Metering box operation state adaptive reporting method and system

Abstract

The application relates to the technical field of electric power communication, and discloses a metering box operation state adaptive reporting method and system, which comprises the following steps: obtaining to-be-reported data and a sampling sequence, a spatial topology structure, a queue length and a channel occupancy rate; calculating a local fluctuation rate and a mutation energy characteristic value to construct a weighted time characteristic vector as a node characteristic, combining an original edge weight to construct a space-time network heterogeneous graph tensor; calculating a congestion penalty coefficient according to the queue length and the channel occupancy rate to correct the original edge weight; processing the graph tensor by using a graph attention mechanism containing a queue delay penalty term, calculating a reporting priority weight and generating a hierarchical reporting strategy; verifying data consistency by using a historical error model to execute distribution reporting, and updating a weight matrix according to actual packet loss rate and communication delay feedback. The application improves the adaptive scheduling efficiency and data transmission reliability of the electric power metering network under a complex concurrent scene.

Description

Technical Field

[0001] This application relates to the field of power communication technology, and in particular to an adaptive reporting method and system for the operating status of a metering box. Background Technology

[0002] With the rapid development of smart grids, a large number of distributed monitoring and metering boxes are widely deployed at the end of the power grid to collect key electrical physical quantities such as voltage and current in real time. These operational status data are periodically aggregated and reported to the cloud master station through the underlying wireless communication network, providing important data support for load forecasting, fault diagnosis and dispatching decisions of the power system.

[0003] However, with the exponential growth in the number of connected devices and the sharp increase in business concurrency, traditional fixed-period or simple threshold-triggered reporting mechanisms have gradually revealed serious limitations. In complex actual operating environments, aging or failure of underlying sensors or transient strong electromagnetic interference can easily lead to abnormal changes in collected data. Existing data reporting strategies often separate the operating status of underlying devices from the carrying capacity of the communication network. This information barrier prevents the system from responding flexibly to sudden surges in business traffic or deterioration of wireless channel quality. When a large number of metering boxes detect anomalies at the same time and report them concurrently, it can easily cause severe network congestion, leading to overflow of memory queues at critical nodes, resulting in large-scale data packet loss and communication delays. At the same time, static routing scheduling and priority allocation mechanisms are difficult to adapt to the slow drift of underlying network topology and external interference in the long term, and cannot proactively avoid congested nodes. In addition, if unverified false high-priority mutation data is directly connected to the network, it will not only ineffectively occupy valuable emergency communication slots, but also cause the cloud master station to make incorrect scheduling decisions, further exacerbating the strain on network resources.

[0004] Therefore, how to break down the barriers between heterogeneous data, achieve joint perception of multi-dimensional physical states and network parameters, comprehensively assess the urgency of electrical anomalies of each node and the congestion status of surrounding communication links in massive concurrent scenarios, and then rationally allocate limited wireless communication resources to ensure the reliable arrival of key data, while building a model dynamic update mechanism with a real network feedback loop, has become a core technical problem that urgently needs to be solved in the field of power Internet of Things communication scheduling. Summary of the Invention

[0005] To address the aforementioned technical problems, this application provides an adaptive reporting method and system for the operating status of metering boxes. This system is used to solve the technical problem in existing power metering networks where information barriers exist between the operating status of underlying equipment and the carrying capacity of the communication network. This leads to data congestion, large-scale packet loss, and false and abnormal data preempting emergency communication resources during sudden surges in business traffic or deterioration of channel quality.

[0006] Firstly, this application provides an adaptive reporting method for the operating status of a metering box, the method comprising: S1. Obtain the data to be reported from the monitoring meter box and the corresponding sampling sequence, including the spatial topology structure containing global steady-state wireless channel quality indicators, the queue length of the access node, and the channel occupancy rate of the access node in different frequency bands. S2. By calculating the local volatility and mutation energy characteristic values ​​of the sampled sequence, a weighted time feature vector is constructed. The weighted time feature vector is used as the node feature, and combined with the connectivity between nodes and the original edge weights, a spatiotemporal network heterogeneous graph tensor is constructed. S3. Construct a dynamic queue tensor based on the queue length and the channel occupancy rate of the access node in different frequency bands, calculate the congestion penalty coefficient corresponding to the dynamic queue tensor, and use the congestion penalty coefficient to correct the original edge weights of the spatiotemporal network heterogeneous graph tensor to obtain the corrected edge weights. S4. Introduce a graph attention mechanism that includes a queue delay penalty term constructed based on the corrected edge weights. Use the graph attention mechanism to process the spatiotemporal network heterogeneous graph tensor, calculate the reporting priority weight of each monitoring meter box, and generate a hierarchical reporting strategy based on the reporting priority weight. S5. Verify the consistency of the data to be reported by the monitoring meter box using a preset historical error model. If the data to be reported fails the consistency verification, trigger the preset downgrade reporting strategy. If the data to be reported passes the consistency verification, detect the transient physical channel status of the target communication link in real time, and perform the allocation and reporting of the data to be reported according to the hierarchical reporting strategy and the transient physical channel status. Update the feature weight matrix inside the graph attention mechanism based on the actual packet loss rate and the actual communication delay value after the reporting is completed.

[0007] Secondly, this application provides an adaptive reporting system for the operating status of a metering box, the system comprising: The data acquisition module is used to acquire the data to be reported from the monitoring meter box and the corresponding sampling sequence, the spatial topology including global steady-state wireless channel quality indicators, the queue length of the access node, and the channel occupancy rate of the access node in different frequency bands. The heterogeneous graph module is used to construct a weighted temporal feature vector by calculating the local volatility and mutation energy feature values ​​of the sampled sequence. The weighted temporal feature vector is used as the node feature, and combined with the connectivity between nodes and the original edge weights, a spatiotemporal network heterogeneous graph tensor is constructed. The weight adjustment module is used to construct a dynamic queue tensor based on the queue length and the channel occupancy rate of the access node in different frequency bands, calculate the congestion penalty coefficient corresponding to the dynamic queue tensor, and use the congestion penalty coefficient to correct the original edge weights of the spatiotemporal network heterogeneous graph tensor to obtain the corrected edge weights. The hierarchical decision module is used to introduce a graph attention mechanism that includes a queue delay penalty term constructed based on the modified edge weight. The graph attention mechanism is used to process the heterogeneous graph tensor of the spatiotemporal network, calculate the reporting priority weight of each monitoring meter box, and generate a hierarchical reporting strategy based on the reporting priority weight. The verification feedback module is used to verify the consistency of the data to be reported by the monitoring meter box using a preset historical error model. If the data to be reported fails the consistency verification, a preset downgrade reporting strategy is triggered. If the data to be reported passes the consistency verification, the transient physical channel status of the target communication link is detected in real time, and the data to be reported is allocated and reported according to the hierarchical reporting strategy and the transient physical channel status. The feature weight matrix inside the graph attention mechanism is updated based on the actual packet loss rate and the actual communication delay after the reporting is completed.

[0008] Compared with the prior art, the beneficial effects of the present invention are at least as follows: 1. This application eliminates the perception blind spot where service data and communication carrying capacity are separated in traditional scheduling mechanisms by synchronously collecting electrical service data containing continuous sampling sequences and underlying network status containing link topology, buffer queues, and spectrum occupancy. It also eliminates dimensional differences by normalizing weighting and combines multi-dimensional electrical features with the inversely mapped link costs to construct a spatiotemporal network heterogeneous graph tensor. Based on this, this application integrates node queue depth and multi-band spectrum occupancy and dynamically corrects the underlying communication network link costs through nonlinear exponential mapping. This truly and comprehensively quantifies the latency resistance faced by communication nodes and overcomes the lag of traditional routing mechanisms that rely solely on fixed thresholds to determine congestion.

