CN120224463A

CN120224463A - Communication network resource optimization method and system based on artificial intelligence

Info

Publication number: CN120224463A
Application number: CN202510356291.6A
Authority: CN
Inventors: 赵文涛
Original assignee: Beijing News Wind Times Software Development Co ltd
Current assignee: Beijing News Wind Times Software Development Co ltd
Priority date: 2025-03-25
Filing date: 2025-03-25
Publication date: 2025-06-27
Anticipated expiration: 2045-03-25
Also published as: CN120224463B

Abstract

The invention provides a communication network resource optimizing method and system based on artificial intelligence, which relates to the technical field of communication optimization, by collecting historical network data, utilizing time sequence decomposition and LSTM algorithm to extract characteristics, combining attention mechanism to calculate the weight of performance index, the network flow is predicted by adopting a multi-scale prediction mechanism, and dynamic calibration is carried out based on the historical error, so that the network flow change trend can be accurately predicted, the prediction precision is improved, the reasonable distribution and the optimal configuration of network resources are realized, and the network operation efficiency is effectively improved.

Description

Communication network resource optimization method and system based on artificial intelligence

Technical Field

The invention relates to the technical field of communication optimization, in particular to a communication network resource optimization method and system based on artificial intelligence.

Background

With the rapid development and the deep advancement of digital transformation of the 5G network, the service types and data traffic carried by the communication network are explosive growth, and the efficient network resource management and optimization have important significance for guaranteeing the network service quality and improving the user experience;

The traditional communication network resource optimization method mainly depends on manual experience and fixed rules, is difficult to adapt to complex and changeable network environments and dynamic changing business requirements, and is currently generally adopted to predict network traffic and optimize resources based on a statistical analysis method, and a simple time sequence model or regression analysis is used to model and analyze historical data;

the existing network resource optimization technology still focuses on the prediction of a single time scale, and is difficult to simultaneously consider the change characteristics of network traffic in different time periods, so that the prediction accuracy is not ideal enough, the relevance between the performance index of network equipment and the traffic change is not fully considered, the influence of the network state on the traffic prediction is neglected, an effective prediction result calibration mechanism is lacking, and the problems that a prediction model cannot be dynamically adjusted and optimized according to the actual network running condition are solved;

accordingly, there is a need for a solution to the problems of the prior art.

Disclosure of Invention

The embodiment of the invention provides a communication network resource optimization method and a communication network resource optimization system based on artificial intelligence, which at least can solve part of problems in the prior art.

In a first aspect of the embodiment of the present invention, there is provided a communication network resource optimization method based on artificial intelligence, including:

Collecting historical network flow data, network equipment performance data and user service demand data in a communication network, decomposing the historical network flow data into trend items, period items and residual items through a time sequence decomposition algorithm, and calculating bandwidth occupancy rate, processor occupancy rate, memory occupancy rate and transmission delay in the network equipment performance data;

Forming an input vector by a trend item, a period item and a residual item, adding the input vector into a long-short time memory network, performing gating operation on the input vector through a forgetting gate, an input gate and an output gate, updating a unit state to obtain a hidden state, and performing linear transformation on the hidden state to obtain a time sequence code;

the bandwidth occupancy rate, the processor occupancy rate, the memory occupancy rate and the transmission delay are formed into a performance index vector, an attention mechanism is introduced to calculate the performance index vector to obtain an importance weight, and the time sequence code and the importance weight are subjected to feature fusion to obtain fusion features;

processing the fusion characteristics by adopting a multi-scale prediction mechanism, respectively predicting short-term, medium-term and long-term network flow values, setting self-adaptive weights based on prediction errors under each time scale, and carrying out weighted combination on the prediction results of a plurality of time scales to output initial network flow prediction values;

And constructing a correction model according to the statistical characteristics of the historical prediction errors, dynamically calibrating an initial network flow predicted value by using an exponential smoothing method, and correcting the predicted value by combining with the mutation detection result of the network performance index to obtain a network flow predicted result.

In an alternative embodiment of the present invention,

The method comprises the steps of collecting historical network flow data, network equipment performance data and user service demand data in a communication network, decomposing the historical network flow data into trend items, period items and residual items through a time sequence decomposition algorithm, and calculating bandwidth occupancy rate, processor occupancy rate, memory occupancy rate and transmission delay in the network equipment performance data, wherein the steps of:

Collecting historical network flow data in a communication network in a preset sampling period, wherein the historical network flow data is the number of flow bytes on an internal network link in unit time, and meanwhile, collecting network equipment performance data which reflects the working state of the network link and collects user service demand data;

Inputting the historical network flow data into a time sequence decomposition algorithm, extracting a fluctuation rule of the historical network flow data along with time change, and obtaining a trend item reflecting a long-term change trend, a periodic item reflecting periodic change and a residual item reflecting random fluctuation;

And carrying out statistical calculation on the network equipment performance data, and carrying out moving average on the network equipment performance data according to a preset time window to obtain bandwidth occupancy rate, processor occupancy rate, memory occupancy rate and transmission delay.

In an alternative embodiment of the present invention,

Forming an input vector by a trend item, a period item and a residual item, adding the input vector into a long-short time memory network, performing gating operation on the input vector through a forgetting gate, an input gate and an output gate, updating a unit state to obtain a hidden state, and performing linear transformation on the hidden state to obtain a time sequence code, wherein the method comprises the following steps of:

sequentially splicing the trend item, the period item and the residual item in the time dimension to form an input vector, wherein the input vector comprises historical values in different time steps;

Based on the input vector and the pre-acquired historical state information, randomly forgetting the historical information through a forgetting gate, calculating current updating information through the input gate and generating candidate unit states, combining a reserved part of the historical information with the current updating information to update the unit states, and performing gating operation on the unit states through the output gate to obtain hidden states;

and performing linear transformation on the hidden state to obtain time sequence coding.

In an alternative embodiment of the present invention,

The method for combining bandwidth occupancy rate, processor occupancy rate, memory occupancy rate and transmission delay into a performance index vector, introducing an attention mechanism to calculate the performance index vector to obtain importance weight, and carrying out feature fusion on time sequence codes and the importance weight to obtain fusion features comprises the following steps:

Collecting performance parameters of the network equipment at a plurality of continuous time points, wherein the performance parameters comprise bandwidth occupancy rate, processor occupancy rate, memory occupancy rate and transmission delay, and the performance parameters collected at each time point form a performance index vector;

Inputting the performance index vector into a query matrix, a key matrix and a value matrix based on an attention mechanism to obtain a query vector, a key vector and a value vector, obtaining a correlation score between the performance indexes based on dot product operation of the query vector and the key vector, and carrying out softmax normalization processing on the correlation score to obtain importance weights, wherein the importance weights represent the influence degree of each performance index on network performance;

and carrying out weighted superposition on the time sequence codes and the importance weights to obtain fusion characteristics.

