A Method and System for 5G Wireless Public Network Data Completion Based on Spatiotemporal Graph Neural Network

Abstract

This invention relates to the field of data completion technology, and in particular to a method and system for 5G wireless public network data completion based on spatiotemporal graph neural networks. This invention constructs a hybrid graph integrating business logic and geographic location, and uses the gated recurrent unit of the spatiotemporal graph neural network to extract long-range dependency features of operational time-series data in the time dimension; it also uses the graph attention network of the spatiotemporal graph neural network to extract the spatial correlation features of operational time-series data determined by both business attributes and geographic location, enabling in-depth mining of the temporal and spatial dependencies between devices; it utilizes the homogeneity of similar business attributes for collaborative inference, effectively solving the problem of large-area data loss caused by network fluctuations or business concurrency; and it combines self-supervised mask training and business rule verification mechanisms to ensure the robustness of the model in unlabeled environments and the physical consistency of the output data.

Description

Technical Field

[0001] This invention relates to the field of data completion technology, and in particular to a method and system for 5G wireless public network data completion based on spatiotemporal graph neural networks. Background Technology

[0002] With the continuous advancement of the power Internet of Things and the construction of new power systems, fifth-generation mobile communication technology (5G) has been widely applied in power business scenarios such as distribution automation, precise load control, and electricity consumption information collection. However, the operating environment of 5G wireless public networks in the power industry is complex. Affected by factors such as power base station channel congestion, channel interference, terminal equipment failure, or edge computing node anomalies, the operational data such as traffic, signal strength, and service load uploaded by power terminal equipment inevitably have varying degrees of missing information.

[0003] Existing missing data completion techniques mainly employ methods such as mean imputation, linear interpolation, or traditional statistical models. These methods are effective when handling low missing data rates and simple missing patterns. However, they struggle to effectively capture the complex nonlinear variations inherent in power 5G wireless public network data, which exhibits high dimensionality, nonlinearity, and long-range dependencies. Furthermore, existing technologies typically process data uploaded by each power terminal device in isolation, performing completion processing separately for each device's data sequence. This approach focuses solely on the temporal variation patterns of data from a single power terminal device, resulting in high computational complexity and low processing efficiency. Especially when facing group data loss or batch data loss across multiple power terminal devices due to concentrated congestion of a particular service, regional interference, or network fluctuations, existing methods fail to meet practical application requirements in terms of speed and accuracy, hindering effective data recovery. Summary of the Invention

[0004] This invention aims to provide a 5G wireless public network data completion method and system based on spatiotemporal graph neural networks, in order to solve the technical problem that the data completion speed and accuracy of existing methods are difficult to meet the actual application requirements and cannot effectively achieve data recovery when facing the problem of collective data loss and batch data loss of multiple power terminal devices.

[0005] To achieve the above objectives, the first aspect of the present invention provides a 5G wireless public network data completion method based on a spatiotemporal graph neural network, comprising the following steps: For several device nodes in the power 5G wireless public network, obtain the operation time sequence data, service attribute data and geospatial data of each device node; Based on the operational time-series data, business attribute data, and geospatial data of each device node, an operational time-series data tensor, a business attribute data tensor, and a geospatial data tensor are constructed respectively. Identify missing data points in the operational time series data tensor, and then construct a binary mask tensor of the operational time series data tensor based on the missing data points; Construct a business attribute association graph about several device nodes based on the business attribute data tensor; Construct a geospatial adjacency graph of several device nodes based on the geospatial data tensor; The business attribute association map and the geospatial adjacency map are merged into a hybrid map; The operational time series data and the binary mask tensor are input into the gated recurrent unit of the pre-trained spatiotemporal graph neural network to obtain the hidden state feature vector matrix. The hybrid graph and the hidden state feature vector matrix are input into the graph attention network of a pre-trained spatiotemporal graph neural network to obtain the spatial aggregation feature vector matrix. The hidden state feature vector matrix and the spatial aggregation feature vector matrix are fused into a high-dimensional feature tensor. Then, the high-dimensional feature tensor is decoded into a completion tensor through a fully connected layer of a pre-trained spatiotemporal graph neural network. The completion tensor contains the completion value for each missing data point.

[0006] The aforementioned 5G wireless public network data completion method based on spatiotemporal graph neural networks is used to perform batch data completion on multiple operational time-series data uploaded by multiple power terminal devices. To address the technical problem of existing technologies failing to effectively recover data when faced with group data loss and batch data loss from multiple power terminal devices, this invention first constructs a business attribute association graph using the business attribute data of multiple power terminal devices. Through this graph, the potential collaborative changes between device nodes with the same business attributes due to business logic are captured. Specifically, operational data uploaded by power outage devices with the same business attributes exhibit similar data characteristics, and when a group of terminal uploads is missing, the data performance of terminal devices with similar business attributes shows high homogeneity. Therefore, when faced with group data loss from uploaded terminal devices, the business attribute association graph can be used to collaboratively shorten and batch complete the missing data from multiple terminal devices.

[0007] Secondly, this invention constructs a geospatial adjacency map using geospatial data from multiple power terminal devices. This geospatial adjacency map represents the correlation of missing data between geographically proximate device nodes due to similar physical environmental interference, thus linking power terminal devices susceptible to power supply anomalies in the same distribution area or severe weather in the same region. The missing data values ​​uploaded by these device nodes are typically homogeneous.

[0008] By fusing business attribute association graphs and geospatial adjacency graphs into a hybrid graph, this invention characterizes the intrinsic connections between different power terminal devices from two dimensions: business logical association and physical spatial proximity. This breaks through the shortcomings of existing technologies that perform data completion processing separately for data sequences of individual power terminal devices and only focus on the time-dimensional change patterns of data of individual power terminal devices for data completion. This enables the collaborative inference and completion of missing data by using spatiotemporal graph neural networks (ST-GNN) with information from associated device nodes when facing the problem of group data loss and batch data loss of multiple power terminal devices.

[0009] Specifically, this invention combines the Gated Recurrent Unit (GRU) of a spatiotemporal graph neural network to extract long-range dependency features of operational time-series data in the time dimension; and combines the Graph Attention Network (GAT) of the spatiotemporal graph neural network to extract the spatial correlation features of operational time-series data determined by business attributes and geographical location. Finally, by fusing the hidden state feature vector matrix and the spatial aggregation feature vector matrix to obtain a high-dimensional feature tensor, and using a fully connected layer for decoding, a completion tensor containing the imputation values ​​of all missing data points can be generated. This process utilizes the spatiotemporal complementarity of batch data loss, and can significantly improve the speed and accuracy of data completion when facing large-area batch data loss caused by regional interference or concentrated business congestion, effectively solving the problem that existing technologies cannot effectively recover data when faced with batch data loss.

[0010] Furthermore, obtaining the operational timing data of each of the device nodes includes: For any device node: Obtain several operational data sequences uploaded by the device node within a preset historical time window; Based on a preset time granularity, several operational data sequences are mapped to the same sampling frequency to obtain several aligned data sequences; Each of the aligned data sequences is subjected to data normalization processing to obtain several normalized data sequences; The normalized data sequences are integrated into the operational time-series data of the device node.

[0011] In this implementation, multiple operational data sequences of the device nodes within a preset historical time window are first obtained, such as uplink / downlink traffic values, network signal received power (RSRP), signal-to-noise ratio (SINR), and round-trip time (RTT). These operational data sequences may be missing to varying degrees due to factors such as power base station channel congestion, channel interference, terminal equipment failure, or edge computing node anomalies.

