Search

CN-122028118-A - Adaptive bandwidth prediction allocation method based on multidimensional link state sensing

CN122028118ACN 122028118 ACN122028118 ACN 122028118ACN-122028118-A

Abstract

The invention relates to a self-adaptive bandwidth prediction allocation method based on multidimensional link state sensing, and belongs to the technical field of communication networks. The method comprises the steps of obtaining multi-dimensional operation characteristics of a link, carrying out weighted fusion to generate comprehensive state characteristics of the link, constructing a time-space association characteristic sequence, executing time sequence characteristic extraction and dynamic weight distribution, predicting future time slot bandwidth requirements, outputting predicted values and credibility, establishing distribution constraint conditions according to the predicted values, the credibility and the link bearing capacity, executing layering decision and dynamic weight adjustment to generate a bandwidth distribution result, obtaining actual operation data in real time, generating prediction deviation, adjusting time-space association weights and distribution decision parameters based on the deviation, and realizing iterative updating of a prediction mechanism and a distribution mechanism. The bandwidth prediction accuracy and the resource allocation rationality are improved through multidimensional sensing, prediction reliability guiding and dynamic feedback adjustment, and the method is suitable for dynamic bandwidth management in a complex network environment.

Inventors

  • WANG RUIJIN
  • SU BENJIE

Assignees

  • 上海逸苑信息科技有限公司

Dates

Publication Date
20260512
Application Date
20260320

Claims (12)

