An underground pipe network partitioning metering monitoring and regulation system based on an internet of things

Abstract

This invention discloses an IoT-based underground pipe network zonal metering monitoring and control system, comprising the following modules: a zonal monitoring module for collecting zonal flow, pressure, water quality, and valve status data, and for collecting upstream flow and pressure data; a data preprocessing module for preprocessing the zonal monitoring data; a time-series feature modeling module for outputting zonal prediction data using an improved TimesNet model; an anomaly detection module for generating zonal anomaly detection results; a control distribution generation module for outputting zonal control distribution based on an improved generative flow network reinforcement learning model; a control parameter generation module for obtaining zonal control parameters and updating model parameters; and a control execution module for adjusting valve opening, water supply distribution ratio, and pressure setpoint. This invention integrates time-series feature modeling and generative flow network reinforcement learning to achieve refined monitoring and intelligent control of underground pipe network zonal areas.

Description

Technical Field

[0001] This invention relates to the field of Internet of Things (IoT) monitoring and intelligent control technology, and in particular to an IoT-based underground pipeline network zone metering monitoring and control system. Background Technology

[0002] With the increasing demand for smart water management and refined urban pipe network management, IoT-based underground pipe network monitoring and control technologies have gradually attracted widespread attention. Existing pipe network zoning metering and control systems mostly rely on single sensor data or empirical models based on historical statistics to estimate and schedule the pipe network's operating status. However, in practical applications, the following problems commonly exist:

[0003] The collected multi-source data, including flow rate, pressure, water quality, and valve status, suffer from inconsistent sampling frequencies, large timestamp deviations, and strong noise interference. This makes it difficult to achieve cross-channel synchronization and reliable alignment of zone monitoring data, affecting monitoring accuracy. Most existing prediction models use traditional time series analysis methods or shallow neural networks, which are unable to effectively capture multi-scale time dependencies and topological constraints during pipeline operation, resulting in insufficient zone prediction accuracy. For anomaly detection, existing methods mostly use static threshold discrimination or simple statistical analysis, lacking dynamic adaptability and sensitivity to nonlinear abrupt changes, making them prone to missed detections and misjudgments. In terms of generating control strategies, most existing pipeline control systems use experience-based rule control or traditional optimization algorithms, which are difficult to generate and adaptively optimize dynamic control distributions under complex topological constraints, risk disturbances, and multi-objective conditions. This leads to unreasonable valve regulation and pressure settings, affecting the overall stability and energy efficiency of water supply.

[0004] Therefore, how to provide an IoT-based underground pipeline network zone metering monitoring and control system is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an Internet of Things-based underground pipe network zone metering monitoring and control system. This invention comprehensively utilizes multi-source data acquisition, time-series feature modeling, anomaly detection, and generator flow network reinforcement learning techniques to describe in detail the process of fine-grained monitoring, prediction, and intelligent control of pipe network flow, pressure, water quality, and valve status at the zone level. It has the advantages of high monitoring accuracy, sensitive anomaly identification, and adaptive optimization of control strategies.

[0006] According to an embodiment of the present invention, an underground pipeline network zone metering monitoring and control system based on the Internet of Things includes the following modules:

[0007] The zone monitoring module is used to divide the underground pipe network into multiple zones, collect zone flow, pressure, water quality data and valve status data, and collect upstream flow and pressure data to generate zone monitoring data and water supply capacity data.

[0008] The data preprocessing module is used to preprocess the partition monitoring data and generate standardized partition monitoring data.

[0009] The temporal feature modeling module is used to extract multi-scale temporal features from standardized partition monitoring data using an improved TimesNet model and perform time-dependent modeling, outputting partition prediction data.

[0010] The anomaly detection module is used to calculate partition deviation data based on partition prediction data and standardized partition monitoring data, extract partition anomaly data, and generate partition anomaly detection results.

[0011] The regulation distribution generation module is used to input the zonal prediction data, zonal anomaly discrimination results and water supply capacity data into the improved generative flow network reinforcement learning model, perform topology coding, risk calibration and constraint sampling, and output the zonal regulation distribution.

[0012] The regulation parameter generation module is used to obtain the partition regulation parameters based on the partition regulation distribution sampling, and update the parameters of the improved generative flow network reinforcement learning model based on feedback.

[0013] The execution control module is used to adjust valve opening, water supply distribution ratio, and pressure setpoint using zone control parameters.

[0014] Optionally, the process of dividing the underground pipe network into multiple zones, collecting zone flow, pressure, water quality data, and valve status data, and collecting upstream flow and pressure data to generate zone monitoring data and water supply capacity data specifically includes:

[0015] Establish the underground pipeline network topology and divide the area according to the pipe segment connection method;

[0016] Collect zone monitoring data, which includes instantaneous and cumulative flow of inlet and outlet water pipe sections, real-time pressure and pressure fluctuation amplitude of pipe section nodes, turbidity, residual chlorine and conductivity of water samples in the pipe section, and valve opening percentage and switching status.

[0017] Collect water supply capacity data, which includes the total outflow and water supply pressure of upstream water plants or pumping stations.

[0018] Optionally, the preprocessing of the partition monitoring data specifically includes:

[0019] Time synchronization processing is performed on the zone monitoring data, and the collection timestamps of flow rate, pressure, water quality and valve status information are aligned according to a unified time base.

[0020] Noise suppression is achieved by using filtering methods on the zoned monitoring data;

[0021] For the zonal monitoring data, interpolation methods or the average of adjacent time periods are used to replace missing data to fill in missing values.

