Method and system for centralized reading management of water meters in water affairs

Abstract

The invention relates to the technical field of intelligent water affairs, and discloses a water meter centralized reading management method and system.The water meter centralized reading management method for the water affairs comprises the steps that a multi-layer space-time diagram model is constructed, and a water meter centralized reading network is abstracted into a multi-layer space-time diagram structure; extracting network topology features by applying a graph neural network, wherein the network topology features comprise processing a spatial dependency relationship by applying a graph convolutional network and processing a time dependency relationship by applying a time convolutional network; calculating attention weights among the nodes, and modeling an interdependence relationship among the nodes; executing multi-scale network state prediction, and predicting the future state of the network on a plurality of time scales; an abnormal event response mechanism is realized, and emergencies in a water meter centralized reading network are detected and coped with; according to the invention, the adaptive adjustment of the network topology is realized through the multi-level space-time diagram model, the network structure can be dynamically optimized according to the actual operation environment change and the network load condition, and the data packet loss rate is effectively reduced.

Description

Technical Field

[0001] The present invention relates to the field of smart water technology, and more specifically, to a method and system for managing water meter reading in water services. Background Art

[0002] With the rapid development of smart city construction, the level of intelligent water management continues to improve. As a key component of smart water management, water meter centralized reading systems enable real-time monitoring and management of water consumption through automated data collection and transmission. However, existing water meter centralized reading management technologies still have many technical flaws and shortcomings in practical applications.

[0003] In existing technologies, centralized water meter reading networks typically use a fixed topology and static routing strategies. This traditional approach, once deployed, is difficult to dynamically adjust to changes in the actual operating environment and network load. Specifically, existing technologies have the following major issues: First, the network topology lacks adaptive capabilities. Once the traditional water meter reading system is deployed, its network topology is basically fixed. During peak water usage, some network nodes are under excessive data transmission pressure, resulting in a data packet loss rate of up to 15%-25%, seriously affecting the accuracy and completeness of data collection. At the same time, due to the use of fixed communication paths, some nodes in the network consume energy too quickly, while other nodes have low energy utilization rates, resulting in serious imbalance in energy consumption between network nodes and reducing the service life of the entire network; secondly, there is a lack of the ability to predict the evolution trend of network topology. In the urban water environment, the network status will change due to a variety of factors, including changes in seasonal water use patterns, changes in the communication environment caused by urban construction, and performance degradation caused by equipment aging. Existing technologies can only passively adapt to these changes and cannot predict and respond in advance. The lack of corresponding preparations leads to a significant lag in network adaptability, affecting the reliability and stability of data collection; secondly, the spatiotemporal dependency modeling is not accurate enough. When dealing with the spatiotemporal dependencies in the water meter reading network, the existing technology usually adopts a simplified processing method with unified weights, which does not fully consider the importance differences of characteristics at different spatiotemporal scales. This processing method ignores the changing characteristics of the network topology at different time scales, and cannot accurately capture and model complex spatiotemporal dependencies, resulting in limited network optimization effects; finally, the emergency response capability is insufficient. For the drastic changes in network topology caused by emergencies (such as pipeline bursts, equipment failures, communication interference, etc.), the existing technology reacts slowly and cannot adjust the network configuration and routing strategy in time. This lag causes the network to be unable to quickly restore normal operation when an emergency occurs, affecting the continuity and reliability of data collection.

[0004] Furthermore, while some existing technical solutions attempt to incorporate machine learning methods to optimize network management, most are limited to single-time-scale analysis and lack comprehensive consideration of multi-scale spatiotemporal characteristics. When dealing with complex urban water management environments, these methods often fail to fully exploit the spatiotemporal evolution of network topology, resulting in suboptimal optimization results.

[0005] Therefore, there is an urgent need for a water meter centralized reading management method that can dynamically adapt to changes in the network environment, has multi-scale prediction capabilities, can accurately model spatiotemporal dependencies, and quickly respond to emergencies, so as to improve the intelligence level and operational efficiency of water management. Summary of the Invention

[0006] The present invention provides a method and system for water meter centralized reading management in water affairs, which solves the technical problems in related technologies such as fixed water meter centralized reading network topology, lack of prediction ability, inaccurate spatiotemporal dependency modeling, and slow response to emergencies.

[0007] The present invention provides a method for managing water meter reading in water services, comprising: Construct a multi-level spatiotemporal graph model and abstract the water meter reading network into a multi-level spatiotemporal graph structure; Based on the multi-level spatiotemporal graph structure, graph neural networks are used to extract network topology features, including the use of graph convolutional networks to process spatial dependencies and temporal convolutional networks to process temporal dependencies. Based on the network topology characteristics, the attention weights between nodes are calculated to model the interdependence between nodes; Based on the attention weights between nodes, multi-scale network state prediction is performed to predict the future state of the network at multiple time scales; Based on the output of multi-scale network status prediction, an abnormal event response mechanism is implemented to detect and respond to emergencies in the water meter reading network.

[0008] Furthermore, the step of constructing a multi-level spatiotemporal graph model includes: Number all nodes in the water meter centralized reading network to form a node set; Based on the communication connection relationship between nodes, a time-varying edge set is constructed; Convert the edge set into a time-varying adjacency matrix; Construct a multi-level spatiotemporal graph, where each layer represents the network topology characteristics at different spatiotemporal scales; Establish inter-layer connections and connect the same nodes at different levels through cross-layer edge sets.

[0009] Furthermore, the step of applying graph neural network to extract network topology features includes: For each layer in the multi-level spatiotemporal graph, a graph convolutional network is applied to extract topological features; Combined with time convolutional networks to capture the evolution of network topology over time; For cross-layer connections in multi-level spatiotemporal graphs, a cross-layer feature fusion algorithm is applied to integrate information at different spatiotemporal scales; Generate a node feature matrix that reflects the complete spatiotemporal characteristics of the water meter reading network.