In an alternative embodiment of the present invention,

Processing the fusion characteristic by adopting a multi-scale prediction mechanism, respectively predicting short-term, medium-term and long-term network flow values, setting self-adaptive weights based on prediction errors under each time scale, and performing weighted combination on the prediction results of the time scales to output initial network flow prediction values, wherein the steps comprise:

Respectively constructing three time scale prediction processing units of short term, medium term and long term by adopting a double-layer long-short time memory network structure, processing the fusion characteristics by the double-layer long-short time memory network structure, and outputting initial network flow prediction values under each time scale;

Continuously collecting network flow data of each time scale in a preset sampling time window, calculating the change rate of the network flow data under each time scale, and calculating the time sequence association degree between adjacent time scales and the fluctuation degree of each time scale according to the change rate;

Dividing a history sample into a plurality of continuous training sequence segments according to preset time intervals, calculating a difference value between a network flow true value and an initial network flow predicted value under a corresponding time scale in each training sequence segment, carrying out cumulative summation operation on the difference value, calculating an average value, and generating a predicted error under each time scale;

extracting regular characteristics of time sequence association degree between time scales along with time change according to a graph structure network, establishing a constraint relation between adjacent time scale predicted values according to the regular characteristics, correcting an initial network flow predicted value based on the constraint relation and a federal learning frame, generating a corrected predicted value considering time sequence association, and adjusting an initial self-adaptive weight based on the corrected predicted value to obtain a corrected weight considering time sequence association, wherein the corrected weight corresponding to the time scale with larger fluctuation degree is relatively increased;

and carrying out weighted summation operation on the corrected predicted value and the corresponding corrected weight, and outputting a final network flow predicted value.

In an alternative embodiment of the present invention,

Extracting regular features of time sequence association degree between time scales along with time change according to a graph structure network, establishing a constraint relation between adjacent time scale predicted values according to the regular features, correcting an initial network flow predicted value based on the constraint relation and a federal learning frame, and generating a corrected predicted value considering the time sequence association comprises:

Constructing a dynamic heterogeneous graph network structure, connecting network flow nodes with different time scales through dynamic edges, wherein node attributes comprise flow values, statistical features and variation trends, and the edge attributes are initialized to basic time sequence association degrees between adjacent time scales;

Based on the updated node state vector and the updated edge association strength, constructing a time sequence association degree extraction model, capturing the change characteristics of network traffic on different time scales through a sliding time window, and identifying the dominant factors of the time sequence association degree by combining a reverse fact reasoning method to extract the rule characteristics of the time sequence association degree changing along with time;

Constructing a conditional time sequence pattern diagram according to the extracted regular features, mapping time sequence association patterns under different conditions into nodes in the diagram, establishing node connection based on evolution relations among the patterns, describing constraint relations among adjacent time scale predicted values by using the conditional time sequence pattern diagram, integrating constraint relation information of a plurality of network areas by using a distributed federal learning framework, adjusting calculation parameters of time sequence association in real time based on local prediction deviation, constructing a self-adaptive probability diagram model by using an extraction mode of continuously optimizing the regular features by using global prediction performance indexes, and dynamically associating update frequency of model parameters with fluctuation degree of network flow;

According to node state distribution and transition probability in the self-adaptive probability graph model, constructing an energy function by combining constraint relation provided by a conditional time sequence mode graph, taking constraint relation strength, a conditional time sequence rule and probability distribution as component parts of the energy function, wherein the weight coefficient of each component part is self-adaptively adjusted along with the change of prediction precision, correcting an initial network flow predicted value by iteratively optimizing the energy function until all time sequence constraint conditions are met, and outputting a corrected predicted value considering time sequence relevance.

In an alternative embodiment of the present invention,

Constructing a correction model according to the statistical characteristics of the historical prediction errors, dynamically calibrating an initial network flow prediction value by using an exponential smoothing method, and correcting the prediction value by combining with a sudden change detection result of a network performance index, wherein the obtaining of the network flow prediction result comprises the following steps:

Obtaining a historical prediction error sequence, wherein the historical prediction error sequence is obtained by subtracting a predicted network flow value from an actual network flow value, calculating an error mean value and an error standard deviation based on the historical prediction error sequence, constructing an autocorrelation function of the error sequence, and calculating an exponential weighted movement variance, wherein the exponential weighted movement variance is determined by the weighted sum of the mean square error of the error at the current moment and the mean value and the exponential weighted movement variance at the previous moment;

Constructing a correction model based on the error mean value, the error standard deviation, the autocorrelation function and the exponentially weighted moving variance, wherein the correction model comprises an error distribution correction term, a time sequence correlation correction term and a fluctuation correction term, the error distribution correction term, the time sequence correlation correction term and the fluctuation correction term respectively correspond to different self-adaptive weight coefficients, and the self-adaptive weight coefficients are updated through a gradient descent method;

According to an exponential smoothing method, a historical information storage matrix is constructed, an initial network flow predicted value is dynamically calibrated through an adaptive smoothing factor and an attention mechanism to obtain a first corrected predicted value, wherein the adaptive smoothing factor is determined by the weighted sum of the previous time smoothing factor and the absolute value of the current time historical predicted error, and the weighted coefficient is dynamically adjusted through the change quantity of an exponential weighted movement variance;

constructing a multi-dimensional network performance index vector, wherein the multi-dimensional network performance index vector comprises values of a plurality of network performance indexes at the current moment, and carrying out accumulated summation operation based on time sequence differences of the multi-dimensional network performance index vector to obtain a mutation detection value, wherein the mutation detection value is compared with the difference between the mutation detection value at the previous moment and a detection threshold value;

When the mutation detection value is larger than a mutation judgment threshold value, calculating the relative change rate of each network performance index in the multi-dimensional network performance index vector, multiplying the relative change rate of each network performance index by a corresponding weight coefficient, summing to obtain a mutation degree, outputting the product of the first correction prediction value and the mutation degree as a network flow prediction result, and when the mutation detection value is not larger than the mutation judgment threshold value, outputting the first correction prediction value as a network flow prediction result.

In an alternative embodiment of the present invention,

According to the exponential smoothing method, constructing a historical information storage matrix, and dynamically calibrating an initial network flow predicted value through an adaptive smoothing factor and an attention mechanism to obtain a first corrected predicted value comprises the following steps:

Carrying out multi-layer division on the network traffic time sequence according to a time scale to obtain a plurality of time scale layers, and respectively constructing independent self-adaptive smoothing factors for each time scale layer;

calculating the fluctuation rate according to the change trend of the network traffic time sequence, dynamically adjusting the size of a historical data window based on the fluctuation rate, wherein the size of the historical data window is determined by the product of the size of a basic window and the fluctuation rate, and acquiring the smoothing result of each time scale layer in the range of the historical data window;

Constructing a feature matrix of the smoothing result according to a time scale layer, calculating attention weights based on the feature matrix and a query vector at the current moment, wherein the attention weights represent the importance degree of different time scale features on the current prediction, and carrying out self-adaptive fusion on the features of different time scales by utilizing the attention weights;

constructing a history information storage matrix, wherein the history information storage matrix comprises a history network flow value, the prediction error and the self-adaptive smoothing factor, extracting history information related to the current state from the history information storage matrix by adopting an attention mechanism, and determining a dynamic learning rate based on the change amount of the prediction error, wherein the dynamic learning rate is attenuated along with the increase of the change amount of the prediction error;

And introducing a residual connection structure into the initial network flow predicted value, carrying out weighted summation on the correction function output of each time scale and the self-adaptive smoothing factor according to the attention weight to obtain a residual item, and adding the residual item and the initial network flow predicted value to obtain a first correction predicted value.

