a spatiotemporal alignment and data governance unit, a topological semantic graph construction unit, a physical mechanism twin unit, a data-driven prediction unit, a confidence assessment and virtual sensing unit, a partitioned collaborative optimization unit, an event-driven edge execution unit, a security degradation control unit, and a recharge self-learning and evidence recording unit. The multi-source data access unit is used to acquire real-time data from heat sources, primary pipe networks, heat exchange nodes, heat meters, controlled temperature sampling points, outdoor meteorological services, and business operation and maintenance management systems; the spatiotemporal alignment and data governance unit is used to perform protocol adaptation, clock alignment, deduplication, and quality marking on the real-time data, and output data frames with a unified time base; the topology semantic graph construction unit is used to abstract the thermal network into a thermal semantic graph with attributes; the physical mechanism twin unit is used to construct equivalent physical models of pipe segments, heat exchange nodes, and load groups based on the thermal semantic graph and output physical estimation states; the data-driven prediction unit is used to output the virtual sensing estimation states of key nodes and the future heat load prediction sequence; the confidence assessment and virtual sensing unit is used to normalize the residuals according to the historical error variance of each estimation channel, and fuse and output a reliable state vector and regional confidence. The partitioned collaborative optimization unit is used to solve the rolling time-domain collaborative optimization problem containing dimensionless diagonal weighted matrix and confidence modulation control increment terms, and outputs heterogeneous target control quantities; the event-driven edge execution unit is used to determine whether to update and issue the heterogeneous target control quantities with bounded steps according to the trigger indication; the security degradation control unit is used to switch to a conservative security curve mode based on multi-level degradation environmental temperature input when the communication interruption exceeds a set threshold or the confidence of the critical area is lower than the security threshold; the refeed self-learning and evidence recording unit is used to record the data of the whole process and update the model parameters and trigger thresholds in a rolling manner.

7. The industrial thermal network control system based on cyber-physical fusion according to claim 6, characterized in that: The multi-source data access unit sets up a local physical environment temperature sensor as a redundant acquisition channel on the edge gateway side, and periodically caches the outdoor weather forecast curves for the next few hours issued by the central platform on the edge side, so that it can still provide usable environmental temperature input to the safety degradation control unit during the period of disconnection from the central platform; the data collected by the multi-source data access unit includes at least the temperature of the heat source supply side, the supply side temperature and return side temperature of the trunk key nodes, the flow rate, pressure difference, circulation pump frequency, electric valve opening degree, heat meter reading, the controlled temperature of the sampling room or load group, outdoor dry bulb temperature and wind speed, time calendar characteristics, and repair and complaint information.

8. The industrial thermal network control system based on cyber-physical fusion according to claim 6, characterized in that: The physical mechanism twin unit characterizes the heat transfer and heat dissipation process of the heat medium for each pipe segment using an equivalent time delay and heat loss model, describes the heat transfer relationship between the primary and secondary sides using a heat transfer efficiency model for heat exchange nodes, and characterizes the thermal inertia of the building envelope using an equivalent first-order inertial element for the load group. The response of indoor air temperature to the secondary supply temperature is characterized as a calculable hysteresis-inertial process. When the pipe segment transmission delay is not an integer multiple of the unified time base sampling period, the physical mechanism twin unit uses a linear interpolation method between adjacent discrete sampling points to obtain the historical value of the inlet temperature.

9. The industrial thermal network control system based on cyber-physical fusion according to claim 6, characterized in that: When the safety degradation control unit switches to the safety curve mode, it generates a conservative curve of the heat source supply side temperature from the ambient temperature using the following formula: ; in The current available ambient temperature, , Based on the basic curve parameters, To address the bias in the current control area, the safety degradation control unit simultaneously tightens the maximum adjustment step size, freezes the direct adjustment weights for low-confidence variables, and only allows conservative actions to be performed at the levels of heat source temperature and primary circulation pump frequency, while maintaining the secondary side pressure difference and valve opening at the previous steady-state value. When the communication link is restored and the confidence level of the critical area rises above the safety threshold, it re-enters the trusted state generation stage, first reconstructs the trusted state vector, and then makes a decision.

10. The industrial thermal network control system based on cyber-physical fusion according to claim 6, characterized in that: The objects periodically updated by the reinjection self-learning and evidence-tracking unit include at least the equivalent heat loss parameters of the pipe section, the equivalent thermal inertia time constant of the building complex, the virtual sensing error statistics and physical estimation error statistics, the area weight, the trigger threshold, and the comfort preference parameters. The data tracked by the reinjection self-learning and evidence-tracking unit includes at least the trusted state snapshot, the optimization target value, the constraint activity status, the state difference before and after the instruction is issued, the fault label, the manual intervention record, and the energy efficiency comparison baseline. The tracked data is used for online model identification, operation auditing, and post-event traceability.

Images