[0010] Furthermore, the step of calculating the attention weights between nodes includes: Based on the extracted node feature matrix, the graph attention network is applied to calculate the attention weights between nodes; Apply a multi-head attention mechanism to simultaneously calculate multiple sets of independent attention weights and then merge the results; For different levels in the multi-level spatiotemporal graph, attention weights are calculated separately to form a hierarchical set of attention weights; According to the attention weights, a weighted adjacency matrix is ​​generated.

[0011] Furthermore, the step of performing multi-scale network state prediction includes: Build a multi-scale prediction model to simultaneously predict network status changes at four time scales: hourly, daily, weekly, and monthly; For prediction tasks at different time scales, a hierarchical time series prediction architecture is adopted; Introducing a multi-task learning framework to jointly optimize prediction tasks at four time scales; Based on the prediction results, a prediction-driven network parameter optimization strategy is calculated.

[0012] Furthermore, the steps of implementing the abnormal event response mechanism include: Build an abnormal event detector to identify abnormal events by monitoring sudden changes in network topology; Rapidly locate and assess the impact of detected abnormal events; Design a rapid response strategy based on the anomaly-affected area and prioritize adjusting network parameters in the affected area; For different types of abnormal events, special processing strategies are applied.

[0013] Furthermore, the graph convolutional network includes: Input layer, which receives the initial features of the nodes; Multiple graph convolutional layers update node representations by aggregating neighbor node information; Output layer, generates the final node feature representation; The temporal convolutional network includes a causal convolutional layer, a dilated convolutional layer and a residual connection.

[0014] Furthermore, the graph attention network includes: Feature transformation layer, which performs linear transformation on input node features; Attention calculation layer, which calculates the attention coefficient between node pairs; Feature aggregation layer, which aggregates neighbor node information based on weighted attention coefficients; The multi-head attention mechanism simultaneously calculates 8 to 16 attention heads, each of which focuses on a different feature subspace.

[0015] Furthermore, the hierarchical time series prediction architecture includes: The short-term prediction layer uses gated recurrent units to predict hourly network state changes; The medium-term prediction layer uses long short-term memory networks to predict daily and weekly network state changes; The long-term prediction layer uses an attention-enhanced temporal convolutional network to predict monthly network state changes; The prediction-driven network parameter optimization strategy includes routing strategy optimization, energy management optimization and communication frequency optimization.

[0016] The present invention provides a water meter centralized reading management system for water services, which is used to execute the above-mentioned water meter centralized reading management method for water services, including: A multi-level spatiotemporal graph construction module is used to abstract the water meter reading network into a multi-level spatiotemporal graph structure; Graph neural network feature extraction module, used to extract network topology features; Attention weight calculation module, used to model the interdependence between nodes; Multi-scale network state prediction module, used to predict the future state of the network at multiple time scales; The abnormal event response module is used to detect and respond to emergencies in the water meter reading network.

[0017] The beneficial effects of the present invention are: adaptive adjustment of network topology is achieved through a multi-level spatiotemporal graph model, the network structure can be dynamically optimized according to actual operating environment changes and network load conditions, and the data packet loss rate can be effectively reduced; It has the ability to predict network status at multiple scales, predicting changes in network status at different time scales. It can respond to various changes in advance through prediction-driven optimization strategies, avoiding performance degradation caused by passive adaptation. The introduction of the attention mechanism can automatically learn and quantify the influence weights between network nodes, accurately capturing the importance differences of features at different spatiotemporal scales; We have built an event-driven rapid response system that can quickly detect abnormal events, accurately locate the scope of impact, and take targeted response measures; Through intelligent routing strategies and load balancing mechanisms, the problem of uneven energy consumption between nodes is improved, the network service life is extended, and the overall system throughput is improved, data transmission efficiency is increased, and operation and maintenance costs are reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a flow chart of a method for centralized water meter reading management in water affairs according to the present invention; Figure 2 It is a line graph comparing the data packet loss rate of the water meter centralized reading network before and after applying the method of the present invention at different time points; Figure 3 It is a bar graph comparing the data transmission success rate before and after applying the method of the present invention under different network load conditions; Figure 4 This is a radar chart comparing the performance of the solution of the present invention and the traditional water meter centralized reading management solution; Figure 5 It is a pie chart showing the proportion of nodes with different energy consumption levels in the network before and after applying the method of the present invention; Figure 6 It is a line graph of the accuracy of network status prediction at different time scales. DETAILED DESCRIPTION

[0019] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0020] At least one embodiment of the present invention discloses a method for managing water meter reading in water services, such as Figure 1 Shown, including: Step 1: Construct a multi-level spatiotemporal graph model and abstract the water meter reading network into a multi-level spatiotemporal graph structure. This step uses a multi-level spatiotemporal graph construction algorithm to abstract the water meter reading network into a multi-level spatiotemporal graph structure to comprehensively capture the spatiotemporal characteristics of the network topology.

[0021] Step 1.1, build a node set; According to one embodiment of the present application, all nodes in the water meter collection network (including smart water meters, data collectors, relay nodes and data aggregators, etc.) are first numbered to form a node set. ,Each node has attribute information such as location coordinates, ,functional type, and energy status.

[0022] Step 1.2, construct a time-varying edge set; In addition, based on the communication connection relationship between nodes, a time-varying edge set is constructed When there is a communication connection between two nodes, at the corresponding time Create an edge and record edge attributes such as communication quality and data transmission rate.

[0023] Step 1.3, forming a time-varying adjacency matrix; Set the edges Convert to a time-varying adjacency matrix , where the matrix elements Representation node With node In time The connection strength can be a binary value (0 or 1) indicating whether it is connected, or a continuous value indicating the connection strength.