[0012] Since data from different sources have different sampling frequencies, such as traffic data every 15 minutes and signal strength every hour, this invention maps raw data with different sampling frequencies to the same sampling frequency based on a preset time granularity. This eliminates the time scale differences caused by inconsistent collection frequencies of different data sources and ensures the consistency and synchronization of input data for time series analysis.

[0013] Furthermore, this invention normalizes the aligned data, mapping operational data of different dimensions and orders of magnitude to the same numerical range. This eliminates the impact of different dimensions and value ranges on model training, enabling the spatiotemporal graph neural network model to learn the intrinsic features of the data more stably and efficiently, and preventing the model from biasing towards features with larger values ​​due to excessive differences in numerical ranges. Finally, the various normalized data sequences are integrated into operational time-series data, forming a comprehensive data tensor containing multi-dimensional information. This provides the spatiotemporal graph neural network with rich, aligned, and standardized input, helping to improve the convergence speed and prediction accuracy of the neural network model in capturing complex nonlinear changes and spatiotemporal dependencies in the data.

[0014] Furthermore, after fusing the hidden state feature vector matrix and the spatial aggregation feature vector matrix into a high-dimensional feature tensor, and then decoding the high-dimensional feature tensor into a completed tensor through a fully connected layer of a pre-trained spatiotemporal graph neural network, the process further includes: The completed tensor is subjected to inverse data normalization to obtain an inverse normalized tensor, which contains the completed value for each missing data point.

[0015] In this implementation, the input data is normalized during the training and inference phases to optimize learning performance, mapping all feature data distributions to a standard normal distribution with a mean of 0 and a variance of 1. Therefore, the completed tensor directly output by the model also falls within this normalized numerical range, rather than the original actual physical units. By performing inverse normalization, the normalized completed values ​​output by the model can be restored to operational data with actual physical meaning and units, thus obtaining the final completed values ​​that meet the requirements of actual engineering projects. These values ​​can be directly used in subsequent practical scenarios such as power system business analysis, billing statistics, or fault monitoring.

[0016] Furthermore, the business attribute data includes a business type identifier for the corresponding device node; the step of constructing a business attribute association graph about several device nodes based on the business attribute data tensor includes: By traversing several device nodes, logical connection edges are established between any two device nodes that have the same business type identifier, thereby obtaining a business attribute association graph of several device nodes.

[0017] In this implementation, a business attribute association graph is constructed using business type identifiers. This graph traverses all device nodes and establishes logical connections between any two nodes with the same business type identifier, accurately depicting the business logic relationships between power terminal devices. By constructing this graph, the graph neural network can treat device nodes with the same business attributes as related neighbors during subsequent spatial aggregation, guiding the model to use data features of similar normal devices for association inference. This allows the model to capture data change patterns at the business level, improving its ability to complete missing data with business homogeneity in batches of missing data. This effectively solves the problem of batch data loss in the power industry's 5G wireless public network.

[0018] Furthermore, the geospatial data includes the longitude, latitude, and altitude of the corresponding device nodes; the construction of a geospatial adjacency graph about several device nodes based on the geospatial data tensor includes: Calculate the spatial distance between two device nodes based on their longitude, latitude, and altitude. By traversing a number of device nodes, logical connection edges are established between any two device nodes whose spatial distance is less than a preset spatial distance threshold, thereby obtaining a geospatial adjacency map of the number of device nodes.

[0019] In this implementation, firstly, the precise spatial distance between any two device nodes is calculated based on their longitude, latitude, and altitude data, accurately reflecting the geographical proximity of power terminal devices in the power 5G wireless public network. Then, by setting a preset spatial distance threshold, a logical connection is established between two device nodes only when the spatial distance between them is less than this threshold. This construction method effectively filters out geographically proximate device nodes, which are more likely to face similar physical environmental changes or share the same network infrastructure in actual operation, such as being powered by the same transformer substation or affected by the same regional weather. Therefore, missing data in their uploaded operational data shows a significant correlation. This graph structure allows the model to utilize the data characteristics of spatially adjacent nodes to assist in recovering missing data, effectively addressing data loss caused by regional interference.

[0020] Furthermore, the step of inputting the hybrid graph and the hidden state feature vector matrix into the graph attention network of the pre-trained spatiotemporal graph neural network to obtain the spatial aggregation feature vector matrix includes: Based on the hybrid graph, identify all neighboring nodes corresponding to each device node; For any device node, the attention coefficients between the device node and any corresponding neighbor node are calculated based on the hidden state feature vector matrix and the graph attention network, thereby obtaining several attention coefficients. The spatial aggregated feature vector matrix is ​​obtained based on several attention coefficients.

[0021] In this implementation, firstly, based on the logical connections in the hybrid graph, all neighboring nodes with business logic or spatial proximity relationships are identified for each device node, thus clarifying the scope of spatial information propagation. Next, for any target device node, the model utilizes a graph attention network to dynamically calculate the attention coefficient between the target node and each of its neighboring nodes based on the hidden state feature vector matrices of the target node and its neighbors. Finally, the features of all neighboring nodes are weighted and summed based on the calculated attention coefficients to obtain the spatial aggregated feature vector matrix of the target node. This process enables the model to accurately capture the spatiotemporal correlations between data, significantly improving the accuracy of spatial feature extraction.

[0022] Furthermore, the training process of the pre-trained spatiotemporal graph neural network includes: For several device nodes in the power 5G wireless public network, obtain the historical operation data of each device node; Construct a historical operation data tensor based on the operation time-series data of each of the aforementioned device nodes; Randomly sample training data points from the operational time series data tensor according to a preset mask rate, and then construct a training mask tensor for the historical operational data tensor based on the training data points. The original spatiotemporal graph neural network is obtained, and then iteratively trained based on the historical operation data tensor and the training mask tensor to obtain the pre-trained spatiotemporal graph neural network.

[0023] In this implementation, a random mask is used to simulate data loss scenarios for model training. Specifically, historical operational data of device nodes is first acquired as training samples to construct a historical operational data tensor. Then, data points are randomly sampled according to a preset mask rate, and a training mask tensor is constructed, randomly creating "missing values" on the complete historical data. This simulates the diversity and randomness of data loss in real-world scenarios, allowing the model to fully learn the inherent distribution patterns, temporal dependencies, and spatial correlation patterns between devices in the operational data. The preset mask rate setting enables the model to adapt to data scenarios with different degrees of loss, ensuring that the pre-trained spatiotemporal graph neural network has good generalization ability and completion effect when facing complex real-world loss situations.

[0024] Further, after fusing the hidden state feature vector matrix and the spatial aggregation feature vector matrix into a high-dimensional feature tensor, and then decoding the high-dimensional feature tensor into a completed tensor through a fully connected layer of a pre-trained spatiotemporal graph neural network, the method further includes: Boundary constraint verification is performed on the completion value of each missing data point in the completion tensor to obtain several outlier values ​​that do not meet the preset boundary constraint conditions. Then, each outlier value is updated to a correction value according to the boundary constraint conditions.

[0025] In this implementation, to address the possibility that the neural network model might output values ​​deviating from the normal physical range under certain extreme conditions, a boundary constraint verification mechanism is introduced to check the physical rationality of the completed values. For the completed value of each missing data point in the completed tensor, based on the physical characteristics or business rules of its corresponding operational data type, it is determined whether the value conforms to a preset reasonable value range and is corrected accordingly. This effectively eliminates completed results that do not conform to physical common sense or power business logic, thereby ensuring the rationality and credibility of the completed data and preventing abnormal completed data from misleading subsequent power business analysis.

