Bandwidth prediction and scheduling method and system for edge network cost optimization
By employing an adaptive anomaly repair and multi-granularity feature extraction bandwidth prediction model, combined with peak-valley asymmetric loss function and marginal cost-sensitive scheduling, the problem of accurate bandwidth demand prediction and cost optimization in edge networks is solved, achieving efficient real-time scheduling decisions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUIZHOU INST OF TECH
- Filing Date
- 2026-05-29
- Publication Date
- 2026-06-26
AI Technical Summary
Existing technologies fail to effectively predict outliers and temporal patterns of bandwidth demand in edge networks, resulting in insufficient prediction accuracy, neglecting cost risks during peak periods, and traditional scheduling strategies fail to achieve global transmission cost minimization and dynamic adjustment of service quality.
An adaptive anomaly repair method is used to extract multi-granularity statistical features, and a bandwidth demand prediction model based on temporal convolutional networks and bidirectional gated recurrent units is constructed. A peak-valley asymmetric weighted loss function is designed, and combined with a marginal cost-sensitive online scheduling optimization model, real-time scheduling is achieved through a greedy heuristic algorithm.
It improves the accuracy of bandwidth demand forecasting, reduces scheduling cost risks caused by insufficient resource preparation, ensures that tasks are assigned to nodes that meet quality standards, meets the low latency requirements of edge networks, and effectively reduces total transmission costs.
Smart Images
Figure CN122293532A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of network resource scheduling technology, and more specifically to a bandwidth prediction and scheduling method and system for cost optimization of edge networks. Background Technology
[0002] With the rapid development of mobile internet and IoT technologies, data-intensive applications such as video streaming, online games, and content delivery networks are being deployed more and more widely in edge computing networks. In such scenarios, network bandwidth cost is one of the key factors affecting service operation costs. An unreasonable bandwidth scheduling scheme can lead to overload of some edge nodes and a surge in transmission costs, while other nodes remain idle, seriously affecting service quality and system economy.
[0003] Current technologies still have shortcomings in dealing with the complex scenarios unique to edge networks. Most existing prediction methods ignore outliers and temporal patterns in bandwidth demand data, resulting in low-quality basic data and insufficient prediction accuracy. Traditional loss functions treat all time periods the same, failing to distinguish the huge cost risks caused by underestimating peak periods. Existing scheduling strategies usually only consider the current load of nodes, ignoring the nonlinear characteristics of marginal costs increasing with load, making it difficult to minimize global transmission costs. Most methods do not incorporate real-time dynamic changes in service quality into scheduling decisions, which may assign tasks to nodes with poor network connectivity.
[0004] Therefore, there is a need for a real-time online scheduling method that can accurately predict future bandwidth demand based on high-quality data and perform marginal cost-sensitive and service quality-aware scheduling accordingly, so as to effectively reduce the total transmission cost of edge networks. Summary of the Invention
[0005] The technical problem to be solved by the present invention is to address the shortcomings of the prior art by providing a bandwidth prediction and scheduling method and system for cost optimization of edge networks.
[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0007] A bandwidth prediction and scheduling method for edge network cost optimization includes the following steps:
[0008] Step S1: Collect historical bandwidth demand data of task nodes in the edge network, perform adaptive anomaly repair on the historical bandwidth sequence, extract multi-granularity statistical features, and construct a standardized input feature tensor.
[0009] Step S2: Construct a bandwidth demand prediction model based on a hybrid of temporal convolutional network and bidirectional gated recurrent unit. Train the bandwidth demand prediction model using the standardized input feature tensor and output the bandwidth demand prediction value for the future prediction step size.
[0010] Step S3: Design a peak-valley asymmetric weighted loss function to train the bandwidth demand prediction model, and assign corresponding penalty weights according to whether the prediction point is in the peak period of bandwidth demand and the direction of prediction error.
[0011] Step S4: Based on the bandwidth demand forecast value of the future prediction step size output by the bandwidth demand forecast model, construct a marginal cost sensitive online scheduling optimization model with the objective of minimizing the total transmission cost in the prediction time domain.
[0012] Step S5: Use a greedy heuristic algorithm based on marginal cost increment sorting to solve the scheduling optimization model online, and output a bandwidth allocation scheme that satisfies the service quality constraints and capacity constraints.
[0013] Furthermore, step S1 specifically includes the following steps:
[0014] Step S1.1: For the historical bandwidth sequence of a single task node, use local weighted regression to smoothly decompose it into trend components and residual components;
[0015] Step S1.2: Calculate the joint anomaly score for the residual components. When the joint anomaly score exceeds the adaptive threshold, the corresponding time point is determined to be an anomaly point. The adaptive threshold is the sum of the mean of all joint anomaly scores and three times the standard deviation.
[0016] Step S1.3: Extract a context segment with a preset neighborhood radius from the anomaly point, use dynamic time warping to retrieve a preset number of candidate segments with the most similar shape to the context segment from the historical normal records, weight and aggregate the corresponding position values of the candidate segments to obtain the repair value, and output the repaired bandwidth sequence.