[0024] In some embodiments, the adjacency matrix It can be expanded into a multi-dimensional tensor, containing multiple types of connection relationships. For example, multiple connection types such as physical connection, logical connection and data flow connection can be considered at the same time to form a multi-relation adjacency tensor. ,in Indicates the relationship type.

[0025] Step 1.4, construct a multi-level space-time graph; The expression for constructing a multi-level space-time graph is: ; in, represents a multi-level spatiotemporal graph collection, 、 、 Represents the first layer, the second layer, and the Layer space-time diagram, is the total number of layers of the space-time graph.

[0026] Each layer of the spatiotemporal graph represents the network topology characteristics at different spatiotemporal scales: : hourly spatiotemporal graph, capturing network topology changes in the short term; : Daily-level spatiotemporal graph, capturing the network topology change pattern within a day; : Weekly spatiotemporal graph, capturing the periodic changes in network topology; : Monthly-level spatiotemporal graph, capturing long-term network topology evolution trends.

[0027] Optionally, the division of spatiotemporal scales can be adjusted according to the needs of specific application scenarios. For example, a minute-level spatiotemporal map can be added in areas with drastic changes in water consumption, or a quarter-level spatiotemporal map can be added in areas with obvious seasonal changes.

[0028] Step 1.5, forming a complete multi-level space-time graph; Establish inter-layer connections, connect the same nodes at different levels, and form a complete multi-level space-time graph , where the time-varying graph structure is used to model the characteristics of network topology changing over time: ; in Indicates time The time-varying graph structure of Represents a collection of nodes, Indicates time The edge set of Indicates time The adjacency matrix of .

[0029] In some embodiments, cross-layer connections may adopt a non-uniform connection strategy, assigning different cross-layer connection weights to different nodes based on the importance or activity of the nodes, thereby enabling key nodes to transmit information more effectively at different time scales.

[0030] Through the above sub-steps, this application abstracts the complex water meter reading network into a mathematical, multi-layered spatiotemporal graph model, laying the foundation for subsequent analysis and optimization based on graph neural networks. It should be noted that this multi-layered spatiotemporal graph model can comprehensively capture the evolution of network topology at different time scales, providing a solid theoretical foundation for accurate network status prediction and intelligent management.

[0031] Step 2: Based on the multi-level spatiotemporal graph structure, apply graph neural networks to extract network topology features, including applying graph convolutional networks to process spatial dependencies and temporal convolutional networks to process temporal dependencies; This step applies the graph neural network to the multi-level spatiotemporal graph constructed in step 1 to extract the topological features of the water meter reading network for subsequent network status analysis and optimization.

[0032] Step 2.1, extract topological features; According to one embodiment of the present application, a graph convolutional network (GCN) is applied to extract topological features for each layer in a multi-layer spatiotemporal graph. The mathematical expression of the graph convolution operation is: ; in Indicates the The node feature matrix of the layer; Indicates the The node feature matrix of the layer, that is, the output feature after a graph convolution operation; yes The degree matrix of is a diagonal matrix, whose diagonal elements are the degrees of the corresponding nodes; Degree matrix The negative power of half is used for normalization operation; It is The trainable weight matrix of the layer; is the activation function; is the adjacency matrix with self-connection added, expressed as: ; in is the original adjacency matrix, is the identity matrix; Optionally, in some implementations, different graph convolution variants, such as GraphSAGE or GraphIsomorphismNetwork (GIN), can be used to adapt to different types of network topologies. For example, when the node degree distribution in the network is extremely uneven, the sampling aggregation method in GraphSAGE can be used to avoid excessive influence of high-degree nodes on the convolution operation.

[0033] The graph convolutional network used in this application includes an input layer, multiple graph convolutional layers, and an output layer. Among them: The input layer receives the initial features of the nodes, such as node type, location, communication capabilities and other attributes; The graph convolution layer updates the node representation by aggregating neighbor node information, and each layer contains trainable weight parameters; The output layer generates the final node feature representation.

[0034] In the water meter reading network scenario, graph convolutional networks are particularly suitable for capturing the spatial dependencies between water meter nodes. For example, they can identify that some water meter nodes may have similar data transmission patterns due to their geographical proximity.

[0035] Step 2.2, capturing the evolutionary pattern of network topology over time; In addition, a temporal convolutional network (TCN) is combined to capture the evolution of network topology over time. The temporal convolution operation applies a one-dimensional convolution on the time dimension, processing node feature sequences containing historical information and generating feature representations that take into account time dependencies.

[0036] The temporal convolutional network structure used in this application includes: Causal convolutional layer: ensures the model does not use “future” information; Dilated convolutional layer: captures long-range temporal dependencies by increasing the receptive field; Residual connection: helps train deeper networks and alleviates the vanishing gradient problem.

[0037] In some embodiments, different time convolution parameters can be configured for different levels of spatiotemporal graphs according to the characteristics of different time scales. For example, for hourly spatiotemporal graphs, smaller convolution kernels and smaller dilation rates can be used to capture fine-grained temporal patterns; while for monthly spatiotemporal graphs, larger convolution kernels and larger dilation rates can be used to capture long-term temporal dependencies.

[0038] In practical applications, the time convolution network can learn seasonal patterns (such as water usage peak periods) and periodic changes (such as differences between weekdays and weekends) from historical data, thereby predicting possible future network load changes.

[0039] Step 2.3, integrate information of different spatiotemporal scales; In addition, for cross-layer connections in multi-level spatiotemporal graphs, a cross-layer feature fusion algorithm is applied to integrate information of different spatiotemporal scales. The specific implementation is to design a cross-layer aggregation function to weight and fuse node features of different levels.

[0040] Optionally, the cross-layer aggregation function can take different forms of implementation, such as weighted average, attention weighting or gating mechanism, etc. In some embodiments, an adaptive weight learning mechanism can be introduced to dynamically adjust the fusion weights of features of different levels according to the current network state, so that the system can better adapt to the dynamic changes of the network environment.