[0009] 2. Furthermore, this application introduces a graph attention mechanism that integrates queue delay penalty terms, enabling the model to actively reduce its dependence on highly congested links when aggregating node features. This achieves deep joint perception of electrical anomalies and communication network congestion, improving the accuracy of priority assessment and congestion resistance. Simultaneously, by using dual thresholds, global nodes are precisely divided into three non-overlapping reporting queues, and differentiated communication resources such as contention-free time slots, contention access, and sleep waiting are matched. This not only ensures the absolute priority and high reliability of emergency anomaly data transmission, but also effectively alleviates the collision conflicts and spectrum resource pressure of the overall network, improving the adaptive scheduling efficiency of the power metering network in complex concurrent scenarios.

[0010] 3. Finally, this application proposes a data consistency verification degradation mechanism based on time-series predictive networks, which accurately identifies and intercepts false mutation data caused by sensor failures or electromagnetic interference. Combined with a forced degradation strategy, it prevents invalid dirty data from preempting emergency communication resources. Furthermore, by extracting the real timestamp and packet loss sequence number from the cloud confirmation message, the actual communication delay and packet loss rate are quantitatively calculated and used as the weight matrix of the supervision signal backpropagation update graph attention mechanism. This constructs a complete closed loop from policy issuance to effect feedback, enabling the reporting scheduling model to continuously learn and adapt to the dynamic fading and topology changes of the underlying physical channel, significantly enhancing the robustness of the system during long-term operation. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart illustrating the steps of the adaptive reporting method for the operating status of the metering box in this embodiment of the application. Figure 2 This is a schematic diagram illustrating the generation of a hierarchical reporting strategy based on graph attention mechanism in an embodiment of this application. Figure 3 This is a structural diagram of the adaptive reporting system for the metering box operating status in this embodiment of the application. Detailed Implementation

[0013] This application provides an adaptive reporting method and system for the operating status of a metering box. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0014] Example 1: For ease of understanding, the specific process of the embodiments of this application is described below, such as... Figure 1The adaptive reporting method for the operating status of a metering box shown in this embodiment of the application includes: Specifically, in order to address the technical problem in existing power metering networks where there is an information barrier between the operating status of underlying equipment and the carrying capacity of the communication network, leading to a rigid reporting mechanism and making it prone to data congestion and large-scale packet loss during sudden surges in business traffic or deterioration of wireless channel quality, this embodiment breaks down the barriers between heterogeneous data by simultaneously performing multi-dimensional physical state and network parameter sensing on both the power grid side and the grid side. This provides a precise underlying data foundation for the subsequent construction of a network adaptive scheduling graph model with spatiotemporal correlation evolution characteristics.

[0015] S1. Obtain the data to be reported from the monitoring meter box and the corresponding sampling sequence, the spatial topology including global steady-state wireless channel quality indicators, the queue length of the access node, and the channel occupancy rate of the access node in different frequency bands.

[0016] In actual power Internet of Things (IoT) operation scenarios, monitoring and metering boxes distributed at the end of the power grid are equipped with high-precision metering chips and sensors. During operation, they continuously collect basic electrical physical quantities such as voltage and current at a preset fixed frequency. First, the microprocessor inside the monitoring and metering box extracts the data to be reported, which is packaged and prepared for transmission to the cloud within the current business cycle. In order to obtain the dynamic evolution pattern behind this data, the system simultaneously extracts the discrete measurement values ​​continuously acquired by the monitoring and metering box in the period before the generation of the data to be reported, and arranges and combines them strictly in chronological order to generate the corresponding sampling sequence. This sampling sequence constitutes the basic time series foundation for subsequent steps to capture transient abnormal fluctuations of the equipment and extract time-domain abrupt change characteristics.

[0017] While acquiring physical data from the power grid end, the underlying communication module is driven to synchronously perceive the complex external wireless network environment. By collecting physical connection handshake information between each distributed monitoring metering box and the aggregation gateway, a spatial topology containing global steady-state wireless channel quality indicators is obtained. Specifically, this spatial topology is logically represented as a network connectivity graph with device terminals as nodes and wireless communication links as connecting edges. The wireless channel quality indicators specifically include underlying parameters such as received signal strength indication and signal-to-noise ratio. These indicators are attached as physical attributes to each communication link in the topology to objectively and statically quantify the basic wireless attenuation cost of data transmission between nodes.

[0018] Furthermore, to comprehensively grasp the real-time dynamic congestion load status of network nodes, the queue length of access nodes is polled and obtained in real time through the underlying network management protocol. These access nodes refer to edge gateways or base station devices in the network responsible for aggregating and uploading data from various monitoring and metering boxes. The physical meaning of their queue length represents the total number of uplink data packets currently queued and waiting to be processed in the node's internal memory buffer, a core indicator directly measuring the bottleneck of the device's internal communication processing. In addition, through the carrier sensing mechanism of the underlying radio frequency chip, the channel occupancy rate of access nodes in different frequency bands within the set frequency band whitelist is scanned and obtained in real time. This channel occupancy rate refers to the percentage of time a specific wireless frequency band is actually occupied by external radio frequency energy carriers within a unit observation period, indicating a busy state. By simultaneously obtaining the queue length and multi-band channel occupancy rate, the real-time dynamic carrying capacity of the current communication network is accurately characterized from two dimensions: the hardware buffer backlog pressure within the node and the spatial competition degree of the external wireless spectrum.

[0019] The above content proposes a joint perception mechanism for the underlying heterogeneous state, which integrates the electrical operation sequence of the end side with the multi-dimensional spatiotemporal communication quality indicators of the network side. By synchronously collecting electrical service data containing continuous sampling sequences and underlying network state including link topology, buffer queues, and spectrum occupancy, it eliminates the perception blind spot where service data and communication carrying capacity are separated in the traditional scheduling mechanism. This comprehensive multi-source heterogeneous data acquisition method not only anchors the static physical attenuation cost of the link in the spatial dimension, but also captures the dynamic evolution trend of network congestion and the transient processing bottleneck of the equipment in the temporal dimension. This lays a comprehensive and accurate data foundation for the subsequent system to map multi-dimensional physical parameters into a unified spatiotemporal heterogeneous graph model, and then achieve highly reliable adaptive reporting to actively avoid congested nodes.

[0020] Specifically, after obtaining the continuous discrete sampling sequence and multidimensional communication state parameters of the power grid end side through the above steps, in order to solve the technical problem that the system is unable to directly discover the deep coupling law between electrical transient anomalies and dynamic congestion of the communication network due to the complexity of the underlying physical sensing data types, different dimensions and spatiotemporal independence, this embodiment introduces a spatiotemporal graph tensor modeling mechanism to transform isolated heterogeneous parameter mapping into a feature data base of tightly correlated evolutionary topology.

[0021] S2. By calculating the local volatility and mutation energy characteristic values ​​of the sampled sequence, a weighted time feature vector is constructed. The weighted time feature vector is used as the node feature, and combined with the connectivity between nodes and the original edge weights, a spatiotemporal network heterogeneous graph tensor is constructed.