[0017] Step S1.4: Extract multi-granularity statistical features from the repaired bandwidth sequence, and construct a standardized input feature tensor after normalization.
[0018] Further, in step S1.2, the joint anomaly score is a weighted sum of the outlier score and the mutation score, wherein the outlier score is the result of the local outlier factor value of the residual component at the corresponding time after max-min normalization; the mutation score is the result of the absolute value of the local Z-score of the first difference of the residual component after max-min normalization; the joint anomaly score is equal to the sum of the outlier score and the mutation score multiplied by their respective weight coefficients.
[0019] Further, in step S1.3, the weighted aggregation of the corresponding position values of the candidate fragments to obtain the repair value specifically includes:
[0020] The distance between segments is a dynamic time-warped distance with a trend penalty, which is a weighted sum of the dynamic time-warped distance of the original sequence and the dynamic time-warped distance of the first-order difference sequence.
[0021] The bandwidth observations of the candidate segments, after dynamic time warping and alignment, and the corresponding positions of the outliers are weighted and aggregated using the exponential decay function of distance as the weight to obtain the repair value. The temperature coefficient is the median of the distances of each candidate segment.
[0022] After repair, the second-order difference mean of the sequence in the neighborhood of the outlier is calculated as the smoothness index. If the smoothness index exceeds the preset quantile limit of the smoothness of the historical normal neighborhood, the temperature coefficient is reduced and the weighted aggregation is re-executed until the smoothness requirement is met or the preset maximum number of iterations is reached.
[0023] Further, in step S1.4, the multi-granularity statistical features specifically include: the original repair value; short-term fluctuation features, which are the bandwidth standard deviation and skewness calculated within a preset first window length; long-term trend features, which are the slope and goodness of fit obtained by least-squares linear fitting within a preset second window length, where the second window length is greater than the first window length; periodic reference features, which are the median bandwidth of the weekday and hour attributes of the current moment in the historical records of this node; and spatial correlation features, which are the current moment bandwidth values of a preset number of neighboring nodes with the highest Spearman correlation coefficient with the current node's bandwidth sequence.
[0024] Furthermore, in step S2, the bandwidth demand prediction model is composed of a temporal convolutional network module, a bidirectional gated recurrent unit module, and an output layer connected sequentially.
[0025] The temporal convolutional network module is composed of a preset number of residual blocks stacked together. The inflation rate of each residual block increases exponentially. Each residual block contains a causal dilated convolution, weight normalization, activation function and dropout layer, and maintains gradient propagation through residual connections.
[0026] The bidirectional gated recurrent unit module takes the output of the temporal convolutional network module as the input sequence, extracts bidirectional temporal dependencies through stacked bidirectional gated recurrent unit layers, and concatenates the forward hidden state and the backward hidden state at each time step for output.
[0027] The output layer performs average pooling along the time axis on the sequence output by the bidirectional gated recurrent unit module to obtain a fixed-length feature vector, which is then mapped by the fully connected layer to the bandwidth requirement prediction value for the future prediction step size.
[0028] Furthermore, in step S3, the peak-valley asymmetric weighted loss function is the mean of the weighted squared errors over all prediction steps;
[0029] The weighting rules are as follows: when the actual bandwidth value is greater than the preset percentile of the historical bandwidth sequence of the node and the prediction error is negative, the weight is a preset penalty value greater than 1; otherwise, the weight is 1.
[0030] Furthermore, in step S4, the marginal cost-sensitive online scheduling optimization model takes minimizing the total transmission cost in the prediction time domain as its objective function;
[0031] The constraints include: demand integrity constraint, which means that the prediction bandwidth requirement of each task node in each prediction step must be fully allocated to each edge node; node capacity constraint, which means that the total load of each edge node in each prediction step does not exceed its maximum carrying bandwidth limit; allocation non-negative constraint, which means that the allocation ratio is non-negative; and service quality constraint, which means that task nodes can only be allocated to edge nodes that meet the preset minimum service quality threshold.
[0032] Furthermore, step S5 specifically includes the following steps:
[0033] Step S5.1: Sort the task nodes in descending order of their predicted bandwidth requirements for the nearest future time.
[0034] Step S5.2: Process each task node in the order of arrangement and filter the set of candidate edge nodes that meet the service quality constraints in all prediction steps;
[0035] Step S5.3: Calculate the total marginal cost increment for each node in the candidate edge node set after accepting all predicted requirements of the current task;
[0036] Step S5.4: Select the candidate edge node with the smallest total marginal cost increment as the target node, assign all current tasks to the target node, and update the load of the target node;
[0037] Step S5.5: If the target node load reaches the maximum carrying bandwidth after the update, remove the target node from the candidate edge node set of subsequent task nodes.
[0038] Step S5.6: Initiate downgrade processing for task nodes with an empty candidate edge node set.
[0039] A bandwidth prediction and scheduling system for edge network cost optimization, used to implement the bandwidth prediction and scheduling method for edge network cost optimization as described in any one of the claims, including:
[0040] The data preprocessing module is used to adaptively repair anomalies in the collected historical bandwidth requirement data of task nodes, extract multi-granularity statistical features, and construct standardized input feature tensors.