[0041] Step 2.4, generate node feature matrix; Therefore, through the above processing, a node feature matrix reflecting the complete spatiotemporal characteristics of the water meter collection network is generated , where each row corresponds to a network node and each column corresponds to a feature dimension. These features include node communication patterns, load characteristics, energy consumption characteristics, and other key information.

[0042] It should be understood that through this step, the system of the present application can learn the communication patterns and load characteristics between network nodes, providing data support for subsequent network optimization. It should be noted that the application of graph neural networks enables the system to automatically extract key features contained in complex network topologies without the need for manual design of feature extraction rules.

[0043] Step 3, based on the network topology features, calculate the attention weights between nodes and model the interdependence between nodes; This step uses attention mechanism to calculate the influence weight between each node in the water meter set copying network, which is used to accurately model the mutual dependence between nodes.

[0044] Step 3.1, calculate the attention weight between nodes; According to an embodiment of the present application, based on the node feature matrix extracted in step 2 , the graph attention network (GAT) is applied to calculate the attention weight between nodes. For each pair of connected nodes in the network and , the attention weight calculation formula is: ; Where represents the attention weight of node to node , and represent the source node and the target node respectively; , and represent the feature vectors of nodes , and respectively; is a trainable weight matrix used for linear transformation of node features; is an attention vector used to calculate the similarity score between node pairs; represents the vector concatenation operation, which connects two vectors into a longer vector; represents the neighbor node set of node , represents the center node; represents the LeakyReLU activation function, which is used to introduce nonlinearity and avoid gradient vanishing; represents the exponential function, which is used to convert the score to a positive value; represents the sum of all neighbor nodes of node , which is used for normalization; represents the vector transpose operation; Optionally, in some embodiments, the attention calculation can consider the feature information of the edge, which is modified to include the edge feature vector containing communication quality, bandwidth, stability and other attributes. This edge-enhanced attention mechanism can more comprehensively consider the communication characteristics between nodes, further improving the accuracy of the attention weight.

[0045] The graph attention network structure adopted in the present application mainly includes the following components: Feature transformation layer: performs linear transformation on input node features to enhance expression capabilities; Attention calculation layer: calculates the attention coefficient between node pairs; Feature aggregation layer: Aggregate neighbor node information based on weighted attention coefficients.

[0046] In water metering network scenarios, graph attention networks can identify asymmetric influences between different nodes. For example, relay nodes on critical paths have a much greater impact on the entire network than edge nodes. This asymmetric influence relationship can be automatically learned and quantified using the attention mechanism.

[0047] In addition, in order to capture different types of dependencies, a multi-head attention mechanism is applied, that is, This mechanism combines multiple sets of results by averaging and concatenating the attention weights, enabling the model to simultaneously focus on dependencies in different feature subspaces.

[0048] The multi-head attention mechanism used in this application usually sets 8 to 16 attention heads (i.e. Each attention head focuses on a different feature subspace, enabling the capture of diverse inter-node relationships. For example, in a water meter reading network, different attention heads might focus on different aspects of inter-node relationships, such as geographic distance, energy consumption, and data transmission volume.

[0049] Step 3.2: Form a hierarchical set of attention weights; Next, for different levels in the multi-level spatiotemporal graph, the attention weights are calculated separately to form a hierarchical set of attention weights.

[0050] Optionally, in some implementations, an inter-layer attention mechanism can be introduced to calculate attention weights between different layers. This inter-layer attention calculation measures the degree of association between nodes at different layers using a similarity function and normalizes the result into a probability distribution. This inter-layer attention mechanism can more finely model dependencies between different time scales, improving the system's ability to capture multi-scale temporal patterns.

[0051] Step 3.3, generate weighted adjacency matrix; Therefore, according to the attention weights, a weighted adjacency matrix is ​​generated , where the matrix elements are: ; in represents the adjacency matrix weighted by attention weights; Represents the adjacency matrix weighted by attention weights Rank Elements of the column, corresponding to nodes To Node The weight of Representation node For Node The attention weight of It should be noted that through this step, the system of this application can automatically learn the influence weights between network nodes, accurately capture the interdependencies between different nodes, and provide an important reference for subsequent network optimization. It should be understood that compared with the traditional method of using fixed weights, the attention mechanism adopted in this application can dynamically adjust the influence weights between nodes according to the actual network state, greatly improving the model's ability to express complex dependencies in network topology.

[0052] Step 4: Based on the attention weights between nodes, perform multi-scale network state prediction to predict the future state of the network at multiple time scales; Based on the features extracted and attention weights calculated in the previous steps, this step predicts the future state of the water meter reading network at multiple time scales, providing a basis for the active optimization of network parameters.

[0053] Step 4.1, construct a multi-scale prediction model; According to one embodiment of the present application, a multi-scale prediction model is constructed to simultaneously predict network status changes at four time scales: hour, day, week, and month. , build the corresponding prediction model : ; in is the time scale Future Moment The predicted network status; Indicates the current moment; is the time scale The corresponding prediction step size is, Represents a specific time scale (hourly scale is 1 hour, daily scale is 1 day, weekly scale is 1 week, and monthly scale is 1 month); Representing time scale The corresponding prediction function is, Represents a forecast model for a specific time scale; Indicates from time arrive The historical network state sequence, Indicates the length of the historical time window.

[0054] Optionally, in some implementations, the prediction model can employ a probabilistic forecasting approach, outputting a distribution of future network states rather than a single-point prediction. This probabilistic forecasting approach can estimate the uncertainty of the prediction, providing more comprehensive information for subsequent decision-making and is particularly suitable for highly dynamic and uncertain network environments.