[0022] The monitoring data of different zones are normalized to map the values ​​of different dimensions to a unified range.

[0023] Optionally, the improved TimesNet model specifically includes:

[0024] The improved TimesNet model includes an input encoding unit, an adaptive multi-scale convolutional branch, a topologically constrained temporal attention unit, a physically consistent residual unit, an anomaly detection unit, and an output mapping unit.

[0025] The input encoding unit concatenates the zone flow rate sequence, zone pressure sequence, zone water quality sequence and valve status sequence in the channel dimension, and uses the water supply capacity data as the adjustment vector, multiplying it element by element with the concatenated input tensor to obtain the adjusted input representation;

[0026] The adaptive multi-scale convolution branch dynamically sets the kernel size and dilation rate according to the adjustment vector, performs multi-scale convolution operations on the input representation in parallel, and outputs a multi-scale temporal feature matrix.

[0027] The topology-constrained temporal attention unit constructs an adjacency weight matrix using pipe segment connection relationships, performs weighted calculations on the multi-scale temporal feature matrix, and outputs a temporal feature representation of topology constraints.

[0028] The physical consistency residual unit compares the time feature representation with the physical constraint results based on the relationship between flow rate and pressure difference to obtain the residual vector, and then adds the residual vector to the time feature representation element by element to obtain the physical consistency enhancement feature;

[0029] The anomaly perception unit generates an anomaly indication vector through a reconstruction mechanism, performs gating and filtering of physical consistency enhancement features, and outputs the feature representation after anomaly perception.

[0030] The output mapping unit takes the feature representation after anomaly detection and inputs it into the fully connected layer, performs linear transformation and normalization operations, and outputs a fixed-dimensional vector as partition prediction data.

[0031] Optionally, the step of calculating partition deviation data based on partition prediction data and standardized partition monitoring data, extracting partition anomaly data, and generating partition anomaly discrimination results specifically includes:

[0032] Subtracting the zonal prediction data from the standardized zonal monitoring data point by point along the time dimension yields the flow deviation sequence, pressure deviation sequence, water quality deviation sequence, and valve status deviation sequence, which are then concatenated along the feature dimension to form the zonal deviation data.

[0033] Within a sliding time window, the mean, standard deviation, and difference scores between adjacent time points are calculated for the partitioned deviation data, and anomaly candidate feature vectors containing stability, volatility, and abrupt change characteristics are constructed.

[0034] K-means clustering is used on the abnormal candidate feature vectors to divide all feature vectors into normal clusters and abnormal clusters, and the distance between each sample and the cluster center is calculated using Euclidean distance.

[0035] Quantile statistics are performed on the distance distribution to determine the dynamic threshold interval. When the distance of a certain deviation point is greater than the upper quantile threshold, it is marked as an anomaly. When it is in the upper or lower neighborhood of the threshold, a second comparison is performed. Points that continuously exceed the window number limit are uniformly marked as anomalies, thus obtaining partitioned anomaly data.

[0036] The partition anomaly data is input into the multilayer perceptron, and linear mapping, nonlinear activation and normalization operations are performed in sequence to output the partition anomaly discrimination result.

[0037] Optionally, the improved generative stream network reinforcement learning model specifically includes:

[0038] The improved generative flow network reinforcement learning model includes a topology coding unit, a conditional flow transformation branch, a risk calibration unit, and a constraint sampling unit.

[0039] The topology coding unit concatenates the partition prediction data, partition anomaly discrimination results and water supply capacity data to form a joint state vector. It performs normalization processing on the joint state vector, uses the directed adjacency matrix of the underground pipe network to perform matrix multiplication on the joint state vector to generate topology propagation results, and obtains the topology embedding representation by adding and normalizing elements one by one.

[0040] The conditional flow transformation branch receives the topological embedding representation as conditional input, performs invertible affine transformations on the latent variables layer by layer, including element-wise scaling and element-wise translation, and connects nonlinear activation and batch normalization operations after each layer transformation to output the candidate partition control distribution.

[0041] The risk calibration unit generates a risk weight vector based on the partition anomaly discrimination result, multiplies the variance vector of the candidate partition control distribution element by element with the risk weight vector, performs numerical compression on the variance elements of partitions identified as anomalous, keeps the original values ​​of the variance elements of partitions identified as normal, and renormalizes the variance vector to obtain the candidate distribution after risk calibration.

[0042] The constraint sampling unit sets upper and lower limits for zone flow, pressure and valve opening based on water supply capacity data. It performs random sampling from the candidate distribution after risk calibration to obtain an initial control sample set. It performs constraint tests on the initial samples point by point, removes samples that exceed the upper and lower limits, and performs L2 projection or boundary truncation correction on the samples that fail to pass, and outputs the zone control distribution that meets the constraint conditions.

[0043] Optionally, the step of obtaining the partition control parameters based on the partition control distribution sampling and updating the parameters of the improved generative flow network reinforcement learning model based on feedback specifically includes:

[0044] Random sampling is performed from the zone control distribution to generate multiple zone control parameter sets according to the preset sampling quantity. Each zone control parameter set includes valve opening, water supply distribution ratio and pressure set value.