[0041] The bandwidth demand prediction module has a built-in bandwidth demand prediction model based on a hybrid of temporal convolutional networks and bidirectional gated recurrent units. It is used to output the bandwidth demand prediction value for the future prediction step size based on the standardized input feature tensor.
[0042] The model training module is used to train the bandwidth demand prediction model using a peak-valley asymmetric weighted loss function.
[0043] The online scheduling optimization module is used to construct a marginal cost-sensitive online scheduling optimization model based on the predicted bandwidth demand. It uses a greedy heuristic algorithm based on marginal cost increment sorting to solve the problem and output a bandwidth allocation scheme.
[0044] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0045] 1. This invention utilizes an adaptive anomaly repair method based on trend residual decomposition and dynamic time warping similar segment retrieval to fully exploit the temporal morphological characteristics of bandwidth demand data, improve the quality of basic data, and provide reliable input for accurate prediction.
[0046] 2. This invention uses a peak-valley asymmetric weighted loss function to train the prediction model, which imposes a higher penalty on the underestimation error during peak periods, forcing the model to prioritize the sufficiency of predictions during peak bandwidth demand periods, effectively reducing the scheduling cost risk caused by insufficient resource preparation.
[0047] 3. This invention constructs an online scheduling optimization model sensitive to marginal costs. It uses a nonlinear cost index to characterize the characteristic that the marginal cost per unit bandwidth increases with the increase of load, and uses a service quality index to quantify the connection quality between task nodes and edge nodes, ensuring that tasks are only assigned to nodes that meet the quality standards.
[0048] 4. This invention uses a greedy heuristic algorithm based on marginal cost incremental sorting to achieve real-time scheduling, which can complete the optimal allocation decision in milliseconds, meet the low latency requirements of online scheduling of edge networks, and effectively reduce the total transmission cost of the system. Attached Figure Description
[0049] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:
[0050] Figure 1 This is a flowchart illustrating an embodiment of the present invention;
[0051] Figure 2 This is a schematic diagram of the system structure according to an embodiment of the present invention;
[0052] Figure 3 This is a schematic diagram of the bandwidth demand prediction model structure according to an embodiment of the present invention;
[0053] Figure 4 This is a schematic diagram of the online scheduling solution algorithm according to an embodiment of the present invention. Detailed Implementation
[0054] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
[0055] like Figure 1 As shown, the bandwidth prediction and scheduling method for edge network cost optimization includes the following steps:
[0056] Step S1: Collect historical bandwidth demand data of task nodes in the edge network, perform adaptive anomaly repair on the historical bandwidth sequence, extract multi-granularity statistical features, and construct a standardized input feature tensor.
[0057] Step S2: Construct a bandwidth demand prediction model based on a hybrid of temporal convolutional network and bidirectional gated recurrent unit. Train the bandwidth demand prediction model using the standardized input feature tensor and output the bandwidth demand prediction value for the future prediction step size.
[0058] Step S3: Design a peak-valley asymmetric weighted loss function to train the bandwidth demand prediction model, and assign corresponding penalty weights according to whether the prediction point is in the peak period of bandwidth demand and the direction of prediction error.
[0059] Step S4: Based on the bandwidth demand forecast value of the future prediction step size output by the bandwidth demand forecast model, construct a marginal cost sensitive online scheduling optimization model with the objective of minimizing the total transmission cost in the prediction time domain.
[0060] Step S5: Use a greedy heuristic algorithm based on marginal cost increment sorting to solve the scheduling optimization model online, and output a bandwidth allocation scheme that satisfies the service quality constraints and capacity constraints.
[0061] Step S1 specifically includes the following steps:
[0062] Step S1.1: For the historical bandwidth sequence of a single task node, use local weighted regression to smoothly decompose it into trend components and residual components;
[0063] Step S1.2: Calculate the joint anomaly score for the residual components. When the joint anomaly score exceeds the adaptive threshold, the corresponding time point is determined to be an anomaly point. The adaptive threshold is the sum of the mean of all joint anomaly scores and three times the standard deviation.
[0064] Step S1.3: Extract a context segment with a preset neighborhood radius from the anomaly point, use dynamic time warping to retrieve a preset number of candidate segments with the most similar shape to the context segment from the historical normal records, weight and aggregate the corresponding position values of the candidate segments to obtain the repair value, and output the repaired bandwidth sequence.
[0065] Step S1.4: Extract multi-granularity statistical features from the repaired bandwidth sequence, and construct a standardized input feature tensor after normalization.
[0066] In step S1.2, the joint anomaly score is a weighted sum of the outlier score and the mutation score, wherein the outlier score is the result of the local outlier factor value of the residual component at the corresponding time after max-min normalization; the mutation score is the result of the absolute value of the local Z-score of the first difference of the residual component after max-min normalization; the joint anomaly score is equal to the sum of the outlier score and the mutation score multiplied by their respective weight coefficients.
