Heat supply two-network dynamic balance intelligent regulation and control method based on fluid mechanics and AI coupling

Abstract

The invention provides a heat supply secondary network dynamic balance intelligent regulation and control method based on fluid mechanics and AI coupling, and relates to the technical field of data processing, and the method comprises the steps: collecting the pressure, temperature and flow data of all nodes in real time through an Internet of Things sensor network deployed in a heat supply secondary network, and constructing a multi-dimensional spatial-temporal characteristic data set; constructing a two-network hydraulic working condition space-time coupling dynamic model through a fluid mechanics continuity equation and a Bernoulli dynamic equation based on the constructed multi-dimensional space-time characteristic data set; and based on the established two-network hydraulic working condition space-time coupling dynamic model, training and optimizing the model through a machine learning algorithm to obtain an optimized model for performing second-level high-precision dynamic deduction on the pressure gradient and the flow distribution of the pipe network. According to the invention, full-chain intellectualization from sensing, deduction to regulation of the heat supply two networks is realized.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a dynamic balance intelligent control method for a secondary heating network based on the coupling of fluid mechanics and AI. Background Technology

[0002] In the field of urban centralized heating secondary pipeline network regulation, traditional regulation methods are limited by the dual technical bottlenecks of reliance on manual experience and static hydraulic models, making it difficult to adapt to the dynamic complexity of pipeline network operation. Taking a certain community as an example, during the 2024-2025 heating season, due to the lack of intelligent automatic control devices and data-driven analysis capabilities in the courtyard network, staff had to manually measure temperature, read pressure and flow data on-site every day, and achieve basic regulation by manually adjusting the balancing valve. A complete commissioning cycle lasted as long as 3 weeks. Moreover, due to the limitations of manual experience, the regulation effect could only be maintained for 2-3 days. Soon, a significant thermal imbalance problem appeared, with the room temperature in buildings near the heat source reaching as high as 26°C, requiring windows to be opened for heat dissipation, while the room temperature in buildings far away was only 16°C. More seriously, when a pipeline leak occurred, the traditional method relied entirely on manual inspection and investigation, which could not quickly capture the subtle parameter changes caused by the leak. This once led to a 10-hour heating outage in 3 buildings, seriously affecting the heating experience of users.

[0003] The core shortcomings of traditional technical systems are twofold. First, they lack dynamic modeling capabilities based on fluid mechanics. Traditional steady-state models cannot reflect the dynamic characteristics of pipeline pressure and flow changes over time and user load, and the overestimation error of heat loss can reach 27.4%, making it difficult to support precise control. Second, they lack the support of AI technology, especially neural network algorithms. They cannot perform in-depth analysis of multi-dimensional spatiotemporal data collected by sensors at all nodes of the pipeline network. Traditional methods are unable to extract subtle features of abnormal operating conditions such as sudden changes in user-side flow and pipeline leaks through algorithms such as convolutional neural networks. They cannot build data-driven anomaly identification logic based on neural networks and can only rely on manual trial-and-error adjustment and periodic inspections. This not only results in low control efficiency and insufficient accuracy, but also fails to achieve real-time identification and dynamic response to abnormal operating conditions, leading to repeated thermal imbalances and serious energy waste, making it difficult to meet the core requirements of modern heating for precise control and efficient energy-saving operation. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a dynamic balance intelligent control method for secondary heating networks based on the coupling of fluid mechanics and AI, which realizes the intelligent control of the entire chain of secondary heating networks from perception and deduction to control.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a dynamic balance intelligent control method for a secondary heating network based on the coupling of fluid mechanics and AI, the method comprising: By deploying an IoT sensor network in the secondary heating network, pressure, temperature and flow data of all nodes are collected in real time to construct a multi-dimensional spatiotemporal feature dataset. Based on the constructed multi-dimensional spatiotemporal feature dataset, a spatiotemporal coupled dynamic model of the hydraulic conditions of the two-network system is constructed through the fluid mechanics continuity equation and Bernoulli dynamic equation. Based on the established spatiotemporal coupled dynamic model of the hydraulic conditions of the two-network system, the model is trained and optimized through machine learning algorithms to obtain an optimized model that can perform second-level high-precision dynamic extrapolation of the pressure gradient and flow distribution of the pipeline network. The optimized model is used as the intelligent hub to perform real-time hydraulic state simulation calculations of the pipeline network and obtain real-time simulation results. Based on the real-time simulation results, the operating data is analyzed in depth through neural network algorithms to intelligently identify sudden changes in flow on the user side and abnormal operating conditions of pipeline leakage, so as to obtain real-time control strategies. The real-time control strategy is converted into control commands to drive the variable frequency pump and electric regulating valve actuator to perform actions, and the pipeline status feedback data after the actuators perform actions is obtained. By collecting pipeline status feedback data after the actuator's actions, the feedback data is used to learn and optimize the control parameters of the neural network algorithm to obtain optimized control parameters. The optimized control parameters are then fed back to the identification and strategy generation process to achieve adaptive closed-loop control.

[0006] Furthermore, by deploying an IoT sensor network in the secondary heating network, pressure, temperature, and flow data from all nodes are collected in real time, constructing a multi-dimensional spatiotemporal feature dataset, including: It receives a full node dataset containing pressure, temperature, flow data and the three-dimensional spatial coordinates of each sensor uploaded by an IoT sensor network deployed in the secondary heating network; based on the spatial topology of the secondary heating network, it abstracts the pipeline structure into a three-dimensional spatial topology geometry, and performs position verification and data validity verification of the sensor spatial coordinates according to the three-dimensional spatial point-line relationship judgment algorithm to obtain a verified valid full node dataset. The validated full-node data is spatiotemporally aligned and cleaned to obtain regular spatiotemporal sequence data; based on the regular spatiotemporal sequence data, the pressure gradient, flow rate change rate and temperature distribution characteristics of the pipeline network are extracted. The extracted pressure gradient, flow rate change rate, and temperature distribution features are fused to obtain a multi-dimensional spatiotemporal feature dataset.

[0007] Furthermore, based on the constructed multi-dimensional spatiotemporal feature dataset, a spatiotemporal coupled dynamic model of the secondary network hydraulic conditions is constructed using the fluid dynamics continuity equation and Bernoulli's dynamic equation. Based on this established spatiotemporal coupled dynamic model, machine learning algorithms are used to train and optimize the model, resulting in an optimized model capable of performing high-precision dynamic extrapolation of the network pressure gradient and flow distribution within seconds. This optimized model includes: Based on the obtained multi-dimensional spatiotemporal feature dataset, an initial spatiotemporal coupled dynamic model of the two-network hydraulic working condition is constructed through the fluid dynamics continuity equation and Bernoulli dynamic equation. The initial spatiotemporal coupled dynamic model is constructed and iteratively trained using historical data from a multi-dimensional spatiotemporal feature dataset as training samples through machine learning algorithms to obtain an optimized model that performs second-level dynamic extrapolation of pipeline pressure gradient and flow distribution. Based on the optimized model, the optimized model is validated and its parameters are fine-tuned using real-time collected data in order to maintain the accuracy of model inference.

[0008] Furthermore, the optimized model is used as the intelligent hub to perform real-time hydraulic state simulation calculations of the pipeline network, obtaining real-time simulation results, including: The optimized model is then loaded into the intelligent hub to obtain the loaded optimized model. The real-time collected multi-dimensional spatiotemporal feature dataset is input into the loaded optimized model; The optimized model performs second-level hydraulic calculations on the input real-time data to deduce real-time hydraulic state data including the pressure gradient and flow distribution of the entire pipe network. Based on the real-time hydraulic state data obtained from the simulation, abnormal operating condition features are extracted and analyzed to generate real-time simulation results that include the real-time hydraulic state of the pipeline network and the results of anomaly identification.

[0009] Furthermore, based on real-time simulation results, a neural network algorithm is used to perform in-depth analysis of the operational data, intelligently identifying sudden changes in user-side flow and abnormal pipeline leakage conditions to derive real-time control strategies, including: The received real-time simulation results are used for abnormal operating condition feature analysis to extract feature vectors of user-side flow mutations and pipeline leakage. The extracted feature vectors are input into a pre-trained neural network anomaly recognition model, which intelligently diagnoses the specific abnormal operating conditions and severity levels, and obtains the anomaly diagnosis results. Based on the obtained anomaly diagnosis results, combined with the reinforcement learning control strategy library, a preliminary set of control instructions for variable frequency pumps and electric regulating valves is generated. The generated preliminary control instruction set is subjected to hydraulic safety and stability verification, and the final executable real-time control strategy is output.