[0022] S2 includes: calculating the ratio of the standard deviation to the mean of the sampled sequence within a set time window to obtain the local volatility; inputting the sampled sequence into a preset mutation energy function to output mutation energy feature values; weighting and concatenating the local volatility and mutation energy feature values ​​to generate a weighted time feature vector, and using the weighted time feature vector as the node feature matrix; extracting the original edge weights representing the link cost based on the spatial topology containing global steady-state wireless channel quality indicators, and generating an adjacency matrix representing connectivity; fusing the queue length of the access node with the original edge weights to map it into the edge feature matrix of the communication link, and extracting the original edge weights from the edge feature matrix; and performing tensor joint stacking of the node feature matrix, edge feature matrix, and adjacency matrix along the time series dimension to construct a spatiotemporal network heterogeneous graph tensor.

[0023] The process of generating a weighted time feature vector includes: dividing the sampling sequence into a sliding time window; calculating the standard deviation and mean of the sampling sequence within the sliding time window, and using the ratio of the standard deviation to the mean as the local volatility; calculating the sum of squares of the first-order differences of adjacent sampling points within the sliding time window, and using the sum of squares of the first-order differences as the mutation energy feature value; obtaining the first weight coefficient corresponding to the local volatility and the second weight coefficient corresponding to the mutation energy feature value; and concatenating the product of the local volatility and the first weight coefficient, and the product of the mutation energy feature value and the second weight coefficient, to generate the weighted time feature vector.

[0024] Specifically, in the implementation process, the acquired sampling sequence is first subjected to time-domain feature extraction. The sampling sequence is divided according to a preset fixed time span to form a sliding time window that evolves in a time sequence. Within each specific sliding time window, the standard deviation and mean of all data points in the sampling sequence are calculated, and the ratio of the calculated standard deviation to the mean is used as the local volatility. The local volatility physically represents the relative coefficient of variation, which can effectively eliminate baseline differences caused by different benchmark workloads of different monitoring and metering boxes, and objectively reflect the slow drift and relative dispersion of the equipment during steady-state operation. At the same time, the sampling sequence is input into a preset mutation energy function. The specific calculation process of the mutation energy function is as follows: calculate the square of the difference between the values ​​of all adjacent sampling points in the current sliding time window, then sum all the squared differences to output the first-order difference sum of squares, and use this first-order difference sum of squares as the mutation energy feature value. This mutation energy feature value can non-linearly amplify the drastic jumps in data between adjacent time moments through squaring operations, and is used to keenly capture transient electrical anomalies that may occur in the metering box equipment.

[0025] After obtaining the two time-domain features mentioned above, considering that local volatility is a dimensionless ratio coefficient while the dimension of the mutation energy feature value is the square of the original physical voltage or current unit, directly performing numerical fusion would lead to excessively large differences in absolute numerical magnitudes, thus causing the fundamental problem of small-scale features being obscured by calculation. Therefore, before performing weighted calculation, a maximum-minimum normalization method is used to linearly map the local volatility and mutation energy feature values ​​to a unified dimensionless numerical range of 0 to 1, respectively. Subsequently, a value used to measure the importance of the two types of features is obtained by pre-training and optimization using a large amount of historical sample data. The first weight coefficient and the corresponding second weight coefficient are used to calculate the product of the normalized local volatility and the first weight coefficient, and the product of the normalized mutation energy feature value and the second weight coefficient. Then, the two product results are concatenated and spliced ​​in the feature dimension to generate a weighted time feature vector that includes both steady-state drift attributes and transient anomaly attributes. The weighted time feature vectors generated for each monitoring and metering box in the entire business network are used as node features and combined and stacked in the order of node number to construct a node feature matrix describing the electrical status of the equipment end side of the entire network.

[0026] After completing the matrix-based feature representation of node attributes, the focus shifts to joint modeling of the spatial communication network topology. Based on the acquired spatial topology containing global steady-state wireless channel quality indices, the physical transmit / receive connection states between network nodes are extracted, thereby generating an adjacency matrix consisting only of zeros and ones to characterize the rigid connectivity of the network. Simultaneously, since wireless channel quality indices such as signal-to-noise ratio (SNR) or received signal strength are positive gain physical quantities characterizing communication capability, while link cost represents transmission obstacles, reciprocal operations or negative exponential mapping algorithms are used to numerically transform the wireless channel quality indices, calculating and extracting the original edge weights that are inversely proportional to communication quality and can truly represent static physical transmission attenuation. Furthermore, since communication bottlenecks depend not only on static radio frequency losses but also on node buffering... To address the limitations of storage capacity, the queue length of the destination access node pointed to by the communication link is extracted. After normalizing the queue length, it is combined and concatenated with the corresponding original edge weights in a vector dimension to generate an edge feature matrix that comprehensively reflects the physical attenuation of the link plus the queuing pressure of the endpoints. To meet the needs of subsequent dynamic adjustment of link costs, a matrix dimension slicing extraction operation is used to independently extract and retain one-dimensional original edge weights from the fused and concatenated edge feature matrix for use in subsequent steps. Finally, along the evolution dimension of the time series, the node feature matrix, edge feature matrix, and adjacency matrix continuously generated under different sliding time windows are jointly stacked as tensors to construct a spatiotemporal network heterogeneous graph tensor that can comprehensively characterize the three-dimensional evolution state of the power metering network.

[0027] The above content proposes a spatiotemporal tensor construction method that eliminates dimensional differences through normalized weighting and jointly integrates multidimensional electrical features with the link cost of reverse mapping. By introducing the ratio of the first-order sum of squared differences to the mean of standard deviation, the steady-state drift and transient anomaly features of the device are accurately decoupled synchronously from the discrete sampling sequence, effectively avoiding misjudgment of the state caused by a single feature. By cost mapping of the forward channel index and cross-border fusion of the queuing depth of the target node into the edge features of the communication link, the barriers between network layers are completely broken down. Finally, a spatiotemporal network heterogeneous graph tensor is constructed by stacking multidimensional matrices in the time domain, which lays a solid foundation for the subsequent introduction of graph neural networks to realize refined congestion avoidance and adaptive reporting strategies, and is a heterogeneous model that is fully aligned in structure and time.

[0028] Specifically, after constructing a spatiotemporal network heterogeneous graph tensor containing static physical link costs through the above steps, considering that in actual power business concurrency scenarios, the congestion of the underlying communication network is usually determined by the intertwined factors of the accumulation of internal memory queue depth and the preemption of external available radio frequency resources, in order to solve the technical problem that relying solely on the original edge weights that characterize the static channel physical attenuation cannot accurately quantify the dynamic transmission bottlenecks caused by sudden traffic surges and external interference, and thus easily leads to the missending of data to fully loaded and congested nodes during subsequent adaptive scheduling, causing data stagnation or even overflow and packet loss, this embodiment introduces a cost correction mechanism that fuses multi-dimensional congestion parameters to dynamically and adaptively update the topological edge costs in the graph model.

[0029] S3. Construct a dynamic queue tensor based on the queue length and the channel occupancy rate of the access node in different frequency bands, calculate the congestion penalty coefficient corresponding to the dynamic queue tensor, and use the congestion penalty coefficient to correct the original edge weights of the spatiotemporal network heterogeneous graph tensor to obtain the corrected edge weights.

[0030] S3 includes: extracting the target channel occupancy rate of the access node in each frequency band within a set of frequency bands from the acquired channel occupancy rates; calculating the product of the queue length and the target channel occupancy rate to obtain the frequency band congestion feature value of the access node in each frequency band; combining the frequency band congestion feature values ​​along the frequency band dimension to construct a dynamic queue tensor; inputting the dynamic queue tensor into a preset nonlinear activation function to calculate the congestion penalty coefficient mapped to a set numerical range; extracting the original edge weights corresponding to the access nodes in the spatiotemporal network heterogeneous graph tensor; multiplying the original edge weights by the congestion penalty coefficient to obtain the corrected edge weights; and replacing the original edge weights with the corrected edge weights.

[0031] In the specific implementation process, the data is first filtered from the multi-dimensional communication status parameters of the underlying network collected in the early stage to accurately extract the target channel occupancy rate of the network access node in each frequency band within the set frequency band. Here, the set frequency band refers to the list of all legal radio frequency working channels that are pre-divided and authorized for business data interaction between the underlying metering terminal and the edge gateway during the planning stage of the power Internet of Things communication system. Then, the access node queue length, which has been normalized and converted to a dimensionless value of zero to one in the previous processing, is retrieved, and the product of the length of this single queue and the target channel occupancy rate of the node in each frequency band is calculated. Since the channel occupancy rate itself is a small percentage representing the proportion of busy time of the frequency band, the calculation is not performed in this way. The aforementioned product operation can directly and realistically characterize, from both mathematical and physical mechanisms, the combined queuing and media access delay resistance faced by a gateway node when a specific external wireless operating frequency band is busy. This is achieved through the multiplication and coupling of cross-dimensional physical quantities, calculating the frequency band congestion characteristic value of the access node in each specific frequency band. Subsequently, the multiple independent frequency band congestion characteristic values ​​calculated in parallel for the same access node across all set frequency bands are combined along the dimension of the frequency band sequence to construct a dynamic queue tensor that comprehensively describes the multi-dimensional resource-constrained state of the access node.

[0032] To transform the aforementioned high-dimensional dynamic queue tensor into a scalar multiplier that can directly affect the graph topology cost, the constructed dynamic queue tensor is input into a preset nonlinear activation function. The output is then calculated and mapped to a congestion penalty coefficient within a defined numerical range. The specific algorithm execution process of this nonlinear activation function is as follows: First, the arithmetic mean of the congestion feature values ​​of all frequency bands within the dynamic queue tensor is calculated, thereby reducing the dimensionality of the multidimensional tensor data into a congestion scalar measure to characterize the overall congestion level of nodes. Then, this congestion scalar measure is used as an exponential term and input into a preset exponential activation function with the natural constant as the base for nonlinear mapping calculation. The calculated exponential result is then used as the congestion penalty coefficient. Since the normalized queue length and channel occupancy rate involved in the calculation are both positive decimals between 0 and 1, the mean of their product also strictly falls within this numerical range. This ensures that the congestion penalty coefficient output by the activation function is smoothly and rigorously constrained to a range from one to the natural constant. Within a set numerical range, this exponential mapping-based calculation mechanism ensures that: when a communication node is in a free-flowing idle state, the lower bound of the penalty coefficient approaches one; and as the congestion metric scalar increases, the penalty coefficient exhibits non-linear accelerated scaling; finally, using the graph topology traversal addressing method, the original edge weights of all communication links pointing to the access node in the spatial dimension are extracted from the constructed spatiotemporal network heterogeneous graph tensor; since the original edge weights have been reverse-mapped to represent the cost values ​​of basic physical transmission resistance and attenuation in the previous steps, the mathematical product of the extracted original edge weights and the above-mentioned congestion penalty coefficient is directly calculated. This product operation uses a penalty coefficient greater than or equal to one to directly amplify the original physical link attenuation resistance in the mathematical topology space. The result of the multiplication is used as the corrected edge weight, and the corrected edge weight is used to accurately replace the original edge weight at the corresponding position in the data structure of the spatiotemporal network heterogeneous graph tensor, thereby completing the closed-loop correction process of the dynamic congestion cost of the entire network topology.

[0033] The above content proposes a method that integrates node queue depth and multi-band spectrum occupancy rate and dynamically corrects the link cost of the underlying communication network through nonlinear exponential mapping. By modeling the queue length, which reflects the internal backlog status of the equipment, and the channel occupancy rate, which reflects the severity of the external radio electromagnetic environment, the method realistically, comprehensively, and dimensionlessly quantifies the delay resistance faced by communication nodes, overcoming the lag of traditional routing mechanisms that rely solely on fixed thresholds to determine congestion. By calculating the arithmetic mean and introducing a nonlinear exponential activation function with the natural constant as the base, the dynamically changing multidimensional congestion characteristics are strictly mapped to a continuous congestion penalty coefficient with a lower bound of 1. This directly amplifies and corrects the link cost in the graph network structure, enabling the originally static data transmission path to keenly reflect instantaneous local traffic peaks. This prompts subsequent scheduling algorithms to automatically detect high-impedance links, significantly improving the high connectivity and high reliability of the entire process of reporting operational status data at the power grid end.

[0034] Specifically, after completing the dynamic congestion correction of the topological edge cost in the heterogeneous graph tensor of the spatiotemporal network through the above steps, although the system has accurately quantified the real-time transmission resistance of the underlying communication network, in the face of the complex scenario of concurrent reporting by a large number of monitoring and metering boxes, how to comprehensively evaluate the urgency of the electrical anomalies of each node and the congestion status of the surrounding communication links, and then rationally allocate limited wireless communication resources to ensure the reliable arrival of key data, has become an urgent technical problem to be solved. To this end, this embodiment introduces a graph attention mechanism that integrates queue delay penalty terms, and generates an adaptive hierarchical reporting strategy by performing deep feature aggregation and priority evaluation on the heterogeneous graph tensor.

[0035] S4. Introduce a graph attention mechanism that includes a queue delay penalty term constructed based on the modified edge weights. Use the graph attention mechanism to process the heterogeneous graph tensor of the spatiotemporal network, calculate the reporting priority weight of each monitoring meter box, and generate a hierarchical reporting strategy based on the reporting priority weight.

[0036] In S4, a graph attention mechanism is used to process the spatiotemporal network heterogeneous graph tensor to calculate the reporting priority weight of each monitoring meter box. This includes: extracting the target node feature vector of the target monitoring meter box and the neighboring node feature vectors of the adjacent nodes connected to the target monitoring meter box from the spatiotemporal network heterogeneous graph tensor; performing feature concatenation and linear transformation on the target node feature vector and the neighboring node feature vector to calculate the basic attention coefficient; obtaining the corrected edge weight between the target monitoring meter box and the neighboring nodes, and calculating the queue delay penalty term based on the corrected edge weight and the preset penalty scaling factor; subtracting the queue delay penalty term from the basic attention coefficient to obtain the corrected attention coefficient, and normalizing the corrected attention coefficient using a normalization function to obtain the attention weight of the target monitoring meter box relative to the neighboring nodes; using the attention weight to perform weighted summation and aggregation of the neighboring node feature vectors to obtain the aggregated node state vector of the target monitoring meter box; inputting the aggregated node state vector into a preset fully connected network layer for dimensionality compression mapping, calculating the output scalar score, and using the scalar score as the reporting priority weight of each monitoring meter box.

[0037] Specifically, S4 generates a hierarchical reporting strategy based on reporting priority weights, including: obtaining the reporting priority weights of all monitoring and metering boxes in the system, arranging the reporting priority weights in descending order according to their numerical values ​​to generate a global priority sequence; obtaining preset emergency communication thresholds and regular communication thresholds, using the emergency communication thresholds and regular communication thresholds to divide the global priority sequence into intervals, and assigning the monitoring and metering boxes to emergency reporting queues, regular reporting queues, and delayed reporting queues; allocating dedicated contention-free time slots and maximum transmit power to the monitoring and metering boxes in the emergency reporting queue; allocating carrier-sensing multiple access channels to the monitoring and metering boxes in the regular reporting queue; instructing the monitoring and metering boxes in the delayed reporting queue to enter a dormant waiting state; and combining the resource allocation instructions corresponding to the emergency reporting queue, regular reporting queue, and delayed reporting queue to generate a hierarchical reporting strategy.

[0038] Specifically, a graph attention mechanism is employed to process the updated spatiotemporal network heterogeneous graph tensor. This mechanism, a spatial domain-based graph neural network model, works by calculating attention coefficients between nodes and adaptively assigning different feature aggregation weights to different neighboring nodes. This extracts deep feature representations containing local topology and node attributes. The model's input consists of the node feature matrix and corrected edge weights from the spatiotemporal network heterogeneous graph tensor, and its output is the aggregated node state vector for each node. During offline training, a large amount of graph tensor sample data collected over historical periods and manually labeled optimal reporting priority tags are used as the training set. The model output is then calculated. The mean squared error loss between the prediction priority and the true label is calculated, and the gradient descent algorithm is used to continuously backpropagate and update the graph attention mechanism and the learnable parameter matrix in the subsequent fully connected network layers until the loss function converges, thus obtaining a mature model. In the online inference stage, the feature vector of the target node corresponding to the target monitoring meter box to be evaluated and the feature vectors of all neighboring nodes that have a direct physical connection with the target monitoring meter box in the network spatial topology are extracted from the spatiotemporal network heterogeneous graph tensor. The target node feature vector and the feature vector of each neighboring node are concatenated and concatenated in the feature dimension, and the concatenated high-dimensional vector is input. The single-layer feedforward neural network of the input attention mechanism undergoes a linear transformation, namely multiplying by a pre-trained shared weight matrix and processing it through an activation function, to calculate the basic attention coefficient, which characterizes the basic correlation strength between nodes. Simultaneously, the corrected edge weights between the target monitoring meter and its corresponding adjacent nodes are obtained. These corrected edge weights are multiplied by a preset penalty scaling factor to calculate the queue delay penalty term. This penalty scaling factor is a dimensionless adjustment coefficient set according to the overall network bandwidth capacity, used to unify the numerical magnitude between the basic attention coefficient and the corrected edge weights. Subtracting the corresponding queue delay penalty term from the calculated basic attention coefficient yields the corrected attention coefficient. The force coefficient, in terms of physical mechanism, implements a system where the greater the congestion cost of the link where the neighboring node is located, the heavier the penalty is imposed on it, thereby actively reducing the attention to the congested path. The Softmax normalization function is used to exponentially normalize the modified attention coefficients corresponding to all neighboring nodes, and the attention weights with a sum of one are calculated. Using the calculated attention weights, scalar and vector multiplication operations are performed on the feature vectors of the corresponding neighboring nodes, and all product results are summed by vector addition to complete the feature aggregation operation, thereby obtaining the aggregated node state vector of the target monitoring meter box that integrates the state of the surrounding neighborhood and the link congestion situation.The aggregation node state vector is input into a pre-defined fully connected network layer. This fully connected network layer performs feature dimensionality reduction and spatial compression mapping on the high-dimensional aggregation node state vector through multiplication of weight matrices and addition of biases across multiple neurons. Finally, it calculates and outputs a one-dimensional scalar score, which is directly used as the reporting priority weight to characterize the urgency of the current data transmission for the target monitoring meter box.

[0039] After obtaining the reporting priority weights of all monitoring and metering boxes within the system, the reporting priority weights are sorted in descending order of value to generate a one-dimensional global priority sequence. Pre-set emergency communication thresholds and regular communication thresholds are obtained, with the emergency communication threshold being strictly greater than the regular communication threshold. These two thresholds are used to divide the global priority sequence into non-overlapping intervals. Monitoring and metering boxes with reporting priority weights greater than or equal to the emergency communication threshold are placed in the emergency reporting queue; those with reporting priority weights less than the emergency communication threshold but greater than or equal to the regular communication threshold are placed in the regular reporting queue; and those with reporting priority weights strictly less than the regular communication threshold are placed in the delayed reporting queue. This strict unidirectional closed interval division method, including the equals sign, effectively avoids logical conflicts where a single value falls into two different queues simultaneously. After queuing, differentiated resource allocation instructions are executed for different queues. A dedicated, contention-free time slot is allocated to the monitoring and metering boxes in the emergency reporting queue, and the maximum allowed transmit power is configured to ensure that nodes experiencing severe electrical anomalies can penetrate network congestion and complete data reporting with extremely low latency and high reliability. For monitoring and metering boxes in the regular reporting queue, a multiple access channel based on carrier sense multiple access (CMA) is allocated, allowing them to compete for and transmit regular status data when the channel is idle. For monitoring and metering boxes in the delayed reporting queue, due to their steady-state state and limited network resources, they are instructed to enter a sleep waiting state, suspending data transmission for the current period to release wireless spectrum resources. All resource allocation instructions generated for the emergency reporting queue, regular reporting queue, and delayed reporting queue are encapsulated and combined using protocols to generate the final hierarchical reporting strategy, such as... Figure 2 The diagram shows a hierarchical reporting strategy based on graph attention mechanism.

[0040] The above content proposes a graph attention mechanism that integrates a queue delay penalty term. It evaluates node priority by aggregating graph tensor features and generates an adaptive hierarchical reporting strategy based on dual threshold partitioning. By introducing a queue delay penalty term based on modified edge weights into the graph attention mechanism, the model actively reduces its dependence on highly congested links when aggregating node features. This achieves deep joint perception of electrical anomalies and communication network congestion, improving the accuracy of priority evaluation and congestion resistance.

[0041] By using dual thresholds to precisely divide global nodes into three non-overlapping reporting queues and matching them with differentiated communication resources such as contention-free time slots, contention access, and sleep waiting, the absolute priority and high reliability of emergency abnormal data are guaranteed. At the same time, by actively suppressing the sending behavior of low-priority nodes, the collision conflicts and spectrum resource pressure of the overall network are effectively alleviated, and the adaptive scheduling efficiency of the power metering network in complex concurrent scenarios is improved.

[0042] Specifically, after evaluating the reporting priority of the monitoring metering boxes and generating a hierarchical reporting strategy through the above steps, although the system has allocated corresponding communication resources to nodes with different levels of urgency, in the actual power Internet of Things operating environment, aging failures of underlying sensor devices or transient strong electromagnetic interference can easily cause abnormal changes in the collected data. If these false high-priority abnormal data are directly connected to the network, not only will they ineffectively occupy valuable emergency communication time slots, but they will also cause the cloud master station to make incorrect scheduling decisions. At the same time, if the weight parameters of the graph attention mechanism in the previous steps remain static, it will be difficult to adapt to the slow drift of the underlying network topology and external interference in the long term. In order to solve the technical problems of false abnormal data occupying resources and the lack of real network feedback closed loop in the model, this embodiment introduces a data consistency verification mechanism based on the historical error model and a model dynamic update mechanism based on real communication costs.

[0043] S5. Verify the consistency of the data to be reported by the monitoring meter box using a preset historical error model. If the data to be reported fails the consistency verification, trigger the preset downgrade reporting strategy. If the data to be reported passes the consistency verification, detect the transient physical channel status of the target communication link in real time, and perform the allocation and reporting of the data to be reported according to the hierarchical reporting strategy and the transient physical channel status. Update the feature weight matrix inside the graph attention mechanism based on the actual packet loss rate and the actual communication delay value after the reporting is completed.

[0044] In S5, a preset historical error model is used to verify the consistency of the data to be reported by the monitoring meter box. If the data to be reported fails the consistency verification, a preset downgrade reporting strategy is triggered, including: extracting the historical feature sequence continuously collected by the monitoring meter box within a preset historical time window; inputting the historical feature sequence into a pre-trained time series prediction network to calculate and output the prediction benchmark value corresponding to the data to be reported; calculating the absolute difference between the data to be reported and the prediction benchmark value, and calculating the dynamic confidence interval threshold by combining the preset confidence scaling factor and the local standard deviation of the historical feature sequence, and comparing the absolute difference with the dynamic confidence interval threshold; if the absolute difference is greater than the dynamic confidence interval threshold, it is determined that the data to be reported has failed the consistency verification; when the data to be reported fails the consistency verification, a sensor anomaly identifier is added to the data to be reported, and the scheduling status of the monitoring meter box is forcibly reset; the monitoring meter box is removed from the current emergency reporting queue or the current regular reporting queue, and the monitoring meter box is reassigned to the delayed reporting queue, thus triggering the downgrade reporting strategy.

[0045] In S5, if the data to be reported passes the consistency verification, the transient physical channel state of the target communication link is detected in real time. Based on the hierarchical reporting strategy and the transient physical channel state, the data to be reported is allocated and reported. This includes: if the absolute difference is less than or equal to the dynamic confidence interval threshold, the data to be reported is determined to have passed the consistency verification; when the data to be reported passes the consistency verification, the instantaneous effective signal-to-noise ratio and link occupancy rate of the target communication link near the transmission time are obtained as the transient physical channel state; the modulation and coding scheme level adapted to the target communication link is determined based on the transient physical channel state; according to the resource allocation instruction corresponding to the emergency reporting queue or regular reporting queue to which the monitoring meter box belongs in the hierarchical reporting strategy, the data to be reported is protocol-encapsulated in conjunction with the modulation and coding scheme level to generate a reporting data frame; and the reporting data frame is sent to the cloud receiving node according to the transmission power and access mechanism indicated in the resource allocation instruction.

[0046] In S5, the feature weight matrix within the graph attention mechanism is updated based on the actual packet loss rate and actual communication delay after the reporting is completed. This includes: receiving the downlink communication confirmation message for the reported data frame returned by the cloud receiving node; extracting the receiving timestamp and missing data frame sequence number contained in the downlink communication confirmation message; calculating the actual communication delay and actual packet loss rate for the current reporting period by combining the sending time of the reported data frame and the missing data frame sequence number; inputting the actual communication delay and actual packet loss rate into a preset network performance cost function to calculate the overall error loss value corresponding to the actual network state; calculating the loss gradient of the overall error loss value relative to the feature weight matrix of the current network node in the graph attention mechanism with the goal of minimizing the overall error loss value; and updating the feature weight matrix within the graph attention mechanism by using the loss gradient along the backpropagation path.

[0047] Specifically, a pre-set historical error model is used to verify the consistency of the data to be reported. This model employs a pre-trained time-series prediction network to capture the evolution of power metering parameters over time. During offline training, a large amount of historical time-series data from the monitoring metering boxes under normal operating conditions is collected as training samples. The mean squared error is used as the loss function, and the weight parameters within the network are continuously optimized through backpropagation until the model can accurately fit the normal fluctuation trend of the data. In the online verification phase, historical feature sequences continuously collected from the monitoring metering boxes within a pre-set historical time window are extracted and input into the pre-trained time-series prediction network. Forward propagation is used to calculate and output the prediction benchmark value corresponding to the data to be reported, and the local standard deviation of the historical feature sequence within the pre-set historical time window is calculated simultaneously. The actual data to be reported is then calculated. The system calculates the dynamic confidence interval threshold at the current moment by multiplying the local standard deviation by the predicted baseline value and the predicted baseline value by obtaining a preset confidence scaling factor. The system then compares the absolute difference with the dynamic confidence interval threshold. If the calculated absolute difference is greater than the dynamic confidence interval threshold, it indicates that the sudden change in the current data significantly exceeds the normal evolution pattern of the physical equipment, and the data to be reported fails the consistency verification. When the data to be reported fails the consistency verification, a sensor anomaly flag is added to the data to be reported, and the scheduling status of the monitoring meter box in the underlying communication module is forcibly reset. The monitoring meter box is removed from the current emergency reporting queue or the current regular reporting queue, and the monitoring meter box is reassigned to the delayed reporting queue, thus triggering the downgraded reporting strategy and effectively preventing invalid erroneous data from preempting high-priority communication resources.

[0048] If the calculated absolute difference is less than or equal to the dynamic confidence interval threshold, the data to be reported is determined to have passed the consistency verification. When the data to be reported passes the consistency verification, the instantaneous effective signal-to-noise ratio and link occupancy rate of the target communication link near the transmission time are obtained as the transient physical channel state. The modulation and coding scheme level adapted to the target communication link is determined according to the transient physical channel state. For example, when the signal-to-noise ratio is high and the link occupancy rate is low, a higher-order modulation is selected to improve the rate. According to the resource allocation instructions corresponding to the emergency reporting queue or regular reporting queue to which the monitoring meter box belongs in the hierarchical reporting strategy, the data to be reported is protocol encapsulated in combination with the determined modulation and coding scheme level to generate a reporting data frame. The reporting data frame is sent to the cloud receiving node according to the transmission power and access mechanism indicated in the resource allocation instructions.

[0049] After the reporting action is completed, the downlink communication confirmation message for the reported data frame returned by the cloud receiving node is received. The receiving timestamp and the sequence number of the missing data frame contained in the downlink communication confirmation message are extracted. Combining the sending time of the reported data frame and the sequence number of the missing data frame, the actual communication delay value of the current reporting period is calculated by subtracting the sending timestamp from the receiving timestamp. The actual packet loss rate value is calculated by statistically analyzing the proportion of missing data frame sequence numbers. Since the actual communication delay value and the actual packet loss rate value differ in physical dimensions, a normalization method is used to map them to a dimensionless interval. The normalized actual communication delay value and the actual packet loss rate value are input into a preset network performance cost function. A linear weighted sum is performed using preset weighting coefficients to calculate the overall error loss value corresponding to the actual network state. With minimizing the overall error loss value as the optimization objective, the loss gradient of the overall error loss value relative to the feature weight matrix of the current network node in the graph attention mechanism is calculated using the chain rule. The loss gradient is used to update the feature weight matrix inside the graph attention mechanism along the backpropagation path, enabling the model to continuously adaptively correct the attention allocation strategy for network congestion.

[0050] The above content proposes a data consistency verification degradation mechanism based on time series prediction networks, and a method for updating graph attention network feature weights based on feedback closed loop of real communication delay and packet loss rate. By introducing time series prediction networks for consistency verification, false mutation data caused by sensor failure or electromagnetic interference can be accurately identified and intercepted. Combined with a forced degradation strategy, invalid dirty data is prevented from occupying emergency communication resources, thereby improving the success rate of reporting high-value data and the effective utilization of network bandwidth.

[0051] By extracting the real timestamp and packet loss sequence number from the cloud confirmation message, the actual communication delay and packet loss rate are quantitatively calculated and used as the weight matrix of the supervision signal backpropagation update graph attention mechanism. This constructs a complete closed loop from policy issuance to effect feedback, enabling the reporting scheduling model to continuously learn and adapt to the dynamic fading and topology changes of the underlying physical channel, thereby enhancing the robustness and scheduling accuracy of the system during long-term operation.

[0052] Through the coordination of steps S1 to S5 above, this application realizes deep joint perception and dynamic closed-loop scheduling of electrical abnormality and communication network congestion, effectively improving the anti-congestion capability of the power metering network in massive concurrent scenarios, the success rate of reporting high-value data, and the robustness of the system in long-term operation.

[0053] Example 2: The above describes an adaptive reporting method for the operating status of a metering box in an embodiment of this application. The following describes an adaptive reporting system for the operating status of a metering box in an embodiment of this application, such as... Figure 3 As shown in the figure, an adaptive reporting system for the operating status of a metering box in this application embodiment includes: The data acquisition module is used to acquire the data to be reported from the monitoring meter box and the corresponding sampling sequence, the spatial topology including global steady-state wireless channel quality indicators, the queue length of the access node, and the channel occupancy rate of the access node in different frequency bands.

[0054] The heterogeneous graph module is used to construct a weighted temporal feature vector by calculating the local volatility and mutation energy feature values ​​of the sampled sequence. The weighted temporal feature vector is used as the node feature, and combined with the connectivity between nodes and the original edge weights, a spatiotemporal network heterogeneous graph tensor is constructed.

[0055] The weight adjustment module is used to construct a dynamic queue tensor based on the queue length and the channel occupancy rate of the access node in different frequency bands, calculate the congestion penalty coefficient corresponding to the dynamic queue tensor, and use the congestion penalty coefficient to correct the original edge weights of the spatiotemporal network heterogeneous graph tensor to obtain the corrected edge weights.

[0056] The hierarchical decision module introduces a graph attention mechanism that includes a queue delay penalty term constructed based on modified edge weights. It uses the graph attention mechanism to process the heterogeneous graph tensor of the spatiotemporal network, calculates the reporting priority weight of each monitoring meter box, and generates a hierarchical reporting strategy based on the reporting priority weight.

[0057] The verification feedback module is used to verify the consistency of the data to be reported by the monitoring meter box using a preset historical error model. If the data to be reported fails the consistency verification, a preset downgrade reporting strategy is triggered. If the data to be reported passes the consistency verification, the transient physical channel status of the target communication link is detected in real time, and the data to be reported is allocated and reported according to the hierarchical reporting strategy and the transient physical channel status. The feature weight matrix inside the graph attention mechanism is updated based on the actual packet loss rate and the actual communication delay after the reporting is completed.

[0058] Through the collaborative efforts of the aforementioned components, this application further enhances the congestion resistance of the power metering network in high-concurrency scenarios and the success rate of reporting high-value data.

[0059] In summary, the adaptive reporting method and system for metering box operation status disclosed in this application, through multi-dimensional perception of end-side electrical status and network-side communication quality, constructs a spatiotemporal heterogeneous network graph tensor, and uses a dynamic queue tensor to calculate congestion penalty coefficients to correct link costs, breaking down the information barrier between service data and communication carrying capacity. Based on this, a graph attention mechanism incorporating queue delay penalty terms is introduced to accurately evaluate node reporting priorities and generate a dual-threshold hierarchical reporting strategy, rationally allocating differentiated communication resources, ensuring reliable transmission of emergency data and mitigating network conflicts. Furthermore, data consistency verification is performed through time-series prediction networks to intercept false and sudden data changes to prevent resource preemption, and a feedback loop is constructed using real communication delays and packet loss rates to continuously update model weights. This application not only achieves adaptive learning of the dynamic fading of the underlying physical channel but also significantly improves the overall scheduling efficiency and absolute reliability of data reporting in the power Internet of Things under complex concurrent environments.

[0060] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0061] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0062] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. An adaptive reporting method for the operating status of a metering box, characterized in that, The method includes: S1. Obtain the data to be reported from the monitoring meter box and the corresponding sampling sequence, including the spatial topology structure containing global steady-state wireless channel quality indicators, the queue length of the access node, and the channel occupancy rate of the access node in different frequency bands. S2. By calculating the local volatility and mutation energy characteristic values ​​of the sampled sequence, a weighted time feature vector is constructed. The weighted time feature vector is used as the node feature, and combined with the connectivity between nodes and the original edge weights, a spatiotemporal network heterogeneous graph tensor is constructed. S3. Construct a dynamic queue tensor based on the queue length and the channel occupancy rate of the access node in different frequency bands, calculate the congestion penalty coefficient corresponding to the dynamic queue tensor, and use the congestion penalty coefficient to correct the original edge weights of the spatiotemporal network heterogeneous graph tensor to obtain the corrected edge weights. S4. Introduce a graph attention mechanism that includes a queue delay penalty term constructed based on the corrected edge weights. Use the graph attention mechanism to process the spatiotemporal network heterogeneous graph tensor, calculate the reporting priority weight of each monitoring meter box, and generate a hierarchical reporting strategy based on the reporting priority weight. S5. Verify the consistency of the data to be reported by the monitoring meter box using a preset historical error model. If the data to be reported fails the consistency verification, trigger the preset downgrade reporting strategy. If the data to be reported passes the consistency verification, detect the transient physical channel status of the target communication link in real time, and perform the allocation and reporting of the data to be reported according to the hierarchical reporting strategy and the transient physical channel status. Update the feature weight matrix inside the graph attention mechanism based on the actual packet loss rate and the actual communication delay value after the reporting is completed.

2. The adaptive reporting method for the operating status of the metering box according to claim 1, characterized in that, S2 include: The local volatility is obtained by calculating the ratio of the standard deviation to the mean of the sampled sequence within a set time window; the sampled sequence is input into a preset mutation energy function to output mutation energy feature values; the local volatility and mutation energy feature values ​​are weighted and concatenated to generate a weighted time feature vector, and the weighted time feature vector is used as the node feature matrix. Based on the spatial topology containing global steady-state wireless channel quality indicators, the original edge weights representing link costs are extracted, and an adjacency matrix representing connectivity is generated. The queue length of the access node is fused with the original edge weights and mapped to the edge feature matrix of the communication link. The original edge weights are then extracted from the edge feature matrix. Tensors are stacked together along the time series dimension to construct a spatiotemporal network heterogeneous graph tensor.

3. The adaptive reporting method for the operating status of the metering box according to claim 2, characterized in that, Generate weighted time feature vectors, including: The sampling sequence is divided into a sliding time window. The standard deviation and mean of the sampling sequence within the sliding time window are calculated, and the ratio of the standard deviation to the mean is used as the local volatility. The sum of squares of the first-order differences of the sampling sequence between adjacent sampling points within the sliding time window is calculated, and the sum of squares of the first-order differences is used as the mutation energy characteristic value. Obtain the first weighting coefficient corresponding to the local volatility and the second weighting coefficient corresponding to the mutation energy feature value; concatenate the product of the local volatility and the first weighting coefficient, and the product of the mutation energy feature value and the second weighting coefficient, to generate a weighted time feature vector.

4. The adaptive reporting method for the operating status of the metering box according to claim 1, characterized in that, S3 includes: From the obtained channel occupancy rates, extract the target channel occupancy rates of the access nodes in each frequency band within the set of frequency bands; calculate the product of the queue length and the target channel occupancy rate to obtain the frequency band congestion feature values ​​of the access nodes in each frequency band; combine the frequency band congestion feature values ​​along the frequency band dimension to construct a dynamic queue tensor; Input the dynamic queue tensor into a preset nonlinear activation function, calculate the congestion penalty coefficient that maps the output to a set numerical range; extract the original edge weights corresponding to the access nodes in the spatiotemporal network heterogeneous graph tensor, multiply the original edge weights by the congestion penalty coefficient to obtain the corrected edge weights, and replace the original edge weights with the corrected edge weights.