[0067] In step S1.3, the weighted aggregation of the corresponding position values of the candidate fragments to obtain the repair value specifically includes:
[0068] The distance between segments is a dynamic time-warped distance with a trend penalty, which is a weighted sum of the dynamic time-warped distance of the original sequence and the dynamic time-warped distance of the first-order difference sequence.
[0069] The bandwidth observations of the candidate segments, after dynamic time warping and alignment, and the corresponding positions of the outliers are weighted and aggregated using the exponential decay function of distance as the weight to obtain the repair value. The temperature coefficient is the median of the distances of each candidate segment.
[0070] After repair, the second-order difference mean of the sequence in the neighborhood of the outlier is calculated as the smoothness index. If the smoothness index exceeds the preset quantile limit of the smoothness of the historical normal neighborhood, the temperature coefficient is reduced and the weighted aggregation is re-executed until the smoothness requirement is met or the preset maximum number of iterations is reached.
[0071] In step S1.4, the multi-granularity statistical features specifically include: the original repair value; short-term fluctuation features, which are the bandwidth standard deviation and skewness calculated within a preset first window length; long-term trend features, which are the slope and goodness of fit obtained by least-squares linear fitting within a preset second window length, where the second window length is greater than the first window length; periodic reference features, which are the median bandwidth of the weekday and hour attributes of the current moment in the historical records of this node; and spatial correlation features, which are the current moment bandwidth values of a preset number of neighboring nodes with the highest Spearman correlation coefficient with the current node's bandwidth sequence.
[0072] Historical bandwidth requirement data are collected from M task nodes in the edge network. The original bandwidth sequence of the m-th node is as follows: ,in This represents the actual bandwidth value of node m at sampling time t. This represents the total number of historical sampling points, and performs adaptive anomaly repair and multi-granularity feature construction on the sequence of a single node;
[0073] Locally weighted regression smoothing is used to decompose the sequence into trend and residual components. For each time t, a joint anomaly score is calculated using the following formula:
[0074]
[0075] in, Indicates joint anomaly score, The outlier score is represented by the Local Outlier Factor (LOF) value of the residual sequence at time t, which is then normalized to the interval [0, 1] using a maximal-minimum normalization method. The mutation degree score is obtained by normalizing the absolute value of the Z-score of the first difference of the residual to [0, 1] within a local neighborhood of 36 with a window length centered at t.
[0076] When joint anomaly score When the threshold is exceeded, the corresponding time point is determined to be an anomaly. The adaptive threshold is the sum of the mean of all joint anomaly scores and three times the standard deviation.
[0077] For anomaly time t, anomaly context segments with a neighborhood radius of 10 are extracted, resulting in a total segment length of 21. Within historical anomaly-free time periods, all possible continuous segments are extracted with the same length to form a historical candidate segment set. The three segments most morphologically similar to the anomaly context segments are retrieved. To measure the morphological similarity between two segments, a weighted distance between segments is defined as a similarity comparison metric, with the specific formula as follows:
[0078]
[0079] in, Indicates the weighted distance between segments. This represents the i-th candidate fragment in the set of historical candidate fragments. This represents an exception context segment centered on t. This represents the dynamic time-warped distance between two original sequence segments. Let these represent the first-order difference sequences of the candidate fragment and the anomalous context fragment, respectively. Represents the dynamic time-normalized distance between two difference sequences;
[0080] calculate With each of the historical candidate fragment sets The weighted distance is used to select the three segments with the smallest distance as the result for weighted aggregation repair;
[0081] The repair value is obtained by weighted aggregation of the values at the corresponding positions at time t from the three candidate segments after dynamic time warping and alignment. The specific formula is as follows:
[0082]
[0083] in, This represents the anomaly repair value at time t. This represents the bandwidth observation value corresponding to the current time t in the i-th candidate segment, determined by the dynamic time warping alignment path. Represents the alignment offset of the i-th candidate segment, which is an integer, representing the displacement added to the original time index t. The weight assigned to the i-th candidate segment by exponential decay based on distance is given by the following formula:
[0084]
[0085] in, Indicates the index of the candidate fragment. Represents the temperature coefficient, usually taken as The median of the three distance values;
[0086] The smoothness is calculated within a neighborhood centered at t with a radius of 12 using the following formula:
[0087]
[0088] in, This represents the neighborhood smoothness index, which is the mean of the absolute values of the second-order differences within the neighborhood. This represents the second difference of the repair sequence at time k. Indicates the neighborhood radius;
[0089] Simultaneously, calculate all historical normal neighborhoods of the sequence, i.e., those without outliers. The set of values, taking its 95th quantile as the upper limit, if If the value is less than or equal to the upper limit, the smoothness after repair is considered to meet the standard, and the repair value is accepted. Otherwise, the temperature coefficient is reduced by 20% (i.e., multiplied by 0.8) and the weighted aggregation is re-executed. The process is repeated a maximum of 2 times, and the final repaired sequence is output.