[0055] The multi-scale prediction model of this application adopts a hierarchical design, including the following key components: Feature extractor: extracts spatiotemporal features from the original network state; Time scale encoder: encodes information of different time scales into a unified representation; Predictive decoder: Decodes and generates predictions for different time scales.

[0056] In a centralized water metering network, multi-scale prediction models can simultaneously capture short-term fluctuations (such as morning and evening peaks) and long-term trends (such as seasonal variations), enabling comprehensive predictions of network state changes. For example, the model can predict the peak water usage expected in a particular area within the next few hours, while also predicting overall shifts in water usage patterns due to seasonal changes over the next few weeks.

[0057] Step 4.2, hierarchical time series forecasting architecture; For forecasting tasks at different time scales, a hierarchical time series forecasting architecture is adopted, including: Short-term prediction layer: Uses gated recurrent units (GRUs) to predict hourly network state changes. Medium-term prediction layer: Uses Long Short-Term Memory (LSTM) networks to predict daily and weekly network state changes. Long-term prediction layer: Use attention-enhanced temporal convolutional networks to predict month-level network state changes.

[0058] In some implementations, external factors such as weather data, holiday information, and city events can be introduced as auxiliary inputs to further improve prediction accuracy. For example, an external factor encoder can be designed to encode this information into a feature vector, which is then fused with network state features and fed into the prediction model.

[0059] The Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM) structures in this application are specifically optimized to adapt to the characteristics of water meter reading network data: GRU unit: uses update gate and reset gate to control information flow, suitable for processing short-term fluctuating data; LSTM unit: contains input gate, forget gate and output gate, which can better preserve long-term dependency information; Temporal attention layer: introduces attention mechanism in long sequence prediction, focusing on key patterns in historical data.

[0060] Step 4.3, jointly optimize the prediction tasks of four time scales; In addition, in order to improve the prediction accuracy, a multi-task learning framework is introduced to jointly optimize the prediction tasks of four time scales. The loss function is designed as the weighted sum of the prediction loss of each time scale, and the weight coefficient of each time scale can be pre-set or automatically adjusted through learning.

[0061] Optionally, in some embodiments, an adaptive weight allocation strategy can be used to dynamically adjust the weight coefficients in the loss function according to the learning difficulty and current performance of different time scale prediction tasks. This adaptive weight allocation mechanism can automatically balance the optimization difficulty of different prediction tasks during training, avoiding the dominance of certain tasks in the learning process of the entire model.

[0062] The multi-task learning framework of the present application improves the overall prediction accuracy by sharing the underlying feature representation and utilizing the complementary information between different time scale tasks. For example, daily scale prediction can learn fine-grained fluctuation patterns from hourly scale prediction, while hourly scale prediction can obtain long-term trend information from monthly scale prediction.

[0063] Step 4.4, network parameter optimization strategy; Therefore, based on the prediction results , the prediction-driven network parameter optimization strategy is calculated, including: Routing strategy optimization: adjusting data transmission path according to predicted network load; Energy management optimization: adjusting node sleep and wake-up strategy according to predicted node energy consumption state; Communication frequency optimization: adjusting data acquisition and transmission frequency according to predicted data importance.

[0064] It should be understood that through this step, the system of the present application can predict the state change of the network on multiple time scales, and based on the prediction results, the network parameters can be adjusted in advance to realize prediction-driven proactive optimization, effectively responding to changes in network environment and load fluctuations. It should be noted that compared with traditional passive response optimization methods, the prediction-driven optimization method of the present application can respond to network changes in advance, greatly reducing the performance loss caused by network adaptation lag.

[0065] Step 5, based on the output of multi-scale network state prediction, an abnormal event response mechanism is realized to detect and respond to sudden events in the water meter collection network; This step builds an event-driven rapid response system to detect and respond to emergencies in the water meter reading network and achieve real-time optimization of the network topology.

[0066] Step 5.1, build an abnormal event detector; According to one embodiment of the present application, an abnormal event detector is constructed to identify abnormal events by monitoring sudden changes in network topology. The detection function is defined as: ; in Represents the detection function, used to determine the current moment and the previous moment Is there any abnormality in the network status? Indicates the current time The network diagram status, Indicates the current moment; Indicates the previous moment The network diagram status, Indicates the previous moment; and are the adjacency matrices of the current moment and the previous moment, respectively, indicating the corresponding moments; Represents the Frobenius norm of the matrix, F represents the Frobenius norm, which is used to measure the difference between the adjacency matrices at two moments; It is a preset threshold. When the difference exceeds this threshold, an abnormal event response is triggered.

[0067] Optionally, in some implementations, anomaly detection can consider a longer historical window, using a time series anomaly detection approach. This approach detects anomalies by comparing the difference between the currently observed network state and the predicted normal network state based on historical data. When the difference exceeds a preset threshold, it is considered an anomaly event. This prediction-based anomaly detection approach can capture more complex anomaly patterns, such as gradually evolving anomalies or anomalies within a cyclical pattern.

[0068] The abnormal event detection system of this application includes the following key components: Real-time monitoring module: continuously monitors changes in network topology status; Difference calculation module: calculates the topological difference measure within a continuous time window; Threshold adaptive adjustment module: dynamically adjusts the anomaly detection threshold based on historical data; Event classification module: classifies detected abnormal events into different types (such as node failure, link interruption, etc.).

[0069] In the water meter reading network scenario, the abnormal event detector can quickly identify various emergencies, such as drastic changes in network topology caused by pipe bursts and node failures caused by equipment failures, thereby triggering corresponding processing processes.

[0070] Step 5.2: Rapid location and impact assessment; For detected anomalies, we quickly locate and assess the impact. We first calculate the anomaly impact area, which is the set of nodes in the network directly affected by the anomaly. We then determine which nodes are affected by calculating the degree of change in each node's connection status. When the change in a node's connection exceeds a node-level threshold, it is marked as affected.

