Communication base station flow prediction management system based on deep learning

Abstract

The invention relates to the technical field of communication management, and provides a communication base station traffic prediction management system based on deep learning, which comprises a multi-source data acquisition module used for acquiring base station space-time traffic data, user behavior data, network state data, external influence factors and data set slice service parameters; the spatio-temporal feature processing module is used for performing spatio-temporal alignment, noise filtering and slice feature coding on the multi-source data; and the dynamic model prediction module is used for constructing a multi-modal prediction network of a space-time Transform, a graph neural network and a slice exclusive sub-model. A base station association graph based on geographical distance and service correlation is constructed, spatial features are extracted through a multilayer graph convolutional network, slice features and spatial-temporal features are fused by using a gating mechanism, a Pareto optimal strategy is generated by using a multi-objective optimization algorithm, and a strategy library is updated in combination with a forgetting factor, so that the probability of forgetting is reduced. The response time of the system in an abnormal scene is shortened, and the strategy optimization efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of communication management, and in particular, to a communication base station traffic prediction management system based on deep learning. Background Art

[0002] In the field of communication networks, base station traffic management faces challenges such as strong spatio-temporal dynamics, complex multi-source data fusion, insufficient anomaly detection accuracy, and low resource allocation efficiency. On the one hand, traditional methods mostly use simple coordinate mapping or fixed grid division, without fully considering the geometric characteristics of the base station spatial distribution and the temporal dynamic association of traffic data, resulting in insufficient spatio-temporal feature fusion accuracy and difficulty in effectively capturing the spatio-temporal laws of traffic fluctuations.

[0003] On the other hand, traditional methods mostly rely on fixed thresholds or simple statistical methods to detect abnormal traffic, lack the adaptive ability to the dynamic changes of data distribution, and the outlier correction strategy does not fully combine historical data with local features, easily leading to data distortion or untimely correction. Summary of the Invention

[0004] Aiming at the problems existing in the prior art, the purpose of the present invention is to provide a communication base station traffic prediction management system based on deep learning to solve the problems proposed in the above background art.

[0005] To achieve the above purpose, the present invention provides a communication base station traffic prediction management system based on deep learning, including:

[0006] A multi-source data acquisition module, used to collect base station spatio-temporal traffic data, user behavior data, network status data, external influence factors, and dataset slicing service parameters;

[0007] A spatio-temporal feature processing module, used to perform spatio-temporal alignment, noise filtering, and slice feature encoding on multi-source data;

[0008] A dynamic model prediction module, used to construct a multi-modal prediction network of spatio-temporal Transformer, graph neural network, and slice-specific sub-models, and output base station-level and slice-level traffic prediction results;

[0009] An edge-cloud collaboration module, used for hierarchical collaboration of lightweight prediction at edge nodes and global optimization at the cloud.

[0010] Preferably, the spatio-temporal alignment method of the spatio-temporal feature processing module includes the following steps:

[0011] S21. Map the base station longitude and latitude coordinates to the hexagonal network coordinates through projective transformation , construct a mapping matrix to assist in the location information processing of the communication base station traffic prediction management system, and the formula is:

[0012]

[0013]

[0014] In the formula, is the mapping matrix, which is optimized by minimizing the Wasserstein distance ; is the original coordinate distribution, is the target grid distribution;

[0015] S22. Approximately solve the Wasserstein distance through the Sinkhorn algorithm to obtain the optimal transport matrix , which provides a basis for the optimal allocation and transmission strategy formulation of traffic resources. The formula is:

[0016]

[0017] In the formula, is the Gibbs kernel, and , is the cost matrix, is the regularization parameter, and are both probability orientations that satisfy the marginal constraints;

[0018] S23. Calculate the DTW distance between the time series and . Use path constraints to reduce the computational complexity and provide support for the similarity measurement and dynamic trend analysis of historical and predicted traffic time series. The formula is:

[0019]

[0020] In the formula, is the dynamic time warping distance between the -th point of the time series and the -th point of the time series . is the data point difference degree between the -th point of the time series and the -th point of the time series . is the set of allowed path step sizes, and , is the path step;

[0021] S24. Align the spatial features and the time features Through the fusion of the attention mechanism, it helps with traffic prediction. The formula is:

[0022]

[0023]

[0024] In the formula, is the fused feature vector, is the attention weight, is the sigmoid function, is the learnable weight vector;

[0025] S25. Perform noise filtering and slicing on the data after feature fusion to purify the data and extract effective segments, improving the accuracy and management efficiency of traffic prediction.

[0026] Preferably, in step S25, performing noise filtering and slicing on the data after feature fusion includes the following steps:

[0027] S251. Train the GI-GAN generator to repair missing data, and the discriminator to distinguish real data from generated data . The formula for the generator loss function is:

[0028]

[0029] In the formula, is the noise vector, is the gradient penalty coefficient, is the reconstruction loss weight, is the gradient operator;

[0030] S252. Set traffic constraints to identify and correct abnormal traffic values, filter out abnormal fluctuations in slice-level traffic data, and improve the accuracy of prediction and management;

[0031] S253. Combine statistical methods and deep learning to detect anomalies, perform multi-scale anomaly detection, and calculate the Z-score and autoencoder reconstruction error. The formula is:

[0032]

[0033]

[0034] In the formula, is the traffic of slice at time , is the traffic of slice at time The Z-score value, is the average flow of the slice ; is the standard deviation of the flow of the slice ; is the slice at the moment the reconstruction error of the autoencoder, is the slice at the moment of the original flow data, is the slice at the moment the flow data after reconstruction by the autoencoder;

[0035] S254. Dynamically update the anomaly threshold according to the flow fluctuation , adaptively adjust the anomaly detection criterion, and the formula is:

[0036]

[0037] In the formula, and are the updated and pre-updated anomaly thresholds respectively, is the learning rate.