[0090] For the repaired sequence, a feature vector is constructed for each time step, specifically including: the original repaired value; short-term fluctuation features, i.e., the standard deviation and skewness calculated within a sliding window of length 12; long-term trend features, i.e., least-squares linear fitting performed within a sliding window of length 168, extracting the slope as the trend strength, and the goodness-of-fit R. 2 As a trend stability indicator; periodic reference feature, namely the historical bandwidth median of the grouping of the day of the week + hour of time t; spatial correlation feature, namely the Spearman correlation coefficient of the bandwidth sequence between nodes, and the three-dimensional vector formed by the bandwidth values of the three neighboring nodes with the highest correlation at time t.
[0091] All eigenvalues are normalized to [0, 1] by min-max and then concatenated into a single eigenvector. The vector is then aligned by time step and node to obtain the input tensor.
[0092] In step S2, the bandwidth demand prediction model is composed of a temporal convolutional network module, a bidirectional gated recurrent unit module, and an output layer connected sequentially.
[0093] The temporal convolutional network module is composed of a preset number of residual blocks stacked together. The inflation rate of each residual block increases exponentially. Each residual block contains a causal dilated convolution, weight normalization, activation function and dropout layer, and maintains gradient propagation through residual connections.
[0094] The bidirectional gated recurrent unit module takes the output of the temporal convolutional network module as the input sequence, extracts bidirectional temporal dependencies through stacked bidirectional gated recurrent unit layers, and concatenates the forward hidden state and the backward hidden state at each time step for output.
[0095] The output layer performs average pooling along the time axis on the sequence output by the bidirectional gated recurrent unit module to obtain a fixed-length feature vector, which is then mapped by the fully connected layer to the bandwidth requirement prediction value for the future prediction step size.
[0096] like Figure 3 As shown, the bandwidth demand prediction model consists of a temporal convolutional network module, a bidirectional gated recurrent unit module, and an output layer connected sequentially.
[0097] The temporal convolutional network module consists of four stacked residual blocks with dilation rates of 1, 2, 4, and 8, respectively. Each residual block contains a causal dilated convolution with a kernel size of 5 and 64 output channels, weight normalization, ReLU activation function, a dropout layer with a dropout rate of 0.2, and a causal dilated convolution with a kernel size of 55 and 64 output channels, weight normalization, ReLU activation function, and a dropout layer with a dropout rate of 0.2. The residual connections use 1×1 convolutions to match dimensions, merging the spatial dimension and feature dimension of the input tensor as the number of channels. After processing by the temporal convolutional network module, the output is a temporal feature sequence with a dimension of 64.
[0098] The output of the temporal convolutional network module is fed into a bidirectional gated recurrent unit with a stack of 2 layers and 128 hidden units. The hidden states of the forward gated recurrent unit and the backward gated recurrent unit are concatenated at each time step to obtain an output sequence with a dimension of 256.
[0099] The 256-dimensional sequence output by the bidirectional gated recurrent unit module is averaged along the time axis to obtain a fixed-length feature vector. This vector is then passed through a fully connected layer, a ReLU activation function, a Dropout layer with a dropout rate of 0.2, and another fully connected layer. Finally, it is shaped into a prediction tensor, which outputs the predicted bandwidth requirements of each node for the next 24 time steps.
[0100] During training, the optimizer used was Adam, with an initial learning rate of 1×10⁻⁶. -3 A cosine annealing strategy was used to decay the material to 1×10⁻⁶ within 50 epochs. -5 Batch size B=64; maximum training time is 200 epochs. If the validation loss does not decrease for 15 consecutive epochs, training will stop early.
[0101] In step S3, the peak-valley asymmetric weighted loss function is the mean of the weighted squared errors of all prediction steps;
[0102] The weighting rules are as follows: when the actual bandwidth value is greater than the preset percentile of the historical bandwidth sequence of the node and the prediction error is negative, the weight is a preset penalty value greater than 1; otherwise, the weight is 1.
[0103] Let the error of node m in the u-th training sample at the v-th prediction step be:
[0104]
[0105] in, This represents the error of task node m in the v-th prediction step in the u-th training sample. This represents the predicted bandwidth requirement of task node m in the u-th training sample at the v-th prediction step. This represents the true bandwidth requirement of task node m in the u-th training sample at the v-th prediction step. Indicates the prediction step index;
[0106] The loss function is the mean of the weighted squared errors over all prediction steps, and the specific formula is as follows:
[0107]
[0108] in, Represents the loss function. Indicates the number of samples in the batch. Indicates the total number of task nodes. Indicates the prediction step size. The weight function is determined based on the actual bandwidth value and the prediction error. Specifically, the weight is 3 when the actual bandwidth value is greater than the 80th percentile of the historical bandwidth sequence of task node m and the prediction error is negative; otherwise, the weight is 1.
[0109] In step S4, the marginal cost sensitive online scheduling optimization model takes minimizing the total transmission cost in the prediction time domain as its objective function.
[0110] The constraints include: demand integrity constraint, which means that the prediction bandwidth requirement of each task node in each prediction step must be fully allocated to each edge node; node capacity constraint, which means that the total load of each edge node in each prediction step does not exceed its maximum carrying bandwidth limit; allocation non-negative constraint, which means that the allocation ratio is non-negative; and service quality constraint, which means that task nodes can only be allocated to edge nodes that meet the preset minimum service quality threshold.
[0111] During scheduling The trained bandwidth demand prediction model is used to obtain the predicted bandwidth demand of each task node in the next H steps, and decision variables are defined. This represents the proportion of the transmission tasks of node m allocated to edge node n in step h, which must satisfy the requirement integrity constraint. The specific formula is:
[0112]
[0113] The optimization objective is to minimize the total transmission cost over the next H steps, as shown in the following formula:
[0114]
[0115] in, This represents the total transmission cost of the system within H steps in the time domain. Indicates the number of edge nodes. This represents the unit bandwidth transmission cost coefficient for edge node n. Indicates the predicted time At that time, the existing basic load on edge node n is obtained by accumulating the scheduled but not yet completed transmission tasks. Indicates that task node m at time... The predicted bandwidth demand, This represents a non-linear cost exponent, with a value of 1.5, which causes the marginal cost per unit bandwidth to increase as the load increases. This represents the proportion of bandwidth allocated from task node m to edge node n in step h.
[0116] The service quality index is calculated by comprehensively analyzing the historical network performance data between task node m and edge node n. The specific formula is as follows:
[0117]
[0118] in, Indicates service quality indicators. This indicates the relationship between task node m and edge node n at time [time value missing]. The predicted average network latency is obtained by using an exponentially weighted moving average to predict the historical latency series. This indicates the relationship between task node m and edge node n at time [time value missing]. The predicted latency jitter is obtained by using an exponentially weighted moving average of historical latency sequences. This represents the jitter penalty coefficient, with a value of 0.5. It is used to consider both latency magnitude and latency stability in service quality assessment. If the system requires the average latency between task nodes and edge nodes to not exceed 100ms and the maximum allowable jitter is about 20ms, then the corresponding service quality value is 0.0091. Therefore, the preset minimum service quality threshold can be preset to 0.01.
[0119] Task node m is only allowed to be assigned to edge node n when the service quality index is not lower than the preset minimum service quality threshold, which is the service quality constraint.
[0120] like Figure 4 As shown, step S5 specifically includes the following steps:
[0121] Step S5.1: Sort the task nodes in descending order of their predicted bandwidth requirements for the nearest future time.
[0122] Step S5.2: Process each task node in the order of arrangement and filter the set of candidate edge nodes that meet the service quality constraints in all prediction steps;
[0123] Step S5.3: Calculate the total marginal cost increment for each node in the candidate edge node set after accepting all predicted requirements of the current task;
[0124] Step S5.4: Select the candidate edge node with the smallest total marginal cost increment as the target node, assign all current tasks to the target node, and update the load of the target node;
[0125] Step S5.5: If the target node load reaches the maximum carrying bandwidth after the update, remove the target node from the candidate edge node set of subsequent task nodes.
[0126] Step S5.6: Initiate downgrade processing for task nodes with an empty candidate edge node set.
[0127] Press all task nodes That is, task node m at time The predicted bandwidth requirements are sorted in descending order from largest to smallest to generate a task processing queue.
[0128] For the current task node m, filter out nodes that satisfy the following conditions in all prediction steps. Edge nodes that are greater than or equal to the preset minimum service quality threshold constitute a candidate edge node set.
[0129] For each edge node n in the candidate edge node set, calculate the total marginal cost increment after accepting all predicted demands from task node m, using the following formula:
[0130]
[0131] in, This represents the total marginal cost increment resulting from allocating task node m to edge node n, which is the increase in the objective function. The first term in the square brackets is the total cost function value of edge node n after acceptance, and the second term is the basic cost function value before acceptance. The difference between the two is the marginal cost brought by the task.
[0132] choose The smallest edge node is used as the target node. All task nodes m are assigned to the target node, and the load of the target node is updated. The specific formula is as follows:
[0133]
[0134] in, For the target node;
[0135] If, after the update, a prediction step h causes the target node's load to reach its maximum bandwidth, then the edge node... Remove from the set of candidate edge nodes for all subsequent task nodes;
[0136] If the candidate edge node set for a task node is empty, a degradation process is initiated, which first temporarily reduces the service quality threshold by 10%, i.e., multiplies it by 0.9, and then re-filters. If there is still no solution, the transmission task of the task node is delayed to the next scheduling cycle or split according to predefined rules.
[0137] After all task nodes have been processed, the output bandwidth allocation scheme and predicted total cost are calculated. The algorithm complexity is O(n log n). It can make rapid decisions under the scale of a conventional system, meeting the real-time online scheduling requirements of edge networks.
[0138] like Figure 2 As shown, a bandwidth prediction and scheduling system for edge network cost optimization is used to implement a bandwidth prediction and scheduling method for edge network cost optimization, including:
[0139] The data preprocessing module is used to adaptively repair anomalies in the collected historical bandwidth requirement data of task nodes, extract multi-granularity statistical features, and construct standardized input feature tensors.
[0140] The bandwidth demand prediction module has a built-in bandwidth demand prediction model based on a hybrid of temporal convolutional networks and bidirectional gated recurrent units. It is used to output the bandwidth demand prediction value for the future prediction step size based on the standardized input feature tensor.