[0071] In some implementations, impact assessment can utilize a graph diffusion model to simulate the propagation of anomalies within a network and predict areas potentially affected by chain reactions. This approach calculates the probability of additional nodes being affected based on a known set of affected nodes. When the probability exceeds a preset threshold, these nodes are included in the potential impact range. This propagation-based impact assessment can proactively identify potential risk areas and enable more proactive anomaly response.

[0072] This application's impact assessment algorithm, based on a graph propagation model, determines the potential impact range by calculating the multi-hop neighbors of an anomaly node. For anomalies of varying severity, the system uses impact assessments with varying radiuses to ensure accurate capture of the diffusion effect of anomalies.

[0073] Step 5.3, rapid response strategy; Next, based on the affected area, a rapid response strategy is designed to prioritize adjusting network parameters in the affected area. The response process includes: Emergency routing reconstruction: building temporary communication paths for affected areas; Emergency resource dispatch: temporarily deploy communication resources from surrounding areas to the affected area; Data collection strategy adjustment: Adjust data collection frequency and priority based on abnormal situations.

[0074] Alternatively, in some implementations, a reinforcement learning-based response strategy generation method can be employed to learn the optimal response strategy through interaction with the environment. This method uses the current network state as the state input and possible response actions as the action space, selecting the optimal strategy by evaluating the long-term benefits of different actions. This reinforcement learning method can learn the optimal response strategy for different types of anomalies through continuous trial and error.

[0075] The rapid response system in this application uses a priority scheduling mechanism to ensure that the most critical affected nodes receive priority resource support. At the same time, the system maintains a backup resource pool for quickly deploying additional resources when abnormal events occur, improving the network's resilience.

[0076] Step 5.4, apply the processing strategy; For different types of abnormal events, apply specialized handling strategies: Node failure event: data transmission is guaranteed through redundant paths and load balancing strategies; Communication interference events: Avoid interference by dynamically adjusting communication frequency and power; Network congestion events: Alleviate congestion through flow control and priority queuing mechanisms; Security threat events: Ensure network security through secure communication channels and abnormal traffic isolation.

[0077] In some implementations, the system can establish an abnormal event knowledge base to record historical abnormal events and their effective handling solutions. When encountering similar abnormal situations, it can quickly retrieve and apply proven effective response strategies to achieve knowledge-driven abnormal handling.

[0078] The exception handling strategy library of this application contains processing templates for various types of abnormal events. The system will select the most suitable processing solution from the strategy library based on the abnormal type detected in real time, and adjust the parameters according to the specific situation to realize intelligent processing of abnormal events.

[0079] It should be noted that, through this step, the system of the present application can trigger a rapid re-optimization process when a sudden change in the network topology is detected, prioritize the adjustment of the affected areas, achieve real-time response to emergencies, and ensure the stable operation of the water meter reading network under various abnormal conditions. It should be understood that compared with traditional systems, the abnormal event response mechanism of the present application can not only quickly detect anomalies, but also accurately locate the affected areas and take targeted measures, greatly improving the system's ability to respond to emergencies.

[0080] A water meter centralized reading management system for water services, used to execute the above-mentioned water meter centralized reading management method for water services, comprising: A multi-level spatiotemporal graph construction module is used to abstract the water meter reading network into a multi-level spatiotemporal graph structure; Graph neural network feature extraction module, used to extract network topology features; Attention weight calculation module, used to model the interdependence between nodes; Multi-scale network state prediction module, used to predict the future state of the network at multiple time scales; The abnormal event response module is used to detect and respond to emergencies in the water meter reading network.

[0081] like Figures 2 to 6 As shown, respectively, are a line graph comparing the data packet loss rate of the water meter centralized reading network before and after the application of the method of the present invention at different time points; a bar graph comparing the data transmission success rate before and after the application of the method of the present invention under different network load conditions; a radar chart comparing the performance of the solution of the present invention and the traditional water meter centralized reading management solution; a pie chart showing the proportion of nodes with different energy consumption levels in the network before and after the application of the method of the present invention (the inner circle is the proportion of nodes with different energy consumption levels in the network before the application of the method of the present invention, and the outer circle is the proportion of nodes with different energy consumption levels in the network after the application of the method of the present invention); and a line graph showing the accuracy of network status prediction at different time scales.

[0082] Here, the present invention provides an implementation example: The application scenario of this implementation example is a water meter reading network in the central area of ​​a city, covering an area of ​​approximately 20 square kilometers, including more than 15,000 smart water meter nodes, 200 data collectors, 50 relay nodes, and 10 data aggregators. This area has the following characteristics: High building density, complex communication environment, and multiple interference sources; Water use patterns are diverse, including residential, commercial, and some industrial areas; Seasonal water use varies significantly, with summer water use increasing by approximately 40% compared to winter; There are frequent municipal construction projects in the area, which causes frequent changes in the communication environment. Before applying this application method, the water meter reading network in this area faced problems such as high data packet loss rate, uneven energy consumption, and poor network adaptability. Especially during peak water consumption periods, the data transmission success rate dropped significantly, affecting the accuracy of water management decisions.

[0083] In the actual deployment, the system first collected three months of historical water meter reading network data in the area, including: Node location information: geographic coordinates of each water meter, collector, relay node, and aggregator; Communication connection records: connection status, signal strength, transmission rate, etc. between nodes; Energy status information: the percentage of remaining power in battery-powered nodes; Data collection records: data collection volume for each period, transmission success rate, etc.

[0084] Based on this data, the system builds a four-layer spatiotemporal graph model: Hourly spatiotemporal graphs: Capture short-term changes within a 24-hour period, such as differences in network load during peak hours in the morning and evening. Daily-level spatiotemporal graph: Captures differences in usage patterns between weekdays and weekends; Weekly spatiotemporal graph: captures cyclical changes within a week; Monthly spatiotemporal map: Capturing seasonal trends.