[0010] Furthermore, the real-time control strategy is converted into control commands to drive the variable frequency pump and electric regulating valve actuators to perform actions, obtaining pipeline network status feedback data after the actuators' actions, including: The received real-time control strategy is parsed and converted into standardized control commands that can be recognized by the edge controller; The generated standardized control commands are sent to the corresponding variable frequency pumps and electric regulating valve actuators; The system monitors the execution status of control commands issued in real time and collects pressure, temperature and flow data of the pipeline network after the actuator moves as pipeline network status feedback data. The collected pipeline status feedback data is summarized and preprocessed to obtain standardized feedback data for control parameter self-learning.

[0011] Furthermore, by collecting pipeline status feedback data after the actuator's actions; using the feedback data to self-learn and optimize the control parameters of the neural network algorithm to obtain optimized control parameters; and feeding the optimized control parameters back to the recognition and strategy generation process to achieve adaptive closed-loop control, including: Based on the received normalized feedback dataset, the control parameters of the neural network anomaly recognition model and the reinforcement learning control strategy library are optimized through self-learning to obtain the optimized control parameter set. The optimized control parameter set is fed back to the abnormal operating condition feature extraction and analysis process and the neural network anomaly recognition model to obtain updated control parameters. Based on updated control parameters, the control strategy is continuously and adaptively optimized during real-time simulation and strategy generation, forming a closed-loop control.

[0012] Secondly, a dynamic balance intelligent control system for a secondary heating network based on the coupling of fluid mechanics and AI includes: The acquisition module is used to collect pressure, temperature and flow data of all nodes in real time through an IoT sensor network deployed in the secondary heating network, and to build a multi-dimensional spatiotemporal feature dataset. The module is used to construct a spatiotemporally coupled dynamic model of the hydraulic conditions of the two-network system based on the constructed multi-dimensional spatiotemporal feature dataset and through the fluid dynamics continuity equation and Bernoulli dynamic equation. Based on the established spatiotemporally coupled dynamic model of the hydraulic conditions of the two-network system, the model is trained and optimized through machine learning algorithms to obtain an optimized model that performs high-precision dynamic extrapolation of the pressure gradient and flow distribution of the pipeline network in seconds. The optimization module is used to use the optimized model as an intelligent hub to perform real-time hydraulic state simulation calculations of the pipeline network and obtain real-time simulation results. Based on the real-time simulation results, the module performs in-depth analysis of the operating data through neural network algorithms to intelligently identify sudden changes in flow on the user side and abnormal operating conditions such as pipeline leakage, so as to obtain real-time control strategies. The real-time control strategies are converted into control commands to drive the variable frequency pump and electric regulating valve actuators to perform actions, and obtain pipeline state feedback data after the actuators perform actions. The processing module is used to collect pipeline status feedback data after the actuators have acted; use the feedback data to learn and optimize the control parameters of the neural network algorithm to obtain optimized control parameters; and feed the optimized control parameters back to the recognition and strategy generation process to achieve adaptive closed-loop control.

[0013] Thirdly, a computing device includes: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0014] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0015] The above-described solution of the present invention has at least the following beneficial effects: An IoT sensor network was deployed to collect pressure, temperature, and flow data from all nodes of the secondary heating network. After spatiotemporal alignment and cleaning, a multi-dimensional spatiotemporal feature dataset was constructed. An initial hydraulic condition model was built by combining the fluid mechanics continuity equation and Bernoulli's dynamic equation. Through iterative training and optimization using machine learning algorithms, a model capable of predicting network pressure gradients and flow distributions with high accuracy down to the second was obtained. Using this optimized model as the intelligent hub, a neural network algorithm was used to extract abnormal features and diagnose user-side flow mutations, network leaks, and other conditions. Combined with a reinforcement learning control strategy library, control strategies were generated and verified. Finally, the strategies were parsed into standardized commands to drive the variable frequency pumps and electric regulating valves. Simultaneously, based on the network status feedback data after execution, the neural network and reinforcement learning control were self-learned and optimized. This technology, which uses parameters to form a closed-loop control system, overcomes the technical problems of traditional secondary heating network control, such as reliance on experience and lag in manual adjustment, inability of static models to accurately reflect the dynamic hydraulic characteristics of the network, difficulty in timely identification of abnormal conditions, and lack of continuous adaptive optimization mechanisms leading to thermal imbalance and low energy efficiency. It achieves second-level accurate simulation of the network's hydraulic state, reduces the temperature difference at the user end from ±5℃ to within ±1℃, ensures 90% of users maintain a stable indoor temperature within the comfortable range of 22-24℃, enables second-level identification of abnormal conditions and reconstruction of control schemes, achieves a response time of less than 15 seconds for electric regulating valves, and achieves an average energy saving rate of over 12% for the entire heating season. This promotes the transformation of the heating system from passive operation and maintenance to proactive intelligent control, improving heating quality, user comfort and fairness, and system operating energy efficiency. Attached Figure Description

[0016] Figure 1 This is a schematic flowchart of an intelligent control method for dynamic balance of a secondary heating network based on the coupling of fluid mechanics and AI, provided by an embodiment of the present invention.

[0017] Figure 2 This is a schematic diagram of a dynamic balance intelligent control system for a secondary heating network based on the coupling of fluid mechanics and AI, provided by an embodiment of the present invention. Detailed Implementation

[0018] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0019] like Figure 1 As shown, an embodiment of the present invention proposes a dynamic balance intelligent control method for a secondary heating network based on the coupling of fluid mechanics and AI. The method includes the following steps: Step 1: By deploying an IoT sensor network in the secondary heating network, pressure, temperature and flow data of all nodes are collected in real time to construct a multi-dimensional spatiotemporal feature dataset. Step 2: Based on the constructed multi-dimensional spatiotemporal feature dataset, a spatiotemporal coupled dynamic model of the two-network hydraulic conditions is constructed using the fluid mechanics continuity equation and Bernoulli's dynamic equation. Based on the established spatiotemporal coupled dynamic model of the two-network hydraulic conditions, the model is trained and optimized using machine learning algorithms to obtain an optimized model that can perform second-level high-precision dynamic extrapolation of the pressure gradient and flow distribution of the pipeline network. Step 3: Use the optimized model as the intelligent hub to perform real-time hydraulic state simulation calculations of the pipeline network and obtain real-time simulation results; based on the real-time simulation results, perform in-depth analysis of the operating data through neural network algorithms to intelligently identify user-side flow mutations and abnormal pipeline leakage conditions in order to obtain real-time control strategies. Step 4: Convert the real-time control strategy into control commands to drive the variable frequency pump and electric regulating valve actuators to perform actions, and obtain pipeline status feedback data after the actuators perform actions; Step 5: Collect pipeline status feedback data after the actuator's actions; use the feedback data to learn and optimize the control parameters of the neural network algorithm to obtain optimized control parameters; feed the optimized control parameters back to the identification and strategy generation process to achieve adaptive closed-loop control.

[0020] In this embodiment of the invention, an IoT sensor network is deployed to collect real-time pressure, temperature, and flow data of all nodes in the secondary heating network and construct a multi-dimensional spatiotemporal feature dataset. An initial hydraulic condition model of the secondary network is built based on the fluid mechanics continuity equation and Bernoulli's dynamic equation. This model is then trained and optimized using machine learning algorithms to obtain a model capable of predicting network pressure gradients and flow distribution with high precision down to the second. This optimized model serves as the intelligent central hub for predicting real-time hydraulic conditions. A neural network algorithm is used to identify sudden changes in flow on the user side and abnormal network leakage conditions, generating real-time control strategies. These strategies are then converted into control commands to drive the variable frequency pumps and electric regulating valves. Simultaneously, feedback on the network status after execution is collected. The technology of using self-learning neural network control parameters based on feedback data to achieve adaptive closed-loop regulation overcomes the technical problems of traditional secondary heating network regulation, such as manual adjustment relying on experience and having a lag in response, static models failing to accurately reflect the dynamic hydraulic characteristics of the network, difficulty in timely identification of abnormal conditions, and lack of continuous adaptive optimization mechanisms leading to thermal imbalance and low energy efficiency. This results in second-level accurate simulation of the network's hydraulic state, rapid identification of abnormal conditions and dynamic generation of control schemes, efficient response of electric regulating valves, reduction of temperature differences at the user end, and improved operational energy efficiency. This promotes the transformation of heating from passive operation and maintenance to proactive intelligent control, while simultaneously improving heating quality and the fairness of user comfort.