[0045] The control parameter sets of each zone and the underground pipe network topology data are input into the calculation process. The flow rate of each node is calculated according to the node flow balance equation, the pressure of each node is calculated according to the node energy equation, the head loss of each pipe segment is calculated according to the pipe segment resistance formula, and the decay of water quality concentration of each pipe segment over time is calculated according to the water quality transmission model. At the node, the water quality concentration is mixed according to the flow weighted average rule to obtain the simulated flow rate data, simulated pressure data and simulated water quality data at each time step.

[0046] For each set of zonal control parameters, calculate the point-by-point difference between the simulated flow data, simulated pressure data, simulated water quality data and preset target data, generate residual signals, call the energy consumption calculation module to obtain the corresponding energy consumption data, call the zonal anomaly discrimination results to generate anomaly weights, call the zonal control distribution output to output the corresponding distribution variance, and calculate the single-step distribution reward based on the residual signals, energy consumption data, anomaly weights and distribution variances.

[0047] In the time dimension, the single-step distributed reward of all time steps is accumulated, the residual signal of all time steps is accumulated, and the product of the anomaly weight and the distribution variance of all time steps is accumulated to generate trajectory-level distributed reward.

[0048] The gradient signal is obtained by performing element-wise multiplication between the trajectory-level distribution return and the logarithmic probability of the partitioned control distribution.

[0049] The gradient signal is input into the backpropagation process to iteratively update the layer-by-layer affine transformation parameters, scaling parameters, translation parameters, and nonlinear activation function parameters in the improved generative flow network reinforcement learning model.

[0050] The beneficial effects of this invention are:

[0051] This invention proposes an IoT-based underground pipe network zonal metering monitoring and control system. By deploying multi-source sensors at the zonal level to collect flow, pressure, water quality, and valve status data, and combining this data with upstream water supply capacity information to form zonal monitoring data, it solves the problems of single data source and difficulty in synchronizing different channels in existing systems. Through preprocessing methods such as time synchronization, missing value imputation, and normalization, this invention ensures the integrity and consistency of zonal monitoring data. An improved TimesNet model is used to model multi-scale temporal features and capture time dependencies of standardized data, and combined with physical consistency residuals and topological constraint mechanisms, effectively improving prediction accuracy and stability. Based on the dynamic differences between predicted and measured data, this invention constructs a sliding window statistical and clustering discrimination mechanism, which can sensitively identify abnormal fluctuations and nonlinear disturbances, overcoming the shortcomings of high false negative and false positive rates in existing static threshold methods. At the control level, this invention introduces an improved generative flow network reinforcement learning model, jointly modeling zonal prediction data, anomaly discrimination results, and water supply capacity data. Topological coding, risk calibration, and constraint sampling are used to generate a control distribution that meets operational constraints, and adaptive optimization of the strategy is achieved through sampling and feedback iteration. The final output control parameters can be directly applied to valve opening, water supply distribution ratio, and pressure setting, achieving dynamic, precise, and efficient regulation of the underground pipe network. This invention offers the advantages of high monitoring accuracy, sensitive anomaly identification, intelligent control strategies, and significantly improved water supply operation efficiency. Attached Figure Description

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

[0053] Figure 1 This is a schematic diagram of the structure of an Internet of Things-based underground pipeline network zoning metering monitoring and control system proposed in this invention;

[0054] Figure 2 This is a schematic diagram of the process of an Internet of Things-based underground pipeline network zoning metering monitoring and control system proposed in this invention.

[0055] Figure 3 This is a schematic diagram of an improved TimesNet model in an Internet of Things-based underground pipeline network zoning metering monitoring and control system proposed in this invention.

[0056] Figure 4 This is a schematic diagram of an improved generative flow network reinforcement learning model in an IoT-based underground pipeline network zonal metering monitoring and control system proposed in this invention. Detailed Implementation

[0057] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0058] refer to Figure 1-4 An Internet of Things-based underground pipeline network zone metering monitoring and control system includes the following modules:

[0059] The zone monitoring module is used to divide the underground pipe network into multiple zones, collect zone flow, pressure, water quality data and valve status data, and collect upstream flow and pressure data to generate zone monitoring data and water supply capacity data.

[0060] The data preprocessing module is used to preprocess the partition monitoring data and generate standardized partition monitoring data.

[0061] The temporal feature modeling module is used to extract multi-scale temporal features from standardized partition monitoring data using an improved TimesNet model and perform time-dependent modeling, outputting partition prediction data.

[0062] The anomaly detection module is used to calculate partition deviation data based on partition prediction data and standardized partition monitoring data, extract partition anomaly data, and generate partition anomaly detection results.

[0063] The regulation distribution generation module is used to input the zonal prediction data, zonal anomaly discrimination results and water supply capacity data into the improved generative flow network reinforcement learning model, perform topology coding, risk calibration and constraint sampling, and output the zonal regulation distribution.

[0064] The regulation parameter generation module is used to obtain the partition regulation parameters based on the partition regulation distribution sampling, and update the parameters of the improved generative flow network reinforcement learning model based on feedback.

[0065] The execution control module is used to adjust valve opening, water supply distribution ratio, and pressure setpoint using zone control parameters.

[0066] In this embodiment, the process of dividing the underground pipe network into multiple zones, collecting zone flow, pressure, water quality data, and valve status data, and collecting upstream flow and pressure data to generate zone monitoring data and water supply capacity data specifically includes:

[0067] Establish the underground pipeline network topology and divide the area according to the pipe segment connection method;

[0068] Collect zone monitoring data, which includes instantaneous and cumulative flow of inlet and outlet water pipe sections, real-time pressure and pressure fluctuation amplitude of pipe section nodes, turbidity, residual chlorine and conductivity of water samples in the pipe section, and valve opening percentage and switching status.

[0069] Collect water supply capacity data, which includes the total outflow and water supply pressure of upstream water plants or pumping stations.

[0070] This implementation method establishes the underground pipe network topology and divides it into zones based on the pipe segment connection method. In each zone, it collects multi-dimensional monitoring data such as instantaneous flow rate, cumulative flow rate, real-time pressure, pressure fluctuation amplitude, turbidity, residual chlorine, conductivity, and valve opening and switching status. At the same time, it combines the total outflow and supply pressure of the upstream water plant or pumping station to generate water supply capacity data, thereby achieving an accurate characterization of the hydraulic conditions and water quality of the zones. This provides comprehensive, accurate, and high-precision data support for subsequent time series prediction and anomaly identification.

[0071] In this embodiment, the preprocessing of the zonal monitoring data specifically includes:

[0072] Time synchronization processing is performed on the zone monitoring data, and the collection timestamps of flow rate, pressure, water quality and valve status information are aligned according to a unified time base.

[0073] Noise suppression is achieved by using filtering methods on the zoned monitoring data;

[0074] For the zonal monitoring data, interpolation methods or the average of adjacent time periods are used to replace missing data to fill in missing values.

[0075] The monitoring data of different zones are normalized to map the values ​​of different dimensions to a unified range.

[0076] This implementation method performs time synchronization, noise suppression, missing value imputation, and normalization on the zonal monitoring data, ensuring the consistency of multi-source flow, pressure, water quality, and valve status data in the time dimension and the comparability in the numerical dimension. This significantly reduces the uncertainty caused by sensor bias and environmental interference, enabling subsequent models to perform feature extraction and modeling on a unified and stable data basis, thereby improving the reliability and accuracy of zonal prediction and anomaly detection.

[0077] In this embodiment, the improved TimesNet model specifically includes:

[0078] The improved TimesNet model includes an input encoding unit, an adaptive multi-scale convolutional branch, a topologically constrained temporal attention unit, a physically consistent residual unit, an anomaly detection unit, and an output mapping unit.

[0079] The input encoding unit concatenates the zone flow rate sequence, zone pressure sequence, zone water quality sequence and valve status sequence in the channel dimension, and uses the water supply capacity data as the adjustment vector, multiplying it element by element with the concatenated input tensor to obtain the adjusted input representation;

[0080] The adaptive multi-scale convolution branch dynamically sets the kernel size and dilation rate according to the adjustment vector, performs multi-scale convolution operations on the input representation in parallel, and outputs a multi-scale temporal feature matrix.

[0081] The topology-constrained temporal attention unit constructs an adjacency weight matrix using pipe segment connection relationships, performs weighted calculations on the multi-scale temporal feature matrix, and outputs a temporal feature representation of topology constraints.

[0082] The physical consistency residual unit compares the time feature representation with the physical constraint results based on the relationship between flow rate and pressure difference to obtain the residual vector, and then adds the residual vector to the time feature representation element by element to obtain the physical consistency enhancement feature;

[0083] The anomaly perception unit generates an anomaly indication vector through a reconstruction mechanism, performs gating and filtering of physical consistency enhancement features, and outputs the feature representation after anomaly perception.

[0084] The output mapping unit takes the feature representation after anomaly detection and inputs it into the fully connected layer, performs linear transformation and normalization operations, and outputs a fixed-dimensional vector as partition prediction data.

[0085] This implementation introduces units such as input encoding, adaptive multi-scale convolutional branching, topological constraint time attention, physical consistency residual, anomaly perception, and output mapping into the improved TimesNet model, enabling dynamic modeling of multi-dimensional sequences such as zone flow, pressure, water quality, and valve status. It can effectively capture complex dependencies across multiple time scales and improve prediction stability by combining topological structure and physical constraints, significantly enhancing the prediction accuracy and anomaly sensitivity of zone operation status.

[0086] In this embodiment, the step of calculating partition deviation data based on partition prediction data and standardized partition monitoring data, extracting partition anomaly data, and generating partition anomaly discrimination results specifically includes:

[0087] Subtracting the zonal prediction data from the standardized zonal monitoring data point by point along the time dimension yields the flow deviation sequence, pressure deviation sequence, water quality deviation sequence, and valve status deviation sequence, which are then concatenated along the feature dimension to form the zonal deviation data.

[0088] Within a sliding time window, the mean, standard deviation, and difference scores between adjacent time points are calculated for the partitioned deviation data, and anomaly candidate feature vectors containing stability, volatility, and abrupt change characteristics are constructed.

[0089] K-means clustering is used on the abnormal candidate feature vectors to divide all feature vectors into normal clusters and abnormal clusters, and the distance between each sample and the cluster center is calculated using Euclidean distance.

[0090] Quantile statistics are performed on the distance distribution to determine the dynamic threshold interval. When the distance of a certain deviation point is greater than the upper quantile threshold, it is marked as an anomaly. When it is in the upper or lower neighborhood of the threshold, a second comparison is performed. Points that continuously exceed the window number limit are uniformly marked as anomalies, thus obtaining partitioned anomaly data.

[0091] The partition anomaly data is input into the multilayer perceptron, and linear mapping, nonlinear activation and normalization operations are performed in sequence to output the partition anomaly discrimination result.