[0141] The model training module is used to train the bandwidth demand prediction model using a peak-valley asymmetric weighted loss function.
[0142] The online scheduling optimization module is used to construct a marginal cost-sensitive online scheduling optimization model based on the predicted bandwidth demand. It uses a greedy heuristic algorithm based on marginal cost increment sorting to solve the problem and output a bandwidth allocation scheme.
[0143] The bandwidth demand prediction module includes:
[0144] Temporal convolutional network units are used to extract multi-scale local temporal patterns from input feature tensors through stacked causal dilated convolutional residual blocks. The dilation rate increases exponentially, and the output is a temporal feature sequence.
[0145] The bidirectional gated recurrent unit is used to take the output of the temporal convolutional network unit as the input sequence, extract the bidirectional temporal dependencies through stacked bidirectional gated recurrent layers, and concatenate the forward hidden state and the backward hidden state at each time step.
[0146] The output mapping unit is used to perform average pooling on the sequence output by the bidirectional gated recurrent unit along the time axis to obtain a fixed-length feature vector, which is then mapped by the fully connected layer to the bandwidth requirement prediction value for the future prediction step size.
[0147] The online scheduling optimization module includes:
[0148] The task sorting unit is used to arrange all task nodes in descending order of their predicted bandwidth requirements at the nearest future time, thus generating a task processing queue.
[0149] The candidate node filtering unit is used to filter the current task node into a set of candidate edge nodes that satisfy the quality of service constraints in all prediction steps.
[0150] The marginal cost calculation unit is used to calculate the total marginal cost increment for each node in the candidate edge node set after accepting all the predicted demands of the current task node.
[0151] The allocation and load update unit is used to select the candidate edge node with the smallest total marginal cost increment as the target node, allocate all current task nodes to the target node, and update the load of the target node in each prediction step.
[0152] The capacity maintenance unit is used to remove the target node from the candidate set of subsequent task nodes when the target node load reaches the maximum carrying bandwidth.
[0153] The degradation processing unit is used to initiate degradation processing for task nodes whose candidate edge node set is empty, including relaxing the service quality threshold, delaying scheduling, or splitting tasks.
[0154] The examples described herein are merely preferred embodiments of the invention and are not intended to limit the concept and scope of the invention. Any modifications and improvements made by those skilled in the art to the technical solutions of the invention without departing from the design concept of the invention should fall within the protection scope of the invention.
Claims
1. A bandwidth prediction and scheduling method for cost optimization in edge networks, characterized in that, Includes the following steps: Step S1: Collect historical bandwidth demand data of task nodes in the edge network, perform adaptive anomaly repair on the historical bandwidth sequence, extract multi-granularity statistical features, and construct a standardized input feature tensor. Step S2: Construct a bandwidth demand prediction model based on a hybrid of temporal convolutional network and bidirectional gated recurrent unit. Train the bandwidth demand prediction model using the standardized input feature tensor and output the bandwidth demand prediction value for the future prediction step size. Step S3: Design a peak-valley asymmetric weighted loss function to train the bandwidth demand prediction model, and assign corresponding penalty weights according to whether the prediction point is in the peak period of bandwidth demand and the direction of prediction error. Step S4: Based on the bandwidth demand forecast value of the future prediction step size output by the bandwidth demand forecast model, construct a marginal cost sensitive online scheduling optimization model with the objective of minimizing the total transmission cost in the prediction time domain. Step S5: Use a greedy heuristic algorithm based on marginal cost increment sorting to solve the scheduling optimization model online, and output a bandwidth allocation scheme that satisfies the service quality constraints and capacity constraints.
2. The method according to claim 1, characterized in that, Step S1 specifically includes the following steps: Step S1.1: For the historical bandwidth sequence of a single task node, use local weighted regression to smoothly decompose it into trend components and residual components; Step S1.2: Calculate the joint anomaly score for the residual components. When the joint anomaly score exceeds the adaptive threshold, the corresponding time point is determined to be an anomaly point. The adaptive threshold is the sum of the mean of all joint anomaly scores and three times the standard deviation. Step S1.3: Extract a context segment with a preset neighborhood radius from the anomaly point, use dynamic time warping to retrieve a preset number of candidate segments with the most similar shape to the context segment from the historical normal records, weight and aggregate the corresponding position values of the candidate segments to obtain the repair value, and output the repaired bandwidth sequence. Step S1.4: Extract multi-granularity statistical features from the repaired bandwidth sequence, and construct a standardized input feature tensor after normalization.
3. The method according to claim 2, characterized in that, In step S1.2, the joint anomaly score is a weighted sum of the outlier score and the mutation score, wherein the outlier score is the result of the local outlier factor value of the residual component at the corresponding time after max-min normalization; the mutation score is the result of the absolute value of the local Z-score of the first difference of the residual component after max-min normalization; the joint anomaly score is equal to the sum of the outlier score and the mutation score multiplied by their respective weight coefficients.