[0021] In a preferred embodiment of the present invention, step 1 above may include: Step 1.1: Receive the full node dataset uploaded by the IoT sensor network deployed in the secondary heating network, which includes pressure, temperature, flow data, and the three-dimensional spatial coordinates of each sensor. Based on the spatial topology of the secondary heating network, the pipeline structure is abstracted into a three-dimensional spatial topology geometry. The sensor spatial coordinates are then verified for position and data validity using a three-dimensional spatial point-line relationship judgment algorithm to obtain a verified valid full node dataset. Specifically, this includes: receiving data uploaded by sensors deployed at key locations in the secondary heating network. The sensors need to be pre-installed at nodes, user entrances, and pipeline branches in the network. The uploaded data includes real-time collected pressure, temperature, and flow data, as well as the three-dimensional spatial coordinates of each sensor to identify its specific installation location in the network. All data is transmitted to the edge controller in real time via the wireless or wired communication link of the sensor network to ensure the continuity and timeliness of data reception. Next, the abstraction of the pipeline spatial topology was carried out. First, the actual construction drawings and design documents of the secondary heating network were retrieved to clarify the pipeline route, pipe diameter specifications, pipeline connection relationship and installation position of each equipment. The starting point, ending point and branch point of each pipeline in the network were abstracted into three-dimensional spatial nodes, and the pipelines connecting the nodes were abstracted into line segments in three-dimensional space. At the same time, the pipeline parameters corresponding to each line segment were marked to construct a three-dimensional spatial topology geometry consistent with the actual pipeline network structure. In the three-dimensional spatial topology geometry, the binding position corresponding to the three-dimensional coordinates of the sensor was reserved. Then, the sensor spatial coordinate position verification is performed, and the three-dimensional spatial point-line relationship determination algorithm is activated. The algorithm reads the three-dimensional spatial coordinates of each sensor sequentially according to the sensor's unique identifier, including the X, Y, and Z axis coordinate values, and retrieves the pipeline segment information corresponding to the sensor's theoretical installation location. For example, if a certain segment of the main pipeline in the east unit of Building 3 should be installed, its starting coordinates are X1, Y1, Z1, and its ending coordinates are X2, Y2, Z2. Next, through spatial geometric calculations, the perpendicular distance from the sensor coordinate point to the line segment defined by the starting and ending points is calculated. If the distance is less than a preset threshold set according to the pipeline diameter and installation allowable error, such as 0.3 meters, to ensure that the sensor coordinates fall within the spatial range of the pipeline entity, it is preliminarily determined that the sensor coordinates are on the corresponding pipeline line. Within the spatial coverage area of ​​the segment; simultaneously, extract the three-dimensional coordinates of adjacent sensors, determine the upstream and downstream or adjacent sensors in the same area according to the pipeline network topology, calculate the three-dimensional straight-line distance between the current sensor and each adjacent sensor, and compare it with the actual installation spacing of adjacent sensors recorded in the pipeline network design document. If the deviation between the calculated distance and the design spacing is within the allowable range, such as ±0.5 meters, considering construction and installation errors, the sensor position verification is further confirmed to be successful; if the distance from the sensor coordinates to the corresponding pipeline segment exceeds the preset threshold, or the deviation from the spacing of adjacent sensors exceeds the allowable range, the algorithm automatically marks the sensor as having an abnormal position and records the abnormality type, such as coordinates deviating from the pipeline segment or spacing not matching that of adjacent sensors; Next, data validity verification is performed, taking into account the normal operating parameter range of the secondary heating network. For example, the normal pressure range is usually 0.2MPa to 0.6MPa, the normal water supply temperature range is usually 45℃ to 60℃, and the normal flow rate range is determined according to the network design load. The data uploaded by each sensor that has passed the verification at each location is checked one by one. If the data is within the normal parameter range, and there is no logical contradiction between the data collected by different sensors at the same node, such as when the pressure and temperature data are collected simultaneously, or when the temperature is within the normal range but the pressure is zero, then the data is considered valid. If the data exceeds the normal parameter range, such as when the pressure is below 0.1MPa or above 0.8MPa, or when there is a logical contradiction, then the data is considered invalid and marked. Finally, all sensor data that passed location verification and were valid were integrated, categorized and organized according to the unique identifier of the sensor, to form a verified and valid full-node dataset containing pressure data, temperature data, flow data, and corresponding three-dimensional spatial coordinates and topological location information of each valid sensor.

[0022] Step 1.2 involves performing spatiotemporal alignment and data cleaning on the verified valid full-node data to obtain a regularized spatiotemporal sequence data. Based on this regularized spatiotemporal sequence data, the pressure gradient, flow rate change rate, and temperature distribution characteristics of the pipeline network are extracted. Specifically, this includes: performing spatiotemporal alignment on the verified valid full-node data; determining the corresponding spatial location of the pipeline network based on the unique identifier of each sensor; and then, based on the timestamp of data acquisition, mapping the data collected by different sensors at the same time to the specific spatial location of the pipeline network to ensure data consistency in both time and space, eliminating data deviations caused by asynchronous sensor acquisition times and inconsistent location correspondences. Then... Data cleaning was performed to filter out and remove abnormal data that exceeded the normal heating parameter range. For data gaps caused by temporary sensor malfunctions, normal data trends from adjacent time periods were used to supplement the missing data, resulting in a regularized spatiotemporal sequence. Features were then extracted from this regularized spatiotemporal sequence to calculate the pressure difference between two adjacent nodes, thus obtaining the pressure gradient of the pipeline network. The ratio of the flow rate change to the time interval for the same node in two adjacent data collection periods was calculated to obtain the flow rate change rate. Spatial interpolation was performed on the temperature data of each node to generate the temperature distribution of the entire secondary heating network, thereby extracting the network's temperature distribution characteristics.

[0023] Step 1.3 involves fusing the extracted pressure gradient, flow rate change rate, and temperature distribution features to obtain a multi-dimensional spatiotemporal feature dataset. Specifically, this includes: integrating the extracted pressure gradient, flow rate change rate, and temperature distribution features, using the time sequence of data collection as the time dimension clue, and associating the pressure gradient, flow rate change rate, and pipeline temperature distribution features of each node at each time point; simultaneously, combining the spatial location information of the pipeline network, binding the feature data of different spatial nodes with the corresponding pipeline network location, so that the integrated data contains both temporal change information and spatial distribution information, ultimately forming a multi-dimensional spatiotemporal feature dataset.

[0024] In this embodiment of the invention, pressure, temperature, and flow data of all nodes are received in real time from an IoT sensor network deployed in the secondary heating network. The received data is spatiotemporally aligned and cleaned to obtain regular spatiotemporal sequence data. Then, the network pressure gradient, flow rate of change, and temperature distribution features are extracted based on the regular data. Finally, the extracted features are fused to form a multi-dimensional spatiotemporal feature dataset. This overcomes the technical problems of traditional secondary heating network data acquisition, such as sensor data being prone to spatiotemporal deviations, raw data containing noise or outliers leading to low data quality, and the inability to meet the input requirements of subsequent models due to the lack of key feature extraction relying solely on raw data. This achieves the goal of ensuring spatiotemporal consistency of data, eliminating invalid interference data, and enriching the core hydraulic and thermal features of the network.