  1. 1. The adaptive bandwidth prediction allocation method based on multidimensional link state sensing is characterized by comprising the following steps of: S1, acquiring a multi-dimensional basic feature of a link, calculating information entropy of the basic feature of the link based on an entropy weight method, and generating an initial weight coefficient of each dimensional feature; acquiring link state parameters, adjusting the initial weight coefficient based on a preset weight correction rule, and carrying out weighted combination on each dimensional characteristic based on the corrected weight coefficient to generate a link comprehensive state characteristic; s2, constructing a time-space association feature sequence, executing time sequence feature extraction and dynamic weight distribution operation, and predicting the bandwidth demand of a future time slot based on the extracted time sequence feature and the distributed weight to generate a bandwidth demand predicted value and prediction reliability; S3, establishing an allocation constraint condition based on a link transmission performance core index according to the bandwidth demand predicted value, the prediction reliability and the link bearing capacity, executing layering decision and dynamic weight adjustment operation, and generating a bandwidth allocation result corresponding to each service; and S4, acquiring actual link operation data after bandwidth allocation in real time, comparing actual bandwidth use data with the bandwidth demand predicted value to generate a predicted deviation, adjusting the space-time association weight and allocation decision parameter based on the predicted deviation, and updating a prediction mechanism and an allocation mechanism.
  2. 2. The method according to claim 1, wherein the calculating the information entropy of the link basic feature based on the entropy weight method specifically comprises: Carrying out standardized pretreatment on the obtained multi-dimensional basic characteristics of the link, aiming at the standardized basic characteristics of the link, obtaining the characteristic values of all samples in the corresponding dimension, and calculating the cumulative sum of the characteristic values of all samples in the corresponding dimension; calculating the duty ratio of the characteristic value of the sample in the cumulative sum of the characteristic values of all samples in the dimension based on the characteristic value of each sample and the cumulative sum; according to the duty ratio of each sample and the number of samples in corresponding dimensions, calculating the information entropy of the link basic features, traversing the link basic features in all dimensions, and generating the information entropy respectively corresponding to the link basic features in each dimension.
  3. 3. The method according to claim 1, wherein the dynamically adjusting the initial weight coefficient based on a preset weight correction rule specifically comprises: Acquiring real-time running state parameters of a link, and presetting influence coefficients of the real-time running state parameters on the basis of the association degree of the basic features of the links in different dimensions and the transmission performance of the links; correspondingly matching the acquired real-time running state parameters with the link basic characteristics of each dimension to generate real-time running state parameters and influence coefficients corresponding to the link basic characteristics; And calculating a weight correction coefficient of the basic characteristic according to the real-time value of the real-time running state parameter and the corresponding influence coefficient, and carrying out combined operation on the weight correction coefficient and the corresponding initial weight coefficient to obtain the corrected weight coefficient of the basic characteristic of each dimension link.
  4. 4. The method according to claim 1, wherein the construction of the spatio-temporal correlation feature sequence is performed by: Acquiring comprehensive state characteristics of a link, and constructing a characteristic time dimension sequence by combining historical operation characteristic data of the link within a preset duration; synchronously collecting real-time operation characteristics of different nodes of a link, and constructing a characteristic space dimension sequence; and combining the time dimension sequence and the space dimension sequence based on a preset space-time association mapping rule to generate a space-time association characteristic sequence.
  5. 5. The method according to claim 1, wherein the performing of the timing feature extraction and dynamic weight distribution operations is performed by: Based on the space-time correlation feature sequence, performing multi-scale decomposition on the space-time correlation feature sequence, extracting short-term fluctuation components and long-term trend components, and constructing a multi-scale time sequence feature set; inputting the multi-scale time sequence feature set into a pre-trained time sequence feature extraction network to generate time sequence hidden state features; Introducing an attention mechanism, calculating contribution degrees of each historical time step and each scale feature in the time sequence hidden state feature to current prediction, and generating a corresponding attention weight; And carrying out weighted aggregation on the time sequence hidden state features based on the attention weight to generate a time sequence feature weight distribution result.
  6. 6. The method according to claim 5, wherein the calculating the contribution of each historical time step and each scale feature in the time sequence hidden state feature to the current prediction generates a corresponding attention weight, and the specific method is as follows: acquiring hidden state vectors of each historical time step in the time sequence hidden state characteristics; Calculating similarity scores between each hidden state vector and a preset query vector; and carrying out normalization processing on the similarity score to generate attention weights corresponding to the historical time steps, wherein the numerical values of the attention weights represent the importance degree of hidden state vectors corresponding to the historical time steps in bandwidth demand prediction.
  7. 7. The method according to claim 1, wherein the predicting the bandwidth requirement of the future time slot is performed by: Acquiring the multi-scale time sequence feature set and a time sequence feature weight distribution result, and carrying out weighted combination on the multi-scale time sequence feature set to generate a fused time sequence feature vector; Inputting the time sequence feature vector into a multitask prediction network, generating a point predicted value of the bandwidth demand through a first output branch, and generating a predicted interval of the bandwidth demand through a second output branch; and calculating prediction reliability according to the prediction interval, and outputting the point predicted value and the corresponding prediction reliability.
  8. 8. The method according to claim 1, wherein the establishing the allocation constraint condition based on the core index of link transmission performance comprises: The method comprises the steps of obtaining bandwidth demand predicted values and prediction credibility of each service to be allocated, and the total amount of residual bandwidth resources and a link state parameter set of a current link, presetting threshold coefficients of each core index according to the prediction credibility, respectively calculating upper limit thresholds of bandwidth resource utilization rate based on the threshold coefficients and lower limit thresholds of service transmission performance corresponding to each parameter in the link state parameter set, and constructing a constraint condition set of bandwidth allocation.
  9. 9. The method according to claim 1, wherein the performing hierarchical decision and dynamic weight adjustment operations comprises: The method comprises the steps of obtaining constraint condition sets of bandwidth allocation and priority levels of all services, carrying out hierarchical division on all services according to the priority levels to generate bandwidth allocation priorities of all the levels, allocating initial bandwidth quota to all the levels by combining bandwidth demand predicted values, prediction credibility and allocation constraint conditions, introducing a level weight dynamic adjustment factor, carrying out dynamic iteration adjustment on level weights and bandwidth quota of all the services according to actual bandwidth demand feedback of all the levels of the services and link transmission performance core index standard reaching conditions, and outputting bandwidth allocation results corresponding to all the services.
  10. 10. The method according to claim 9, wherein the introducing the dynamic adjustment factor of the hierarchical weight is specifically: acquiring link real-time operation data, extracting weight adjustment core parameters, and carrying out standardized processing on the weight adjustment core parameters to generate standardized parameter values; Based on a preset weight distribution rule, distributing corresponding weight coefficients to each standardized parameter value; carrying out weighted summation on each standardized parameter value and the corresponding weight coefficient, and calculating to generate a dynamic adjustment factor of the hierarchical weight; And correcting the dynamic adjustment factor of the hierarchical weight according to the real-time standard condition of the link transmission performance core index, and generating the corrected dynamic adjustment factor of the hierarchical weight.
  11. 11. The method according to claim 1, wherein the generating the prediction bias is performed by: acquiring actual bandwidth use data of each time slot and bandwidth demand predicted values corresponding to each time slot, aligning according to time sequence, and constructing a comparison sample pair of the predicted values and the actual values; Calculating absolute difference values and relative difference values of the comparison sample pairs, and carrying out weighted combination on the absolute difference values and the relative difference values to generate instantaneous prediction deviation; And carrying out smoothing processing on the instantaneous prediction deviations of a plurality of time slots in a preset sliding window to generate a globally stable prediction deviation.
  12. 12. The method according to claim 1, wherein the updating of the prediction mechanism and the allocation mechanism is performed by: The method comprises the steps of obtaining a global stable prediction deviation, carrying out grading judgment on the prediction deviation by combining with link real-time operation data, adjusting the construction parameters of a time-space correlation characteristic sequence and training parameters of a time-sequence characteristic extraction network according to a prediction mechanism, synchronously adjusting weight distribution logic of an attention mechanism, correcting output deviation of a bandwidth demand prediction model according to a deviation grading result and link transmission performance core index standard reaching condition according to the distribution mechanism, adjusting the priority weight of a layering decision and the calculation parameters of a hierarchical weight dynamic adjustment factor according to the deviation grading result, and updating the distribution rule of an initial bandwidth quota.