In a second aspect of the embodiments of the present invention, there is provided an artificial intelligence based communication network resource optimization system, comprising:

The first unit is used for collecting historical network flow data, network equipment performance data and user service demand data in the communication network, decomposing the historical network flow data into a trend item, a period item and a residual item through a time sequence decomposition algorithm, and calculating bandwidth occupancy rate, processor occupancy rate, memory occupancy rate and transmission delay in the network equipment performance data;

The second unit is used for forming an input vector by a trend item, a period item and a residual item, adding the input vector into a long-short time memory network, performing gating operation on the input vector through a forgetting gate, an input gate and an output gate, updating the state of the unit to obtain a hidden state, and performing linear transformation on the hidden state to obtain a time sequence code;

the third unit is used for forming a performance index vector from the bandwidth occupancy rate, the processor occupancy rate, the memory occupancy rate and the transmission delay, introducing an attention mechanism to calculate the performance index vector to obtain an importance weight, and carrying out feature fusion on the time sequence code and the importance weight to obtain fusion features;

A fourth unit, configured to process the fusion feature by using a multi-scale prediction mechanism, predict short-term, medium-term and long-term network traffic values respectively, set adaptive weights based on prediction errors under each time scale, and perform weighted combination on the prediction results of the multiple time scales to output an initial network traffic prediction value;

And a fifth unit, configured to construct a correction model according to the statistical characteristics of the historical prediction error, dynamically calibrate the initial network traffic prediction value by using an exponential smoothing method, and correct the prediction value by combining with the mutation detection result of the network performance index to obtain the network traffic prediction result.

In the invention, historical network flow data is processed by combining time sequence decomposition and long-short time memory networks, the time sequence characteristics and long-term dependency of the data are fully considered, the accuracy and reliability of network flow prediction are improved, a more accurate decision basis is provided for network resource optimization configuration, an attention mechanism is introduced to carry out importance weighting on network performance indexes and fuse with the time sequence characteristics, comprehensive perception and characteristic extraction of network states are realized, the adaptability of a model to network environment changes is enhanced, a prediction result is more in line with actual network operation conditions, a multi-scale prediction mechanism and a dynamic calibration method are adopted, prediction results of different time scales are combined through self-adaptive weights, and the correction is carried out by combining historical error statistical characteristics, so that the prediction deviation is effectively reduced, the prediction performance of the model under different time periods is improved, and more accurate prediction support is provided for intelligent scheduling and optimization of network resources.

Drawings

FIG. 1 is a flow chart of an artificial intelligence based communication network resource optimization method according to an embodiment of the invention;

FIG. 2 is a graph showing comparison of prediction accuracy under different load conditions corresponding to an artificial intelligence-based communication network resource optimization method according to an embodiment of the present invention;

fig. 3 is a diagram of prediction error data under different fluctuation scenarios corresponding to an artificial intelligence-based communication network resource optimization method according to an embodiment of the present invention.

Detailed Description

For the purpose of making the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, and it is apparent that the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

The technical scheme of the invention is described in detail below by specific examples. The following embodiments may be combined with each other, and some embodiments may not be repeated for the same or similar concepts or processes.

Fig. 1 is a flow chart of an artificial intelligence based communication network resource optimization method according to an embodiment of the present invention, as shown in fig. 1, the method includes:

In an alternative embodiment of the present invention,

And deploying flow acquisition equipment in the network, and setting the sampling period to be 5 minutes. The collector will record the number of bytes of traffic through the network link at each sampling point in time. For example, for a 10Gbps network link, the collector records the actual traffic data on the link every 5 minutes, including the information of the number of bytes, the number of data packets, etc. of the uplink and the downlink. These raw traffic data are stored in real time in a database, forming continuous time series data.

Performance data of the network device is collected. The network equipment is accessed through SNMP protocol, and the running state parameters are obtained. The collector sends an SNMP inquiry request to the equipment every 5 minutes, and reads the CPU utilization rate, the memory occupation condition, the port state and other performance indexes of the equipment. These performance data reflect the current operating state and load conditions of the network link.

And collecting service demand data of the user, wherein the service demand data comprises information such as bandwidth application records, service types, service quality requirements and the like submitted by the user. These data are used to learn the source and law of variation of the network load.

The collected historical network flow data is processed by using a time sequence decomposition algorithm, continuous flow data are arranged in time sequence, the data are decomposed into three components by using a seasonal decomposition method, trend items reflect long-term change trends of the flow, such as ascending or descending of the flow, periodic items reflect periodic change rules of the flow, such as peak and valley of each day, weekend difference of each week and the like, and residual items reflect random short-term fluctuation.

And carrying out statistical calculation on the performance data of the network equipment by adopting a sliding time window. A fixed size time window (e.g., 30 minutes) is set, and the average value of each performance index is calculated in the window. The window slides over time, continuously updating the calculation results.

Illustratively, monitoring is performed for a 10Gbps backbone link. The flow collection device records data every 5 minutes, namely 8:00 to 3.2Gbps,8:05 to 3.5Gbps and 8:10 to 3.8Gbps. The collected data of 24 continuous hours are decomposed through a time sequence, a trend item shows a situation of 100Mbps increase per hour, a period item shows that the flow reaches 5Gbps every day in a peak period of 9:00-11:00 and 14:00-16:00, the flow is reduced to 1Gbps in an early-to-valley period of 2:00-5:00, and a residual item shows that the flow has random fluctuation of +/-500 Mbps.

The statistical result of the performance data shows that in a 30-minute sliding window, the average bandwidth occupancy rate of a link is 35%, the average occupancy rate of a processor is 45%, the memory occupancy rate is 60%, and the average transmission delay is 15 milliseconds. The user traffic demand data shows that the link mainly carries data center synchronization traffic (2 Gbps) and video transport traffic (3 Gbps).

In the embodiment, by performing time sequence decomposition on the historical flow data, the change rule of the network flow is accurately grasped, an important basis is provided for network planning and optimization, the utilization efficiency of network resources is improved, the performance index is statistically analyzed by adopting a sliding time window method, the running state of network equipment is comprehensively grasped, the performance bottleneck is timely found, the stable running of the network is ensured, the on-demand distribution and dynamic adjustment of the network resources are realized by combining the user service demand data, the user experience quality is improved, and the network operation and maintenance cost is reduced.

In an alternative embodiment of the present invention,

A trend term, a period term, and a residual term are obtained. Taking daily average temperature data in a certain area within one year as an example, a trend item reflecting long-term change trend, a period item reflecting seasonal change and a residual item reflecting random fluctuation can be obtained through time sequence decomposition.