[0025] In a preferred embodiment of the present invention, step 2 above may include: Step 2.1: Based on the obtained multi-dimensional spatiotemporal feature dataset, an initial spatiotemporal coupled dynamic model of the secondary heating network's hydraulic operating conditions is constructed using the fluid dynamics continuity equation and Bernoulli's dynamic equation. Specifically, this includes: first, organizing the basic information of the pipe network contained in the multi-dimensional spatiotemporal feature dataset, including the pipe length, pipe diameter, pipe material roughness, and fluid physical property data of the secondary heating network; simultaneously, extracting pressure gradients, flow rate changes, and temperature distribution characteristics at different time periods and nodes in the dataset to ensure data coverage of various common operating conditions of the pipe network; subsequently, using the fluid dynamics continuity equation... Reflecting the flow conservation relationship and Bernoulli's dynamic equation in the pipeline network The core concept reflects the relationship of fluid energy changes, among which, For traffic volume, For pressure, For speed, For fluid density, For height, For viscous stress, As the source and sink terms, the well-organized basic information of the pipeline network and multi-dimensional spatiotemporal characteristic data are substituted into these two equations. In the modeling process, the focus is on the dynamic changes of pipeline network operation over time, such as parameter changes caused by fluctuations in user heating demand at different times, and the spatial distribution of nodes, such as the hydraulic correlation of nodes at different locations. An initial spatiotemporal coupled dynamic model of the hydraulic conditions of the secondary network is constructed, which can simultaneously reflect the coupling relationship between the time and spatial dimensions. This enables the model to initially describe the changes of pipeline network hydraulic parameters with time and space, breaking the limitation of traditional models that rely solely on steady-state models and cannot reflect real-time changes.

[0026] Step 2.2 involves iteratively training the constructed initial spatiotemporal coupled dynamic model using a machine learning algorithm with historical data from a multi-dimensional spatiotemporal feature dataset as training samples. This yields an optimized model capable of dynamically extrapolating the pressure gradient and flow distribution of the pipe network at the second level. Specifically, this includes: selecting historical operational data from the multi-dimensional spatiotemporal feature dataset for multiple complete heating cycles, including actual pressure, flow, and temperature values ​​for different dates and times at each node, as well as corresponding records of the actual hydraulic state of the pipe network. This historical data is then categorized according to operating conditions, such as weekday morning peak heating, nighttime low-load operation, and operation during extreme low-temperature weather, and used as training samples for the machine learning algorithm. The initial spatiotemporal coupled dynamic model is then used to perform second-level dynamic extrapolation of the pipe network pressure gradient and flow distribution. The empty-coupled dynamic model and training samples are imported into the machine learning platform to start the iterative training process. First, the model outputs the inferred pressure gradient and flow distribution results based on the input data in the training samples. Then, the inferred results are compared with the corresponding actual pressure gradient and flow distribution data in the training samples, and the deviation value between the two is calculated. Based on the deviation value, the parameters in the model, such as the fluid resistance coefficient and energy loss coefficient, are adjusted. The above process of model inference, result comparison and parameter adjustment is repeated until the deviation between the pressure gradient and flow distribution results output by the model and the actual data stabilizes at a low level. The calculation is completed instantly after receiving the input data, realizing the second-level dynamic inference of the pressure gradient and flow distribution of the pipeline network, and obtaining the optimized model.

[0027] Step 2.3: Based on the optimized model, the optimized model is verified and its parameters are fine-tuned using real-time collected data to maintain the accuracy of the model's extrapolation. Specifically, this includes: continuously acquiring the latest pressure, temperature, and flow data collected in real-time by the IoT sensor network during the daily operation of the secondary heating network; inputting the real-time data into the optimized model at fixed time intervals; allowing the model to output the extrapolation results of the network pressure gradient and flow distribution at the current moment; then comparing the model's extrapolation results with the actual network pressure gradient and flow distribution data collected by the sensors at the same moment; calculating the error between the two; if the error is within a preset reasonable range, such as less than 3.2%, the current parameters of the model are maintained; if the error exceeds the reasonable range, the cause of the error is analyzed, such as a sudden change in the external ambient temperature leading to a significant change in user heating demand, or temporary adjustments to the heating load of some network nodes, etc., and relevant parameters in the model are fine-tuned accordingly, such as adjusting the coefficient reflecting the influence of ambient temperature and optimizing the calculation parameters of fluid flow resistance in the pipeline. Through this continuous real-time data verification and parameter fine-tuning, the optimized model can adapt to various dynamic changes during pipeline operation and maintain high projection accuracy of pipeline pressure gradient and flow distribution in the long term.

[0028] In this embodiment of the invention, based on a multi-dimensional spatiotemporal feature dataset, an initial spatiotemporal coupled dynamic model of the secondary heating network's hydraulic conditions is constructed using the fluid mechanics continuity equation and Bernoulli's dynamic equation. This initial model is then iteratively trained using historical data from the multi-dimensional spatiotemporal feature dataset as training samples, employing machine learning algorithms to obtain an optimized model capable of dynamically extrapolating the network's pressure gradient and flow distribution within seconds. Finally, the optimized model is validated and its parameters are fine-tuned based on real-time collected data. This overcomes the technical problems of traditional secondary heating network control, such as the lack of dynamic modeling methods based on fluid mechanics mechanisms, the inability of steady-state models to reflect real-time dynamic changes in network pressure gradients and flow distribution, and the low extrapolation accuracy and difficulty in adapting to actual network operating conditions due to unoptimized training. This achieves the construction of a secondary heating network hydraulic condition model that combines mechanistic rigor with data-driven adaptability, enabling second-level dynamic extrapolation of network pressure gradients and flow distribution, and maintaining high extrapolation accuracy through real-time validation and fine-tuning.

[0029] In a preferred embodiment of the present invention, step 3 above may include: Step 3.1: Using the obtained optimized model, load the optimized model into the intelligent hub to obtain the loaded optimized model. Specifically, this includes: first, determining the intelligent hub for heating. The intelligent hub is a central control platform with data processing and model running capabilities, compatible with the operating environment of the optimized model. Before loading, the optimized secondary network hydraulic condition model is adapted for compatibility to ensure that the model's parameter format, calculation logic, and hardware computing power and software of the intelligent hub match, avoiding operational conflicts. Subsequently, the complete file of the optimized model is transmitted to the intelligent hub through a high-speed data transmission channel. After the transmission is completed, the model loading program is started. The intelligent hub automatically verifies the model's core algorithm, parameter library, and inference logic to confirm that the model is undamaged and can be called normally. After the verification is passed, the optimized model is in a real-time standby state in the intelligent hub, ready to receive input data and perform inference calculations at any time, thereby replacing the traditional mode without an intelligent core and relying on manual judgment of the pipeline status.

[0030] Step 3.2 involves inputting the real-time collected multi-dimensional spatiotemporal feature dataset into the loaded optimized model. Specifically, during the operation of the secondary heating network, the IoT sensor network continuously collects pressure, temperature, and flow data from all nodes. After spatiotemporal alignment and data cleaning, the real-time data forms a regular real-time multi-dimensional spatiotemporal feature dataset. Through IoT wireless communication, the real-time dataset is automatically transmitted to the intelligent hub at fixed short time intervals, such as every 1 second. Data compression and verification technologies are used during transmission to ensure that the data is not lost or damaged during transmission. After receiving the real-time dataset, the intelligent hub automatically converts the data format to match the input format required by the loaded optimized model. Then, according to the model's input requirements, key information such as pressure gradient, flow rate of change, and temperature distribution characteristics in the real-time dataset are input into the optimized model one by one.

[0031] Step 3.3 involves performing second-level hydraulic calculations on the input real-time data using the optimized model to deduce real-time hydraulic state data, including the pressure gradient and flow distribution of the entire pipe network. Specifically, this includes: the optimized model, after being loaded by the intelligent central control, starts the hydraulic calculation program. The model calls its core algorithm and, combined with the input real-time multi-dimensional spatiotemporal feature data, calculates the hydraulic state of the entire pipe network. During the calculation, the model fully utilizes the parallel computing power of the intelligent central control and simultaneously performs rapid calculations on the pressure transmission relationship and flow distribution law of each node in the pipe network. Without manual intervention, it can complete the pressure gradient calculation and flow distribution deduction of all nodes in the entire pipe network within 1 second. After the calculation is completed, the model automatically generates real-time hydraulic state data, which includes the pressure gradient value of each branch of the pipe network, the specific flow value of each node, and the distribution trend of pressure and flow within the spatial range of the pipe network. This clearly presents the current hydraulic operation of the entire pipe network and solves the problem that traditional static models cannot reflect the dynamic changes of pipe network pressure and flow in real time.