5. The adaptive reporting method for the operating status of the metering box according to claim 1, characterized in that, In S4, a graph attention mechanism is used to process the spatiotemporal network heterogeneous graph tensor to calculate the reporting priority weight for each monitoring meter box, including: The target node feature vector and the neighbor node feature vector of the adjacent node connected to the target monitoring meter box are extracted from the spatiotemporal network heterogeneous graph tensor respectively; the target node feature vector and the neighbor node feature vector are concatenated and linearly transformed to calculate the basic attention coefficient. Obtain the corrected edge weights between the target monitoring meter box and its adjacent nodes. Calculate the queue delay penalty term based on the corrected edge weights and the preset penalty scaling factor. Subtract the queue delay penalty term from the basic attention coefficient to obtain the corrected attention coefficient. Then, normalize the corrected attention coefficient using a normalization function to obtain the attention weight of the target monitoring meter box relative to its adjacent nodes. The feature vectors of adjacent nodes are weighted and aggregated using attention weights to obtain the aggregated node state vector of the target monitoring meter box. The aggregated node state vector is then input into a preset fully connected network layer for dimensionality compression mapping, and the output scalar score is calculated. The scalar score is used as the reporting priority weight for each monitoring meter box.

6. The adaptive reporting method for the operating status of the metering box according to claim 1, characterized in that, S4 generates a tiered reporting strategy based on reporting priority weights, including: Obtain the reporting priority weights of all monitoring and metering boxes in the system, sort the reporting priority weights in descending order according to their numerical values, and generate a global priority sequence; obtain the preset emergency communication threshold and regular communication threshold, use the emergency communication threshold and regular communication threshold to divide the global priority sequence into intervals, and divide the monitoring and metering boxes into emergency reporting queue, regular reporting queue and delayed reporting queue respectively. Allocate dedicated contention-free time slots and maximum transmit power to monitoring and metering boxes in the emergency reporting queue; allocate carrier-sensing-based multiple access channels to monitoring and metering boxes in the regular reporting queue; instruct monitoring and metering boxes in the delayed reporting queue to enter a dormant waiting state; combine the resource allocation instructions corresponding to the emergency reporting queue, regular reporting queue, and delayed reporting queue to generate a hierarchical reporting strategy.

7. The adaptive reporting method for the operating status of the metering box according to claim 6, characterized in that, S5 uses a preset historical error model to verify the consistency of the data to be reported from the monitoring meter box. If the data to be reported fails the consistency verification, a preset downgrade reporting strategy is triggered, including: Extract historical feature sequences continuously collected by the monitoring and metering box within a preset historical time window; input the historical feature sequences into a pre-trained time series prediction network to calculate and output the prediction benchmark value corresponding to the data to be reported; calculate the absolute difference between the data to be reported and the prediction benchmark value, and calculate the dynamic confidence interval threshold by combining the preset confidence scaling factor and the local standard deviation of the historical feature sequences, and compare the absolute difference with the dynamic confidence interval threshold; if the absolute difference is greater than the dynamic confidence interval threshold, it is determined that the data to be reported has failed the consistency verification; when the data to be reported fails the consistency verification, add a sensor anomaly flag to the data to be reported, and forcibly reset the scheduling status of the monitoring and metering box; Remove the monitoring meter box from the current emergency reporting queue or the current regular reporting queue, and reassign the monitoring meter box to the delayed reporting queue to complete the triggering of the degraded reporting strategy.

8. The adaptive reporting method for the operating status of the metering box according to claim 7, characterized in that, In S5, if the data to be reported passes the consistency verification, the transient physical channel state of the target communication link is detected in real time, and the data to be reported is allocated and reported according to the hierarchical reporting strategy and the transient physical channel state, including: If the absolute difference is less than or equal to the dynamic confidence interval threshold, the data to be reported is determined to have passed the consistency verification. When the data to be reported passes the consistency verification, the instantaneous effective signal-to-noise ratio and link occupancy rate of the target communication link near the transmission time are obtained as the transient physical channel state. The modulation and coding scheme level adapted to the target communication link is determined according to the transient physical channel state. According to the resource allocation instructions corresponding to the emergency reporting queue or regular reporting queue to which the monitoring meter box belongs in the hierarchical reporting strategy, the data to be reported is protocol encapsulated in combination with the modulation and coding scheme level to generate a reporting data frame. The reporting data frame is sent to the cloud receiving node according to the transmission power and access mechanism indicated in the resource allocation instructions.

9. The adaptive reporting method for the operating status of the metering box according to claim 8, characterized in that, In S5, the feature weight matrix within the graph attention mechanism is updated based on the actual packet loss rate and actual communication delay values ​​after the reporting is completed. This includes: The system receives downlink communication acknowledgment messages for reported data frames returned by the cloud receiving node; extracts the receiving timestamp and missing data frame sequence number contained in the downlink communication acknowledgment message; combines the sending time of the reported data frame and the missing data frame sequence number to calculate the actual communication delay and actual packet loss rate for the current reporting period; inputs the actual communication delay and actual packet loss rate into a preset network performance cost function to calculate the overall error loss value corresponding to the actual network state; with minimizing the overall error loss value as the optimization objective, it calculates the loss gradient of the overall error loss value relative to the feature weight matrix of the current network node in the graph attention mechanism; and uses the loss gradient to update the feature weight matrix inside the graph attention mechanism along the backpropagation path.

10. A metering box operating status adaptive reporting system, used to implement the metering box operating status adaptive reporting method as described in any one of claims 1-9, characterized in that, The system includes: The data acquisition module is used to acquire the data to be reported from the monitoring meter box and the corresponding sampling sequence, the spatial topology including global steady-state wireless channel quality indicators, the queue length of the access node, and the channel occupancy rate of the access node in different frequency bands. The heterogeneous graph module is used to construct a weighted temporal feature vector by calculating the local volatility and mutation energy feature values ​​of the sampled sequence. The weighted temporal feature vector is used as the node feature, and combined with the connectivity between nodes and the original edge weights, a spatiotemporal network heterogeneous graph tensor is constructed. The weight adjustment module is used to construct a dynamic queue tensor based on the queue length and the channel occupancy rate of the access node in different frequency bands, calculate the congestion penalty coefficient corresponding to the dynamic queue tensor, and use the congestion penalty coefficient to correct the original edge weights of the spatiotemporal network heterogeneous graph tensor to obtain the corrected edge weights. The hierarchical decision module is used to introduce a graph attention mechanism that includes a queue delay penalty term constructed based on the modified edge weight. The graph attention mechanism is used to process the heterogeneous graph tensor of the spatiotemporal network, calculate the reporting priority weight of each monitoring meter box, and generate a hierarchical reporting strategy based on the reporting priority weight. The verification feedback module is used to verify the consistency of the data to be reported by the monitoring meter box using a preset historical error model. If the data to be reported fails the consistency verification, a preset downgrade reporting strategy is triggered. If the data to be reported passes the consistency verification, the transient physical channel status of the target communication link is detected in real time, and the data to be reported is allocated and reported according to the hierarchical reporting strategy and the transient physical channel status. The feature weight matrix inside the graph attention mechanism is updated based on the actual packet loss rate and the actual communication delay after the reporting is completed.

Images