[0092] This implementation method constructs a deviation sequence based on zonal prediction data and standardized monitoring data, and combines sliding window statistics, clustering operations, and dynamic threshold determination to achieve sensitive detection of abnormal fluctuations and sudden events. It also utilizes a multilayer perceptron to output zonal anomaly discrimination results, effectively avoiding the problems of high false positive rate and serious missed detection in traditional static threshold methods. This improves the accuracy and real-time performance of pipeline anomaly identification and ensures the reliability of zonal operation status monitoring.

[0093] In this embodiment, the improved generative stream network reinforcement learning model specifically includes:

[0094] The improved generative flow network reinforcement learning model includes a topology coding unit, a conditional flow transformation branch, a risk calibration unit, and a constraint sampling unit.

[0095] The topology coding unit concatenates the partition prediction data, partition anomaly discrimination results and water supply capacity data to form a joint state vector. It performs normalization processing on the joint state vector, uses the directed adjacency matrix of the underground pipe network to perform matrix multiplication on the joint state vector to generate topology propagation results, and obtains the topology embedding representation by adding and normalizing elements one by one.

[0096] The conditional flow transformation branch receives the topological embedding representation as conditional input, performs invertible affine transformations on the latent variables layer by layer, including element-wise scaling and element-wise translation, and connects nonlinear activation and batch normalization operations after each layer transformation to output the candidate partition control distribution.

[0097] The risk calibration unit generates a risk weight vector based on the partition anomaly discrimination result, multiplies the variance vector of the candidate partition control distribution element by element with the risk weight vector, performs numerical compression on the variance elements of partitions identified as anomalous, keeps the original values ​​of the variance elements of partitions identified as normal, and renormalizes the variance vector to obtain the candidate distribution after risk calibration.

[0098] The constraint sampling unit sets upper and lower limits for zone flow, pressure and valve opening based on water supply capacity data. It performs random sampling from the candidate distribution after risk calibration to obtain an initial control sample set. It performs constraint tests on the initial samples point by point, removes samples that exceed the upper and lower limits, and performs L2 projection or boundary truncation correction on the samples that fail to pass, and outputs the zone control distribution that meets the constraint conditions.

[0099] This implementation method sets up topology coding, conditional flow transformation, risk calibration and constraint sampling units in the improved generative flow network reinforcement learning model, and jointly models the zonal prediction data, anomaly discrimination results and water supply capacity data. This not only ensures that the control parameters meet the network topology and physical constraints, but also improves the robustness and risk prevention and control capabilities under abnormal conditions. As a result, dynamic optimization and reliable generation of zonal control distribution are achieved, ensuring the stability and safety of the water supply system operation.

[0100] In this embodiment, the step of obtaining the partition control parameters based on the partition control distribution sampling and updating the parameters of the improved generative flow network reinforcement learning model based on feedback specifically includes:

[0101] Random sampling is performed from the zone control distribution to generate multiple zone control parameter sets according to the preset sampling quantity. Each zone control parameter set includes valve opening, water supply distribution ratio and pressure set value.

[0102] The control parameter sets for each zone and the underground pipe network topology data are input into the calculation process. The flow at each node is calculated according to the node flow balance equation, which is as follows:

[0103] ;

[0104] in, For inflow node Traffic, outflow node Traffic, This refers to the node's required quantity.

[0105] The pressure at each node is calculated according to the nodal energy equation, which is as follows:

[0106] ;

[0107] in, For nodes Water head, For nodes pressure, The specific weight of water. For nodes Elevation;

[0108] The head loss of each pipe section is calculated according to the pipe section resistance formula, which is:

[0109] ;

[0110] in, For the head loss of the pipeline section, The drag coefficient, For the length of the pipe section, For the flow rate of the pipe section, It is the acceleration due to gravity. Pipe diameter;

[0111] The decay of water concentration over time in each pipe section is calculated based on the water quality transport model, which is as follows:

[0112] ;

[0113] in, Indicates the current pipe segment in the next time step Water quality concentration data, For a moment Water quality concentration, The attenuation coefficient;

[0114] The water quality concentration was mixed at the nodes according to the flow weighted average rule to obtain the simulated flow rate data, simulated pressure data and simulated water quality data at each time step;

[0115] For each set of zonal control parameters, calculate the point-by-point difference between the simulated flow data, simulated pressure data, simulated water quality data and preset target data, generate residual signals, call the energy consumption calculation module to obtain the corresponding energy consumption data, call the zonal anomaly discrimination results to generate anomaly weights, call the zonal control distribution output to output the corresponding distribution variance, and calculate the single-step distribution reward based on the residual signals, energy consumption data, anomaly weights and distribution variances.

[0116] In the time dimension, the single-step distributed reward of all time steps is accumulated, the residual signal of all time steps is accumulated, and the product of the anomaly weight and the distribution variance of all time steps is accumulated to generate trajectory-level distributed reward.

[0117] The gradient signal is obtained by performing element-wise multiplication between the trajectory-level distribution return and the logarithmic probability of the partitioned control distribution.

[0118] The gradient signal is input into the backpropagation process to iteratively update the layer-by-layer affine transformation parameters, scaling parameters, translation parameters, and nonlinear activation function parameters in the improved generative flow network reinforcement learning model.

[0119] This implementation method generates residual signals and energy consumption data by sampling from the zoned control distribution and combining it with hydraulic simulation to calculate flow rate, pressure and water quality. Then, it integrates anomaly weights and distribution variance to calculate distribution returns and uses backpropagation to iteratively update the parameters of the generated flow network. This achieves adaptive optimization of the control strategy, which can continuously improve the rationality of valve opening, water supply ratio and pressure setpoints while meeting physical constraints, thereby ensuring the efficiency and stability of pipeline network control.

[0120] Example 1:

[0121] To verify the feasibility of this invention in practice, it was applied to a scenario of zoned metering monitoring and control of underground pipe networks in a central urban area of ​​a city. This urban area's water supply covers a population of approximately 800,000, with a total underground pipe network length exceeding 350 kilometers and an average daily water supply of approximately 420,000 cubic meters, meeting the needs of residential, commercial, and some industrial water use. For a long time, the operation of the urban pipe network has faced several typical problems: first, uneven water metering between zones, with insufficient monitoring points in some older areas, making it difficult to detect abnormal water usage in a timely manner; second, some areas experience large pressure fluctuations and declining water quality at the end of the pipe network during morning and evening peak hours; and third, traditional manual control methods rely on experience, lacking scientific prediction and dynamic optimization methods, easily leading to increased energy consumption and water supply service complaints.

[0122] In this embodiment, the urban pipe network is divided into 12 zones according to the pipe segment connections. Each zone is equipped with flow meters, pressure sensors, water quality monitoring sensors (mainly monitoring residual chlorine and turbidity), and valve status acquisition devices. All monitoring points upload data to the monitoring center in real time via an IoT platform. The data acquisition frequency is once per minute, and the total outflow and pressure data are simultaneously collected at the upstream pumping station to form unified water supply capacity data.

[0123] First, the zonal monitoring module completes the raw data acquisition. Considering the potential clock deviations between different channels in multi-source data, the system adopts a unified clock synchronization protocol to align all sensors. Simultaneously, noise suppression and missing value imputation are performed on the acquired data; for example, median filtering is used to smooth sudden abnormal peaks, and the average of adjacent time periods is used to replace missing points. Finally, a normalization method is used to map data of different dimensions such as flow rate, pressure, and water quality to a unified numerical range, resulting in standardized zonal monitoring data.

[0124] Subsequently, standardized data is input into an improved TimesNet model for modeling. This model not only incorporates multi-scale convolutional kernels to extract features at different time granularities but also introduces a topological constraint attention mechanism and physically consistent residual units, enabling the prediction results to simultaneously reflect the pipeline network topology and hydraulic constraints. During operation, the model predicts the zone flow and pressure for the next hour and outputs the prediction sequence. By comparing the prediction results with actual data, the system calculates the deviation value and, combined with sliding time window statistical features and cluster analysis, generates zone anomaly detection results. This process allows the system to detect anomalies in certain zones in real time, such as pipeline leaks, localized water supply shortages, or water quality deterioration.

[0125] During the regulation phase, zonal prediction data, zonal anomaly detection results, and upstream water supply capacity data are input into an improved generative flow network reinforcement learning model. This model maintains the coupling between zonals through topological encoding and compresses the regulation variance of anomaly zonals through risk calibration, ensuring priority is given to maintaining stable water supply when anomalies occur. Simultaneously, it uses constraint sampling to eliminate invalid solutions that do not meet valve opening, pressure limits, or water supply capacity boundaries. Finally, the model outputs a zonal regulation distribution that satisfies both physical and operational constraints. Based on this, the system performs sampling and feedback iterations to generate the final zonal regulation parameters, including valve opening adjustments, water supply allocation ratios, and pressure setpoints.

[0126] Table 1. Comparison of the control effects of traditional manual control and the system of this invention (partial division)

[0127] ;

[0128] Table 1 shows the results of traditional manual control on the left and the results of the system of this invention on the right. As can be seen from the data in Table 1, the present invention performs superiorly in several key indicators. Taking zone 6 as an example, under traditional control, pressure fluctuations reached 41 kPa, residual chlorine concentration was only 0.17 mg / L, and there were as many as 10 user complaints. After implementing the system of this invention, pressure fluctuations decreased to 26 kPa, residual chlorine increased to 0.25 mg / L, the number of complaints decreased to 3, and energy consumption was reduced by about 4%. Zone 9 performed reasonably well under the traditional mode, but still experienced unstable pressure and 5 user complaints. Under the mode of this invention, the residual chlorine concentration significantly increased to 0.29 mg / L, the number of complaints decreased to 1, and energy consumption decreased by 25 kWh, indicating that the reinforcement learning model can achieve energy consumption optimization while ensuring water supply quality.

[0129] Further analysis reveals that this invention comprehensively improves various characteristics. First, pressure stability is significantly improved. Under traditional methods, pressure fluctuations in all four zones exceeded 37 kPa, with some zones reaching over 40 kPa. In contrast, under this invention's system, fluctuations are controlled within 25 kPa, a reduction of over 30%. Second, water quality remains more stable. Under traditional methods, residual chlorine concentrations are often at the edge of the 0.18-0.21 mg / L range, posing a risk of substandard water quality. This invention stabilizes the concentration in the 0.25-0.29 mg / L range, closer to the ideal value. Third, user experience is significantly improved. The number of complaints across the four zones decreased from 31 to 9, a reduction of over 70%, demonstrating that this invention effectively alleviates user concerns caused by unstable pressure or water quality issues. Finally, energy consumption is optimized. Under traditional methods, the total energy consumption across the four zones is 1720 kWh, while the total energy consumption under this invention is 1638 kWh, resulting in an overall energy saving of approximately 4.8%, with even more significant energy-saving potential in large-scale, long-term operation.