[0026] Further, after fusing the hidden state feature vector matrix and the spatial aggregation feature vector matrix into a high-dimensional feature tensor, and then decoding the high-dimensional feature tensor into a completed tensor through a fully connected layer of a pre-trained spatiotemporal graph neural network, the method further includes: Based on the completion value of each missing data point in the completed tensor, the variance of each missing data point in the completed tensor is obtained; For any missing data point in the completed tensor, if the variance of the missing data point exceeds a preset safety threshold, the completed value of the missing data point is updated to a corrected value according to a linear interpolation algorithm or a historical average algorithm.

[0027] In this implementation, after obtaining the imputation tensor, the reliability of the imputation value is evaluated by analyzing the variance of the imputation value for each missing data point within its neighboring time window. Specifically, if the variance calculated for the imputation value of a missing data point exceeds a preset safety threshold, it indicates a significant difference between the imputation value and its neighboring data, with excessive fluctuations, potentially indicating significant uncertainty or error in the model's prediction at that point. In this case, instead of directly using the model's output, traditional but highly stable methods such as linear interpolation or historical averages are employed to perform a secondary correction on the imputation value. This avoids prediction biases generated by the model in regions with scarce samples or ambiguous features, improving the reliability of the final output when the confidence level of the model's output imputation value for missing data is insufficient.

[0028] A second aspect of the present invention provides a 5G wireless public network data completion system based on a spatiotemporal graph neural network, comprising: The data acquisition module is used to acquire the operation time-series data, service attribute data and geospatial data of each of the several device nodes in the power 5G wireless public network. The data integration module is used to construct an operational time-series data tensor, a business attribute data tensor, and a geospatial data tensor based on the operational time-series data, business attribute data, and geospatial data of each device node, respectively. The mask tensor construction module is used to identify missing data points in the operational time series data tensor, and then construct a binary mask tensor of the operational time series data tensor based on the missing data points. A business attribute association graph construction module is used to construct a business attribute association graph about several device nodes based on the business attribute data tensor; A geospatial adjacency graph construction module is used to construct a geospatial adjacency graph about several device nodes based on the geospatial data tensor. The map fusion module is used to fuse the business attribute association map and the geospatial adjacency map into a hybrid map; The data completion module is used to perform the following steps: The operational time series data and the binary mask tensor are input into the gated recurrent unit of the pre-trained spatiotemporal graph neural network to obtain the hidden state feature vector matrix. The hybrid graph and the hidden state feature vector matrix are input into the graph attention network of a pre-trained spatiotemporal graph neural network to obtain the spatial aggregation feature vector matrix. The hidden state feature vector matrix and the spatial aggregation feature vector matrix are fused into a high-dimensional feature tensor. Then, the high-dimensional feature tensor is decoded into a completion tensor through a fully connected layer of a pre-trained spatiotemporal graph neural network. The completion tensor contains the completion value for each missing data point. The 5G wireless public network data completion method and system based on spatiotemporal graph neural networks provided by the present invention have at least the following advantages compared with the prior art: This invention employs a 5G wireless public network data completion method and system based on spatiotemporal graph neural networks (ST-GNN), aiming to achieve accurate repair of power 5G wireless public network data under high missing rate conditions. By constructing a multimodal association graph that integrates business logic and physical space, and combining it with a spatiotemporal graph neural network (ST-GNN), the system can deeply explore the temporal and spatial dependencies between devices. Utilizing the homogeneity of similar service devices for collaborative inference, it effectively solves the problem of large-area data loss caused by network fluctuations or service concurrency. Combined with a self-supervised mask training and business rule verification mechanism, it ensures the robustness of the model in an unlabeled environment and the physical consistency of the output data, thereby significantly improving the completeness, accuracy, and service availability of power 5G wireless public network operation data. Attached Figure Description

[0029] Figure 1 This is a flowchart illustrating a 5G wireless public network data completion method based on a spatiotemporal graph neural network provided in an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the principle of a 5G wireless public network data completion method based on a spatiotemporal graph neural network provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a 5G wireless public network data completion system based on a spatiotemporal graph neural network provided in an embodiment of the present invention. Detailed Implementation

[0030] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that the following detailed descriptions are exemplary and intended to provide further detailed explanation of the invention. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs; the terminology used herein in the specification is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings, are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings are used to distinguish different objects, not to describe a particular order.

[0031] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0032] Before describing the embodiments of this application in detail with reference to the accompanying drawings and examples, the terms and application scenarios involved in this application will first be explained.

[0033] With the construction of the power Internet of Things and new power systems, 5G communication technology, due to its high bandwidth, low latency, and wide connectivity, has been widely applied in business scenarios such as distribution automation, precise load control, and electricity consumption information collection. However, the power 5G wireless public network environment is complex. Affected by factors such as base station congestion, channel interference, terminal failure, or edge computing node anomalies, the uploaded operational data such as traffic, signal strength, and service load will inevitably have varying degrees of loss. Existing data completion methods mainly rely on mean filling, linear interpolation, or traditional statistical models. These methods are relatively effective when the data loss rate is low and the patterns are simple, but when faced with the high-dimensional, nonlinear, and long-range dependency characteristics of power 5G wireless public network data, the completion accuracy is difficult to meet the needs of refined operation management. In addition, existing methods often view each terminal device in isolation, lacking effective utilization of the potential service homogeneity and spatial topological correlation between devices, making it difficult to cope with the problem of batch data loss caused by network fluctuations or anomalies in specific service groups.

[0034] In one existing technology, a missing data-aware embedding module is constructed, using a mask matrix to distinguish between observed and missing data, thus avoiding noise interference caused by zero-padding. Simultaneously, a memory-enhanced encoder and a time-dimensional multi-scale encoder are constructed to capture global similarity, inter-variable dependencies, and multi-scale features in the time dimension, respectively, thereby reconstructing and completing the time-series data matrix. However, while this method considers inter-variable dependencies, it is essentially still based on modeling general time-series characteristics, neglecting the topological relationships of power 5G wireless public network devices in physical space and business logic. When faced with group data missingness caused by concentrated congestion of a certain type of service or regional interference, this method lacks collaborative inference capabilities based on homogeneous neighbors.

[0035] In another existing technology, a cloud server is used to perform fault detection on the collected time-series data. Faulty or missing data and its adjacent data are reconstructed using tensors to form tensor-based time-series data. Subsequently, a completion algorithm based on the core tensor norm of tensor decomposition is constructed, and pre-rank calculation is used to assist the HOOI algorithm in updating the factor matrix to achieve data completion. This method mainly targets key equipment within substations, utilizing the low-rank characteristics of the data for recovery. However, the tensor decomposition method has high computational complexity and is usually based on the linear low-rank assumption, making it difficult to capture the complex nonlinear spatiotemporal dynamic changes in power 5G wireless public network traffic data. Furthermore, this method does not address the multi-dimensional "service-space" coupling mechanism between devices in complex public network environments, and cannot leverage the homogeneity of similar service terminals in terms of traffic and behavior patterns to improve completion accuracy.

[0036] This invention aims to achieve intelligent collaborative data completion for missing data in the power 5G wireless public network by integrating the homogeneity association of service attributes with spatiotemporal graph neural network technology. A multimodal association graph of devices integrating service attribute topology and physical spatial location is constructed. Combined with spatiotemporal feature extraction algorithms and a self-supervised mask training mechanism, the spatiotemporal dependencies of power terminals within the same service group and geographical area are deeply mined. Neighbor node features are aggregated through a graph attention mechanism, and missing data is collaboratively inferred using the similarity of traffic patterns of devices with similar service attributes, dynamically addressing the problem of batch data loss caused by base station congestion or network fluctuations.