[0038] Preferably, in the step S252, dealing with the abnormal fluctuation in the slice-level flow data includes the following steps:[[ID=4%]]

[0039] S2521. Determine the constraint condition and judge whether there is an anomaly in the slice-level flow value. The formula is:

[0040]

[0041] In the formula, is the flow at the moment , is the proportionality coefficient, is the slice the maximum flow;

[0042] S2522. Check whether the flow value violates the above constraint. If it violates, regard as an outlier. If it meets the constraint, go to step S253;

[0043] S2523. For the outlier, correct to improve the data quality and assist in reasonable management decisions. The formula is:

[0044]

[0045] In the formula, is the historical weight, For slicing The set of traffic values within the window

[0046] Preferably, the steps for the dynamic model prediction module to construct a multi-modal prediction network are as follows:

[0047] S31. Construct a base station association graph , fuse location and traffic information through node features calculate the edge weights Based on geographical distance and service correlation, construct a graph structure reflecting the spatial association of base stations. The formula is:

[0048]

[0049] In the formula, is the geographical distance, is the traffic correlation, and are both weights;

[0050] S32. Update the node representation of the base station association graph through a multi-layer graph convolutional network combined with residual connection and layer normalization to extract more effective spatial features. The formula is:

[0051]

[0052] In the formula, and are the feature representation vectors of node at the th layer and th layer respectively, is the feature representation vector of node at the th layer, and node is a node in the neighbor set of node , is the neighbor set of node , is the weight matrix of the th layer, is the normalization coefficient, and , is the neighbor set of node , is the activation function;

[0053] S33. Use a gating mechanism to fuse slice features and spatio-temporal features to make the feature representation more comprehensive. The formula is:

[0054]

[0055] ​

[0056] is the slice feature vector, is the spatio-temporal feature vector, is the feature vector obtained after fusing the slice feature and the spatio-temporal feature, is the gating weight, is the activation function, is the weight matrix of the gating;

[0057] S34. According to the predicted traffic and the service quality requirements, dynamically calculate the slice bandwidth allocation to enable the elastic and reasonable allocation of slice resources. The formula is:

[0058]

[0059] In the formula, and are respectively the predicted traffic and the bandwidth allocation amount of the slice , is the total bandwidth, is the slice 's maximum resource share, is the traffic-bandwidth conversion relationship parameter, is the slice 's signal-to-noise ratio;

[0060] S35. Use the proximal policy optimization algorithm to update the policy network and optimize the base station resource scheduling policy. The formula is:

[0061]

[0062]

[0063] In the formula, and are respectively the current policy network and the historical policy network, is the probability ratio, is to measure the degree of advantage of taking the action in the state compared with the average action, is the clipping parameter that limits the transformation range, is the expectation operator, is the objective function for evaluating the performance of the policy network.

[0064] Preferably, in the step S34, after dynamically calculating the slice bandwidth allocation, the following steps are further included:

[0065] S341. Introduce a prediction uncertainty penalty, establish a reward function, comprehensively consider throughput, energy consumption, and interference, provide a quantitative evaluation criterion during bandwidth allocation, and optimize the accuracy of slice bandwidth allocation. The formula is as follows:

[0066]

[0067] In the formula, 、 、 and are the reward value, throughput, energy consumption, and interference at time respectively. is the variance of the predicted traffic, and 、 、 and are the corresponding weight coefficients respectively;

[0068] S342. Establish a revenue function, determine the service quality satisfaction, bandwidth allocation utilization, and interference among slices, and provide a decision-making basis for subsequent optimization of the base station resource scheduling strategy. The formula is:

[0069]

[0070]

[0071] In the formula, is the comprehensive revenue function value for measuring slice , is the quality satisfaction of slice , and are the actually allocated and maximum available bandwidth resources of slice respectively, is the service delay of slice , is the number of service requests of slice ;

[0072] S343. Detect slice anomalies, timely discover abnormal fluctuations in the traffic data of communication base stations, ensure the quality of communication services, and ensure the stable operation of the system.

[0073] Preferably, in step S343, detecting slice anomalies includes the following steps:

[0074] S3431. Calculate the slice traffic prediction residual, standardize the anomaly score, and quantify the traffic prediction deviation. The formula is:

[0075]

[0076] In the formula, For slicing At time The standardized anomaly score, And Are respectively the slicing At time The true and predicted traffic data, Is the sliding window size for defining the time range, For slicing At time The mean predicted residual at that time;

[0077] S3432. Combine the anomalies in the time series, spatial, and inter-slice dimensions, calculate the multi-dimensional comprehensive anomaly index, and accurately locate the anomalies. The formula is:

[0078]

[0079] In the formula, Is to reflect the slicing The overall anomaly degree value under multiple dimensions, For slicing The spatial anomaly situation value, For slicing And slicing The traffic correlation anomaly situation value, , Are respectively the corresponding weights;

[0080] S3433. Obtain the slice status flag, and dynamically adjust the anomaly detection threshold based on historical anomalies, so that the system can adapt to traffic changes and improve the accuracy and effectiveness of anomaly detection. The formula is:

[0081]

[0082] In the formula, And Are respectively the slicing At time And time The anomaly detection thresholds, And Are respectively the learning rate for controlling the threshold adjustment step and the adjustment factor for adjusting the threshold amplitude, For slicing The status flag;

[0083] S3434. Establish a preview of the faults caused by slice anomalies, judge the impact of different slice anomalies on policy allocation, and optimize and update the policy library.