4. The method according to claim 3, characterized in that, In step S1.3, the weighted aggregation of the corresponding position values of the candidate fragments to obtain the repair value specifically includes: The distance between segments is a dynamic time-warped distance with a trend penalty, which is a weighted sum of the dynamic time-warped distance of the original sequence and the dynamic time-warped distance of the first-order difference sequence. The bandwidth observations of the candidate segments, after dynamic time warping and alignment, and the corresponding positions of the outliers are weighted and aggregated using the exponential decay function of distance as the weight to obtain the repair value. The temperature coefficient is the median of the distances of each candidate segment. After repair, the second-order difference mean of the sequence in the neighborhood of the outlier is calculated as the smoothness index. If the smoothness index exceeds the preset quantile limit of the smoothness of the historical normal neighborhood, the temperature coefficient is reduced and the weighted aggregation is re-executed until the smoothness requirement is met or the preset maximum number of iterations is reached.
5. The method according to claim 4, characterized in that, In step S1.4, the multi-granularity statistical features specifically include: the original repair value; short-term fluctuation features, which are the bandwidth standard deviation and skewness calculated within a preset first window length; long-term trend features, which are the slope and goodness of fit obtained by least-squares linear fitting within a preset second window length, where the second window length is greater than the first window length; periodic reference features, which are the median bandwidth of the weekday and hour attributes of the current moment in the historical records of this node; and spatial correlation features, which are the current moment bandwidth values of a preset number of neighboring nodes with the highest Spearman correlation coefficient with the current node's bandwidth sequence.
6. The method according to claim 5, characterized in that, In step S2, the bandwidth demand prediction model is composed of a temporal convolutional network module, a bidirectional gated recurrent unit module, and an output layer connected sequentially. The temporal convolutional network module is composed of a preset number of residual blocks stacked together. The inflation rate of each residual block increases exponentially. Each residual block contains a causal dilated convolution, weight normalization, activation function and dropout layer, and maintains gradient propagation through residual connections. The bidirectional gated recurrent unit module takes the output of the temporal convolutional network module as the input sequence, extracts bidirectional temporal dependencies through stacked bidirectional gated recurrent unit layers, and concatenates the forward hidden state and the backward hidden state at each time step for output. The output layer performs average pooling along the time axis on the sequence output by the bidirectional gated recurrent unit module to obtain a fixed-length feature vector, which is then mapped by the fully connected layer to the bandwidth requirement prediction value for the future prediction step size.
7. The method according to claim 6, characterized in that, In step S3, the peak-valley asymmetric weighted loss function is the mean of the weighted squared errors of all prediction steps; The weighting rules are as follows: when the actual bandwidth value is greater than the preset percentile of the historical bandwidth sequence of the node and the prediction error is negative, the weight is a preset penalty value greater than 1; otherwise, the weight is 1.
8. The method according to claim 7, characterized in that, In step S4, the marginal cost sensitive online scheduling optimization model takes minimizing the total transmission cost in the prediction time domain as its objective function. The constraints include: demand integrity constraint, which means that the prediction bandwidth requirement of each task node in each prediction step must be fully allocated to each edge node; node capacity constraint, which means that the total load of each edge node in each prediction step does not exceed its maximum carrying bandwidth limit; allocation non-negative constraint, which means that the allocation ratio is non-negative; and service quality constraint, which means that task nodes can only be allocated to edge nodes that meet the preset minimum service quality threshold.
9. The method according to claim 8, characterized in that, Step S5 specifically includes the following steps: Step S5.1: Sort the task nodes in descending order of their predicted bandwidth requirements for the nearest future time. Step S5.2: Process each task node in the order of arrangement and filter the set of candidate edge nodes that meet the service quality constraints in all prediction steps; Step S5.3: Calculate the total marginal cost increment for each node in the candidate edge node set after accepting all predicted requirements of the current task; Step S5.4: Select the candidate edge node with the smallest total marginal cost increment as the target node, assign all current tasks to the target node, and update the load of the target node; Step S5.5: If the target node load reaches the maximum carrying bandwidth after the update, remove the target node from the candidate edge node set of subsequent task nodes. Step S5.6: Initiate downgrade processing for task nodes with an empty candidate edge node set.
10. A bandwidth prediction and scheduling system for edge network cost optimization, used to implement the bandwidth prediction and scheduling method for edge network cost optimization as described in any one of claims 1-9, characterized in that, include: The data preprocessing module is used to adaptively repair anomalies in the collected historical bandwidth requirement data of task nodes, extract multi-granularity statistical features, and construct standardized input feature tensors. The bandwidth demand prediction module has a built-in bandwidth demand prediction model based on a hybrid of temporal convolutional networks and bidirectional gated recurrent units. It is used to output the bandwidth demand prediction value for the future prediction step size based on the standardized input feature tensor. The model training module is used to train the bandwidth demand prediction model using a peak-valley asymmetric weighted loss function. The online scheduling optimization module is used to construct a marginal cost-sensitive online scheduling optimization model based on the predicted bandwidth demand. It uses a greedy heuristic algorithm based on marginal cost increment sorting to solve the problem and output a bandwidth allocation scheme.