[0037] Please refer to Figure 1 To achieve the above objectives, the first embodiment of the present invention provides a 5G wireless public network data completion method based on a spatiotemporal graph neural network, comprising the following steps: S1. For several device nodes in the power 5G wireless public network, obtain the operation time sequence data, service attribute data and geospatial data of each device node; S2. Based on the operational time-series data, business attribute data, and geospatial data of each device node, construct the operational time-series data tensor, the business attribute data tensor, and the geospatial data tensor, respectively. S3. Identify missing data points in the operational time series data tensor, and then construct a binary mask tensor of the operational time series data tensor based on the missing data points; S4. Construct a business attribute association graph about several device nodes based on the business attribute data tensor; S5. Construct a geospatial adjacency graph of several device nodes based on the geospatial data tensor; S6. Merge the business attribute association map and the geospatial adjacency map into a hybrid map; S7. Input the operational time series data and the binary mask tensor into the gated recurrent unit of the pre-trained spatiotemporal graph neural network to obtain the hidden state feature vector matrix. S8. Input the hybrid graph and the hidden state feature vector matrix into the graph attention network of the pre-trained spatiotemporal graph neural network to obtain the spatial aggregation feature vector matrix. S9. The hidden state feature vector matrix and the spatial aggregation feature vector matrix are fused into a high-dimensional feature tensor. Then, the high-dimensional feature tensor is decoded into a completion tensor through a fully connected layer of a pre-trained spatiotemporal graph neural network. The completion tensor contains the completion value for each missing data point.

[0038] The aforementioned 5G wireless public network data completion method based on spatiotemporal graph neural networks is used to perform batch data completion on multiple operational time-series data uploaded by multiple power terminal devices. To address the technical problem of existing technologies failing to effectively recover data when faced with group data loss and batch data loss from multiple power terminal devices, this invention first constructs a business attribute association graph using the business attribute data of multiple power terminal devices. Through this graph, the potential collaborative changes between device nodes with the same business attributes due to business logic are captured. Specifically, operational data uploaded by power outage devices with the same business attributes exhibit similar data characteristics, and when a group of terminal uploads is missing, the data performance of terminal devices with similar business attributes shows high homogeneity. Therefore, when faced with group data loss from uploaded terminal devices, the business attribute association graph can be used to collaboratively shorten and batch complete the missing data from multiple terminal devices.

[0039] Secondly, this invention constructs a geospatial adjacency map using geospatial data from multiple power terminal devices. This geospatial adjacency map represents the correlation of missing data between geographically proximate device nodes due to similar physical environmental interference, thus linking power terminal devices susceptible to power supply anomalies in the same distribution area or severe weather in the same region. The missing data values ​​uploaded by these device nodes are typically homogeneous.

[0040] By fusing business attribute association graphs and geospatial adjacency graphs into a hybrid graph, this invention characterizes the intrinsic connections between different power terminal devices from two dimensions: business logical association and physical spatial proximity. This breaks through the shortcomings of existing technologies that perform data completion processing separately for data sequences of individual power terminal devices and only focus on the time-dimensional change patterns of data of individual power terminal devices for data completion. This enables the collaborative inference and completion of missing data by using spatiotemporal graph neural networks (ST-GNN) with information from associated device nodes when facing the problem of group data loss and batch data loss of multiple power terminal devices.

[0041] Specifically, this invention combines the Gated Recurrent Unit (GRU) of a spatiotemporal graph neural network to extract long-range dependency features of operational time-series data in the time dimension; and combines the Graph Attention Network (GAT) of the spatiotemporal graph neural network to extract the spatial correlation features of operational time-series data determined by business attributes and geographical location. Finally, by fusing the hidden state feature vector matrix and the spatial aggregation feature vector matrix to obtain a high-dimensional feature tensor, and using a fully connected layer for decoding, a completion tensor containing the imputation values ​​of all missing data points can be generated. This process utilizes the spatiotemporal complementarity of batch data loss, and can significantly improve the speed and accuracy of data completion when facing large-area batch data loss caused by regional interference or concentrated business congestion, effectively solving the problem that existing technologies cannot effectively recover data when faced with batch data loss.

[0042] Furthermore, obtaining the operational timing data of each of the device nodes includes: For any device node: Obtain several operational data sequences uploaded by the device node within a preset historical time window; Based on a preset time granularity, several operational data sequences are mapped to the same sampling frequency to obtain several aligned data sequences; Each of the aligned data sequences is subjected to data normalization processing to obtain several normalized data sequences; The normalized data sequences are integrated into the operational time-series data of the device node.

[0043] In a preferred embodiment, operational time-series data ( This is obtained by collecting dynamic operating indicators of power terminal equipment within a historical time window. Several types of operational data specifically include: uplink / downlink traffic values, network signal received power (RSRP), signal-to-noise ratio (SINR), round-trip time (RTT), and equipment online / offline status, etc.

[0044] Because data from different sources are sampled at different frequencies—for example, traffic data is sampled every 15 minutes, while signal strength data is sampled every hour—timestamp alignment is required.

[0045] Set a uniform time granularity For example, 15 minutes, for sampling frequencies higher than Data is aggregated using average values; for data with a sampling frequency lower than [missing information], [missing information] is used. The data is preserved using forward fill. Ultimately, all device data is mapped to time. At the same sampling frequency.

[0046] To eliminate the dimensional differences between various operational data such as bandwidth (MB), signal strength (dBm), and latency (ms) and accelerate model convergence, the Z-Score normalization method is used to process each feature dimension. For the first... The first device node Each feature attribute Or rather, the first Type of operational data sequence Its normalized calculation formula is: in, This is the mean of the operational data series over a historical time window. This represents the standard deviation of the operational data series within the historical time window. This is the normalized data sequence. After processing, the distribution of all feature data is mapped to the standard normal distribution interval with a mean of 0 and a variance of 1.

[0047] In this implementation, multiple operational data sequences of the device nodes within a preset historical time window are first obtained, such as uplink / downlink traffic values, network signal received power (RSRP), signal-to-noise ratio (SINR), and round-trip time (RTT). These operational data sequences may be missing to varying degrees due to factors such as power base station channel congestion, channel interference, terminal equipment failure, or edge computing node anomalies.

[0048] Since data from different sources have different sampling frequencies, such as traffic data every 15 minutes and signal strength every hour, this invention maps raw data with different sampling frequencies to the same sampling frequency based on a preset time granularity. This eliminates the time scale differences caused by inconsistent collection frequencies of different data sources and ensures the consistency and synchronization of input data for time series analysis.

[0049] Furthermore, this invention normalizes the aligned data, mapping operational data of different dimensions and orders of magnitude to the same numerical range. This eliminates the impact of different dimensions and value ranges on model training, enabling the spatiotemporal graph neural network model to learn the intrinsic features of the data more stably and efficiently, and preventing the model from biasing towards features with larger values ​​due to excessive differences in numerical ranges. Finally, the various normalized data sequences are integrated into operational time-series data, forming a comprehensive data tensor containing multi-dimensional information. This provides the spatiotemporal graph neural network with rich, aligned, and standardized input, helping to improve the convergence speed and prediction accuracy of the neural network model in capturing complex nonlinear changes and spatiotemporal dependencies in the data.

[0050] Furthermore, after fusing the hidden state feature vector matrix and the spatial aggregation feature vector matrix into a high-dimensional feature tensor, and then decoding the high-dimensional feature tensor into a completed tensor through a fully connected layer of a pre-trained spatiotemporal graph neural network, the process further includes: The completed tensor is subjected to inverse data normalization to obtain an inverse normalized tensor, which contains the completed value for each missing data point.