And the acquired trend item, period item and residual item are formed into an input vector according to time sequence. Taking 30 days as a time window, each time step comprises trend values, period values and residual values of corresponding dates. For example, the input vector on day 1 contains the trend value of 20 degrees, the period value of 2 degrees and the residual value of 0.5 degrees on the day, the input vector on day 2 contains the trend value of 19.8 degrees, the period value of 1.8 degrees and the residual value of-0.3 degrees on the day, and so on to construct an input sequence containing 30 time steps.

The input vector is processed using a long and short term memory network. The network comprises three gating units, namely a forgetting gate, an input gate and an output gate. The forgetting gate calculates the proportion of historical information to be forgotten based on the current input vector and the hidden state at the last moment. Taking temperature prediction as an example, if the difference between the current temperature and the historical temperature is large, the forgetting door can reduce the retention degree of the historical information.

The input gate calculates the information that needs to be updated at the current time. And combining the current input vector and the historical hiding state to generate candidate unit states. The input gate determines how much new information needs to be written into the memory cell. For example, when a temperature jump is detected, the input gate increases the write rate of new information.

And combining the history information filtered by the forgetting gate with the new information updated by the input gate to update the state of the unit. The output gate controls the output degree of the information based on the updated unit state, and the hidden state at the current moment is obtained. And finally, carrying out linear transformation on the hidden state to obtain the coding vector containing the time sequence characteristics.

Illustratively, the trend term at a time is 3.5Gbps, the period term is 1.2Gbps, and the residual term is-0.3 Gbps. The data from the last 6 time steps are spliced to form an input vector of length 18 (6 time steps x 3 components).

The input vector enters the LSTM network, the state of the unit at the last moment is [0.8,0.6,0.4], and the hidden state is [0.7,0.5,0.3]. The forgetting gate calculates a forgetting coefficient [0.4,0.3,0.5] representing the historical information retention ratio. The input gate computes updated information [0.6,0.5,0.4] to generate candidate cell states [0.9,0.7,0.5]. The new cell state is obtained after the combination update 0.85,0.65,0.45. The hidden state [0.75,0.55,0.35] is obtained after the output gate performs the gating operation. The hidden state is mapped to a two-dimensional time series code by linear transformation 0.65,0.45.

In this embodiment, a plurality of feature components of data are extracted through time sequence decomposition, so that a change rule of the data can be more comprehensively captured, accuracy of time sequence feature extraction is improved, a gating mechanism of a long-short-term memory network is adopted, retention degree of historical information and update degree of new information can be adaptively adjusted, long-term dependency relationship is effectively processed, a trend item, a period item and a residual item are combined and modeled, nonlinear feature extraction is performed through a neural network, richer time sequence modes can be learned, and expression capability of a model is enhanced.

In an alternative embodiment of the present invention,

Four performance parameters of the network equipment are collected at continuous time points, namely bandwidth occupancy rate reflects the use condition of a link, processor occupancy rate represents equipment calculation load, memory occupancy rate represents resource consumption state, and transmission delay displays data packet transmission quality. Four parameters acquired at each time point are sequentially arranged to form a performance index vector, and the sampling data at each time point form a group of elements in the vector.

The performance index vector is processed using an attention mechanism. The performance index vector is respectively input into three transformation matrixes, namely a query matrix is used for extracting target features to be analyzed, a key matrix is used for extracting reference features, and a value matrix is used for extracting actual feature values. And obtaining a query vector, a key vector and a value vector after transformation. And calculating the dot product of the query vector and the key vector to obtain a relevance score for representing the relevance degree between different performance indexes. And carrying out softmax normalization processing on the relevance score, and mapping the score to be between 0 and 1 to obtain importance weight. The importance weight reflects the degree of impact of various performance indicators on overall network performance.

And performing weighted superposition operation on the obtained time sequence codes and the importance weights. The time sequence coding comprises time sequence characteristics of network traffic, importance weights show the influence degree of performance indexes, and fusion characteristics obtained by superposition of the time sequence characteristics reflect time sequence change rules and contain performance influence factors.

For example, four performance parameters are collected at a time, namely, bandwidth occupancy rate is 35%, processor occupancy rate is 45%, memory occupancy rate is 60%, and transmission delay is 15ms. The data of 6 time points are sampled consecutively, and a 24-dimensional performance index vector (6 time points×4 parameters) is constituted.

The performance index vector is input into an attention mechanism, and an 8-dimensional query vector [0.4,0.5,0.3,0.6,0.4,0.2,0.5,0.3], an 8-dimensional key vector [0.3,0.6,0.4,0.5,0.3,0.4,0.6,0.2] and an 8-dimensional value vector [0.5,0.4,0.6,0.3,0.5,0.3,0.4,0.6] are obtained through three transformation matrices. The query vector and key vector dot product are calculated to obtain a relevance score [0.8,0.6,0.7,0.5]. Importance weight [0.35,0.20,0.30,0.15] is obtained through softmax normalization, which shows that the bandwidth occupancy rate and the memory occupancy rate have great influence on the network performance.

And the obtained two-dimensional time sequence code [0.65,0.45] and the four-dimensional importance weight [0.35,0.20,0.30,0.15] are subjected to weighted superposition to obtain a fusion characteristic [0.55,0.40] reflecting the time sequence characteristic and the performance influence.

In this embodiment, the importance weight of the performance index is calculated through the attention mechanism, so that the key index with a larger influence on the network performance is accurately identified, the accuracy of network performance evaluation is improved, the characteristic information of the performance index changing along with time is reserved by adopting the time sequence coding, so that the fusion characteristic can reflect the dynamic change rule of the network performance, the time sequence expression capability of the characteristic is enhanced, the importance weight and the time sequence coding are fused, the obtained fusion characteristic has stronger expression capability, the importance degree of the performance index and the time sequence change characteristic can be simultaneously described, and more effective characteristic support is provided for the subsequent network performance prediction and fault diagnosis.

In an alternative embodiment of the present invention,

A two-layer LSTM network structure is constructed, with each time-scale processing unit comprising two LSTM layers. The number of hidden units of the first layer LSTM is set to 128 for extracting the timing pattern in the input features and the number of hidden units of the second layer LSTM is set to 64 focusing on the sequence prediction task. The short-term prediction unit processes the last 24 hours of data, the medium-term prediction unit processes the last 7 days of data, and the long-term prediction unit processes the last 4 weeks of data. The input to each processing unit contains timing encoding and performance weight information in the fusion feature. The LSTM network updates the unit state and the hidden state in each time step through forward propagation, and the hidden state is mapped into a predicted value by the final output layer through the full connection layer;

Flow data was collected every 5 minutes within a 30 minute sampling window. The short-term scale calculates the change rate of adjacent sampling points (current flow-last moment flow)/sampling interval, the medium-term scale calculates the change rate of adjacent hours, and the long-term scale calculates the change rate of adjacent days. For the correlation degree between adjacent time scales, the pearson correlation coefficient is adopted for calculation, namely, the covariance of the two time scale data sequences is calculated after the data sequences are normalized, and then the product of the covariance and the covariance is divided by the product of the standard deviation. The fluctuation degree is obtained by calculating the standard deviation of the flow sequence in the sliding window, and the window sizes are respectively short-term 2 hours, medium-term 2 days and long-term 2 weeks. The temperature coefficient is set by adopting an exponential decay formula, wherein exp (-fluctuation degree);