[0085] For each smart water meter node, the system records its basic attributes such as function type (such as ordinary water meter, key monitoring water meter), installation location characteristics (indoor / outdoor), communication capability parameters, etc. The time-varying adjacency matrix is ​​represented by a continuous value, where the matrix elements are Represents the node at time t With node The communication quality between them ranges from 0 to 1. The larger the value, the better the communication quality.

[0086] In actual applications, the system selects three types of relationships to construct a multi-relationship adjacency tensor: physical connection relationship (based on geographical distance), logical connection relationship (based on network topology) and data flow relationship (based on actual data transmission path), so that the system can fully capture the multi-dimensional characteristics of the network.

[0087] For the water meter reading network in this area, the system was configured with a graph convolutional network with three layers, each containing 64 hidden units. Input features included 10 dimensions, including node type (one-hot encoding), geographic coordinates, remaining power percentage, and average data transmission volume.

[0088] In the actual configuration of the graph convolutional network, the hourly spatiotemporal graph uses an attention-weighted GraphSAGE variant to deal with the extremely uneven node degree distribution; the daily and weekly spatiotemporal graphs use the standard GCN; and the monthly spatiotemporal graph uses the GIN with a larger receptive field to capture long-term evolution patterns.

[0089] In the temporal convolutional network part, the system adopts different configurations for different levels of spatiotemporal graphs: Hourly: Use causal convolution with kernel size 3 and dilation rate 1; Day level: Use causal convolution with kernel size 3 and dilation rate 2; Zhou level: Use causal convolution with kernel size 3 and dilation rate 4; Monthly level: Use causal convolution with kernel size 3 and dilation rate 8.

[0090] Cross-layer feature fusion adopts an attention weighting mechanism. The system calculates the importance of features at different levels to the current prediction task and dynamically adjusts the fusion weight, so that hourly features obtain higher weights in short-term prediction tasks, while monthly features obtain higher weights in long-term prediction tasks.

[0091] In practice, the system uses a graph attention network with 16 attention heads, each with an output dimension of 32, to generate a 512-dimensional node representation. The attention calculation not only considers node features but also incorporates edge features, including communication quality, stability, and bandwidth.

[0092] During a peak period of network operation, the system automatically identified several relay nodes in commercial areas that were carrying an excessively high forwarding load. Using the attention mechanism, the system quantified the impact of these nodes on the entire network. Specifically, the average attention weight for these key relay nodes was 0.18, significantly higher than the 0.03 for ordinary nodes, indicating their crucial role in the network.

[0093] The various heads of the multi-head attention mechanism do indeed focus on different types of node relationships: some focus on pairs of nodes with close geographical proximity; some focus on nodes with similar energy consumption; and still others focus on nodes with similar data transmission patterns. This multi-dimensional focus enables the system to fully understand the complex dependencies in the network.

[0094] The inter-layer attention mechanism performs well in practical applications, especially in capturing correlation patterns across different time scales. For example, the system finds that seasonal variations in monthly spatiotemporal graphs have an impact on daily forecasts, and therefore assigns a higher weight to inter-layer attention in related forecasting tasks.

[0095] In the actual application of the city's water meter centralized reading network, the system uses a specially optimized prediction architecture for different prediction tasks: For short-term forecasts, a 2-layer GRU network is used with a hidden layer size of 128 and an input window length of 24 (corresponding to 24 hours of historical data); for medium-term forecasts, a 2-layer LSTM network is used with a hidden layer size of 256 and input window lengths of 7 (daily forecasts) and 4 (weekly forecasts), respectively; for long-term forecasts, a convolutional network with temporal attention is used with an input window length of 12 (corresponding to 12 months of historical data).

[0096] The system also incorporates external factors as auxiliary inputs, including temperature data, precipitation forecasts, and holiday information. This information is converted into feature vectors using a specialized encoder and then fused with network state features to improve prediction accuracy. For example, by integrating weather forecast data, the system can predict changes in water usage patterns during the rainy season and adjust network parameters accordingly.

[0097] The adaptive weight assignment strategy within the multi-task learning framework has proven successful in practice. Initially, the system assigned equal weights to the four time-scale prediction tasks. However, as training progressed, the system discovered that the daily prediction task was more difficult and automatically increased its weight (from 0.25 to 0.35), balancing the training losses across the tasks.

[0098] During actual operation, the system successfully detected and responded to many abnormal events. Take a pipeline burst incident as an example: At 3:14 AM one day, the system detected a sudden change in the adjacency matrix within the region, with the Frobenius norm exceeding a preset threshold (0.15). The anomaly detector immediately triggered, locating the affected area to include 12 water meter nodes and two relay nodes. The system immediately implemented the following response measures: Emergency routing reconstruction: A temporary communication path was constructed for water meter nodes in the affected area, bypassing damaged relay nodes; Resource scheduling: temporarily allocate the communication resources of three nearby low-load relay nodes to the affected area; Data collection adjustments: Increase the frequency of water meter data collection in affected areas from once an hour to every 10 minutes to closely monitor possible water volume anomalies; These measures were deployed within 30 seconds, ensuring a data collection success rate of over 92% for the entire pipeline repair process (which lasted approximately four hours). This rate is significantly higher than the approximately 60% achieved by traditional systems in similar situations. Furthermore, the system accurately identified the leak location through real-time analysis of collected data, providing precise guidance for repair work.

[0099] Comparison of data packet loss rates before and after applying this application method: Before application: The average packet loss rate during normal hours was 12.3%, the average packet loss rate during peak hours was 20.5%, and the highest rate reached 26.8%; After application: the average packet loss rate during normal hours is 2.1%, the average packet loss rate during peak hours is 2.8%, and the maximum does not exceed 3.5%; At the same time, the real-time performance of data collection has also been improved. The average delay from data collection to aggregation has been reduced from 8.5 minutes to 1.2 minutes, an improvement of approximately 85.9%.