[0051] In this implementation, the input data is normalized during the training and inference phases to optimize learning performance, mapping all feature data distributions to a standard normal distribution with a mean of 0 and a variance of 1. Therefore, the completed tensor directly output by the model also falls within this normalized numerical range, rather than the original actual physical units. By performing inverse normalization, the normalized completed values ​​output by the model can be restored to operational data with actual physical meaning and units, thus obtaining the final completed values ​​that meet the requirements of actual engineering projects. These values ​​can be directly used in subsequent practical scenarios such as power system business analysis, billing statistics, or fault monitoring.

[0052] In a preferred embodiment, identifying missing data points in the operational time-series data tensor and then constructing a binary mask tensor of the operational time-series data tensor based on the missing data points includes: Define the input operational time series data tensor as ,in This represents the number of device nodes. For the time window step, The feature dimension is the type of operational data. Binary mask tensor. The element definition rules are as follows: Finally, this step outputs the processed operational time-series data tensor. (Missing parts are filled with 0) and the corresponding binary mask tensor This serves as the input for subsequent spatiotemporal graph neural network models.

[0053] Furthermore, the business attribute data includes a business type identifier for the corresponding device node; the step of constructing a business attribute association graph about several device nodes based on the business attribute data tensor includes: By traversing several device nodes, logical connection edges are established between any two device nodes that have the same business type identifier, thereby obtaining a business attribute association graph of several device nodes.

[0054] In a preferred embodiment, business attribute data ( This is obtained by acquiring static service attribute information of device nodes, which is used to distinguish the service characteristics of different devices. Specifically, it includes: device service type identifier Type ID, such as: 01-smart meter, 02-distribution transformer terminal and 03-high-definition video surveillance, etc.; communication module type such as NB-IoT, Cat.1 and 5G RedCap, etc.; and service data collection priority.

[0055] Based on business attribute data, construct an adjacency matrix describing the homogeneity of business logic among devices. Because devices belonging to the same service type (same Type ID) have similar traffic characteristics and the same service priority, such as the same data reporting frequency or packet size, when a collective anomaly occurs, such as a concentrated surge of a certain type of service causing congestion, the data performance of devices with the same type of service is highly synchronized. Therefore, if devices... With equipment Logical connection edges are established if they share the same business type identifier. The formula for calculating matrix elements is defined as follows: in, Indicates the first The service type identifier for each device, such as smart meters and video surveillance.

[0056] In this implementation, a business attribute association graph is constructed using business type identifiers. This graph traverses all device nodes and establishes logical connections between any two nodes with the same business type identifier, accurately depicting the business logic relationships between power terminal devices. By constructing this graph, the graph neural network can treat device nodes with the same business attributes as related neighbors during subsequent spatial aggregation, guiding the model to use data features of similar normal devices for association inference. This allows the model to capture data change patterns at the business level, improving its ability to complete missing data with business homogeneity in batches of missing data. This effectively solves the problem of batch data loss in the power industry's 5G wireless public network.

[0057] Furthermore, the geospatial data includes the longitude, latitude, and altitude of the corresponding device nodes; the construction of a geospatial adjacency graph about several device nodes based on the geospatial data tensor includes: Calculate the spatial distance between two device nodes based on their longitude, latitude, and altitude. By traversing a number of device nodes, logical connection edges are established between any two device nodes whose spatial distance is less than a preset spatial distance threshold, thereby obtaining a geospatial adjacency map of the number of device nodes.

[0058] In a preferred embodiment, geospatial data ( This is obtained by acquiring the static geographical location information of the device nodes, including longitude, latitude, altitude, as well as the power supply area number to which the device belongs and the physical base station (Cell ID) identifier to which it is connected.

[0059] Based on the GIS geographic location information of the devices, an adjacency matrix describing the physical proximity relationships between the devices is constructed. Computing devices With equipment The Euclidean distance between them. Set a spatial distance threshold. For example, if the distance between two devices is less than 500 meters, they are considered to be in similar physical environments (e.g., powered by the same transformer or affected by the same weather conditions), and a connection edge is established. The formula for calculating matrix elements is defined as: in, Represents a node and The Euclidean distance between latitude and longitude coordinates.

[0060] In this implementation, firstly, the precise spatial distance between any two device nodes is calculated based on their longitude, latitude, and altitude data, accurately reflecting the geographical proximity of power terminal devices in the power 5G wireless public network. Then, by setting a preset spatial distance threshold, a logical connection is established between two device nodes only when the spatial distance between them is less than this threshold. This construction method effectively filters out geographically proximate device nodes, which are more likely to face similar physical environmental changes or share the same network infrastructure in actual operation, such as being powered by the same transformer substation or affected by the same regional weather. Therefore, missing data in their uploaded operational data shows a significant correlation. This graph structure allows the model to utilize the data characteristics of spatially adjacent nodes to assist in recovering missing data, effectively addressing data loss caused by regional interference.

[0061] In a preferred embodiment, the step of fusing the business attribute association map and the geospatial adjacency map into a hybrid map includes: The two adjacency matrices are weighted and merged, and a self-loop (a connection between a node and itself) is introduced to generate the final multimodal hybrid adjacency matrix. The fusion formula is as follows: in, and These are the fusion weight hyperparameters for the physical space graph and the business homogeneity graph, respectively. They can be set as learnable parameters or preset fixed values, such as... ; for An identity matrix of order 1 is used to ensure that the model can retain the feature information of the nodes themselves during graph convolution.

[0062] The final constructed hybrid map is defined as ,in For a set of device nodes, This is the set of edges after merging.

[0063] In a preferred embodiment, the step of inputting the operational time-series data and the binary mask tensor into a gated recurrent unit of a pre-trained spatiotemporal graph neural network to obtain a hidden state feature vector matrix includes: The Hadamard product of the operational time-series data and the binary mask tensor is obtained to get the incomplete tensor input to the gated recurrent unit, as shown in the following formula: in, X For operational time-series data, M is a binary mask tensor.

[0064] Gated cyclic units (GRUs) are used to capture the long-term trends and short-term fluctuations of data from each power terminal device. The incomplete tensor... Input the GRU layer. GRU effectively mitigates the vanishing gradient problem in long sequences through reset and update gates, and outputs a hidden state feature vector matrix containing time-dependent information. The computational logic is expressed as follows: Among them, the hidden state feature vector matrix , The feature dimension of the model's hidden layer. The incomplete tensor of the input model, Representation matrix The first in row vectors This represents the relevant data for device node i in the incomplete tensor. This step ensures that the model can remember the normal operating patterns of the device over past time periods, such as daily load curves.

[0065] Furthermore, the step of inputting the hybrid graph and the hidden state feature vector matrix into the graph attention network of the pre-trained spatiotemporal graph neural network to obtain the spatial aggregation feature vector matrix includes: Based on the hybrid graph, identify all neighboring nodes corresponding to each device node; For any device node, the attention coefficients between the device node and any corresponding neighbor node are calculated based on the hidden state feature vector matrix and the graph attention network, thereby obtaining several attention coefficients. The spatial aggregated feature vector matrix is ​​obtained based on several attention coefficients.

[0066] In a preferred embodiment, a graph attention network (GAT) is used for message passing on a multimodal association graph, aggregating features of neighboring nodes to aid in the inference of the current node. For a node... and its neighboring nodes This includes geospatial neighbors and business attribute neighbors, and the attention coefficient is calculated. This coefficient represents the contribution of neighboring nodes to the recovery of the current node's state. For example, when a sudden surge in traffic for a certain type of service occurs, the weight of neighbors with the same type of service will automatically increase. The formula for calculating the attention coefficient is: in, The weight matrix is ​​a learnable matrix. This represents vector concatenation. For nodes The neighborhood group, Represents a node The time-series hidden state vector, i.e., the hidden state feature vector matrix. The first in row vectors It is a learnable weight vector (Attention Vector) for a single-layer feedforward neural network.

[0067] Based on the calculated weights, the neighborhood features are weighted and summed to obtain the spatial aggregated feature vector matrix. : This step leverages business homogeneity and spatial correlation, so that even if a device's data is completely lost, the model can infer its probable value based on the status of its similar business neighbors or spatial neighbors.

[0068] In this implementation, firstly, based on the logical connections in the hybrid graph, all neighboring nodes with business logic or spatial proximity relationships are identified for each device node, thus clarifying the scope of spatial information propagation. Next, for any target device node, the model utilizes a graph attention network to dynamically calculate the attention coefficient between the target node and each of its neighboring nodes based on the hidden state feature vector matrices of the target node and its neighbors. Finally, the features of all neighboring nodes are weighted and summed based on the calculated attention coefficients to obtain the spatial aggregated feature vector matrix of the target node. This process enables the model to accurately capture the spatiotemporal correlations between data, significantly improving the accuracy of spatial feature extraction.

[0069] In a preferred embodiment, the step of fusing the hidden state feature vector matrix and the spatial aggregation feature vector matrix into a high-dimensional feature tensor, and then decoding the high-dimensional feature tensor into a completed tensor through a fully connected layer of a pre-trained spatiotemporal graph neural network, includes: The extracted hidden state feature vector matrix Aggregate eigenvector matrix with space The features are concatenated and fused to form a high-dimensional feature tensor containing the complete spatiotemporal context. Then, a decoder consisting of fully connected (FC) layers maps the high-dimensional features back to the dimensions of the original data, generating a completed tensor containing the predicted values. The output formula is: Output dimension and input incomplete tensor The values ​​are consistent, but specific padding values ​​are generated at the locations of the original missing data points (i.e., at the mask).

[0070] Furthermore, the training process of the pre-trained spatiotemporal graph neural network includes: For several device nodes in the power 5G wireless public network, obtain the historical operation data of each device node; Construct a historical operation data tensor based on the operation time-series data of each of the aforementioned device nodes; Randomly sample training data points from the operational time series data tensor according to a preset mask rate, and then construct a training mask tensor for the historical operational data tensor based on the training data points. The original spatiotemporal graph neural network is obtained, and then iteratively trained based on the historical operation data tensor and the training mask tensor to obtain the pre-trained spatiotemporal graph neural network.

[0071] In a preferred embodiment, a self-supervised learning task is constructed to enable the model to learn the dependencies of power 5G wireless public network data in a multi-dimensional "time-space-service" topology without the need for external manual labels. Training is performed by simulating data loss patterns in real-world scenarios, giving the model the ability to reconstruct missing data.

[0072] Specifically, to simulate random packet loss due to channel interference and continuous packet loss due to base station congestion / specific service group anomalies, such as the concurrent online deployment of a large number of devices, in the power 5G wireless public network, a random masking mechanism is used to construct the training set. From the complete historical service operation dataset, a portion of data points are randomly sampled according to a preset masking rate (e.g., 20%), and their corresponding positions in the masking matrix are set to 0, while the remaining positions are set to 1, generating a training masking matrix specifically for training. The training data input to the model is the Hadamard product of the original data and the training mask: The positions set to 0 represent the targets that the model needs to predict using neighbor and historical information. During model training, this generated incomplete tensor... As the input tensor of the gated recurrent unit, let Substitute the values ​​into the calculation. The model first extracts temporal features through a gated recurrent unit, then aggregates spatial features using a graph attention network, and finally outputs a prediction tensor containing the completed values. This is used for subsequent loss function calculations.

[0073] To quantify the model's completion performance, a loss function based on mean squared error (MSE) is constructed. This function only calculates the prediction error at masked locations (i.e., manually removed missing data points), forcing the model to focus on recovering the missing information. The loss function is defined as follows: in: Original, authentic data. The predicted data output by the model. This is a reverse mask used to filter out the positions that are hidden by the mask. This represents the Frobenius norm. This is an L2 regularization term used to prevent model overfitting and enhance generalization ability in different business scenarios and network environments.

[0074] The Adam (Adaptive Moment Estimation) optimizer is used to optimize the loss function. Gradients are calculated using backpropagation to update the weight parameters in the spatiotemporal graph neural network, including the gating weights of the GRU and the attention mapping matrix in the GAT. The training process iterates continuously until the loss function converges or the preset number of epochs is reached. After training, the model acquires the ability to accurately infer the missing value of the current node based on the states of its surrounding business homogeneous neighbors and physical spatial neighbors.

[0075] In this implementation, a random mask is used to simulate data loss scenarios for model training. Specifically, historical operational data of device nodes is first acquired as training samples to construct a historical operational data tensor. Then, data points are randomly sampled according to a preset mask rate, and a training mask tensor is constructed, randomly creating "missing values" on the complete historical data. This simulates the diversity and randomness of data loss in real-world scenarios, allowing the model to fully learn the inherent distribution patterns, temporal dependencies, and spatial correlation patterns between devices in the operational data. The preset mask rate setting enables the model to adapt to data scenarios with different degrees of loss, ensuring that the pre-trained spatiotemporal graph neural network has good generalization ability and completion effect when facing complex real-world loss situations.

[0076] Further, after fusing the hidden state feature vector matrix and the spatial aggregation feature vector matrix into a high-dimensional feature tensor, and then decoding the high-dimensional feature tensor into a completed tensor through a fully connected layer of a pre-trained spatiotemporal graph neural network, the method further includes: Boundary constraint verification is performed on the completion value of each missing data point in the completion tensor to obtain several outlier values ​​that do not meet the preset boundary constraint conditions. Then, each outlier value is updated to a correction value according to the boundary constraint conditions.

[0077] In a preferred embodiment, a set of rule constraints based on data physical properties is constructed to apply to the completed tensor generated by the decoder. Point-by-point verification is performed. If the predicted value exceeds the preset physical boundary, truncation or correction is enforced to prevent the algorithm from outputting values ​​that violate physical principles. The main constraint rules include: Non-negativity constraint: Perform non-negativity checks on physical quantities such as traffic, number of connections, and packet loss. If the predicted value... If the value is negative, it should be corrected to 0 to prevent errors that violate physical principles, such as negative flow.

[0078] Physical limit constraints: For signal strength (RSRP), define a physical effective range (e.g., If the predicted value exceeds this range, it is considered that the model is diverging or abnormal, and the point is marked as a state to be corrected.

[0079] Terminal service capability constraints: combined with the acquired service attribute data ( The theoretical maximum transmission rate threshold is set according to the type of communication module equipped in the device (such as NB-IoT, Cat.1, 5G). For example, the theoretical upper limit of the uplink speed of an NB-IoT water meter is approximately 200kbps. If the model predicts the instantaneous flow rate at a certain moment... This is clearly an algorithmic illusion, so it should be forcibly corrected to... This is to prevent the generation of false data that exceeds the physical transmission capacity of the terminal.

[0080] In this implementation, to address the possibility that the neural network model might output values ​​deviating from the normal physical range under certain extreme conditions, a boundary constraint verification mechanism is introduced to check the physical rationality of the completed values. For the completed value of each missing data point in the completed tensor, based on the physical characteristics or business rules of its corresponding operational data type, it is determined whether the value conforms to a preset reasonable value range and is corrected accordingly. This effectively eliminates completed results that do not conform to physical common sense or power business logic, thereby ensuring the rationality and credibility of the completed data and preventing abnormal completed data from misleading subsequent power business analysis.

[0081] Further, after fusing the hidden state feature vector matrix and the spatial aggregation feature vector matrix into a high-dimensional feature tensor, and then decoding the high-dimensional feature tensor into a completed tensor through a fully connected layer of a pre-trained spatiotemporal graph neural network, the method further includes: Based on the completion value of each missing data point in the completed tensor, the variance of each missing data point in the completed tensor is obtained; For any missing data point in the completed tensor, if the variance of the missing data point exceeds a preset safety threshold, the completed value of the missing data point is updated to a corrected value according to a linear interpolation algorithm or a historical average algorithm.

[0082] In a preferred embodiment, to prevent the model from generating unreliable predictions with excessively large variance in cases of extreme data loss (such as continuous data loss exceeding 24 hours), a confidence assessment and fallback strategy are introduced. The variance or uncertainty score of each data point in the model's output completed tensor is calculated within a local time window. If the variance exceeds a set safety threshold, indicating low confidence in the model's inference for that point, the system automatically downgrades and activates the fallback mechanism, using linear interpolation or historical averages to replace and correct the outlier. in: This is the final output padding value for the missing point; This is the predicted output value generated by the ST-GNN model for this point; The variance of the model's predicted values ​​within a local time window centered on the current time (such as three time steps before and after) is used to quantify the uncertainty of the prediction. This is the preset variance safety threshold (confidence threshold). Indicates the original observation sequence based on the equipment. The linear interpolation result obtained from the known data points is used as a fallback value under low confidence.

[0083] This mechanism ensures the system's robustness in the face of extreme situations such as model failure, and guarantees the basic availability of data.

[0084] In this implementation, after obtaining the imputation tensor, the reliability of the imputation value is evaluated by analyzing the variance of the imputation value for each missing data point within its neighboring time window. Specifically, if the variance calculated for the imputation value of a missing data point exceeds a preset safety threshold, it indicates a significant difference between the imputation value and its neighboring data, with excessive fluctuations, potentially indicating significant uncertainty or error in the model's prediction at that point. In this case, instead of directly using the model's output, traditional but highly stable methods such as linear interpolation or historical averages are employed to perform a secondary correction on the imputation value. This avoids prediction biases generated by the model in regions with scarce samples or ambiguous features, improving the reliability of the final output when the confidence level of the model's output imputation value for missing data is insufficient.

[0085] Please refer to Figure 2In one specific application embodiment, the present invention first collects operational time-series data, business attribute data, and geospatial data from power terminal equipment. After alignment and normalization, standard operational time-series data tensors, business attribute data tensors, and geospatial data tensors are constructed to complete multi-source data fusion and preprocessing. A two-layer heterogeneous hybrid graph is constructed based on physical spatial adjacency and business attribute homogeneity logical clustering to capture the dual influence of the environment and business side, thus constructing a "business-space" two-layer association graph. Subsequently, a spatiotemporal graph neural network (ST-GNN) is used for collaborative inference, through gating... The control loop unit extracts the device's own temporal features and combines them with graph attention network to aggregate the features of service homogeneity and spatial adjacency. It uses the regularity of similar devices to accurately infer missing values ​​and completes the spatiotemporal feature collaborative inference. Finally, the prediction results are validated according to service rules such as nonnegativity and terminal physical rate limits. For outliers with low confidence, linear interpolation is automatically used to correct them, completing the rule correction / confidence fallback for outliers that fail the test. The final output complete time series data is written into the database to achieve high-precision repair and persistent storage of power 5G wireless public network data.

[0086] Please refer to Figure 3 The second embodiment of the present invention provides a 5G wireless public network data completion system based on a spatiotemporal graph neural network, comprising: The data acquisition module 100 is used to acquire the operation time sequence data, business attribute data and geospatial data of each of the several device nodes in the power 5G wireless public network. The data integration module 200 is used to construct an operation time-series data tensor, a business attribute data tensor, and a geospatial data tensor based on the operation time-series data, business attribute data, and geospatial data of each device node, respectively. The mask tensor construction module 300 is used to identify missing data points in the operation time series data tensor, and then construct a binary mask tensor of the operation time series data tensor based on the missing data points. The business attribute association graph construction module 400 is used to construct a business attribute association graph about several device nodes based on the business attribute data tensor; The geospatial adjacency graph construction module 500 is used to construct a geospatial adjacency graph about several device nodes based on the geospatial data tensor. The map fusion module 600 is used to fuse the business attribute association map and the geospatial adjacency map into a hybrid map; Data completion module 700 is used to perform the following steps: The operational time series data and the binary mask tensor are input into the gated recurrent unit of the pre-trained spatiotemporal graph neural network to obtain the hidden state feature vector matrix. The hybrid graph and the hidden state feature vector matrix are input into the graph attention network of a pre-trained spatiotemporal graph neural network to obtain the spatial aggregation feature vector matrix. The hidden state feature vector matrix and the spatial aggregation feature vector matrix are fused into a high-dimensional feature tensor. Then, the high-dimensional feature tensor is decoded into a completion tensor through a fully connected layer of a pre-trained spatiotemporal graph neural network. The completion tensor contains the completion value for each missing data point.

[0087] It should be noted that traditional methods for completing 5G wireless public network data in the power sector typically rely on simple statistical models or univariate time series analysis, using mean imputation, linear interpolation, or rules based on historical data. These methods have drawbacks including: insufficient ability to capture high-dimensional nonlinear features of the data, resulting in low completion accuracy; focusing only on the time-dimensional patterns of the devices themselves, ignoring the strong correlation between power terminals in terms of business logic topology and physical spatial distribution; lacking collaborative inference capabilities when facing batch or continuous data loss due to base station congestion or anomalies in specific service groups, making effective data recovery difficult; and lacking effective constraints on the physical boundaries of the services, resulting in completed data that may violate actual physical principles and lead to poor data usability.

[0088] Building upon this foundation, this invention employs a collaborative missing data completion method based on business homogeneity association and spatiotemporal graph neural networks, aiming to achieve accurate repair of power 5G wireless public network data under high missing rate conditions. By constructing a multimodal association graph that integrates business attributes and physical space, and combining it with a spatiotemporal graph neural network (ST-GNN), the temporal and spatial dependencies between power terminal devices can be deeply explored. Collaborative inference using the homogeneity of similar business devices effectively solves the problem of large-area data loss caused by network fluctuations or business concurrency. Combined with self-supervised mask training and business rule verification mechanisms, the robustness of the model in an unlabeled environment and the physical consistency of the output data are ensured, thereby significantly improving the completeness, accuracy, and service availability of power 5G wireless public network operation data.

[0089] The key technical point of this invention lies in the collaborative data completion method for missing data in power 5G wireless public networks based on service homogeneity association and spatiotemporal graph neural networks. The core key points are the proposed construction of a "service-space" two-layer heterogeneous association graph, collaborative inference of spatiotemporal features, and a closed-loop verification mechanism combining service rules. By constructing a multimodal association graph that integrates physical location and terminal service logic, and combining spatiotemporal graph neural networks to deeply mine the time dependence and service group homogeneity between devices, the method utilizes neighborhood features to collaboratively infer missing data and correct physical boundaries, effectively improving the data completion accuracy and service availability of power 5G wireless public networks in complex network environments.

[0090] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0091] The term "embodiment" as used herein means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; however, any combination of these technical features that does not contradict each other should be considered within the scope of this specification.

[0092] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various improvements and substitutions without departing from the concept of this application, and these improvements and substitutions should also be considered within the scope of protection of this invention. Therefore, the scope of protection of this application should be determined by the appended claims.

Claims

1. A method for 5G wireless public network data completion based on spatiotemporal graph neural networks, characterized in that, include: For several device nodes in the power 5G wireless public network, obtain the operation time sequence data, service attribute data and geospatial data of each device node; Based on the operational time-series data, business attribute data, and geospatial data of each device node, an operational time-series data tensor, a business attribute data tensor, and a geospatial data tensor are constructed respectively. Identify missing data points in the operational time series data tensor, and then construct a binary mask tensor of the operational time series data tensor based on the missing data points; Construct a business attribute association graph about several device nodes based on the business attribute data tensor; Construct a geospatial adjacency graph of several device nodes based on the geospatial data tensor; The business attribute association map and the geospatial adjacency map are merged into a hybrid map; The operational time series data and the binary mask tensor are input into the gated recurrent unit of the pre-trained spatiotemporal graph neural network to obtain the hidden state feature vector matrix. The hybrid graph and the hidden state feature vector matrix are input into the graph attention network of a pre-trained spatiotemporal graph neural network to obtain the spatial aggregation feature vector matrix. The hidden state feature vector matrix and the spatial aggregation feature vector matrix are fused into a high-dimensional feature tensor. Then, the high-dimensional feature tensor is decoded into a completion tensor through a fully connected layer of a pre-trained spatiotemporal graph neural network. The completion tensor contains the completion value for each missing data point.

2. The 5G wireless public network data completion method based on spatiotemporal graph neural network according to claim 1, characterized in that, The acquisition of operational time-series data for each device node includes: For any device node: Obtain several operational data sequences uploaded by the device node within a preset historical time window; Based on a preset time granularity, several operational data sequences are mapped to the same sampling frequency to obtain several aligned data sequences; Each of the aligned data sequences is subjected to data normalization processing to obtain several normalized data sequences; The normalized data sequences are integrated into the operational time-series data of the device node.

3. The 5G wireless public network data completion method based on spatiotemporal graph neural network according to claim 2, characterized in that, After fusing the hidden state feature vector matrix and the spatial aggregation feature vector matrix into a high-dimensional feature tensor, and then decoding the high-dimensional feature tensor into a completed tensor through a fully connected layer of a pre-trained spatiotemporal graph neural network, the method further includes: The completed tensor is subjected to inverse data normalization to obtain an inverse normalized tensor, which contains the completed value for each missing data point.

4. The 5G wireless public network data completion method based on spatiotemporal graph neural network according to claim 1, characterized in that, The business attribute data includes the business type identifier of the corresponding device node; the construction of a business attribute association graph about several device nodes based on the business attribute data tensor includes: By traversing several device nodes, logical connection edges are established between any two device nodes that have the same business type identifier, thereby obtaining a business attribute association graph of several device nodes.

5. The 5G wireless public network data completion method based on spatiotemporal graph neural network according to claim 1, characterized in that, The geospatial data includes the longitude, latitude, and altitude of the corresponding device nodes; the construction of a geospatial adjacency graph of several device nodes based on the geospatial data tensor includes: Calculate the spatial distance between two device nodes based on their longitude, latitude, and altitude. By traversing a number of device nodes, logical connection edges are established between any two device nodes whose spatial distance is less than a preset spatial distance threshold, thereby obtaining a geospatial adjacency map of the number of device nodes.

6. The 5G wireless public network data completion method based on spatiotemporal graph neural network according to claim 1, characterized in that, The step of inputting the hybrid graph and the hidden state feature vector matrix into a pre-trained spatiotemporal graph neural network's graph attention network to obtain a spatial aggregation feature vector matrix includes: Based on the hybrid graph, identify all neighboring nodes corresponding to each device node; For any device node, the attention coefficients between the device node and any corresponding neighbor node are calculated based on the hidden state feature vector matrix and the graph attention network, thereby obtaining several attention coefficients. The spatial aggregated feature vector matrix is ​​obtained based on several attention coefficients.

7. The 5G wireless public network data completion method based on spatiotemporal graph neural network according to claim 1, characterized in that, The training process of the pre-trained spatiotemporal graph neural network includes: For several device nodes in the power 5G wireless public network, obtain the historical operation data of each device node; Construct a historical operation data tensor based on the operation time-series data of each of the aforementioned device nodes; Randomly sample training data points from the operational time series data tensor according to a preset mask rate, and then construct a training mask tensor for the historical operational data tensor based on the training data points. The original spatiotemporal graph neural network is obtained, and then iteratively trained based on the historical operation data tensor and the training mask tensor to obtain the pre-trained spatiotemporal graph neural network.

8. The 5G wireless public network data completion method based on spatiotemporal graph neural network according to claim 1, characterized in that, After fusing the hidden state feature vector matrix and the spatial aggregation feature vector matrix into a high-dimensional feature tensor, and then decoding the high-dimensional feature tensor into a completed tensor through a fully connected layer of a pre-trained spatiotemporal graph neural network, the method further includes: Boundary constraint verification is performed on the completion value of each missing data point in the completion tensor to obtain several outlier values ​​that do not meet the preset boundary constraint conditions. Then, each outlier value is updated to a correction value according to the boundary constraint conditions.

9. The 5G wireless public network data completion method based on spatiotemporal graph neural network according to claim 1, characterized in that, After fusing the hidden state feature vector matrix and the spatial aggregation feature vector matrix into a high-dimensional feature tensor, and then decoding the high-dimensional feature tensor into a completed tensor through a fully connected layer of a pre-trained spatiotemporal graph neural network, the method further includes: Based on the completion value of each missing data point in the completed tensor, the variance of each missing data point in the completed tensor is obtained; For any missing data point in the completed tensor, if the variance of the missing data point exceeds a preset safety threshold, the completed value of the missing data point is updated to a corrected value according to a linear interpolation algorithm or a historical average algorithm.

10. A 5G wireless public network data completion system based on spatiotemporal graph neural networks, characterized in that, include: The data acquisition module is used to acquire the operation time-series data, service attribute data and geospatial data of each of the several device nodes in the power 5G wireless public network. The data integration module is used to construct an operational time-series data tensor, a business attribute data tensor, and a geospatial data tensor based on the operational time-series data, business attribute data, and geospatial data of each device node, respectively. The mask tensor construction module is used to identify missing data points in the operational time series data tensor, and then construct a binary mask tensor of the operational time series data tensor based on the missing data points. A business attribute association graph construction module is used to construct a business attribute association graph about several device nodes based on the business attribute data tensor; A geospatial adjacency graph construction module is used to construct a geospatial adjacency graph about several device nodes based on the geospatial data tensor. The map fusion module is used to fuse the business attribute association map and the geospatial adjacency map into a hybrid map; The data completion module is used to perform the following steps: The operational time series data and the binary mask tensor are input into the gated recurrent unit of the pre-trained spatiotemporal graph neural network to obtain the hidden state feature vector matrix. The hybrid graph and the hidden state feature vector matrix are input into the graph attention network of a pre-trained spatiotemporal graph neural network to obtain the spatial aggregation feature vector matrix. The hidden state feature vector matrix and the spatial aggregation feature vector matrix are fused into a high-dimensional feature tensor. Then, the high-dimensional feature tensor is decoded into a completion tensor through a fully connected layer of a pre-trained spatiotemporal graph neural network. The completion tensor contains the completion value for each missing data point.

Images