Description

Adaptive bandwidth prediction allocation method based on multidimensional link state sensing Technical Field The invention belongs to the technical field of communication networks, and particularly relates to a self-adaptive bandwidth prediction allocation method based on multidimensional link state sensing. Background With the rapid development of 5G/6G communications, cloud computing, industrial internet and real-time interactive applications, network traffic has exhibited explosive growth and has exhibited highly dynamic and heterogeneous features. Emerging services such as high-definition video streaming, augmented reality/virtual reality (AR/VR), telemedicine, large-scale Internet of things and the like provide very stringent requirements for network bandwidth, transmission delay and service quality guarantee. Under the background, how to efficiently and intelligently allocate limited bandwidth resources to meet the differentiated requirements of diversified services becomes a core problem to be solved in the network communication field. Traditional bandwidth allocation methods are mainly divided into two types, namely static allocation and dynamic allocation based on fixed rules. The static allocation method (such as Fixed Bandwidth Allocation (FBA)) allocates a fixed bandwidth quota for each service or user in advance, is simple to implement, but cannot adapt to real-time fluctuation of network traffic, causes idle resources when the service load is light, and cannot guarantee service quality when the service load is heavy. Although Dynamic Bandwidth Allocation (DBA) methods (such as polling, priority queue based, etc.) based on fixed rules can be adjusted to a certain extent according to real-time requirements, the decision is dependent on a preset threshold or simple flow measurement, and it is difficult to cope with bursty traffic and complex and changeable network environments, which easily causes congestion aggravation, delay jitter, and low resource utilization. In recent years, to solve the above-described problems, researchers have begun to introduce machine learning techniques into bandwidth prediction and allocation. By predicting the bandwidth requirements of future time slots, more aggressive scheduling of resources may be achieved. For example, using Artificial Neural Network (ANN), long-short-term memory network (LSTM), or transfomer models to predict traffic and direct bandwidth allocation based on the prediction results has been shown to be effective in improving network performance. However, the existing allocation method based on prediction still has obvious limitations that firstly, most prediction models only depend on historical flow data with single dimension, the comprehensive influence of key information such as real-time state (such as bandwidth utilization rate, queue depth, channel quality) and service attribute of a link is ignored, so that prediction accuracy is limited, secondly, the prediction models are usually trained based on static historical data, in actual deployment, as a network environment dynamically evolves (i.e. concept drifting), the pre-training models can be reduced in performance due to incapability of adapting to a new flow mode and lack of an effective online updating mechanism, and finally, bandwidth allocation decisions are often independent of the credibility of prediction results, when prediction deviation is large, resource allocation mismatch can be caused, and network performance degradation is further aggravated. Therefore, how to construct a closed-loop bandwidth management method which can integrate multidimensional link state sensing, has high-precision adaptive prediction capability and can dynamically optimize allocation strategies according to prediction reliability and real-time feedback has become a technical problem to be solved by those skilled in the art. Disclosure of Invention In order to solve the problems in the prior art, the invention provides a self-adaptive bandwidth prediction allocation method based on multidimensional link state sensing, and the purpose of the invention can be realized by the following technical scheme: S1, acquiring a multi-dimensional basic feature of a link, calculating information entropy of the basic feature of the link based on an entropy weight method, and generating an initial weight coefficient of each dimensional feature; acquiring link state parameters, adjusting the initial weight coefficient based on a preset weight correction rule, and carrying out weighted combination on each dimensional characteristic based on the corrected weight coefficient to generate a link comprehensive state characteristic; s2, constructing a time-space association feature sequence, executing time sequence feature extraction and dynamic weight distribution operation, and predicting the bandwidth demand of a future time slot based on the extracted time sequence feature and the distributed weight to generate a bandwidth