[0032] Step 3.4: Based on the real-time hydraulic state data obtained from the simulation, perform abnormal operating condition feature extraction and analysis to generate real-time simulation results containing the real-time hydraulic state of the pipeline network and anomaly identification results. Specifically, the intelligent central hub calls a preset feature extraction algorithm to analyze the real-time hydraulic state data obtained from the simulation. First, key feature indicators are extracted from the data, including the instantaneous change amplitude of the pressure gradient, the deviation of the flow distribution, and the difference between pressure and flow between different nodes. Then, the feature indicators are compared with the normal operating condition feature thresholds stored in the intelligent central hub. The normal operating condition thresholds are based on historical data. The hydraulic data collected during normal operation, such as the fluctuation range of pressure gradient and the uniformity of flow distribution, are statistically derived. If a certain characteristic index exceeds the normal threshold range, such as a sudden drop in pressure gradient at a certain node or a significant imbalance in flow distribution in a certain area, it is determined that there is an abnormal operating condition in that area. The specific manifestations of the abnormal characteristics are further analyzed to distinguish whether it is a sudden change in flow on the user side or a pipeline leak. Finally, the intelligent hub integrates the real-time hydraulic status data of the entire pipeline network with the identified abnormal operating condition results to generate a real-time inference result that includes the overall operating status of the pipeline network, the location of the abnormal area, and the type of abnormality.

[0033] In this embodiment of the invention, an optimized secondary heating network hydraulic condition model is loaded into the intelligent central hub. A multi-dimensional spatiotemporal feature dataset collected in real time is input into the loaded model. The model performs second-level hydraulic calculations to extrapolate real-time hydraulic state data of the entire network's pressure gradient and flow distribution. Based on this real-time hydraulic state data, abnormal condition features are extracted and analyzed to generate real-time extrapolation results containing the network's real-time hydraulic state and anomaly identification results. This overcomes the technical problems of traditional secondary heating network control, which lacks real-time and accurate full-network hydraulic state extrapolation capabilities and relies on manual inspections to determine operating conditions, leading to delayed anomaly identification and an inability to promptly grasp the overall network operating status. This achieves the technical effect of rapidly acquiring the real-time hydraulic state of the entire network, timely identifying abnormal conditions, and avoiding control delays or failures due to untimely or inaccurate condition judgments.

[0034] In a preferred embodiment of the present invention, step 3 above may include: Step 3.5 involves analyzing the abnormal operating condition characteristics of the received real-time simulation results, extracting feature vectors for user-side flow mutations and pipeline network leaks. Specifically, this includes receiving the generated real-time simulation results, which contain real-time pressure data, flow data, and preliminary information on the overall operating status of the entire pipeline network at each node. The data in the real-time simulation results is then decomposed and analyzed, partitioning each node according to its spatial topology to analyze pressure change trends, flow fluctuation frequencies, and the linkage relationships between parameters of different nodes. For abnormal scenarios involving user-side flow mutations, the focus is on extracting the instantaneous change amplitude, rate of change, and corresponding pressure response of the flow at each user inlet node, such as the specific data characteristics of a sudden increase or decrease in the flow at a building's user inlet within a short period. For abnormal scenarios involving pipeline network leaks, the focus is on extracting the rate of continuous pressure decrease in pipeline sections, abnormal flow loss, and the correlation characteristics of pressure and flow changes between adjacent nodes, such as the data characteristics of a pipeline section where pressure steadily decreases over time and the corresponding area's flow decreases without a reasonable cause. The extracted key data features corresponding to the two types of abnormal scenarios are then organized into structured feature vectors according to a pre-defined unified format.

[0035] Step 3.6: Input the extracted feature vector into the pre-trained neural network anomaly recognition model to intelligently diagnose the specific abnormal operating conditions and severity levels, and obtain the anomaly diagnosis results. Specifically, this includes: collecting operational data from multiple complete heating cycles of the secondary heating network, covering normal operating conditions data under different seasons and user loads, as well as historical abnormal operating conditions data such as sudden changes in user-side flow and pipeline leaks; manually labeling the collected abnormal operating condition data to clarify the anomaly type and severity level corresponding to each data point, forming a labeled training dataset; cleaning the collected raw data, removing invalid data caused by sensor failures, and supplementing the missing data with the normal data trends of adjacent time periods. Then, the data is normalized to convert parameters of different dimensions such as pressure and flow rate into a range of 0 to 1, eliminating the impact of dimensional differences on model training. Features are then extracted from the processed data, such as the instantaneous rate of change of flow rate, the duration of pressure drop, and the correlation between parameters of adjacent nodes, forming input feature data with the same feature vector format as in step 3.5. Based on the requirements of the anomaly recognition task, a neural network containing an input layer, a hidden layer, and an output layer is constructed. Input layer: The number of neurons matches the dimension of the time series data matrix, that is, the input dimension is T×D, where T is the time step and D is the number of features such as pressure, flow, and temperature. It is used to receive the preprocessed time series feature data to ensure that the time series change information of pipeline operation parameters can be completely input. Convolutional layers: Set 2-3 convolutional layers, each with a different kernel size. Extract local features from the time-series data through convolution operations. The convolution calculation process is as follows: Among the rules, It is the ReLU activation function. For convolution kernel weights, For bias terms, For convolution operators, For the representative layer index, This is used as a feature map to enhance the model's ability to capture nonlinear anomalies; Fully connected layer: One or two fully connected layers are set after the convolutional layer to integrate the local features extracted by the convolutional layer into global features, thereby achieving a comprehensive representation of abnormal features; Output layer: Employs a dual-branch structure, both branches are connected via... Output probability values, where, The role of the activation function is to transform the linear output of the fully connected layer into a probability distribution form, so that the output can be directly interpreted as the probability of belonging to a certain type of anomaly. This is an abnormal probability distribution. This is the weight matrix. For bias terms, For feature maps, As a representative layer index, the first branch outputs the probability of two types of anomalies: user-side traffic mutation and pipeline leakage, to determine the anomaly type; the second branch outputs the probability of three levels (mild, moderate, and severe) under the corresponding type, to determine the severity of the anomaly. Finally, the combination of the type and level with the highest probability is output as the model diagnostic result.

[0036] The preprocessed labeled feature data is divided into training and validation sets in a 7:3 ratio. The training set is then input into the designed neural network, and the cross-entropy loss function is applied. , This represents the average cross-entropy loss of the model across all training samples in a single iteration. This indicates the number of training samples input in a single iteration. Indicates the first Cross-entropy loss for each sample, Indicates the first The actual label of each sample Indicates the total number of categories. The model represents the first The sample belongs to the first Predicted probability of class The natural logarithm function is used to calculate the error between the model's prediction and the actual label. Network weights and bias parameters are adjusted iteratively. During training, the model's accuracy is evaluated using a validation set after every 100 iterations. Training stops when the validation set accuracy no longer improves after 10 consecutive iterations, and the current model parameters are saved to avoid overfitting. The trained model is then evaluated using an independent test set. If the anomaly type recognition accuracy is below 90% or the severity level classification accuracy is below 85%, the number of hidden layer neurons is increased or the learning rate is adjusted for retraining. If the evaluation meets the standards, a small amount of new anomaly case data not used in training is selected to fine-tune the model, further improving its adaptability to new conditions. This ultimately forms a pre-trained neural network anomaly recognition model that can be used for real-time anomaly diagnosis. The processed feature vectors are input into the neural network model, and the model will automatically... The system initiates an internal feature matching and analysis process, comparing the input feature vector with various abnormal feature templates stored during model training. Based on the comparison results, the model first determines the specific type of the current anomaly, clarifying whether it is a sudden change in user-side flow or a pipeline leak. Then, it combines key parameters in the feature vector, such as the maximum magnitude of the flow change and the rate of pressure drop caused by the pipeline leak, with a comparison to the pre-set severity classification standards within the model to determine the severity of the anomaly. For example, user-side flow changes are classified into mild, moderate, and severe. Mild is a flow fluctuation within 10% of the normal range, moderate is 10% to 30%, and severe is more than 30%. Pipeline leaks are classified into slow leaks, moderate leaks, and rapid leaks based on the rate of pressure drop. Finally, the model outputs an anomaly diagnosis result that includes the specific type of abnormal condition and the corresponding severity level.