The method comprises the steps of dividing a 30-day historical sample according to fixed intervals, generating 720 training sequence segments by short-term prediction, generating 120 training sequence segments by medium-term prediction, and generating 30 training sequence segments by long-term prediction, wherein the short-term prediction takes 1 hour. And in each sequence segment, accumulating and calculating a prediction error, namely summing the difference value between the predicted value and the true value at each time point, and dividing the sum by the length of the sequence segment to obtain an average error. Multiplying the average error by the temperature coefficient, taking the negative value, and carrying out nonlinear mapping through an exp function to obtain an index mapping value. Adding the index mapping values of the three time scales to obtain normalized denominators, and dividing the respective index mapping values by the denominators to obtain initial self-adaptive weights;

Constructing a three-node graph network, wherein nodes represent predicted values of three time scales, and the weight of edges is time sequence association degree. Each node contains a current predictor and a historical predicted sequence feature. And extracting dynamic association features among nodes through the graph attention layer to generate an edge feature matrix. And establishing a constraint equation based on the edge feature matrix, wherein the difference of the predicted values of the adjacent nodes is inversely proportional to the association degree. Under the federal learning framework, predictive models for each time scale act as federal members, sharing constraint information but not directly exchanging data. And generating a corrected predicted value by iteratively optimizing the predicted value which violates the constraint. Calculating a correction weight by taking the original weight and the fluctuation degree into consideration, wherein the correction weight=the original weight (1+normalized fluctuation degree);

And carrying out weighted summation operation on the corrected predicted values and the corrected weights of the three time scales. The weight value reflects the credibility of the prediction result of each time scale, and the prediction value reflects the correction result after time sequence association is considered. The weighted summation can balance prediction bias of different time scales to obtain a more accurate final predicted value.

The collected flow data illustratively shows a short-term standard deviation of 0.8Gbps, a mid-term 0.5Gbps, and a long-term 0.3Gbps. The temperature coefficients are set to be short-term 0.4, medium-term 0.6 and long-term 0.8.

The double-layer LSTM outputs three time-scale initial predictions, short-term 4.2Gbps, medium-term 4.0Gbps, and long-term 3.8Gbps. The actual flow value is 4.1Gbps, and the prediction error is calculated to be short-term 0.1Gbps, medium-term 0.1Gbps and long-term 0.3Gbps.

The prediction error is multiplied by the temperature coefficient and taken as negative, and the result is substituted into an exponential function to calculate short-term exp (-0.04) =0.96, medium-term exp (0.06) =1.06 and long-term exp (0.24) =1.27. Normalization yields initial adaptive weights of short term 0.29, medium term 0.32, long term 0.39.

The graph structure network extracts short-term-medium-term association degree 0.7 and medium-term-long-term association degree 0.6. The predicted value is corrected based on the constraint relation to be short-term 4.15Gbps, medium-term 4.05Gbps and long-term 3.9Gbps. The adjusted correction weights are short term 0.35, medium term 0.33, long term 0.32.

The weighted summation yields a traffic prediction value of 4.15×0.35+4.05×0.33+3.9×0.32=4.04 Gbps.

In the embodiment, the characteristics of network flow data under different time scales are fully utilized through a multi-scale prediction mechanism, the prediction precision and robustness are improved, the weight is dynamically adjusted according to the prediction error and the data fluctuation degree of each time scale by adopting a self-adaptive weight distribution scheme, so that the prediction result is more accurate and reliable, a graph structure network and a federal learning framework are introduced, the time sequence association characteristics are effectively extracted, the distributed collaborative optimization is realized, and the generalization capability and the practicability of the model are enhanced.

In an alternative embodiment of the present invention,

And constructing a dynamic heterogeneous graph network structure, organizing network traffic data of different time scales into a graph network form, and constructing the network traffic data of each time point into nodes in the graph. Each node contains an actual flow value, statistical characteristic information and change trend information of the time point. The statistical features are obtained through sliding time window calculation and comprise statistics such as average value, standard deviation, skewness of data distribution, kurtosis and the like in a window range, and the change trend information is obtained through analyzing flow changes of adjacent time points and is used for representing the change state of the flow. For nodes adjacent in time, the system establishes dynamic connection edges, and initial attributes of the edges are determined by calculating correlation coefficients of flow data of adjacent time points.

In order to realize dynamic update of node and edge attributes, a space-time attention mechanism is designed, the association degree of each node and surrounding nodes is calculated in a space dimension, and different types of space association features are respectively focused through a multi-head attention mechanism. In the time dimension, a time sequence attention layer is introduced to capture time sequence position information, so that a model can understand the dependency relationship of data in the time dimension;

and constructing a time sequence association degree extraction model based on the updated node state and the side association information. And analyzing the trend, periodicity and burstiness characteristics of the network traffic in each window by adopting a dynamic adjustment sliding window mechanism. The key factors influencing the time sequence relevancy are identified through a counterfactual reasoning method, namely a control experiment is constructed by the system, and the importance of each factor is evaluated by changing the degree of change of the single factor observation time sequence relevancy.

A conditional timing pattern graph is constructed based on the extracted regular features, condition dimensions including network load levels, time periods, etc., are defined, and corresponding pattern nodes are created for each condition combination. And establishing a connection relation between mode nodes by analyzing the evolution rule of the modes in the historical data. And integrating constraint relation information of a plurality of network areas by adopting a distributed federal learning framework, maintaining a local model in each area, periodically exchanging parameters with a central node, and protecting data privacy while realizing knowledge sharing.

And constructing an adaptive probability map model. Firstly, a multi-dimensional state space is established, wherein the state space comprises a flow horizontal dimension (dividing the flow into a plurality of load levels), a change trend dimension (representing the rising, falling or stable state of the flow), and a time attribute dimension (distinguishing workdays, weekends, holidays and the like). The core components of the model comprise a state transition module, an observation mapping module and an adaptive adjustment module.

The state transition module is responsible for learning and updating transition rules between states. And establishing an initial state transition relation by analyzing the change modes of the state sequences in the historical data. With the continuous collection of new data, the system can dynamically update the transfer relations, and importance weights of the new and old data are considered in the updating process, so that the model can adapt to the change of network environment.

The observation mapping module establishes a corresponding relation between states and actual flow values, and statistically analyzes flow value distribution possibly occurring under each state to form an observation probability map. The mapping relation is continuously optimized along with the accumulation of new data, so that the prediction result is more accurate.

The self-adaptive adjusting module is responsible for three-layer dynamic adjustment, namely, adjustment of parameter updating frequency, dynamic change of updating frequency of model parameters by a system according to flow fluctuation degree, adjustment of state transition probability, updating of transition probability through weighting fusion of new and old data, optimization of a state space structure, periodic evaluation and adjustment of granularity of state division, and merging of low-use-rate states or sub-division high-frequency use states.

And constructing an energy function integrating various information, wherein the energy function comprises three parts of constraint relation strength, a conditional time sequence rule and probability distribution. The weight coefficient is adaptively adjusted according to the change of the prediction precision. And correcting the initial predicted value by iterative optimization energy function of the gradient descent algorithm until all time sequence constraint conditions are met, outputting a corrected predicted result considering time sequence relativity, continuously monitoring a predicted precision index in the optimization process, and adjusting model parameters according to performance feedback.