[0100] Comparison of node energy consumption before and after applying this application method: Before the application: The power consumption ratio between the 10% of nodes with the highest energy consumption and the 10% with the lowest energy consumption was 4.7:1, resulting in some nodes requiring battery replacement more than four times as often as other nodes. After application: the ratio is reduced to 1.9:1, the energy consumption difference between nodes is reduced by about 60%, and the overall service life of the network is extended; The average service life of battery-powered nodes has been extended from 10.5 months to 17.8 months, reducing the frequency of maintenance and replacement and lowering operation and maintenance costs.

[0101] It can be seen from the above practical application examples that the water meter reading management method provided in this application can effectively solve the problems existing in the existing technology, improve the reliability, stability and efficiency of the water meter reading network, and provide strong technical support for smart water management.

[0102] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. A water meter centralized reading management method for water services, characterized in that: include: Construct a multi-level spatiotemporal graph model and abstract the water meter reading network into a multi-level spatiotemporal graph structure; Based on the multi-level spatiotemporal graph structure, graph neural networks are used to extract network topology features, including the use of graph convolutional networks to process spatial dependencies and temporal convolutional networks to process temporal dependencies. Based on the network topology characteristics, the attention weights between nodes are calculated to model the interdependence between nodes; Based on the attention weights between nodes, multi-scale network state prediction is performed to predict the future state of the network at multiple time scales; Based on the output of multi-scale network status prediction, an abnormal event response mechanism is implemented to detect and respond to emergencies in the water meter reading network.

2. A water meter centralized reading management method for water services according to claim 1, characterized in that: The steps of constructing a multi-level spatiotemporal graph model include: Number all nodes in the water meter centralized reading network to form a node set; Based on the communication connection relationship between nodes, a time-varying edge set is constructed; Convert the edge set into a time-varying adjacency matrix; Construct a multi-level spatiotemporal graph, where each layer represents the network topology characteristics at different spatiotemporal scales; Establish inter-layer connections and connect the same nodes at different levels through cross-layer edge sets.

3. A water meter centralized reading management method for water services according to claim 1, characterized in that: The step of applying a graph neural network to extract network topology features includes: For each layer in the multi-level spatiotemporal graph, a graph convolutional network is applied to extract topological features; Combined with time convolutional networks to capture the evolution of network topology over time; For cross-layer connections in multi-level spatiotemporal graphs, a cross-layer feature fusion algorithm is applied to integrate information at different spatiotemporal scales; Generate a node feature matrix that reflects the complete spatiotemporal characteristics of the water meter reading network.

4. A water meter centralized reading management method for water services according to claim 1, characterized in that: The step of calculating the attention weights between nodes includes: Based on the extracted node feature matrix, the graph attention network is applied to calculate the attention weights between nodes; Apply a multi-head attention mechanism to simultaneously calculate multiple sets of independent attention weights and then merge the results; For different levels in the multi-level spatiotemporal graph, attention weights are calculated separately to form a hierarchical set of attention weights; According to the attention weights, a weighted adjacency matrix is ​​generated.

5. The method for centralized water meter reading management in water affairs according to claim 1, characterized in that: The step of performing multi-scale network state prediction includes: Build a multi-scale prediction model to simultaneously predict network status changes at four time scales: hourly, daily, weekly, and monthly; For prediction tasks at different time scales, a hierarchical time series prediction architecture is adopted; Introducing a multi-task learning framework to jointly optimize prediction tasks at four time scales; Based on the prediction results, a prediction-driven network parameter optimization strategy is calculated.

6. A water meter centralized reading management method for water services according to claim 1, characterized in that: The steps of implementing the abnormal event response mechanism include: Build an abnormal event detector to identify abnormal events by monitoring sudden changes in network topology; Rapidly locate and assess the impact of detected abnormal events; Design a rapid response strategy based on the anomaly-affected area and prioritize adjusting network parameters in the affected area; For different types of abnormal events, special processing strategies are applied.

7. A water meter centralized reading management method for water services according to claim 3, characterized in that: The graph convolutional network includes: Input layer, which receives the initial features of the nodes; Multiple graph convolutional layers update node representations by aggregating neighbor node information; Output layer, generates the final node feature representation; The temporal convolutional network includes a causal convolutional layer, a dilated convolutional layer and a residual connection.

8. A water meter centralized reading management method for water services according to claim 4, characterized in that: The graph attention network includes: Feature transformation layer, which performs linear transformation on input node features; Attention calculation layer, which calculates the attention coefficient between node pairs; Feature aggregation layer, which aggregates neighbor node information based on weighted attention coefficients; The multi-head attention mechanism simultaneously calculates 8 to 16 attention heads, each of which focuses on a different feature subspace.

9. A water meter centralized reading management method for water services according to claim 5, characterized in that: The hierarchical time series forecasting architecture includes: The short-term prediction layer uses gated recurrent units to predict hourly network state changes; The medium-term prediction layer uses long short-term memory networks to predict daily and weekly network state changes; The long-term prediction layer uses an attention-enhanced temporal convolutional network to predict monthly network state changes; The prediction-driven network parameter optimization strategy includes routing strategy optimization, energy management optimization and communication frequency optimization.

10. A water meter centralized reading and management system for water services, characterized in that: A method for centralized water meter reading management in water services, used for executing any one of claims 1 to 9, comprising: A multi-level spatiotemporal graph construction module is used to abstract the water meter reading network into a multi-level spatiotemporal graph structure; Graph neural network feature extraction module, used to extract network topology features; Attention weight calculation module, used to model the interdependence between nodes; Multi-scale network state prediction module, used to predict the future state of the network at multiple time scales; The abnormal event response module is used to detect and respond to emergencies in the water meter reading network.

Images