[0130] Furthermore, this invention also demonstrates unique advantages in its operational mode. Traditional manual control relies on operators adjusting valve openings based on experience, with response times typically ranging from tens of minutes to several hours. In contrast, this invention's system updates the predicted and controlled distribution every minute, achieving near real-time response and significantly reducing adjustment latency. During the operation of zone 4, a sudden increase in flow occurred at 19:30. This invention's system completed the adjustment of pressure settings and valve regulation in just two sampling cycles (approximately 2 minutes), stabilizing the zone pressure again. In contrast, in the traditional mode, manual intervention was not completed until 20:00, during which time user complaints increased significantly.

[0131] From an application perspective, this invention not only improves the accuracy of monitoring and prediction but also enables dynamic regulation and optimization based on a reinforcement learning model. In particular, through a risk calibration mechanism, this invention prioritizes the stability of water volume and pressure in abnormal zones under abnormal conditions, while maintaining regular optimization for normal zones, thus ensuring overall water supply safety. The constrained sampling mechanism prevents regulation parameters from exceeding operational limits, improving the feasibility of the regulation results.

[0132] In summary, this embodiment fully verifies the feasibility and superiority of the present invention in actual urban underground pipe network scenarios. Through IoT data acquisition, improved TimesNet modeling, and improved generative flow network reinforcement learning control, the present invention achieves high accuracy in zonal monitoring, sensitive prediction and anomaly identification, and adaptive optimization of control strategies. Its beneficial effects are reflected in: pressure stability improved by more than 30%, residual chlorine concentration in water quality increased by approximately 0.07 mg / L, user complaints reduced by more than 70%, energy consumption reduced by approximately 5%, and overall water supply stability and energy efficiency significantly enhanced, providing a practical and feasible technical path for the refined management of smart water affairs.

[0133] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A zoned metering monitoring and control system for underground pipeline networks based on the Internet of Things, characterized in that, Includes the following modules: The zone monitoring module is used to divide the underground pipe network into multiple zones, collect zone flow, pressure, water quality data and valve status data, and collect upstream flow and pressure data to generate zone monitoring data and water supply capacity data. The data preprocessing module is used to preprocess the partition monitoring data and generate standardized partition monitoring data. The temporal feature modeling module is used to extract multi-scale temporal features from standardized partition monitoring data using an improved TimesNet model and perform time-dependent modeling, outputting partition prediction data. The anomaly detection module is used to calculate partition deviation data based on partition prediction data and standardized partition monitoring data, extract partition anomaly data, and generate partition anomaly detection results. The regulation distribution generation module is used to input the zonal prediction data, zonal anomaly discrimination results and water supply capacity data into the improved generative flow network reinforcement learning model, perform topology coding, risk calibration and constraint sampling, and output the zonal regulation distribution. The regulation parameter generation module is used to obtain the partition regulation parameters based on the partition regulation distribution sampling, and update the parameters of the improved generative flow network reinforcement learning model based on feedback. The control module is used to adjust valve opening, water supply distribution ratio and pressure setpoint using zone control parameters. The improved TimesNet model includes an input encoding unit, an adaptive multi-scale convolutional branch, a topologically constrained temporal attention unit, a physically consistent residual unit, an anomaly detection unit, and an output mapping unit. The input encoding unit concatenates the zone flow rate sequence, zone pressure sequence, zone water quality sequence and valve status sequence in the channel dimension, and uses the water supply capacity data as the adjustment vector, multiplying it element by element with the concatenated input tensor to obtain the adjusted input representation; The adaptive multi-scale convolution branch dynamically sets the kernel size and dilation rate according to the adjustment vector, performs multi-scale convolution operations on the input representation in parallel, and outputs a multi-scale temporal feature matrix. The topology-constrained temporal attention unit constructs an adjacency weight matrix using pipe segment connection relationships, performs weighted calculations on the multi-scale temporal feature matrix, and outputs a temporal feature representation of topology constraints. The physical consistency residual unit compares the time feature representation with the physical constraint results based on the relationship between flow rate and pressure difference to obtain the residual vector, and then adds the residual vector to the time feature representation element by element to obtain the physical consistency enhancement feature; The anomaly perception unit generates an anomaly indication vector through a reconstruction mechanism, performs gating and filtering of physical consistency enhancement features, and outputs the feature representation after anomaly perception. The output mapping unit takes the feature representation after anomaly detection and inputs it into the fully connected layer, performs linear transformation and normalization operations, and outputs a fixed-dimensional vector as partition prediction data.

2. The IoT-based underground pipeline network zoned metering monitoring and control system according to claim 1, characterized in that, The process of dividing the underground pipe network into multiple zones, collecting zone flow, pressure, water quality data, and valve status data, and collecting upstream flow and pressure data to generate zone monitoring data and water supply capacity data specifically includes: Establish the underground pipeline network topology and divide the area according to the pipe segment connection method; Collect zone monitoring data, which includes instantaneous and cumulative flow of inlet and outlet water pipe sections, real-time pressure and pressure fluctuation amplitude of pipe section nodes, turbidity, residual chlorine and conductivity of water samples in the pipe section, and valve opening percentage and switching status. Collect water supply capacity data, which includes the total outflow and water supply pressure of upstream water plants or pumping stations.