[0037] Step 3.7: Based on the obtained anomaly diagnosis results, and in conjunction with the reinforcement learning control strategy library, generate a preliminary control instruction set for the variable frequency pump and electric regulating valve. Specifically, this includes: first, retrieving the pre-built reinforcement learning control strategy library, which was formed through simulation training on thousands of different operating conditions in the secondary heating network. The library stores standardized control schemes for different anomaly types and severity levels. Each scheme specifies whether the corresponding control device is a variable frequency pump or an electric regulating valve, the specific control direction, and the initial adjustment range. Based on the obtained anomaly diagnosis results, match the corresponding control scheme template in the reinforcement learning control strategy library. For example, when the diagnosis result is... When a slight flow change occurs at the entrance of Building 3 in a residential area, the matched template adjusts the opening of the electric regulating valve at the building entrance, with an initial adjustment of 5%. When the diagnosis is a moderate leak in the pipeline network, occurring on the main pipeline on the east side of the residential area, the matched template reduces the output pressure of the associated variable frequency pump, with an initial adjustment of 3%, while simultaneously closing the opening of the electric regulating valve downstream of the leak point by 10%. Based on the matched template and combined with the real-time operating parameters of the current pipeline network, such as the current output pressure of the variable frequency pump and the current opening of the electric regulating valve, the control parameters in the template are adaptively fine-tuned to generate a preliminary control instruction set for specific equipment.

[0038] Step 3.8 involves verifying the hydraulic safety and stability of the generated preliminary control command set, and outputting the final executable real-time control strategy. Specifically, this includes: first, starting the hydraulic safety and stability simulation verification platform, which has a built-in hydraulic calculation model of the secondary heating network; then, simulating the overall hydraulic operation state of the network after the input control command set is executed; inputting the generated preliminary control command set into the platform; and simulating the pressure distribution, flow allocation, and operating load of equipment such as variable frequency pumps and electric regulating valves at each node of the entire network after command execution. In the hydraulic safety verification stage, it checks whether the pressure of each pipeline section exceeds the maximum safe pressure designed for the pipeline and whether it is lower than the minimum pressure to ensure normal heating. Simultaneously, it checks whether the operating parameters of each device are within their rated operating range. To prevent problems such as pipeline overpressure rupture and equipment overload damage; in the system stability verification stage, check whether the pressure and flow of each node in the pipeline network can stabilize in a short time after regulation, and whether there will be situations such as drastic pressure fluctuations and serious imbalances in flow distribution, so as to avoid new thermal imbalance problems caused by regulation actions. If the simulation verification results show that there are safety hazards or stability problems, such as the simulated pressure of a certain pipeline section exceeding the safety value, return to step 3.7, adjust the corresponding parameters in the preliminary regulation instruction set according to the problem type, such as reducing the adjustment range of the electric regulating valve or reducing the pressure adjustment value of the variable frequency pump, and input it into the platform again for simulation verification. Repeat the above process until the simulation results fully meet the hydraulic safety and stability requirements, and finally output a real-time regulation strategy that can be directly executed.

[0039] In this embodiment of the invention, the technical means of analyzing the abnormal operating conditions of the received real-time simulation results to extract feature vectors of sudden changes in flow on the user side and leakage in the pipeline network, inputting the feature vectors into a pre-trained neural network anomaly recognition model to intelligently diagnose the specific abnormal operating condition type and severity level, generating a preliminary control instruction set for variable frequency pumps and electric regulating valves based on the anomaly diagnosis results and a reinforcement learning control strategy library, and then verifying the hydraulic safety and stability of the preliminary control instruction set to output the final executable real-time control strategy, overcomes the technical problems of traditional secondary heating network control, such as the reliance on manual inspection to identify anomalies with lag, large errors in manual judgment of anomaly type and severity level, the generation of control instructions based on experience without scientific basis, and the lack of safety and stability verification which can easily lead to system imbalance or failure. Thus, it achieves the technical effects of quickly diagnosing abnormal operating conditions, generating scientifically adapted control instructions, ensuring the hydraulic safety of the pipeline network and the stable operation of the system after the control instructions are executed, and avoiding problems such as heating stoppage and energy waste caused by lag in anomaly identification or improper control.

[0040] In a preferred embodiment of the present invention, step 4 above may include: Step 4.1 involves parsing and converting the received real-time control strategy into standardized control commands recognizable by the edge controller. This includes receiving the output real-time control strategy, which contains pressure adjustment requirements for variable frequency pumps in different areas, opening adjustment parameters for electric regulating valves in different buildings, and specific control time requirements. Then, the command parsing program is started to decompose the various control requirements in the real-time control strategy into independent control tasks. The control equipment type, unique equipment identifier, and specific control parameters corresponding to each task are identified. Next, the command format specification and communication protocol of the edge controller are retrieved, and the decomposed control tasks are converted into standardized commands recognizable by the edge controller according to the specification.

[0041] Step 4.2 involves sending the generated standardized control commands to the corresponding variable frequency pumps and electric regulating valves. Specifically, this includes: matching the corresponding variable frequency pump or electric regulating valve in the heating equipment management list based on the target equipment code in the standardized control command; clarifying the specific location of the actuator for each command, such as the variable frequency pump in the underground pump room of Building 3 in the East Zone and the electric regulating valve in Unit 1 of Building 5 in the West Zone; then activating the edge-side IoT transmission network for heating. The network uses stable wired or wireless communication to achieve real-time command transmission. Standardized control commands are grouped according to the area where the equipment is located and sent to the corresponding actuators one by one through the IoT transmission network. During the sending process, a command packet transmission and verification mechanism is used to ensure that commands are not lost or damaged during transmission. After receiving a command, each actuator automatically sends a command reception confirmation signal back to the edge controller. The edge controller resends commands that have not received a confirmation signal until all commands are successfully received by the corresponding actuators. This replaces the traditional method of manual on-site adjustment with tools, avoiding delays caused by time-consuming manual back-and-forth trips and ensuring that control commands are quickly and accurately transmitted to the actuators.

[0042] Step 4.3 involves real-time monitoring of the execution status of issued control commands and collecting pressure, temperature, and flow data of the pipeline network after the actuators take action as pipeline network status feedback data. Specifically, after the standardized control commands are issued, the edge controller starts the execution status monitoring program to receive real-time operating data from each actuator, such as the actual output pressure of the variable frequency pump, the actual opening degree of the electric regulating valve, and the equipment operating current. This data is used to determine whether the control commands have been executed and whether the execution is in place. For example, if the command requires the electric regulating valve to increase its opening degree by 20%, the actual opening degree is monitored to see if it reaches the target value. If it does not, a second command is triggered to correct the deviation. At the same time, relying on the 3000+ high-precision IoT sensor network deployed above, pressure, temperature, and flow data of all nodes of the heating network are collected in real time after the actuators take action. The collection frequency is consistent with the model prediction frequency to ensure data timeliness. These data cover key nodes such as pipeline branches and user inlets, and can fully reflect the impact of control actions on the overall hydraulic state of the pipeline network. The data is marked as pipeline network status feedback data.

[0043] Step 4.4 involves summarizing and preprocessing the collected pipeline status feedback data to obtain standardized feedback data for control parameter self-learning. Specifically, this includes: firstly, summarizing the collected pipeline status feedback data according to time and space dimensions. In the time dimension, the data is sorted by the timestamp of data collection to form a continuous time-series data sequence; in the spatial dimension, data is associated according to the location of pipeline nodes, such as associating pressure and flow data at different user entrances in the same building with the corresponding electric regulating valve control commands for that building, ensuring the correspondence between data and control actions; then, preprocessing the summarized data, firstly by cleaning the data to remove invalid data caused by temporary sensor malfunctions, and supplementing missing data due to brief sensor offline periods using a trend-based completion method based on concurrent data from adjacent nodes; secondly, data denoising using a sliding window filtering method to process high-frequency interference signals in the pressure and flow data, eliminating the impact of instantaneous fluctuations on data validity; finally, data normalization processing is performed to uniformly convert parameters with different dimensions such as pressure, temperature, and flow to the same numerical range, ensuring that the data format is consistent with the multi-dimensional spatiotemporal feature dataset constructed earlier. After the above processing, standardized feedback data for self-learning of control parameters is obtained.

[0044] In this embodiment of the invention, the received real-time control strategy is parsed and converted into standardized control commands recognizable by the edge controller. The generated standardized control commands are then sent to the corresponding variable frequency pumps and electric regulating valve actuators. The execution status of the sent control commands is monitored in real time, and the pressure, temperature, and flow data of the pipeline network after the actuators act are collected as pipeline network status feedback data. At the same time, the collected pipeline network status feedback data is summarized and preprocessed to obtain standardized feedback data for control parameter self-learning. Therefore, this invention overcomes the technical problems in traditional secondary heating network control, such as poor compatibility between edge controllers and actuators due to the lack of a standardized command conversion mechanism, errors and low efficiency of manual command issuance, inability to confirm whether the control actions have been implemented without real-time execution status monitoring, and disorganized feedback data that is difficult to support subsequent control parameter optimization, resulting in delayed control response and broken cloud-edge collaborative closed loop. This invention achieves the adaptation of control commands to edge-side execution devices, real-time visibility of the execution status of control actions, and standardized processing of feedback data, ensuring the continuity of the entire technology system from data perception to model construction to intelligent control to adaptive optimization.

[0045] In a preferred embodiment of the present invention, step 5 above may include: Step 5.1: Based on the received normalized feedback dataset, the control parameters of the neural network anomaly recognition model and the reinforcement learning control strategy library are optimized through self-learning to obtain an optimized set of control parameters. Specifically, this includes: First, receiving the normalized feedback dataset, which contains time-series data of pressure, temperature, and flow rate at all nodes of the pipeline network after the actuator's action, as well as records of the actual effects of the corresponding control actions. Then, initiating the self-learning optimization program, for the neural network anomaly recognition model, comparing the actual user-side flow surges and pipeline leakage cases in the dataset with the model's previous diagnostic results, calculating the diagnostic bias, and adjusting the convolutional layer weights and fully connected layer parameters of the model based on the bias to make the model more accurate in capturing features of similar anomalies. For the reinforcement learning control strategy library, analyzing the correlation between the control commands in the dataset and the actual changes in the pipeline network state. If thermal imbalance still exists after control under a certain operating condition, adjusting the reward function parameters for the corresponding operating condition in the strategy library, and simultaneously correcting basic control parameters such as the variable frequency pump pressure adjustment amplitude and the electric regulating valve opening adjustment amount. Through multiple rounds of iterative calculation, the adjusted neural network parameters and reinforcement learning strategy parameters are integrated to form the optimized set of control parameters.

[0046] Step 5.2 involves feeding back the optimized control parameter set to the abnormal operating condition feature extraction and analysis process and the neural network anomaly identification model to obtain updated control parameters. Specifically, this includes: classifying and organizing the optimized control parameter set to identify the parameters belonging to the neural network anomaly identification model and those belonging to the abnormal operating condition feature extraction and analysis; then, through the high-speed data transmission channel on the edge side, feeding back the parameters required for the abnormal operating condition feature extraction and analysis to the feature extraction stage in step 3.5 to update the judgment criteria for abnormal features in the feature extraction algorithm. For example, if the feedback data shows that the original pressure gradient threshold is too high, resulting in the failure to extract leakage features, the threshold is lowered to ensure that the pressure change characteristics of pipeline leakage can be captured more promptly in the future; and the updated parameters of the neural network anomaly identification model are directly imported into the pre-trained model to overwrite the old parameters of the original model and re-initialize the model's calculation logic so that the model can perform diagnosis based on the updated parameters when receiving new feature vectors.

[0047] Step 5.3, based on the updated control parameters, achieves continuous adaptive optimization of the control strategy during real-time simulation and strategy generation, forming a closed-loop control. Specifically, this includes: during real-time operation of the pipeline network, when real-time simulation of the pipeline network's hydraulic state is performed, the optimized neural network anomaly identification model parameters are already in effect. The model can more accurately simulate the pressure gradient and flow distribution of the entire pipeline network, reducing simulation errors caused by outdated model parameters; when generating control strategies, the updated reinforcement learning control strategy library parameters come into play. The strategy library can output a more suitable preliminary control instruction set based on real-time operating conditions. For example, in extreme low-temperature weather, it automatically increases the basic pressure parameter of the variable frequency pump, reducing the deviation between the control instructions and actual needs. After the control instructions are executed, the newly collected pipeline network status feedback data will be processed again into a standardized feedback dataset, and cyclically input into Step 5.1 for a new round of parameter optimization, forming a closed-loop process from parameter optimization to simulation accuracy to strategy adaptation to feedback update to further optimization.

[0048] In this embodiment of the invention, based on the received standardized feedback dataset, the control parameters of the neural network anomaly identification model and the reinforcement learning control strategy library are self-learned and optimized to obtain an optimized set of control parameters. The optimized set of control parameters is then fed back to the abnormal operating condition feature extraction and analysis process and the neural network anomaly identification model to obtain updated control parameters. Based on the updated control parameters, the control strategy is continuously adaptively optimized and a closed-loop control is formed during real-time simulation and strategy generation. This overcomes the technical problems of traditional secondary heating network control, such as the lack of a self-learning optimization mechanism, the susceptibility of models and strategies to failure due to changes in pipeline operating conditions, and the inability of control to continuously adapt to the dynamic operation requirements of the pipeline network. This achieves the goal of maintaining the high accuracy of the neural network anomaly identification model, improving the adaptability of the reinforcement learning control strategy library, ensuring the continuous operation of the entire technology system from data perception to model construction to intelligent control to adaptive optimization, and ultimately achieving a secondary heating network control response speed of seconds, a stable household temperature difference within ±1℃, and an average energy saving rate of over 12% for the entire heating season.

[0049] like Figure 2 As shown, embodiments of the present invention also provide an intelligent control system for dynamic balance of a secondary heating network based on the coupling of fluid mechanics and AI, comprising: The acquisition module is used to collect pressure, temperature and flow data of all nodes in real time through an IoT sensor network deployed in the secondary heating network, and to build a multi-dimensional spatiotemporal feature dataset. The module is used to construct a spatiotemporally coupled dynamic model of the hydraulic conditions of the two-network system based on the constructed multi-dimensional spatiotemporal feature dataset and through the fluid dynamics continuity equation and Bernoulli dynamic equation. Based on the established spatiotemporally coupled dynamic model of the hydraulic conditions of the two-network system, the model is trained and optimized through machine learning algorithms to obtain an optimized model that performs high-precision dynamic extrapolation of the pressure gradient and flow distribution of the pipeline network in seconds. The optimization module is used to use the optimized model as an intelligent hub to perform real-time hydraulic state simulation calculations of the pipeline network and obtain real-time simulation results. Based on the real-time simulation results, the module performs in-depth analysis of the operating data through neural network algorithms to intelligently identify sudden changes in flow on the user side and abnormal operating conditions such as pipeline leakage, so as to obtain real-time control strategies. The real-time control strategies are converted into control commands to drive the variable frequency pump and electric regulating valve actuators to perform actions, and obtain pipeline state feedback data after the actuators perform actions. The processing module is used to collect pipeline status feedback data after the actuators have acted; use the feedback data to learn and optimize the control parameters of the neural network algorithm to obtain optimized control parameters; and feed the optimized control parameters back to the recognition and strategy generation process to achieve adaptive closed-loop control.

[0050] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A dynamic balance intelligent control method for secondary heating networks based on the coupling of fluid mechanics and AI, characterized in that, The method includes: By deploying an IoT sensor network in the secondary heating network, pressure, temperature and flow data of all nodes are collected in real time to construct a multi-dimensional spatiotemporal feature dataset. Based on the constructed multi-dimensional spatiotemporal feature dataset, a spatiotemporal coupled dynamic model of the hydraulic conditions of the two-network system is constructed through the fluid mechanics continuity equation and Bernoulli dynamic equation. Based on the established spatiotemporal coupled dynamic model of the hydraulic conditions of the two-network system, the model is trained and optimized through machine learning algorithms to obtain an optimized model that can perform second-level high-precision dynamic extrapolation of the pressure gradient and flow distribution of the pipeline network. The optimized model is used as the intelligent hub to perform real-time hydraulic state simulation calculations of the pipeline network and obtain real-time simulation results. Based on the real-time simulation results, the operating data is analyzed in depth through neural network algorithms to intelligently identify sudden changes in flow on the user side and abnormal operating conditions of pipeline leakage, so as to obtain real-time control strategies. The real-time control strategy is converted into control commands to drive the variable frequency pump and electric regulating valve actuator to perform actions, and the pipeline status feedback data after the actuators perform actions is obtained. By collecting pipeline status feedback data after the actuator's actions, the feedback data is used to learn and optimize the control parameters of the neural network algorithm to obtain optimized control parameters. The optimized control parameters are then fed back to the identification and strategy generation process to achieve adaptive closed-loop control.

2. The intelligent control method for dynamic balance of secondary heating networks based on the coupling of fluid mechanics and AI according to claim 1, characterized in that, By deploying an IoT sensor network in the secondary heating network, pressure, temperature, and flow data from all nodes are collected in real time, constructing a multi-dimensional spatiotemporal feature dataset, including: It receives a full node dataset containing pressure, temperature, flow data and the three-dimensional spatial coordinates of each sensor uploaded by an IoT sensor network deployed in the secondary heating network; based on the spatial topology of the secondary heating network, it abstracts the pipeline structure into a three-dimensional spatial topology geometry, and performs position verification and data validity verification of the sensor spatial coordinates according to the three-dimensional spatial point-line relationship judgment algorithm to obtain a verified valid full node dataset. The validated full-node data is spatiotemporally aligned and cleaned to obtain regular spatiotemporal sequence data; based on the regular spatiotemporal sequence data, the pressure gradient, flow rate change rate and temperature distribution characteristics of the pipeline network are extracted. The extracted pressure gradient, flow rate change rate, and temperature distribution features are fused to obtain a multi-dimensional spatiotemporal feature dataset.

3. The intelligent control method for dynamic balance of secondary heating networks based on the coupling of fluid mechanics and AI according to claim 2, characterized in that, Based on the constructed multi-dimensional spatiotemporal feature dataset, a spatiotemporal coupled dynamic model of the two-network hydraulic conditions is built using the fluid mechanics continuity equation and Bernoulli's dynamic equation. Based on this established model, machine learning algorithms are used to train and optimize the model, resulting in an optimized model that performs high-precision dynamic extrapolation of the network pressure gradient and flow distribution within seconds. This optimized model includes: Based on the obtained multi-dimensional spatiotemporal feature dataset, an initial spatiotemporal coupled dynamic model of the two-network hydraulic working condition is constructed through the fluid dynamics continuity equation and Bernoulli dynamic equation. The initial spatiotemporal coupled dynamic model is constructed and iteratively trained using historical data from a multi-dimensional spatiotemporal feature dataset as training samples through machine learning algorithms to obtain an optimized model that performs second-level dynamic extrapolation of pipeline pressure gradient and flow distribution. Based on the optimized model, the optimized model is validated and its parameters are fine-tuned using real-time collected data in order to maintain the accuracy of model inference.

4. The intelligent control method for dynamic balance of secondary heating networks based on the coupling of fluid mechanics and AI according to claim 3, characterized in that, The optimized model is used as the intelligent hub to perform real-time hydraulic state simulation calculations of the pipeline network, obtaining real-time simulation results, including: The optimized model is then loaded into the intelligent hub to obtain the loaded optimized model. The real-time collected multi-dimensional spatiotemporal feature dataset is input into the loaded optimized model; The optimized model performs second-level hydraulic calculations on the input real-time data to deduce real-time hydraulic state data including the pressure gradient and flow distribution of the entire pipe network. Based on the real-time hydraulic state data obtained from the simulation, abnormal operating condition features are extracted and analyzed to generate real-time simulation results that include the real-time hydraulic state of the pipeline network and the results of anomaly identification.

5. The intelligent control method for dynamic balance of secondary heating networks based on the coupling of fluid mechanics and AI according to claim 4, characterized in that, Based on real-time simulation results, a neural network algorithm is used to perform in-depth analysis of operational data, intelligently identifying abnormal operating conditions such as sudden changes in user-side flow and pipeline leakage, in order to derive real-time control strategies, including: The received real-time simulation results are used for abnormal operating condition feature analysis to extract feature vectors of user-side flow mutations and pipeline leakage. The extracted feature vectors are input into a pre-trained neural network anomaly recognition model, which intelligently diagnoses the specific abnormal operating conditions and severity levels, and obtains the anomaly diagnosis results. Based on the obtained anomaly diagnosis results, combined with the reinforcement learning control strategy library, a preliminary set of control instructions for variable frequency pumps and electric regulating valves is generated. The generated preliminary control instruction set is subjected to hydraulic safety and stability verification, and the final executable real-time control strategy is output.

6. The intelligent control method for dynamic balance of secondary heating networks based on the coupling of fluid mechanics and AI according to claim 5, characterized in that, The real-time control strategy is converted into control commands to drive the variable frequency pump and electric regulating valve actuators to perform actions, obtaining pipeline network status feedback data after the actuators act, including: The received real-time control strategy is parsed and converted into standardized control commands that can be recognized by the edge controller; The generated standardized control commands are sent to the corresponding variable frequency pumps and electric regulating valve actuators; The system monitors the execution status of control commands issued in real time and collects pressure, temperature and flow data of the pipeline network after the actuator moves as pipeline network status feedback data. The collected pipeline status feedback data is summarized and preprocessed to obtain standardized feedback data for control parameter self-learning.

7. The intelligent control method for dynamic balance of secondary heating networks based on the coupling of fluid mechanics and AI according to claim 6, characterized in that, By collecting pipeline status feedback data after the actuator's actions, the feedback data is used to self-learn and optimize the control parameters of the neural network algorithm, resulting in optimized control parameters. The optimized control parameters are fed back to the identification and strategy generation process to achieve adaptive closed-loop control, including: Based on the received normalized feedback dataset, the control parameters of the neural network anomaly recognition model and the reinforcement learning control strategy library are optimized through self-learning to obtain the optimized control parameter set. The optimized control parameter set is fed back to the abnormal operating condition feature extraction and analysis process and the neural network anomaly recognition model to obtain updated control parameters. Based on updated control parameters, the control strategy is continuously and adaptively optimized during real-time simulation and strategy generation, forming a closed-loop control.

8. A dynamic balance intelligent control system for a secondary heating network based on the coupling of fluid mechanics and AI, wherein the system implements the method as described in any one of claims 1 to 7, characterized in that, include: The acquisition module is used to collect pressure, temperature and flow data of all nodes in real time through an IoT sensor network deployed in the secondary heating network, and to build a multi-dimensional spatiotemporal feature dataset. The module is used to construct a spatiotemporally coupled dynamic model of the hydraulic conditions of the two-network system based on the constructed multi-dimensional spatiotemporal feature dataset and through the fluid dynamics continuity equation and Bernoulli dynamic equation. Based on the established spatiotemporally coupled dynamic model of the hydraulic conditions of the two-network system, the model is trained and optimized through machine learning algorithms to obtain an optimized model that performs high-precision dynamic extrapolation of the pressure gradient and flow distribution of the pipeline network in seconds. The optimization module is used to use the optimized model as an intelligent hub to perform real-time hydraulic state simulation calculations of the pipeline network and obtain real-time simulation results. Based on real-time simulation results, the system uses neural network algorithms to perform in-depth analysis of operational data, intelligently identifying abnormal operating conditions such as sudden changes in user-side flow and pipeline leakage, in order to obtain real-time control strategies. The real-time control strategy is converted into control commands to drive the variable frequency pump and electric regulating valve actuator to perform actions, and the pipeline status feedback data after the actuators perform actions is obtained. The processing module is used to collect pipeline status feedback data after the actuators have performed their actions; The feedback data is used to learn and optimize the control parameters of the neural network algorithm, resulting in optimized control parameters. The optimized control parameters are then fed back to the recognition and strategy generation process to achieve adaptive closed-loop control.

9. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images