Illustratively, the initial attributes of the three types of nodes at a certain moment are short-term nodes [4.2Gbps, (4.0,0.3,4.5), (0.1,0.8) ] representing current traffic of 4.2Gbps, mean of 4.0, variance of 0.3, peak of 4.5, growth rate of 0.1, periodic intensity of 0.8, medium-term nodes [4.0Gbps, (3.8,0.2,4.2), (0.05,0.6) ], and long-term nodes [3.8Gbps, (3.6,0.1,4.0), (0.02,0.4) ]. The initial edge association degree is short-term to medium-term of 0.7, and medium-term to long-term of 0.6. After the attention mechanism update, the node state vector changes to short term [4.15,4.0,0.12], medium term [4.05,3.9,0.06], long term [3.85,3.7,0.03], and the edge association strength update is 0.75 and 0.65.

In a 6-hour sliding window, short-term flow amplitude of 1.2Gbps, frequency of 0.2 times/hour and peak time of 10:00 are calculated, medium-term amplitude of 0.8Gbps, frequency of 0.1 times/hour and peak value of 11:00 are calculated, and long-term amplitude of 0.5Gbps, frequency of 0.05 times/hour and peak value of 12:00 are calculated. The counterfactual analysis showed that a short term flow increase of 1Gbps resulted in a mid-term change of 0.6Gbps and a mid-term increase of 1Gbps resulted in a long term change of 0.4Gbps. The feature importance scores are amplitude 0.8, frequency 0.6 and phase 0.4.

The conditional timing modes include peak mode (trigger condition: short term traffic >4.5Gbps, association strength 0.85 for 2 hours), plateau mode (3.5-4.5 Gbps, strength 0.7 for 6 hours), valley mode (< 3.5Gbps, strength 0.6 for 4 hours). Mode transition probabilities are peak to plateau 0.8, plateau to valley 0.6, valley to plateau 0.7. At the current standard deviation of flow of 0.8Gbps, the model updates the parameters every 10 minutes.

In the energy function optimization process, the initial weight coefficients are all about 0.33. The current constraint term value is-0.65, the rule term probability is 0.8, and the probability term density is 0.7. And obtaining a corrected predicted value of short-term 4.12Gbps, medium-term 4.08Gbps and long-term 3.95Gbps through 50 times of iterative optimization, updating weights to be a constraint term of 0.35, a rule term of 0.35 and a probability term of 0.30, and finally reducing the energy value to-0.2 to meet a convergence condition.

In the embodiment, by introducing a dynamic heterogeneous graph network structure and a space-time attention mechanism, the modeling capability of the relevance of the multi-scale time sequence is obviously improved, and the understanding depth of the network flow change rule is enhanced by adopting a method of inverse fact reasoning and a conditional time sequence pattern diagram. Based on the design of the federal learning framework and the self-adaptive probability model, the perception and response capability of the model to network state change are improved, and the accurate correction and self-adaptive adjustment of the prediction result are realized through an optimization mechanism of a unified energy function;

In the prior art, a single time scale prediction model is generally adopted for network traffic prediction, or prediction results of a plurality of time scales are simply weighted and combined, complex time sequence association relations among different time scales are ignored, dynamic evolution characteristics of network traffic on the plurality of time scales are difficult to accurately capture, modeling of time correlation in the prior art is mostly based on static correlation coefficients, correlation strength changes caused by network traffic fluctuation cannot be adapted, accuracy of the prediction results in a severe fluctuation period is obviously reduced, and in addition, the prior art method often adopts a fixed weight system for prediction result fusion and lacks self-adaption capability for network state changes;

According to the embodiment, the heterogeneous graph structure comprising a plurality of time scale nodes is constructed, dynamic update of node attributes and edge association strength is realized by combining a space-time attention mechanism, modeling capability of network flow time sequence characteristics is effectively improved, a back-facts reasoning method is adopted to analyze time sequence association dominant factors, association rules are extracted, a prediction model can better understand a mutual influence mechanism between different time scales, a conditional time sequence pattern diagram is constructed, constraint relations between adjacent time scale prediction values are depicted through the probability graph model, update frequency of model parameters is dynamically associated with flow fluctuation degree, quick response capability in a severe fluctuation period is guaranteed, in sum, adaptation capability of the model to severe fluctuation of network flow is remarkably improved while prediction accuracy is maintained, time sequence consistency of a prediction result is enhanced, requirements on flow prediction accuracy and stability in an actual network environment are better met, and a more reliable decision basis is provided for network resource scheduling and performance optimization.

Fig. 2 is a graph of comparison of prediction accuracy under different load conditions corresponding to an artificial intelligence-based communication network resource optimization method according to an embodiment of the present invention, and as shown in fig. 2, the comparison of prediction accuracy under different network load conditions of three different prediction methods is shown, and as the network load changes from low to high, the prediction accuracy of all the methods shows a decreasing trend, but the present invention maintains the highest accuracy under various load conditions. Under the condition of low load (30%), the accuracy of the traditional neural network is 82.3%, the Graph Neural Network (GNN) model reaches 86.8%, but the accuracy of the technical scheme is as high as 91.2%, 4.4% higher than the graph neural network model, and 8.9% higher than the traditional neural network. Under the condition of medium load (60%), the accuracy rates of the three methods are 77.4% (traditional neural network), 83.2% (graph neural network model) and 88.7% (the technical scheme). Under the condition of high load (80%), the accuracy is further reduced to 71.2% (traditional neural network), 79.6% (graph neural network model) and 86.3% (the technical scheme). The most challenging is the peak load (95%) condition, the traditional neural network accuracy is only 65.1%, the graph neural network model is 74.1%, and the technical scheme still remains at a higher level of 82.9%. Notably, the performance attenuation of the technical scheme is minimum when the load is increased, the accuracy rate is only reduced by 8.3 percent from low load to peak load, and the traditional neural network and the graph neural network model are respectively reduced by 17.2 percent and 12.7 percent, which indicates that the scheme has stronger stability and robustness under the condition of high load.

In an alternative embodiment of the present invention,

And obtaining a historical prediction error sequence by calculating a difference value between the actual network flow value and the predicted network flow value. And carrying out statistical analysis on the error sequence, calculating an error mean value as a measurement index of the system prediction deviation, and calculating the standard deviation of the error to represent the prediction fluctuation degree. An autocorrelation function of the error sequence is constructed, correlation between error values at different time intervals is analyzed, and a time sequence dependence characteristic of a prediction error is captured. And calculating an exponentially weighted moving variance, and carrying out weighted combination on the square variance of the error and the mean value at the current moment and the exponentially weighted moving variance at the previous moment, wherein the weighting coefficient decays with time so as to highlight the influence of recent error fluctuation.