3. The IoT-based underground pipeline network zoned metering monitoring and control system according to claim 1, characterized in that, The preprocessing of the zonal monitoring data specifically includes: Time synchronization processing is performed on the zone monitoring data, and the collection timestamps of flow rate, pressure, water quality and valve status information are aligned according to a unified time base. Noise suppression is achieved by using filtering methods on the zoned monitoring data; For the zonal monitoring data, interpolation methods or the average of adjacent time periods are used to replace missing data to fill in missing values. The monitoring data of different zones are normalized to map the values ​​of different dimensions to a unified range.

4. The Internet of Things-based underground pipeline network zoning metering monitoring and control system according to claim 1, characterized in that, The process of calculating partition deviation data based on partition prediction data and standardized partition monitoring data, extracting partition anomaly data, and generating partition anomaly discrimination results specifically includes: Subtracting the zonal prediction data from the standardized zonal monitoring data point by point along the time dimension yields the flow deviation sequence, pressure deviation sequence, water quality deviation sequence, and valve status deviation sequence, which are then concatenated along the feature dimension to form the zonal deviation data. Within a sliding time window, the mean, standard deviation, and difference scores between adjacent time points are calculated for the partitioned deviation data, and anomaly candidate feature vectors containing stability, volatility, and abrupt change characteristics are constructed. K-means clustering is used on the abnormal candidate feature vectors to divide all feature vectors into normal clusters and abnormal clusters, and the distance between each sample and the cluster center is calculated using Euclidean distance. Quantile statistics are performed on the distance distribution to determine the dynamic threshold interval. When the distance of a certain deviation point is greater than the upper quantile threshold, it is marked as an anomaly. When it is in the upper or lower neighborhood of the threshold, a second comparison is performed. Points that continuously exceed the window number limit are uniformly marked as anomalies, thus obtaining partitioned anomaly data. The partition anomaly data is input into the multilayer perceptron, and linear mapping, nonlinear activation and normalization operations are performed in sequence to output the partition anomaly discrimination result.

5. The Internet of Things-based underground pipeline network zoning metering monitoring and control system according to claim 1, characterized in that, The improved generative stream network reinforcement learning model specifically includes: The improved generative flow network reinforcement learning model includes a topology coding unit, a conditional flow transformation branch, a risk calibration unit, and a constraint sampling unit. The topology coding unit concatenates the partition prediction data, partition anomaly discrimination results and water supply capacity data to form a joint state vector, performs normalization processing on the joint state vector, uses the underground pipe network directed adjacency matrix to perform matrix multiplication on the joint state vector to generate topology propagation results, and obtains the topology embedding representation by element-wise addition and normalization. The conditional flow transformation branch receives the topological embedding representation as conditional input, performs invertible affine transformations on the latent variables layer by layer, including element-wise scaling and element-wise translation, and connects nonlinear activation and batch normalization operations after each layer transformation to output the candidate partition control distribution. The risk calibration unit generates a risk weight vector based on the partition anomaly discrimination result, multiplies the variance vector of the candidate partition control distribution element by element with the risk weight vector, performs numerical compression on the variance elements of partitions identified as anomalous, keeps the original values ​​of the variance elements of partitions identified as normal, and renormalizes the variance vector to obtain the candidate distribution after risk calibration. The constraint sampling unit sets upper and lower limits for zone flow, pressure and valve opening based on water supply capacity data. It performs random sampling from the candidate distribution after risk calibration to obtain an initial control sample set. It performs constraint tests on the initial samples point by point, removes samples that exceed the upper and lower limits, and performs L2 projection or boundary truncation correction on the samples that fail to pass, and outputs the zone control distribution that meets the constraint conditions.

6. The Internet of Things-based underground pipeline network zoning metering monitoring and control system according to claim 1, characterized in that, The step of obtaining partition control parameters based on partition control distribution sampling and updating the parameters of the improved generative flow network reinforcement learning model based on feedback specifically includes: Random sampling is performed from the zone control distribution to generate multiple zone control parameter sets according to the preset sampling quantity. Each zone control parameter set includes valve opening, water supply distribution ratio and pressure set value. The control parameter sets of each zone and the underground pipe network topology data are input into the calculation process. The flow rate of each node is calculated according to the node flow balance equation, the pressure of each node is calculated according to the node energy equation, the head loss of each pipe segment is calculated according to the pipe segment resistance formula, and the decay of water quality concentration of each pipe segment over time is calculated according to the water quality transmission model. At the node, the water quality concentration is mixed according to the flow weighted average rule to obtain the simulated flow rate data, simulated pressure data and simulated water quality data at each time step. For each set of zonal control parameters, calculate the point-by-point difference between the simulated flow data, simulated pressure data, simulated water quality data and preset target data, generate residual signals, call the energy consumption calculation module to obtain the corresponding energy consumption data, call the zonal anomaly discrimination results to generate anomaly weights, call the zonal control distribution output to output the corresponding distribution variance, and calculate the single-step distribution reward based on the residual signals, energy consumption data, anomaly weights and distribution variances. In the time dimension, the single-step distributed reward of all time steps is accumulated, the residual signal of all time steps is accumulated, and the product of the anomaly weight and the distribution variance of all time steps is accumulated to generate trajectory-level distributed reward. The gradient signal is obtained by performing element-wise multiplication between the trajectory-level distribution return and the logarithmic probability of the partitioned control distribution. The gradient signal is input into the backpropagation process to iteratively update the layer-by-layer affine transformation parameters, scaling parameters, translation parameters, and nonlinear activation function parameters in the improved generative flow network reinforcement learning model.

Images