[0084] Preferably, in the step S3434, establishing a preview of the faults caused by slice anomalies includes the following steps:

[0085] S34341. Update the slice digital twin status based on the current status and exception parameters, and provide a basis for simulating fault scenarios and analyzing fault impacts through the state transition function. The formula is:

[0086]

[0087] In the formula, and are the state vectors of the slice at time and time respectively, is the exception parameter vector, is the state transition function;

[0088] S34342. Calculate the slice service degradation rate, quantify the impact degree of exceptions on the slice service quality, and conduct a quantitative assessment of exceptions. The formula is:

[0089]

[0090] In the formula, is the proportion reflecting the decline in the slice service quality caused by exceptions, and are the service qualities of the slice in the normal and abnormal states respectively;

[0091] S34343. Based on multi-objective optimization, calculate the weighted Euclidean distance of the service degradation rate and resource consumption under different strategies, and balance the service quality and resource cost. The formula is:

[0092]

[0093] In the formula, is the Pareto optimal strategy, is the strategy set, is the resource consumption for strategy execution, and are the corresponding weight coefficients respectively;

[0094] S34344. Based on the forgetting factor, perform exponential weighted update on the strategy library to ensure that the strategy library adapts to traffic changes and improve the system management efficiency. The formula is:

[0095]

[0096] In the formula, and are the th strategies in the updated and pre-updated strategy libraries respectively, is the adjusted new strategy, is the update coefficient, is the forgetting factor, It is the duration of policy usage.

[0097] Preferably, the edge-cloud collaboration module's quantization prediction for edge nodes and global optimization in the cloud include the following steps:

[0098] S41. Conduct self-check on edge nodes to generate a resource profile. Meanwhile, the cloud conducts knowledge distillation on the complete prediction model, compresses the model to a scale that can be run on edge nodes, and verifies the deployment success rate.

[0099] S42. Execute the pre-configured lightweight policy according to the prediction results, record the deviation between the execution result and the actual traffic, and generate a local execution log for subsequent evaluation.

[0100] S43. The cloud conducts global optimization to complete load balancing.

[0101] S44. The cloud transfers the processing knowledge of the global model to edge nodes through knowledge distillation. Edge nodes share industry common knowledge through transfer learning to generate a joint response strategy across edge nodes.

[0102] S45. Continuously monitor the prediction accuracy of edge nodes, the global optimization effect in the cloud, task processing latency, and communication overhead, synchronize the execution status and perform incremental model updates to adapt to changes in the network environment.

[0103] Preferably, in step S43, the cloud's global optimization includes the following steps:

[0104] S431. Edge nodes regularly upload prediction results, execution logs, and key status indicators to the cloud to construct a global view, analyze cross-regional traffic trends, interference relationships between slices, and resource allocation balance.

[0105] S432. Based on cross-edge node collaboration, formulate a slice resource reconfiguration plan to generate a global optimal strategy.

[0106] S433. Send the updated model parameters and knowledge distillation guidance signals to edge nodes. Edge nodes report the current CPU / GPU utilization rate, memory occupancy, and the length of the task queue to be processed in real time. Meanwhile, the cloud predicts the computing pressure of each edge node in the future period.

[0107] S434. Allocate tasks, migrate complex prediction tasks to the cloud for execution, and let edge nodes handle simple tasks. When the load on edge nodes is too high, trigger an early warning mechanism and automatically migrate simple tasks to the cloud for execution.

[0108] The beneficial effects of a communication base station traffic prediction and management system based on deep learning provided by the present invention are:

[0109] 1. By optimizing the mapping matrix based on the Wasserstein distance, introducing the DTW distance to measure time series similarity, and combining the attention mechanism to fuse spatiotemporal features, it is possible to dynamically capture the delayed alignment relationship of traffic data in the time dimension and the geometric correlation in the spatial dimension, significantly improving the representation ability of spatiotemporal features. Through adversarial training between the generator and the discriminator, it effectively fills the data gap. Based on the outlier correction strategy of dynamically updated anomaly thresholds and historical weights, it can adaptively adjust the detection standard according to real-time traffic fluctuations, and correct outliers through weighted fusion of historical data and local medians, avoiding the hysteresis of fixed thresholds and the ambiguity of simple mean correction.

[0110] 2. By constructing a base station association graph based on geographical distance and business relevance, extracting spatial features through a multi-layer graph convolutional network, and using a gating mechanism to fuse slice features and spatiotemporal features, the slice bandwidth allocation is dynamically calculated based on the predicted traffic and service quality requirements. The policy network is updated in combination with the proximal policy optimization algorithm. This can achieve elastic resource allocation under multi-objective constraints such as throughput, energy consumption, and interference. By calculating the standardized anomaly score of the slice traffic prediction residual and combining multi-dimensional comprehensive anomaly indicators in the time series, space, and inter-slice dimensions, abnormal slices can be accurately located. Based on the slice anomaly rehearsal mechanism of digital twins, the state transition function is used to simulate fault scenarios, quantify the service degradation rate, and use a multi-objective optimization algorithm to generate a Pareto optimal strategy. Combined with the forgetting factor to update the policy library, the system response time in abnormal scenarios is shortened and the policy optimization efficiency is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0111] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0112] Figure 1 A schematic diagram of the system modules of a communication base station traffic prediction and management system based on deep learning provided by this application;

[0113] Figure 2 A schematic diagram of the operation flow of the spatiotemporal feature processing module of a communication base station traffic prediction and management system based on deep learning provided in this application;

[0114] Figure 3 A schematic diagram of the dynamic model prediction operation flow of a communication base station traffic prediction management system based on deep learning provided in this application;

[0115] Figure 4A schematic diagram of the operation flow of the edge-cloud collaboration module of a communication base station traffic prediction management system based on deep learning provided in this application;

[0116] Figure 5 A schematic diagram of the slice anomaly preview and strategy update operation flow of a deep learning-based communication base station traffic prediction management system provided in this application. DETAILED DESCRIPTION

[0117] The following embodiments of the present invention are described in further detail in conjunction with the accompanying drawings and examples. The following embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0118] like Figures 1 - 5 As shown, this embodiment proposes a communication base station traffic prediction and management system based on deep learning, including:

[0119] Multi-source data acquisition module, used to collect base station spatiotemporal traffic data, user behavior data, network status data, external influencing factors, and data set slice service parameters;

[0120] The spatiotemporal feature processing module is used to perform spatiotemporal alignment, noise filtering, and slice feature encoding on multi-source data;

[0121] The dynamic model prediction module is used to build a multimodal prediction network consisting of spatiotemporal Transformer, graph neural network, and slice-specific sub-models, and output base station-level and slice-level traffic prediction results;

[0122] The edge-cloud collaboration module is used for hierarchical collaboration of lightweight prediction of edge nodes and global optimization in the cloud.

[0123] In this embodiment, the spatiotemporal alignment method of the spatiotemporal feature processing module includes the following steps:

[0124] S21, the base station longitude and latitude coordinates Mapping to hexagonal grid coordinates using a radial transformation , construct a mapping matrix to assist in the location information processing of the communication base station traffic prediction management system. The formula is:

[0125]

[0126]

[0127] In the formula, For the mapping matrix, by minimizing the Wasserstein distance optimization, is the original coordinate distribution, is the target grid distribution;

[0128] S22. Approximately solve the Wasserstein distance through the Sinkhorn algorithm to obtain the optimal transport matrix, providing a basis for the optimal allocation of traffic resources and the formulation of transmission strategies. The formula is: , providing a basis for the optimal allocation of traffic resources and the formulation of transmission strategies. The formula is:

[0129]

[0130] In the formula, is the Gibbs kernel, and , is the cost matrix, is the regularization parameter, and are both probability orientations that satisfy the marginal constraints;

[0131] S23. Calculate the DTW distance between the time series and , and use path constraints to reduce the computational complexity, providing support for the similarity measurement and dynamic trend analysis of historical and predicted traffic time series. The formula is:

[0132]

[0133] In the formula, is the dynamic time warping distance between the th point of the time series and the th point of the time series , is to measure the difference degree of data points between the th point of the time series and the th point of the time series , is the set of allowed path step sizes, and , is the path step size;

[0134] S24. Fuse the aligned spatial features and temporal features through the attention mechanism to assist in traffic prediction. The formula is:

[0135]

[0136]

[0137] In the formula, is the fused feature vector, is the attention weight, is the sigmoid function, is the learnable weight vector;

[0138] S25. Filter and slice the data after feature fusion to purify the data and extract valid segments, improving the accuracy of traffic prediction and management efficiency.

[0139] In this embodiment, in step S25, filtering and slicing the data after feature fusion includes the following steps:

[0140] S251. Train the GI-GAN generator Repair missing data, discriminator Distinguish real data From generated data , and the generator loss function formula is:

[0141]

[0142] In the formula, is the noise vector, is the gradient penalty coefficient, is the reconstruction loss weight, is the gradient operator;

[0143] S252. Set traffic constraints to identify and correct abnormal traffic values, filter out abnormal fluctuations in slice-level traffic data, and improve the accuracy of prediction and management;

[0144] S253. Combine statistical methods and deep learning to detect anomalies, perform multi-scale anomaly detection, and calculate the Z-score and autoencoder reconstruction error. The formula is:

[0145]

[0146]

[0147] In the formula, is the traffic of slice at time , is the Z-score value of slice at time , is the traffic mean of slice , is the traffic standard deviation of slice , is the reconstruction error of the autoencoder for slice at time , is the original traffic data of slice at time , is the traffic of slice at time Traffic data after reconstruction by the autoencoder;

[0148] S254. Dynamically update the anomaly threshold according to traffic fluctuations , and adaptively adjust the anomaly detection criteria. The formula is:

[0149]

[0150] In the formula, and are the updated and pre-updated anomaly thresholds respectively, is the learning rate.

[0151] In this embodiment, in step S252, dealing with the abnormal fluctuations in the slice-level traffic data includes the following steps:

[0152] S2521. Determine the constraint conditions and judge whether there are anomalies in the slice-level traffic value. The formula is:

[0153]

[0154] In the formula, is the traffic at time , is the proportionality coefficient, is the slice 's maximum traffic;

[0155] S2522. Check whether the traffic value violates the above constraints. If it violates, regard as an outlier. If the constraints are met, enter step S253;

[0156] S2523. For the outlier, correct to improve the data quality and assist in reasonable management decisions. The formula is:

[0157]

[0158] In the formula, is the historical weight, is the slice 's traffic value set within the window .

[0159] Specifically, by optimizing the mapping matrix based on the Wasserstein distance, introducing the DTW distance to measure the similarity of time series, and combining the attention mechanism to fuse spatio-temporal features, it can dynamically capture the delayed alignment relationship of traffic data in the time dimension and the geometric correlation in the space dimension, significantly improving the representation ability of spatio-temporal features. Through the adversarial training of the generator and discriminator, it effectively fills the data gap, and based on the anomaly threshold updated dynamically and the anomaly value correction strategy of historical weights, it can adaptively adjust the detection standard according to real-time traffic fluctuations, and correct the anomaly value through the weighted fusion of historical data and local median, avoiding the lag of fixed thresholds and the ambiguity of simple mean correction.

[0160] In this embodiment, the dynamic model prediction module constructs a multi-modal prediction network, including the following steps:

[0161] S31. Construct a base station association graph , through node features fuse location and traffic information, and calculate the edge weights Based on geographical distance and service relevance, construct a graph structure reflecting the spatial association of base stations. The formula is:

[0162]

[0163] In the formula, is the geographical distance, is the traffic relevance, and are both weights;

[0164] S32. Through a multi-layer graph convolutional network combined with residual connection and layer normalization, update the node representation of the base station association graph, and extract more effective spatial features. The formula is:

[0165]

[0166] In the formula, and are the feature representation vectors of node at the th layer and th layer respectively, is the feature representation vector of node at the th layer, and node is a node in the neighbor set of node , is the neighbor set of node , is the weight matrix of the th layer, is the normalization coefficient, and , is node The neighbor set, is the activation function;

[0167] S33. Use the gating mechanism to fuse the slice features and spatio-temporal features to make the feature representation more comprehensive. The formula is:

[0168]

[0169]

[0170] is the slice feature vector, is the spatio-temporal feature vector, is the feature vector obtained after fusing the slice features and spatio-temporal features, is the gating weight, is the activation function, is the weight matrix of the gating;

[0171] S34. Dynamically calculate the slice bandwidth allocation according to the predicted traffic and service quality requirements to achieve elastic and reasonable allocation of slice resources. The formula is:

[0172]

[0173] In the formula, and are the predicted traffic and bandwidth allocation amount of slice respectively, is the total bandwidth, is the maximum resource share of slice , is the traffic-bandwidth conversion relationship parameter, is the signal-to-noise ratio of slice ;

[0174] S35. Use the proximal policy optimization algorithm to update the policy network and optimize the base station resource scheduling policy. The formula is:

[0175]

[0176]

[0177] In the formula, and are the current policy network and the historical policy network respectively, is the probability ratio, is to measure the degree of advantage of taking action compared with the average action in state , is the clipping parameter that limits the transformation range, is the expectation operator, is the objective function for evaluating the performance of the policy network.

[0178] In this embodiment, in step S34, after dynamically calculating the slice bandwidth allocation, the following steps are further included:

[0179] S341. Introduce prediction uncertainty penalty, establish a reward function, comprehensively consider throughput, energy consumption, and interference, provide a quantitative evaluation criterion during bandwidth allocation, and optimize the accuracy of slice bandwidth allocation. The formula is;

[0180]

[0181] In the formula, , , and are the reward value, throughput, energy consumption, and interference at time respectively, is the variance of the predicted traffic, , , and are the corresponding weight coefficients;

[0182] S342. Establish a revenue function, determine the service quality satisfaction, bandwidth allocation utilization, and interference between slices, and provide a decision-making basis for subsequent optimization of the base station resource scheduling strategy. The formula is:

[0183]

[0184]

[0185] In the formula, is the comprehensive revenue function value for measuring slice , is the quality satisfaction of slice , and are the actually allocated and the maximum available bandwidth resources of slice respectively, is the service delay of slice , is the number of service requests of slice ;

[0186] S343. Detect slice anomalies, timely discover abnormal fluctuations in the traffic data of the communication base station, ensure the quality of communication services, and ensure the stable operation of the system.

[0187] In this embodiment, in step S343, detecting slice anomalies includes the following steps:

[0188] S3431. Calculate the slice traffic prediction residual, standardize the anomaly score, and quantify the traffic prediction deviation. The formula is:

[0189]

[0190] In the formula, is the slice at time The standardized anomaly score, and are the true and predicted traffic data of the slice at time respectively. is the size of the sliding window within the limited time range, is the slice at time The mean prediction residual;

[0191] S3432. Considering the anomalies in the time series, space, and slice dimensions comprehensively, calculate the multi-dimensional comprehensive anomaly index to accurately locate the anomalies. The formula is:

[0192]

[0193] In the formula, reflects the overall anomaly degree value of the slice under multiple dimensions, is the spatial anomaly situation value of the slice , is the slice and the slice The traffic correlation anomaly situation value, , are the corresponding weights respectively;

[0194] S3433. Obtain the slice status flag, and dynamically adjust the anomaly detection threshold according to the historical anomaly situation, so that the system can adapt to the traffic changes and improve the accuracy and effectiveness of anomaly detection. The formula is:

[0195]

[0196] In the formula, and are the anomaly detection thresholds of the slice at time and time respectively, and are the learning rate for controlling the threshold adjustment step and the adjustment factor for adjusting the threshold amplitude respectively, is the slice Status flag;

[0197] S3434. Establish a preview of failures caused by slice anomalies, determine the impact of different slice anomalies on policy allocation, and optimize and update the policy library.

[0198] In this embodiment, in step S3434, establishing a preview of failures caused by slice anomalies includes the following steps:

[0199] S34341. Update the slice digital twin state based on the current state and anomaly parameters, and provide a basis for simulating failure scenarios and analyzing failure impacts through the state transition function. The formula is:

[0200]

[0201] In the formula, and are the state vectors of the slice at time and time respectively, is the anomaly parameter vector, is the state transition function;

[0202] S34342. Calculate the slice service degradation rate to quantify the impact of anomalies on the slice service quality and conduct a quantitative assessment of the anomalies. The formula is:

[0203]

[0204] In the formula, is the proportion reflecting the decrease in slice service quality caused by anomalies, and are the service qualities of the slice in normal and abnormal states respectively;

[0205] S34343. Based on multi-objective optimization, calculate the weighted Euclidean distance of the service degradation rate and resource consumption under different policies to balance service quality and resource costs. The formula is:

[0206]

[0207] In the formula, is the Pareto optimal policy, is the policy set, is the resource consumption of policy execution, and are the corresponding weight coefficients respectively;

[0208] S34344. Perform exponential weighted update on the policy library based on the forgetting factor to ensure that the policy library adapts to traffic changes and improve the system management efficiency. The formula is:

[0209]

[0210] In the formula, and are the th policies in the updated and pre-updated policy libraries respectively, is the newly adjusted policy, is the update coefficient, is the forgetting factor, is the usage duration of the policy.

[0211] In this embodiment, the edge-cloud collaboration module's quantization prediction for edge nodes and global optimization for the cloud include the following steps:

[0212] S41. Conduct self-check on edge nodes to generate a resource profile. Meanwhile, the cloud conducts knowledge distillation on the complete prediction model, compresses the model to a scale that can be run on edge nodes, and verifies the deployment success rate;

[0213] S42. Execute the pre-configured lightweight policy according to the prediction result, record the deviation between the execution result and the actual traffic, and generate a local execution log for subsequent evaluation;

[0214] S43. The cloud conducts global optimization to complete load balancing;

[0215] S44. The cloud transfers the processing knowledge of the global model to edge nodes through knowledge distillation. Edge nodes share industry general knowledge through transfer learning to generate a joint response strategy across edge nodes;

[0216] S45. Continuously monitor the prediction accuracy of edge nodes, the global optimization effect of the cloud, task processing latency, and communication overhead, synchronize the execution status and incrementally update the model to adapt to changes in the network environment.

[0217] In this embodiment, in step S43, the cloud's global optimization includes the following steps:

[0218] S431. Edge nodes regularly upload prediction results, execution logs, and key status indicators to the cloud, construct a global view, and analyze cross-regional traffic trends, interference relationships between slices, and resource allocation balance;

[0219] S432. Based on cross-edge node collaboration, formulate a slice resource reconfiguration plan and generate a global optimal policy;

[0220] S433. Send the updated model parameters and knowledge distillation guidance signals to edge nodes. Edge nodes report the current CPU / GPU utilization rate, memory occupancy, and the length of the task queue to be processed in real time. Meanwhile, the cloud predicts the computing pressure of each edge node in the future period;

[0221] S434. Allocate tasks, migrate complex prediction tasks to the cloud for execution, and process simple tasks by edge nodes. When the load on the edge nodes is too high, trigger the warning mechanism and automatically migrate simple tasks to the cloud for execution.

[0222] Specifically, by constructing a base station association graph based on geographical distance and service relevance, extracting spatial features through a multi-layer graph convolutional network, using a gating mechanism to fuse slice features and spatio-temporal features, dynamically calculating slice bandwidth allocation based on predicted traffic and service quality requirements, and updating the policy network in combination with the proximal policy optimization algorithm, it is possible to achieve elastic allocation of resources under the consideration of multi-objective constraints such as throughput, energy consumption, and interference. By calculating the standardized anomaly score of the slice traffic prediction residual and combining multi-dimensional comprehensive anomaly indicators in the time series, space, and inter-slice dimensions, it is possible to accurately locate abnormal slices. Based on the slice anomaly pre-play mechanism of digital twin, simulate the fault scenario through the state transition function, quantify the service degradation rate, and use the multi-objective optimization algorithm to generate the Pareto optimal strategy, and update the policy library in combination with the forgetting factor, so as to shorten the response time of the system in abnormal scenarios and improve the policy optimization efficiency.

[0223] The above embodiments are only used to illustrate the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the embodiments, those of ordinary skill in the art should understand that various combinations, modifications, or equivalent replacements of the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention and should all be covered within the scope of the claims of the present invention.

Claims

1. A communication base station traffic prediction and management system based on deep learning, characterized in that, Including: A multi-source data acquisition module, which is used to acquire base station spatio-temporal traffic data, user behavior data, network status data, external influencing factors, and dataset slicing service parameters; A spatio-temporal feature processing module, which is used to perform spatio-temporal alignment, noise filtering, and slicing feature encoding on multi-source data; A dynamic model prediction module, which is used to construct a multi-modal prediction network of spatio-temporal Transformer, graph neural network, and slice-specific sub-models, and output base station-level and slice-level traffic prediction results; An edge-cloud collaboration module, which is used for hierarchical collaboration of lightweight prediction at edge nodes and global optimization at the cloud.

2. The traffic prediction management system for a communication base station based on deep learning according to claim 1, wherein The spatio-temporal alignment method of the spatio-temporal feature processing module includes the following steps: S21. Map the longitude and latitude coordinates of the base station to the hexagonal network coordinates through projective transformation , construct a mapping matrix to assist in the location information processing of the communication base station traffic prediction management system. The formula is: ， ， In the formula, is the mapping matrix, which is optimized by minimizing the Wasserstein distance ; is the original coordinate distribution, is the target grid distribution; S22. Approximately solve the Wasserstein distance through the Sinkhorn algorithm to obtain the optimal transport matrix , providing a basis for the optimal allocation of traffic resources and the formulation of transmission strategies. The formula is: ， In the formula, is the Gibbs kernel, and , is the cost matrix, is the regularization parameter, and are both probability orientations that satisfy the edge constraints; S23. Calculate the time series and Calculate the DTW distance, use path constraints to reduce the computational complexity, and provide support for the similarity measurement and dynamic trend analysis of historical and predicted traffic time series. The formula is as follows: ， Wherein, is a time series the -th point and the time series the -th point, the dynamic time warping distance between them, is a measure of the difference degree of data points between the -th point of the time series and the -th point of the time series ; is a set of allowed path step lengths, and , is the path step length; S24. Fuse the aligned spatial features and temporal features through the attention mechanism to assist traffic prediction. The formula is as follows: ， ， Wherein, is the fused feature vector, is the attention weight, is the sigmoid function, is the learnable weight vector; S25. Perform noise filtering and slicing processing on the feature-fused data to purify the data and extract effective segments, improving the accuracy and management efficiency of traffic prediction.

3. The traffic prediction management system of a communication base station based on deep learning according to claim 2, characterized in that, In the step S25, performing noise filtering and slicing processing on the feature-fused data includes the following steps: S251. Train the GI-GAN generator Repair the missing data, discriminator Distinguish real data from generated data , and the generator loss function formula is as follows: ， In the formula, is the noise vector, is the gradient penalty coefficient, is the reconstruction loss weight, is the gradient operator; S252. Set traffic constraints to identify and correct abnormal traffic values, filter out abnormal fluctuations in slice-level traffic data, and improve the accuracy of prediction and management; S253. Combine statistical methods and deep learning to detect anomalies, perform multi-scale anomaly detection, and calculate the Z-score and autoencoder reconstruction error. The formula is: ， ， In the formula, is the slice flow rate at time . is the slice Z - score value at time . is the average flow rate of the slice . is the standard deviation of the flow rate of the slice . is the slice reconstruction error of the auto - encoder at time . is the slice original flow rate data at time . is the slice flow rate data after reconstruction by the auto - encoder at time ; S254. Dynamically update the anomaly threshold according to traffic fluctuations , adaptively adjust the anomaly detection criteria, and the formula is: ， In the formula, and are the abnormal thresholds after and before the update respectively, is the learning rate.

4. The traffic prediction management system for communication base stations based on deep learning according to claim 3, wherein In the step S252, processing abnormal fluctuations in slice-level traffic data includes the following steps: S2521. Determine the constraint conditions and judge whether there are anomalies in the slice-level traffic values. The formula is: ， In the formula, is the flow rate at time , is the proportionality coefficient, is the maximum flow rate of the slice; S2522. Check whether the flow value violates the above constraints. If it violates, then regard as an outlier. If the constraints are satisfied, proceed to step S253; S2523. For outliers, correct to improve data quality and facilitate reasonable management decisions. The formula is: ， wherein, is the historical weight, is the slice within the window and is the set of flow values.

5. A traffic prediction and management system for communication base stations based on deep learning according to claim 1, characterized in that, The dynamic model prediction module constructs a multi-modal prediction network including the following steps: S31. Construct a base station association graph , through node features fuse location and traffic information to calculate edge weights Based on geographical distance and service relevance, construct a graph structure reflecting the spatial association of base stations. The formula is as follows: ， Wherein, is the geographical distance, is the traffic correlation, and are both weights; S32. Update the base station association graph node representation through a multi-layer graph convolutional network combined with residual connection and layer normalization to extract more effective spatial features. The formula is: ， In the formula, and are the feature representation vectors of node in the th layer and th layer respectively. is the feature representation vector of node in the th layer, and node is a node in the neighbor set of node . is the neighbor set of node . is the weight matrix of the th layer. is the normalization coefficient, and . is the neighbor set of node . is the activation function. S33. Use a gating mechanism to fuse slice features and spatio-temporal features to make the feature representation more comprehensive. The formula is: ， ， is the slice feature vector, is the spatio-temporal feature vector, is the feature vector obtained by fusing the slice feature and the spatio-temporal feature, is the gating weight, is the activation function, is the weight matrix of the gating; S34. Dynamically calculate slice bandwidth allocation according to the predicted traffic and service quality requirements to achieve flexible and reasonable allocation of slice resources. The formula is: ， In the formula, and are respectively the predicted traffic and the bandwidth allocation amount of the slice , is the total bandwidth, is the maximum resource share of the slice , is the traffic-bandwidth conversion relationship parameter, is the signal-to-noise ratio of the slice . S35. Adopt the proximal policy optimization algorithm to update the policy network and optimize the base station resource scheduling policy. The formula is: ， ， In the formula, and are the current policy network and the historical policy network respectively, is the probability ratio, is used to measure the degree of advantage of taking action over the average action in state ; is the clipping parameter that limits the transformation range, is the expectation operator, is the objective function for evaluating the performance of the policy network.

6. The traffic prediction and management system for a communication base station based on deep learning according to claim 5, characterized in that, In the step S34, after dynamically calculating the slice bandwidth allocation, the following steps are further included: S341. Introduce prediction uncertainty penalty, establish a reward function, comprehensively consider throughput, energy consumption, and interference, provide a quantitative evaluation standard during bandwidth allocation, and optimize the accuracy of slice bandwidth allocation. The formula is; ， wherein, , , and are the reward value, throughput, energy consumption, and interference at time respectively; is the variance of the predicted traffic; , , and are the corresponding weight coefficients respectively; S342. Establish a revenue function, determine service quality satisfaction, bandwidth allocation utilization, and interference between slices, and provide a decision-making basis for subsequent optimization of the base station resource scheduling policy. The formula is: ， ， Wherein, is the comprehensive revenue function value for measuring the slice , is the quality satisfaction of the slice , and are respectively the actually allocated and the maximum obtainable bandwidth resources of the slice , is the service delay of the slice , is the number of service requests of the slice . S343. Detect slice anomalies, timely discover abnormal fluctuations in communication base station traffic data, ensure communication service quality, and ensure the stable operation of the system.

7. The traffic prediction management system for a communication base station based on deep learning according to claim 6, characterized in that In the step S343, detecting slice anomalies includes the following steps: S3431. Calculate the slice traffic prediction residual, standardize the anomaly score, and quantify the traffic prediction deviation. The formula is: ， Wherein, is the slice at the moment the standardized anomaly score, and are respectively the slice at the moment the true and predicted flow data, is the sliding window size for defining the time range, is the slice at the moment the mean of the predicted residuals; S3432. Integrate anomaly situations in the time series, space, and inter-slice dimensions, calculate the multi-dimensional comprehensive anomaly index, and accurately locate the anomaly. The formula is: ， Wherein, is the overall anomaly degree value reflecting the slice under multiple dimensions, is the spatial anomaly situation value of the slice , is the slice and the slice flow correlation anomaly situation value, , are the corresponding weights respectively; S3433. Obtain the slice status identifier and dynamically adjust the anomaly detection threshold based on historical anomalies to enable the system to adapt to traffic changes and improve the accuracy and effectiveness of anomaly detection. The formula is: ， Wherein, and are the anomaly detection thresholds of the slice at time and time respectively, and are the learning rate for controlling the threshold adjustment step and the adjustment factor for adjusting the threshold amplitude respectively, is the status flag of the slice ; S3434. Establish a rehearsal of failures caused by slice anomalies, determine the impact of different slice anomalies on policy allocation, and optimize and update the policy library.

8. The traffic prediction management system for a communication base station based on deep learning according to claim 7, characterized in that, In step S3434, establishing a preview of a failure caused by a slice anomaly includes the following steps: S34341. Update the slice digital twin state based on the current state and abnormal parameters. The state transition function provides a basis for simulating fault scenarios and analyzing fault impacts. The formula is: ， In the formula, and are the state vectors of the slice at time and time respectively, is the abnormal parameter vector, is the state transition function; S34342. Calculate the slice service degradation rate, quantify the impact of abnormalities on slice service quality, and perform quantitative evaluation of abnormalities. The formula is: ， wherein, is the proportion of the slice whose service quality deteriorates due to abnormal reaction, and are the service qualities of the slices in normal and abnormal states respectively; S34343. Based on multi-objective optimization, calculate the weighted Euclidean distance between service degradation rate and resource consumption under different strategies to balance service quality and resource cost. The formula is: ， In the formula, is the Pareto optimal strategy, is the strategy set, is the resource consumption of strategy execution, and are the corresponding weight coefficients respectively; S34344. Perform exponential weighted updates on the policy library based on the forgetting factor to ensure that the policy library adapts to traffic changes and improves system management efficiency. The formula is: ， Wherein, and are the th policies in the updated and pre-updated policy libraries respectively, is the adjusted new policy, is the update coefficient, is the forgetting factor, is the policy usage duration.

9. The traffic prediction management system for a communication base station based on deep learning according to claim 1, characterized in that The edge-cloud collaboration module includes the following steps for edge node quantitative prediction and cloud-cloud global optimization: S41. Perform self-inspection on edge nodes to generate resource profiles. Simultaneously, the cloud performs knowledge distillation on the complete prediction model, compresses the model to a scale that can be run on edge nodes, and verifies the deployment success rate. S42. Execute the pre-configured lightweight strategy based on the prediction results, record the deviation between the execution results and the actual traffic, and generate a local execution log for subsequent evaluation; S43. The cloud optimizes the overall situation and completes load balancing. S44. The cloud transfers the processing knowledge of the global model to the edge nodes through knowledge distillation. The edge nodes share industry-wide knowledge through transfer learning and generate joint response strategies across edge nodes. S45. Continuously monitor the prediction accuracy of edge nodes, global optimization effects in the cloud, task processing delays, and communication overhead to synchronize execution status with incremental model updates and adapt to changes in the network environment.

10. A communication base station traffic prediction and management system based on deep learning according to claim 9, characterized in that In step S43, the cloud performs global optimization, including the following steps: S431. Edge nodes regularly upload prediction results, execution logs, and key status indicators to the cloud to build a global view and analyze cross-regional traffic trends, inter-slice interference relationships, and resource allocation balance. S432. Formulate a slice resource reconfiguration plan based on cross-edge node collaboration and generate a global optimal strategy. S433: Send the updated model parameters and knowledge distillation guidance signals to the edge nodes. The edge nodes report the current CPU / GPU utilization, memory usage, and the length of the pending task queue in real time. At the same time, the cloud predicts the computing pressure of each edge node in the future period. S434. Assign tasks, migrate complex prediction tasks to the cloud for execution, and have simple tasks processed by edge nodes. When the edge node load is too high, trigger the early warning mechanism and automatically migrate simple tasks to the cloud for execution.

Images