And constructing a three-term correction model based on the error statistical characteristics. The error distribution correction term is used for describing the distribution characteristics of the prediction error by utilizing the error mean value and the standard deviation, the time sequence correlation correction term is used for modeling the time sequence dependency relationship of the error sequence based on the autocorrelation function, and the fluctuation correction term is used for reflecting the dynamic fluctuation characteristics of the prediction error through the exponentially weighted moving variance. And (3) distributing initial weight coefficients for the three correction items, and dynamically updating the weight coefficients based on the improvement degree of the prediction performance by adopting a gradient descent method so as to enable the correction effect to be adaptively adjusted along with the change of the prediction error characteristics.

And constructing a historical information storage matrix to store past prediction data, and calibrating an initial prediction value by using an exponential smoothing method. The self-adaptive smoothing factor is determined by the weighted combination of the previous time smoothing factor and the absolute value of the historical prediction error at the current time, the weighting coefficient is dynamically adjusted along with the change of the exponentially weighted moving variance, and the weight of the current error is increased when the fluctuation of the error is aggravated. And calculating the influence degree of the historical data on the current prediction by combining the attention mechanism, and dynamically calibrating the initial predicted value to obtain a first corrected predicted value.

And constructing a multidimensional index vector containing the current values of a plurality of network performance indexes. And calculating the time sequence difference value of the index vector, carrying out accumulation summation to obtain a mutation detection value, comparing the detection value with the difference value between the detection value at the previous moment and a preset threshold value, and judging whether mutation occurs. When the detection value exceeds the mutation judgment threshold value, calculating the relative change rate of each performance index, multiplying the change rate by the corresponding weight coefficient, summing to obtain the mutation degree, adjusting the first correction predicted value by the mutation degree, and outputting a final predicted result. And when the detection value does not exceed the threshold value, directly outputting the first corrected predicted value as a final predicted result.

Illustratively, a historical prediction error sequence of the last 24 hours is calculated, wherein the average error value is 0.2Gbps, and the standard error difference is 0.5Gbps. The autocorrelation function shows a correlation coefficient of 0.6,2 hours to 0.4,3 hours to 0.2 at 1 hour intervals. The mean square error of the current time error and the mean is 0.09, the index weighted moving variance of the previous time is 0.16, the weighting coefficient is 0.3, and the current index weighted moving variance is calculated to be 0.14.

The initial weights of the three correction models are 0.33, and the weights after gradient descent update are respectively 0.35 of error distribution correction term, 0.4 of time sequence related correction term and 0.25 of fluctuation correction term. The historical information storage matrix stores the predicted data of the last 12 hours, the smoothing factor at the previous moment is 0.7, the absolute value of the current error is 0.3, the change quantity of the exponentially weighted moving variance is-0.02, and accordingly the current smoothing factor is determined to be 0.65.

And calculating by an attention mechanism to obtain influence weight distribution of historical data, and calibrating the initial predicted value of 4.5Gbps to obtain a first corrected predicted value of 4.3Gbps. The multidimensional performance index vector comprises 75% of bandwidth occupancy rate, 60% of processor occupancy rate, 50% of memory occupancy rate and 25ms of transmission delay. The calculated mutation detection value 85 is greater than the mutation judgment threshold value 80. The relative change rates of the indexes are respectively 0.2, 0.15, 0.1 and 0.25, the corresponding weights are 0.3, 0.2 and 0.3, and the mutation degree is calculated to be 1.2. Finally, the first correction predicted value 4.3Gbps is multiplied by the mutation degree 1.2, and a predicted result 5.16Gbps is output.

In this embodiment, by comprehensively analyzing the historical prediction error sequence, comprehensively considering the distribution characteristics, time sequence correlation and dynamic fluctuation characteristics of errors, performing prediction value calibration by adopting an exponential smoothing method and an attention mechanism, and dynamically updating the adaptive smoothing factor, the influence degree of historical information can be adjusted in real time according to the fluctuation condition of the prediction error, the response speed of the model to network state change is improved, the timeliness of the prediction result is enhanced, the mutation degree is calculated based on the relative change rate of each performance index, the prediction result is pertinently adjusted, the prediction accuracy of the model in the severe change period of the network state is improved, the precise correction of the network flow prediction result is realized, and an effective solution is provided for improving the accuracy and reliability of network flow prediction.

In an alternative embodiment of the present invention,

And (3) carrying out multi-layer time scale division on the network traffic time sequence, wherein the time scale division is divided into three scale layers of an hour level, a day level and a week level. Each time scale layer is independently configured with an adaptive smoothing factor for smoothing historical data of a corresponding scale. The initial value of the smoothing factor is set according to the span of the time scale, and the larger the scale span is, the smaller the initial smoothing factor is, so that the dependency degree of different time scales on the historical data is reflected.

And calculating the fluctuation rate of the network traffic time sequence, and measuring the fluctuation rate by the ratio of the variation amplitude of the traffic value to the average value at the adjacent time points. And dynamically adjusting the size of a historical data window based on the calculated fluctuation rate, wherein the size of the window is equal to the product of the preset basic window size and the current fluctuation rate. When the flow fluctuation is large, more history information is acquired by increasing the window, and when the fluctuation is small, the window is reduced to highlight the influence of recent data. And in the determined window range, smoothing the historical data by utilizing the adaptive smoothing factors of each time scale layer to obtain smoothing results of different time scales.

The smoothed results for each time scale are organized into a feature matrix, the rows of the matrix representing the different time scales and the columns representing the time steps. And constructing a query vector at the current moment, wherein the query vector comprises a current flow value and recent change trend information. And calculating attention weight based on the query vector and the feature matrix, adopting a scaling dot product attention mechanism, carrying out similarity calculation on each time scale feature in the query vector and the feature matrix, and normalizing by a softmax function to obtain the attention weight. The weight value reflects the importance of different time scale features to the current prediction.

And constructing a historical information storage matrix, and storing a historical flow value, a prediction error and an adaptive smoothing factor. And calculating the similarity between the current state and the historical state by adopting an attention mechanism, and extracting related historical information. And determining a dynamic learning rate according to the variation of the prediction error, and reducing the learning rate to reduce the correction amplitude and improve the prediction stability when the variation of the prediction error is increased.

And correcting the initial predicted value by introducing a residual error connection structure. Firstly, calculating correction function output according to the characteristics of each time scale, multiplying the correction output by a corresponding self-adaptive smoothing factor, and then carrying out weighted summation with attention weight to obtain a residual error item. And finally, adding the residual error item and the initial predicted value to obtain a corrected predicted value considering a plurality of time scale features.

Illustratively, the network traffic time series is divided into three scale layers of hour, day and week, with initial smoothing factors set to 0.8, 0.6 and 0.4, respectively. The current fluctuation rate is calculated to be 1.5, the basic window size is set to be 24 hours, and the actual window size is obtained to be 36 hours. Three time scales were smoothed over a 36 hour window with an hour level smoothing result of [4.2,4.0,3.8..+ -. Gbps, a day level of [4.1,3.9,3.7..] Gbps, and a week level of [4.0,3.8,3.6.] Gbps.

A feature matrix is constructed, the matrix size being 3 x 36, representing smoothed values for three time scales over 36 time steps. The current query vector is [4.5,0.2,0.1] which represents the current flow rate of 4.5Gbps, the recent growth rate of 0.2 and the acceleration of 0.1. The attention weight is calculated to be 0.5 hour, 0.3 day and 0.2 week.

The history information storage matrix records the last 72 hours of data, including the flow value, prediction error, and smoothing factor for each time point. The current prediction error variation is 0.3Gbps, and the dynamic learning rate is calculated according to the current prediction error variation. And extracting historical information with high similarity with the current state for reference.

The initial predicted value is 4.8Gbps, and the correction function output of each time scale is-0.3, -0.2 and-0.1 respectively. The corrected output is multiplied by a corresponding smoothing factor of-0.24, -0.12, -0.04, and then weighted and summed with the attention weight to obtain a residual term of-0.16. And finally, adding the residual error item and the initial predicted value to obtain a corrected predicted value of 4.64Gbps.

In the embodiment, through the multi-layer division of the time sequence and the configuration of the independent smoothing factors, the differentiation processing of the characteristics of different time scales is realized, the data sampling range can be adaptively stretched and contracted according to the flow change condition based on the history data window dynamically adjusted by the fluctuation rate, the pertinence and the effectiveness of the utilization of the history data are improved, the attention mechanism is adopted to adaptively fuse the characteristics of a plurality of time scales, the correlation degree of the characteristics of different scales and the current state is calculated, the dynamic evaluation and the selective fusion of the characteristic importance are realized, the history information storage matrix and the dynamic learning rate mechanism are introduced, the history experience can be fully utilized in the correction process, and the correction strength is flexibly adjusted according to the change of the prediction error;

The existing network flow prediction correction method generally adopts a fixed time window and a unified smoothing factor to process historical data, cannot adapt to dynamic change characteristics of network flow on different time scales, and meanwhile, the characteristics of the different time scales are often simply overlapped or averaged, so that the difference of importance degree of each time scale characteristic on the current prediction is ignored;

According to the embodiment, through dynamic fusion of the multi-scale characteristics and self-adaptive parameter adjustment, the correction process can be better adapted to the change characteristics of network flow, the introduced dynamic window mechanism and residual error connection structure effectively improve the response capability and correction stability to abnormal fluctuation, the self-adaptability and robustness of the correction process are obviously enhanced while the correction effect is ensured, the requirements on flow prediction accuracy in an actual network environment can be better met, and more reliable technical support is provided for efficient scheduling and optimization management of network resources.

Fig. 3 is a graph of prediction error data under different fluctuation scenarios corresponding to an artificial intelligence-based communication network resource optimization method according to an embodiment of the present invention, and as shown in fig. 3, average absolute percentage error comparison of each prediction method under different fluctuation scenarios is shown. The data clearly show that the technical scheme shows the lowest prediction error in all fluctuation scenes. Under the low fluctuation scene (fluctuation rate < 0.5), the average absolute percentage error of the technical scheme is 3.42%, which is reduced by 41.7% compared with 5.87% of the traditional exponential smoothing method and by 30.5% compared with 4.92% of the SARIMA model. Along with the increase of the flow fluctuation rate, the advantages of the technical scheme are more obvious, and particularly under a high fluctuation scene (the fluctuation rate is less than or equal to 1.5 and is less than or equal to 3.0), the average absolute percentage error of the technical scheme is 8.76 percent, which is reduced by 31.9 percent compared with 12.87 percent of a GRU network and is reduced by 23.3 percent compared with 11.42 percent of a TCN model. Under the extreme fluctuation scene (the fluctuation rate is more than or equal to 3.0), the average absolute percentage error of the technical scheme is 17.52 percent, which is reduced by 46.6 percent compared with 32.78 percent of an exponential smoothing method, and is reduced by 24.1 percent compared with 23.08 percent of the nearest TCN model. From the average absolute percentage error, the technical scheme is 8.64%, which is reduced by 22.0% compared with the TCN model (11.08%) with the best performance in the prior art. The effectiveness of the adaptive smoothing factor and the dynamic window adjustment mechanism in the technical scheme when coping with different fluctuation rate scenes is fully proved, and the adaptive smoothing factor and the dynamic window adjustment mechanism have obvious prediction stability advantages especially in high fluctuation and extreme fluctuation scenes.

An artificial intelligence based communication network resource optimization system comprising:

The present invention may be a method, apparatus, system, and/or computer program product. The computer program product may include a computer readable storage medium having computer readable program instructions embodied thereon for performing various aspects of the present invention.

It should be noted that the above embodiments are merely for illustrating the technical solution of the present invention and not for limiting the same, and although the present invention has been described in detail with reference to the above embodiments, it should be understood by those skilled in the art that the technical solution described in the above embodiments may be modified or some or all of the technical features may be equivalently replaced, and these modifications or substitutions do not make the essence of the corresponding technical solution deviate from the scope of the technical solution of the embodiments of the present invention.

Claims

1. The communication network resource optimization method based on artificial intelligence is characterized by comprising the following steps:

2. The method of claim 1, wherein collecting historical network traffic data, network device performance data, and user traffic demand data in the communication network, decomposing the historical network traffic data into trend terms, period terms, and residual terms by a time series decomposition algorithm, and calculating bandwidth occupancy, processor occupancy, memory occupancy, and transmission delay in the network device performance data comprises:

3. The method of claim 1, wherein the forming of the trend term, the period term, and the residual term into the input vector, adding the input vector to the long-short time memory network, performing a gating operation on the input vector through a forgetting gate, an input gate, and an output gate, updating the cell state to obtain the hidden state, and performing a linear transformation on the hidden state to obtain the time sequence code comprises:

4. The method of claim 1, wherein composing the bandwidth occupancy, the processor occupancy, the memory occupancy, and the transmission delay into a performance index vector, introducing an attention mechanism to calculate the performance index vector to obtain an importance weight, and performing feature fusion on the time sequence code and the importance weight to obtain a fusion feature comprises:

5. The method of claim 1, wherein processing the fusion feature using a multi-scale prediction mechanism to predict short-, medium-, and long-term network traffic values, respectively, and setting adaptive weights based on prediction errors at each time scale, and wherein weighting and combining the prediction results at the plurality of time scales to output an initial network traffic prediction value comprises:

6. The method of claim 5, wherein extracting, from the graph structure network, a regular feature of a time-series association between time scales over time, establishing a constraint relationship between adjacent time scale predictors according to the regular feature, correcting an initial network traffic predictor based on the constraint relationship and a federal learning framework, and generating a corrected predictor considering the time-series association comprises:

7. The method of claim 1, wherein constructing a correction model based on statistical characteristics of historical prediction errors, dynamically calibrating an initial network traffic prediction value using an exponential smoothing method, and correcting the prediction value in combination with a sudden change detection result of a network performance index, the obtaining a network traffic prediction result comprises:

8. The method of claim 7 wherein constructing a historical information storage matrix according to an exponential smoothing method, dynamically calibrating the initial network traffic prediction value by an adaptive smoothing factor and an attention mechanism to obtain a first modified prediction value comprises:

9. Communication network resource optimization system based on artificial intelligence for implementing the method according to any of the previous claims 1-7, characterized in that it comprises: