Optical fiber access flow scheduling method based on edge calculation and AI

Abstract

The invention discloses an edge computing and AI-based optical fiber access flow scheduling method, which comprises the following steps of: acquiring service flow, queue state and link state data to form a time sequence data set; preprocessing the time sequence data, and constructing an input feature vector sequence; deploying an improved space-time flow prediction model at an edge computing node, and initializing model parameters; constructing a graph Laplacian matrix based on the topological relation of the optical fiber access network, and executing spectral domain convolution and graph convolution operation to obtain spatial correlation characteristics; splicing the time features output by the two branches in a channel dimension and inputting the time features into a prediction layer to obtain a flow prediction result and a congestion risk index; accessing a target bandwidth requirement of the equipment, and generating a corresponding dynamic bandwidth allocation parameter; and determining the available bandwidth and the priority and the service sequence of each service queue to form an optical fiber access flow scheduling scheme. According to the invention, flow scheduling of optical fiber access is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent traffic scheduling technology, and in particular to a fiber optic access traffic scheduling method based on edge computing and AI. Background Technology

[0002] In existing fiber optic access networks, bandwidth allocation and queue scheduling are typically handled uniformly by the central office optical line terminal (OCT) for a large number of user-side fiber optic access devices. To ensure the bandwidth and latency requirements of different service types, operators often use pre-configured static bandwidth allocation parameters or dynamic bandwidth allocation algorithms based on simple statistics and threshold rules. These algorithms periodically analyze the service traffic, queue length, and link utilization of each access device before adjusting the scheduling strategy. Some solutions have begun to introduce traffic prediction or machine learning models, but these are mostly deployed in the cloud or central location, primarily based on single-node time series data, making it difficult to reflect the instantaneous traffic changes and topology correlations of a large number of users at the edge in a timely manner.

[0003] In attempts to introduce artificial intelligence technology, some solutions have begun to use machine learning or deep learning models to predict network traffic. However, these solutions often focus on time series data from single links or nodes, frequently employing traditional time-series networks such as LSTM and GRU. They fail to fully utilize the topological relationships between access devices in fiber optic access networks, resulting in limited modeling capabilities for spatial correlation and congestion propagation characteristics. Furthermore, existing methods typically deploy AI models in the cloud or centralized control centers, leading to long inference paths and high latency, making it difficult to respond in real-time to rapid traffic fluctuations at the edge.

[0004] Furthermore, existing AI prediction results mostly remain at the level of providing future bandwidth usage or average traffic, rarely outputting risk indicators that can directly reflect the probability of queue backlog and link congestion. The coupling between the prediction module and bandwidth allocation and time slot scheduling is not high, and there is a lack of a complete link that systematically maps the prediction results into dynamic bandwidth allocation parameters and time slot-level scheduling schemes.

[0005] Therefore, how to provide a fiber optic access traffic scheduling method based on edge computing and AI is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a fiber optic access traffic scheduling method based on edge computing and AI. This invention introduces an improved spatiotemporal traffic prediction model at the edge to adaptively characterize the spatial correlation and multi-scale temporal features of each access device in the fiber optic access network, thereby achieving accurate prediction of future traffic and congestion risks. Under the premise of meeting the total bandwidth constraint, it generates refined dynamic bandwidth allocation parameters and scheduling schemes, significantly improving link utilization, reducing latency and packet loss rate, and enhancing the intelligent scheduling capability and service quality of the fiber optic access network in high-concurrency and differentiated service scenarios.

[0007] A fiber optic access traffic scheduling method based on edge computing and AI according to an embodiment of the present invention includes the following steps: The service traffic, queue status and link status data of each user-side fiber access device are acquired at the edge computing node of the fiber access network, and a time series dataset is formed according to the preset sampling period. The time series data is preprocessed, including outlier removal, missing data imputation and numerical normalization, and the input feature vector sequence is constructed by arranging the user-side fiber optic access equipment in chronological order as an index. An improved spatiotemporal traffic prediction model is deployed on edge computing nodes, including a graph adaptive convolutional structure for characterizing the spatial correlation between various user-side fiber access devices and a temporal bi-branch dilated convolutional structure for extracting multi-scale temporal features, and the model parameters are initialized. The feature vector sequence is input into the graph adaptive convolution structure. The graph Laplacian matrix is ​​constructed based on the topology of the fiber access network, and spectral domain convolution and graph convolution operations are performed. The edge weights in the graph adjacency matrix are dynamically updated according to the obtained graph spectral features. Graph convolution is performed on the updated graph to obtain the spatial correlation features of each user-side fiber access device. Spatial correlation features are input into a temporal dual-branch dilated convolutional structure. In the long-period branch, a one-dimensional causal dilated convolution with increasing dilation factor is used, and in the short-period branch, a one-dimensional causal dilated convolution with smaller dilation factor is used. The temporal features output from the two branches are concatenated in the channel dimension and input into the prediction layer to obtain traffic prediction results and congestion risk indicators. Based on traffic prediction results and congestion risk indicators, the target bandwidth requirements of each user-side fiber optic access device are calculated at the edge computing node, and corresponding dynamic bandwidth allocation parameters are generated. Based on the dynamic bandwidth allocation parameters, the available bandwidth of each user-side fiber optic access device in each time slot and the priority and service order of each service queue are determined within the scheduling period, forming a fiber optic access traffic scheduling scheme, which is then sent to the central office optical line terminal equipment for execution.

[0008] Optionally, the formation of the time series dataset specifically includes: At the edge computing node of the fiber optic access network, a unique device identifier is assigned to each of the user-side fiber optic access devices, a device list is generated, the sampling start time and sampling period are configured in the edge computing node, a data communication channel is established with each user-side fiber optic access device, and a periodic data acquisition task is created so that the data acquisition task is automatically triggered and executed according to the preset sampling period after the sampling start time is reached. At each sampling trigger moment, the edge computing node sends data acquisition instructions to each user-side fiber optic access device in sequence according to the device list, and acquires the service traffic data, queue status data and link status data within the current sampling period respectively. It also records the corresponding device identifier and sampling timestamp for each data, and writes them into the cache area to form the original sampling data set at the current sampling moment. After completing continuous sampling, the edge computing node reads the original sampled data set of each sampling time from the cache area, and processes it according to the order of sampling time. For each sampling time, the service traffic data, queue status data and link status data corresponding to each user-side fiber access device in the device list are combined in the order of device identification to obtain the system status record of the current sampling time. Then, multiple system status records are added to the same time series data structure in chronological order to form a time series dataset for subsequent construction of input feature vector sequences.

[0009] Optionally, the construction of the input feature vector sequence specifically includes: Outlier removal and missing data imputation are performed sequentially on the feature data of each user-side fiber optic access device in the time series dataset. Outlier removal involves statistically analyzing the value distribution in the time dimension according to feature type, determining the normal value range, and marking values ​​that deviate significantly from the range as invalid. Missing data imputation involves finding the most recent valid sampling time before and after invalid or missing values, and interpolating and filling the missing data using the valid values ​​before and after the invalid or missing values ​​according to the time position ratio. If there are only valid values ​​on one side, the most recent valid value is used to fill the missing data, resulting in time series feature data that is continuous in time and has no missing features. After outlier removal and missing data imputation, the average level and fluctuation of the entire time series are calculated according to the feature type of the time series feature data. The feature values ​​of each sampling time and each user-side fiber optic access device are normalized to unify features of different dimensions into a comparable standard value range. At the same time, at each sampling time, the normalized features of all devices are arranged in order of device identification to form the feature matrix of the current sampling time, using the user-side fiber optic access device identifier as an index. Each feature matrix is ​​then added to the same time series structure in the order of sampling time to construct the input feature vector sequence.

[0010] Optionally, the deployment process of the improved spatiotemporal traffic prediction model specifically includes: An improved spatiotemporal traffic prediction model based on artificial intelligence deep learning is created in the edge computing nodes of the fiber optic access network. The input feature vector sequence is registered as the data input source. The functional module regions of graph adaptive convolution structure, temporal double-branch dilated convolution structure and prediction output layer are pre-divided in the model instance. Computational and storage resources are allocated to each functional module to form a basic model framework containing three functional modules. The graph adaptive convolution structure is a graph neural network unit of artificial intelligence, and the temporal double-branch dilated convolution structure is a temporal convolutional neural network module of artificial intelligence. Using the basic model framework as the configuration object, the graph adaptive convolutional structure within the model instance is configured in detail in the edge computing node. Based on the device list and fiber access network topology, node indexes and connection relationships are established for each user-side fiber access device. The configuration results are written into the topology configuration unit of the graph adaptive convolutional structure. The number of layers, input channels, and output channels of the graph adaptive convolutional structure are set, and the output interface information of the graph adaptive convolutional structure is registered in the model instance. Using the output interface of the graph adaptive convolutional structure as the input, the temporal bi-branch dilated convolutional structure is configured in the edge computing node. The input of the temporal bi-branch dilated convolutional structure is bound to the output interface of the graph adaptive convolutional structure. The number of convolutional layers, time window, and dilation sampling interval of the long-period branch and the short-period branch are set respectively. The concatenation order of the two branch outputs in the channel dimension is determined, and the concatenated output interface is associated with the input of the prediction output layer. Initialization operations are performed on the weight parameters, bias parameters, and normalization parameters of the temporal bi-branch dilated convolutional structure and the prediction output layer.

[0011] Optionally, the formation of the spatial correlation features specifically includes: In the graph adaptive convolutional structure, the input feature vector sequence is used as input, and each user-side fiber optic access device in the device list is used as a graph node. Based on the physical topology of the fiber optic access network, it is determined whether there is a physical connection between any two devices. The connected device pairs are registered as graph edge information. The number of connections for each device is counted to generate the corresponding node degree information. Inside the graph adaptive convolutional structure, the initial graph Laplacian matrix configuration data is generated based on the node degree information and the graph edge information. The initial graph Laplacian matrix configuration data is loaded into the spectral domain convolution processing unit of the graph adaptive convolution structure. Spectral domain convolution processing is performed sequentially on the feature vector sequence corresponding to each sampling time. During the spectral domain convolution processing, the spectrum range corresponding to the initial graph Laplacian matrix is ​​divided into different frequency bands of low frequency, mid frequency, and high frequency. A set of filter weight parameters is pre-configured for each frequency band. The different frequency bands are weighted and combined to obtain the spectral feature data representing the relationship between each user-side fiber access device in different frequency bands. The spectral features of each frequency band are then concatenated along the feature dimension to generate a comprehensive graph spectral feature record for the current sampling time. In the edge computing node, based on the comprehensive graph spectrum feature record, the graph adjacency matrix inside the graph adaptive convolution structure is dynamically updated. For each pair of user-side fiber access devices that are registered as having a physical connection, the corresponding feature vector of the device in the comprehensive graph spectrum feature is extracted. According to the preset similarity calculation rules, a similarity index reflecting the degree of correlation between the two devices in the spectrum space is obtained. At the same time, the current similarity index and the preset smoothing coefficient are combined in a weighted manner to generate new edge weight values. The new edge weight values ​​are written into the graph adjacency matrix configuration unit to complete the update of the edge weights of all connected devices and form an adaptive graph adjacency matrix. The adaptive graph structure and the feature vector sequence at the corresponding sampling time are input into the graph convolution operation unit of the graph adaptive convolution structure. According to the preset number of graph convolution layers and multi-order neighbor aggregation rules, in each layer, the node features of each user-side fiber access device and the multi-order neighbor node features are weighted, aggregated and linearly transformed based on the adaptive graph structure. The output features of the previous layer are passed as the input features of the next layer to obtain spatial correlation features.

[0012] Optionally, the acquisition of the traffic prediction results and congestion risk indicators specifically includes: Spatial correlation features are sequentially written into the input buffer of the temporal dual-branch dilated convolution structure according to the sampling time, forming a temporally ordered sequence of time windows. At the same time, they are associated with the input ports of the long-period branch and the short-period branch, and the sampling time is used as a unified time index for aligning the subsequent outputs of the two branches. In the long-period branch, for each time window, the periodic indicators of the historical traffic changes of each user-side fiber access device are calculated. The hole factor value used by the current convolutional layer is selected from the preset different hole factor candidate sets. Multiple one-dimensional causal hole convolutional layers are enabled in sequence according to the increasing order of the selected hole factors. Layer-by-layer convolution operation is performed on the spatial correlation features at different sampling times within the same time window to generate a long-period time feature sequence covering a long time range and register it in the output buffer of the long-period branch. In the short-period branch, for each time window, the local fluctuation index of traffic and queue status of each user-side fiber access device is calculated in a smaller time neighborhood. Based on the local fluctuation index, the hole factor and convolution kernel used in the current convolution layer are selected from the preset combination of small hole factor and convolution kernel. Multi-layer one-dimensional causal hole convolution operation is performed on the spatial correlation features of the current sampling time and a small number of adjacent sampling times. In each layer, channel gating weights are generated based on the local fluctuation index and the convolution output is scaled by channel to obtain a short-period time feature sequence covering a shorter time range. At each sampling time, the long-period time feature sequence and the short-period time feature sequence are concatenated in channel order, and the channel weights are calculated according to the preset channel weight generation rules. The concatenated multi-scale time feature matrix is ​​input into the prediction layer, and the corresponding traffic prediction results and congestion risk indicators are output.

[0013] Optionally, the process of generating the dynamic bandwidth allocation parameters specifically includes: The traffic prediction results and congestion risk indicators are written into the prediction data cache according to the user-side fiber access equipment identifier and sampling time, forming a prediction data set indexed by equipment identifier and sampling time, providing a unified input for bandwidth demand calculation; Using the predicted data set as input, the traffic prediction value of each user-side fiber access device in the target scheduling period is read one by one as the basic demand, and the corresponding congestion risk index is read as the demand increment factor. The basic demand is incrementally calculated to form the initial bandwidth demand value, and the initial bandwidth demand value is written into the temporary bandwidth demand list. Read the total allocable bandwidth from the temporary bandwidth demand list, perform total bandwidth constraint processing on all initial bandwidth demands, scale the initial demand values ​​of each device sequentially according to the ratio of total bandwidth to total demand, form dynamic bandwidth allocation parameters that meet the constraints, and register them in the bandwidth demand registration table in the order of device identification.

[0014] Optionally, the formation of the fiber optic access traffic scheduling scheme specifically includes: The dynamic bandwidth allocation parameters are used to construct a scheduling parameter sequence with the scheduling period as the time index; Based on the scheduling parameter sequence, the current total bandwidth is read one time slot at a time within the scheduling period. The available bandwidth in the corresponding time slot is calculated according to the dynamic bandwidth allocation parameters. Combined with the service queue status and service level, the service priority and service order of each service queue in the current time slot are sorted to form a time slot bandwidth allocation list and time slot scheduling information. Based on the time slot bandwidth allocation list and time slot scheduling information, the available bandwidth of all time slots within the scheduling period is combined with the service queue order to form a fiber optic access traffic scheduling scheme. This scheme is then distributed to the central office optical line terminal equipment for execution via edge computing nodes. The central office optical line terminal is a key device in the fiber optic access network, responsible for managing and controlling the data traffic, bandwidth allocation, and signal forwarding between the fiber optic access network and the user-side equipment.

[0015] The beneficial effects of this invention are: This invention deploys an improved spatiotemporal traffic prediction model on edge computing nodes. It utilizes graph adaptive convolutional structures and temporal bi-branch dilated convolutional structures to jointly model the spatial correlation and multi-scale temporal dependence of each user-side fiber optic access device. Compared with existing schemes based on fixed topology or a single time scale, this model can more accurately characterize traffic evolution patterns and congestion propagation characteristics, thereby improving the accuracy of traffic prediction and congestion risk assessment.

[0016] This invention constructs a continuous, gap-free, and dimensionally consistent time series feature vector sequence using device identifiers and a unified time index through a standardized data acquisition and preprocessing process, providing high-quality input for the model. Combined with an adaptive adjacency matrix update mechanism based on graph Laplacian spectral domain convolution, it achieves dynamic adjustment of the correlation between nodes, overcoming the shortcomings of existing technologies where node relationships are fixed and difficult to adaptively update with changes in business.

[0017] Based on the traffic prediction results and congestion risk indicators, this invention calculates the target bandwidth requirements of each access device in detail for the target scheduling period, generates dynamic bandwidth allocation parameters that meet the total bandwidth constraints, and further maps them to available bandwidth, queue priority and service order at the time slot level. Compared with the traditional bandwidth allocation method based on fixed thresholds or simple proportional scaling, this invention takes into account the differentiated service requirements and congestion control requirements of different devices and different service queues. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a fiber optic access traffic scheduling method based on edge computing and AI proposed in this invention; Figure 2 This is a schematic diagram of the algorithm structure of a fiber optic access traffic scheduling method based on edge computing and AI proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1-2 A fiber optic access traffic scheduling method based on edge computing and AI includes the following steps: The service traffic, queue status and link status data of each user-side fiber access device are acquired at the edge computing node of the fiber access network, and a time series dataset is formed according to the preset sampling period. The time series data is preprocessed, including outlier removal, missing data imputation and numerical normalization, and the input feature vector sequence is constructed by arranging the user-side fiber optic access equipment in chronological order as an index. An improved spatiotemporal traffic prediction model is deployed on edge computing nodes, including a graph adaptive convolutional structure for characterizing the spatial correlation between various user-side fiber access devices and a temporal bi-branch dilated convolutional structure for extracting multi-scale temporal features, and the model parameters are initialized. The feature vector sequence is input into the graph adaptive convolution structure. The graph Laplacian matrix is ​​constructed based on the topology of the fiber access network, and spectral domain convolution and graph convolution operations are performed. The edge weights in the graph adjacency matrix are dynamically updated according to the obtained graph spectral features. Graph convolution is performed on the updated graph to obtain the spatial correlation features of each user-side fiber access device. Spatial correlation features are input into a temporal dual-branch dilated convolutional structure. In the long-period branch, a one-dimensional causal dilated convolution with increasing dilation factor is used, and in the short-period branch, a one-dimensional causal dilated convolution with smaller dilation factor is used. The temporal features output from the two branches are concatenated in the channel dimension and input into the prediction layer to obtain traffic prediction results and congestion risk indicators. Based on traffic prediction results and congestion risk indicators, the target bandwidth requirements of each user-side fiber optic access device are calculated at the edge computing node, and corresponding dynamic bandwidth allocation parameters are generated. Based on the dynamic bandwidth allocation parameters, the available bandwidth of each user-side fiber optic access device in each time slot and the priority and service order of each service queue are determined within the scheduling period, forming a fiber optic access traffic scheduling scheme, which is then sent to the central office optical line terminal equipment for execution.

[0021] In this embodiment, the formation of the time series dataset specifically includes: At the edge computing node of the fiber optic access network, a unique device identifier is assigned to each of the user-side fiber optic access devices, a device list is generated, the sampling start time and sampling period are configured in the edge computing node, a data communication channel is established with each user-side fiber optic access device, and a periodic data acquisition task is created so that the data acquisition task is automatically triggered and executed according to the preset sampling period after the sampling start time is reached. At each sampling trigger moment, the edge computing node sends data acquisition instructions to each user-side fiber optic access device in sequence according to the device list, and acquires the service traffic data, queue status data and link status data within the current sampling period respectively. It also records the corresponding device identifier and sampling timestamp for each data, and writes them into the cache area to form the original sampling data set at the current sampling moment. After completing continuous sampling, the edge computing node reads the original sampled data set of each sampling time from the cache area, and processes it according to the order of sampling time. For each sampling time, the service traffic data, queue status data and link status data corresponding to each user-side fiber access device in the device list are combined in the order of device identification to obtain the system status record of the current sampling time. Then, multiple system status records are added to the same time series data structure in chronological order to form a time series dataset for subsequent construction of input feature vector sequences.

[0022] In this embodiment, the construction of the input feature vector sequence specifically includes: Outlier removal and missing data imputation are performed sequentially on the feature data of each user-side fiber optic access device in the time series dataset. Outlier removal involves statistically analyzing the value distribution in the time dimension according to feature type, determining the normal value range, and marking values ​​that deviate significantly from the range as invalid. Missing data imputation involves finding the most recent valid sampling time before and after invalid or missing values, and interpolating and filling the missing data using the valid values ​​before and after the invalid or missing values ​​according to the time position ratio. If there are only valid values ​​on one side, the most recent valid value is used to fill the missing data, resulting in time series feature data that is continuous in time and has no missing features. After outlier removal and missing data imputation, the average level and fluctuation of the entire time series are calculated according to the feature type of the time series feature data. The feature values ​​of each sampling time and each user-side fiber optic access device are normalized to unify features of different dimensions into a comparable standard value range. At the same time, at each sampling time, the normalized features of all devices are arranged in order of device identification to form the feature matrix of the current sampling time, using the user-side fiber optic access device identifier as an index. Each feature matrix is ​​then added to the same time series structure in the order of sampling time to construct the input feature vector sequence.

[0023] In this embodiment, the deployment process of the improved spatiotemporal flow prediction model specifically includes: An improved spatiotemporal traffic prediction model based on artificial intelligence deep learning is created in the edge computing nodes of the fiber optic access network. The input feature vector sequence is registered as the data input source. The functional module regions of graph adaptive convolution structure, temporal double-branch dilated convolution structure and prediction output layer are pre-divided in the model instance. Computational and storage resources are allocated to each functional module to form a basic model framework containing three functional modules. The graph adaptive convolution structure is a graph neural network unit of artificial intelligence, and the temporal double-branch dilated convolution structure is a temporal convolutional neural network module of artificial intelligence. Using the basic model framework as the configuration object, the graph adaptive convolutional structure within the model instance is configured in detail in the edge computing node. Based on the device list and fiber access network topology, node indexes and connection relationships are established for each user-side fiber access device. The configuration results are written into the topology configuration unit of the graph adaptive convolutional structure. The number of layers, input channels, and output channels of the graph adaptive convolutional structure are set, and the output interface information of the graph adaptive convolutional structure is registered in the model instance. Using the output interface of the graph adaptive convolutional structure as the input, the temporal bi-branch dilated convolutional structure is configured in the edge computing node. The input of the temporal bi-branch dilated convolutional structure is bound to the output interface of the graph adaptive convolutional structure. The number of convolutional layers, time window, and dilation sampling interval of the long-period branch and the short-period branch are set respectively. The concatenation order of the two branch outputs in the channel dimension is determined, and the concatenated output interface is associated with the input of the prediction output layer. Initialization operations are performed on the weight parameters, bias parameters, and normalization parameters of the temporal bi-branch dilated convolutional structure and the prediction output layer.

[0024] This invention integrates graph neural network units and temporal convolutional neural network modules into an improved spatiotemporal traffic prediction model framework on edge computing nodes. It achieves collaborative optimization of AI model structure configuration, resource allocation and parameter initialization, reduces the deployment threshold and operation and maintenance cost of complex deep learning models in fiber optic access networks, and provides a stable and reliable computing power foundation for real-time high-precision traffic prediction and intelligent scheduling.

[0025] In this embodiment, the formation of the spatial correlation feature specifically includes: In the graph adaptive convolutional structure, the input feature vector sequence is used as input, and each user-side fiber optic access device in the device list is used as a graph node. Based on the physical topology of the fiber optic access network, it is determined whether there is a physical connection between any two devices. The connected device pairs are registered as graph edge information. The number of connections for each device is counted to generate the corresponding node degree information. Inside the graph adaptive convolutional structure, the initial graph Laplacian matrix configuration data is generated based on the node degree information and the graph edge information. The initial graph Laplacian matrix configuration data is loaded into the spectral domain convolution processing unit of the graph adaptive convolution structure. Spectral domain convolution processing is performed sequentially on the feature vector sequence corresponding to each sampling time. During the spectral domain convolution processing, the spectrum range corresponding to the initial graph Laplacian matrix is ​​divided into different frequency bands of low frequency, mid frequency, and high frequency. A set of filter weight parameters is pre-configured for each frequency band. The different frequency bands are weighted and combined to obtain the spectral feature data representing the relationship between each user-side fiber access device in different frequency bands. The spectral features of each frequency band are then concatenated along the feature dimension to generate a comprehensive graph spectral feature record for the current sampling time. In the edge computing node, based on the comprehensive graph spectrum feature record, the graph adjacency matrix inside the graph adaptive convolution structure is dynamically updated. For each pair of user-side fiber access devices that are registered as having a physical connection, the corresponding feature vector of the device in the comprehensive graph spectrum feature is extracted. According to the preset similarity calculation rules, a similarity index reflecting the degree of correlation between the two devices in the spectrum space is obtained. At the same time, the current similarity index and the preset smoothing coefficient are combined in a weighted manner to generate new edge weight values. The new edge weight values ​​are written into the graph adjacency matrix configuration unit to complete the update of the edge weights of all connected devices and form an adaptive graph adjacency matrix. The adaptive graph structure and the feature vector sequence at the corresponding sampling time are input into the graph convolution operation unit of the graph adaptive convolution structure. According to the preset number of graph convolution layers and multi-order neighbor aggregation rules, in each layer, the node features of each user-side fiber access device and the multi-order neighbor node features are weighted, aggregated and linearly transformed based on the adaptive graph structure. The output features of the previous layer are passed as the input features of the next layer to obtain spatial correlation features.

[0026] This invention utilizes artificial intelligence graph neural networks to uniformly map physical topology and temporal features to the spectral space in a graph adaptive convolutional structure. By extracting correlations at different scales through multi-band spectral domain convolution and dynamically updating the graph adjacency matrix accordingly, an adaptive graph structure capable of self-learning edge weights and aggregating multi-level neighbor features is formed. This significantly enhances the modeling ability for complex spatial correlations and congestion propagation patterns in optical fiber access networks, and improves the accuracy and robustness of subsequent spatiotemporal traffic prediction.

[0027] In this embodiment, the acquisition of the traffic prediction results and congestion risk indicators specifically includes: Spatial correlation features are sequentially written into the input buffer of the temporal dual-branch dilated convolution structure according to the sampling time, forming a temporally ordered sequence of time windows. At the same time, they are associated with the input ports of the long-period branch and the short-period branch, and the sampling time is used as a unified time index for aligning the subsequent outputs of the two branches. In the long-period branch, for each time window, the periodic indicators of the historical traffic changes of each user-side fiber access device are calculated. The hole factor value used by the current convolutional layer is selected from the preset different hole factor candidate sets. Multiple one-dimensional causal hole convolutional layers are enabled in sequence according to the increasing order of the selected hole factors. Layer-by-layer convolution operation is performed on the spatial correlation features at different sampling times within the same time window to generate a long-period time feature sequence covering a long time range and register it in the output buffer of the long-period branch. In the short-period branch, for each time window, the local fluctuation index of traffic and queue status of each user-side fiber access device is calculated in a smaller time neighborhood. Based on the local fluctuation index, the hole factor and convolution kernel used in the current convolution layer are selected from the preset combination of small hole factor and convolution kernel. Multi-layer one-dimensional causal hole convolution operation is performed on the spatial correlation features of the current sampling time and a small number of adjacent sampling times. In each layer, channel gating weights are generated based on the local fluctuation index and the convolution output is scaled by channel to obtain a short-period time feature sequence covering a shorter time range. At each sampling time, the long-period time feature sequence and the short-period time feature sequence are concatenated in channel order, and the channel weights are calculated according to the preset channel weight generation rules. The concatenated multi-scale time feature matrix is ​​input into the prediction layer, and the corresponding traffic prediction results and congestion risk indicators are output.

[0028] This invention utilizes a temporal dual-branch dilated convolutional structure to uniformly incorporate spatial correlation features. The long-period branch adaptively selects different dilation factors to extract large-scale historical trends, while the short-period branch combines local fluctuation indicators and channel gating mechanisms to characterize sudden changes. After multi-scale temporal feature channel splicing and weighted fusion, the output traffic prediction results and congestion risk indicators are obtained, significantly improving the accuracy and generalization ability of characterizing traffic evolution and congestion probability in multi-service scenarios of fiber optic access networks, and providing a more reliable decision-making basis for subsequent bandwidth allocation and scheduling strategies.

[0029] In this embodiment, the process of generating the dynamic bandwidth allocation parameters specifically includes: The traffic prediction results and congestion risk indicators are written into the prediction data cache according to the user-side fiber access equipment identifier and sampling time, forming a prediction data set indexed by equipment identifier and sampling time, providing a unified input for bandwidth demand calculation; Using the predicted data set as input, the traffic prediction value of each user-side fiber access device in the target scheduling period is read one by one as the basic demand, and the corresponding congestion risk index is read as the demand increment factor. The basic demand is incrementally calculated to form the initial bandwidth demand value, and the initial bandwidth demand value is written into the temporary bandwidth demand list. Read the total allocable bandwidth from the temporary bandwidth demand list, perform total bandwidth constraint processing on all initial bandwidth demands, scale the initial demand values ​​of each device sequentially according to the ratio of total bandwidth to total demand, form dynamic bandwidth allocation parameters that meet the constraints, and register them in the bandwidth demand registration table in the order of device identification.

[0030] In this embodiment, the formation of the fiber optic access traffic scheduling scheme specifically includes: The dynamic bandwidth allocation parameters are used to construct a scheduling parameter sequence with the scheduling period as the time index; Based on the scheduling parameter sequence, the current total bandwidth is read one time slot at a time within the scheduling period. The available bandwidth in the corresponding time slot is calculated according to the dynamic bandwidth allocation parameters. Combined with the service queue status and service level, the service priority and service order of each service queue in the current time slot are sorted to form a time slot bandwidth allocation list and time slot scheduling information. Based on the time slot bandwidth allocation list and time slot scheduling information, the available bandwidth of all time slots within the scheduling period is combined with the service queue order to form a fiber optic access traffic scheduling scheme. This scheme is then distributed to the central office optical line terminal equipment for execution via edge computing nodes. The central office optical line terminal is a key device in the fiber optic access network, responsible for managing and controlling the data traffic, bandwidth allocation, and signal forwarding between the fiber optic access network and the user-side equipment.

[0031] Example 1: To verify the feasibility of this invention in practice, it was applied to a typical fiber optic access network of a certain operator. This network consists of a central office optical line terminal and multiple fiber optic access branches. Each branch connects several residential broadband users, small office users, and a small number of shop users, carrying various services such as high-definition video-on-demand, online games, cloud backup, video conferencing, and enterprise leased lines. During long-term operation, it was observed that during peak service periods, the downlink throughput of some access devices decreased significantly. Users experienced video buffering, high game latency, and prolonged cloud backup bandwidth usage leading to congestion in other services. This indicates that the existing scheduling strategy is unable to promptly identify and alleviate impending congestion, often triggering adjustments only after significant queue backlog or even packet loss, resulting in significant fluctuations in user experience.

[0032] Prior to the application of this invention, the network employed a typical existing technical solution: a dynamic bandwidth allocation module built into the optical line terminal (OLT) uniformly scheduled each access device. This primarily allocated uplink and downlink bandwidth based on the average service traffic, queue length, and link utilization obtained from each sampling period, combined with pre-configured service level weights, using threshold judgment and proportional scaling. In some scenarios, simple traffic prediction models were introduced at the central side, but these were mostly based on time series data from a single access port, employing moving averages or basic recurrent neural network structures. They only output an estimated average traffic value for a future period, without considering the topological correlation between different access devices or outputting risk indicators that directly reflect the probability of congestion. Because the model was deployed at the central side, there was a certain delay in sending the prediction results to the edge, and the coupling with the queue scheduling logic within the OLT was weak. Consequently, when users experienced sudden access surges and the service structure changed rapidly, existing technologies often failed to detect local hotspots and potential bottlenecks in advance, and the link bandwidth utilization was difficult to approach saturation.

[0033] When deploying this invention in the network, an edge computing node is set up in an integrated device close to the optical line terminal. The service traffic, queue status, and link status reported by each user-side fiber optic access device are collected at second-level intervals and organized into time-series data according to device identifiers and unified timestamps. The edge computing node uses the data preprocessing method of this invention to perform outlier removal, missing value imputation, and normalization on the raw data, constructing a continuous, gap-free, and dimensionlessly consistent sequence of input feature vectors. Based on this, an improved spatiotemporal traffic prediction model based on artificial intelligence deep learning is created in the edge computing nodes. The graph adaptive convolution structure, the temporal dual-branch dilated convolution structure, and the prediction output layer are divided into functional module regions. Graph nodes and connection relationships are established for each user-side device according to the fiber access network topology. Multi-band spatial correlation features are extracted through Laplacian spectral domain convolution and the graph adjacency matrix is ​​adaptively updated. Then, one-dimensional causal dilated convolution is performed on historical features at different time scales by long-period branches and short-period branches to obtain multi-scale temporal features. After channel splicing and weighted fusion, the traffic prediction results and congestion risk indicators of each user-side device in several future time slices are output.

[0034] In actual operation, the spatiotemporal traffic prediction model of this invention continuously receives the latest feature vector sequence in second-level steps. It uses the predicted traffic of each access device within the target scheduling period as the basic demand and the congestion risk index as the demand increment factor. An initial bandwidth demand list is generated within the edge computing node, and total volume constraint scaling is performed in conjunction with the currently allocable total bandwidth to obtain dynamic bandwidth allocation parameters that meet the constraints. Subsequently, these parameters are mapped to available bandwidth and service queue priority configurations at the time slot level. The scheduling engine generates an optical fiber access traffic scheduling scheme containing bandwidth allocation for each time slot and queue service order, which is then distributed to the optical line terminal for execution, enabling advance mitigation of potential congestion locations in the future.

[0035] To evaluate the effectiveness of this invention compared to the prior art, two access branches with similar structures and user configurations were selected in the same optical fiber access network. Specific experimental data are shown in Table 1. Table 1. Comparison data between the embodiments of the present invention and the prior art

[0036] As shown in Table 1, under the same network environment, the solution of this invention achieves significant improvements in all key indicators compared to existing solutions: the average downlink throughput of users during peak business hours increases from approximately 68 Mbps to approximately 82 Mbps; the 95th percentile of queue latency during peak business hours decreases from approximately 38 ms to approximately 22 ms; the packet loss rate decreases from approximately 0.45% to approximately 0.12%; the average number of daily congestion alarms decreases from approximately 210 to approximately 95; the utilization rate of fiber optic access link bandwidth increases from approximately 68% to approximately 83%; and the service level breach rate for high-value users decreases from approximately 4.7% to approximately 1.3%. These data demonstrate that this invention can effectively reduce latency and packet loss, decrease the frequency of congestion events, and significantly enhance the ability to guarantee the quality of services for critical users while improving link resource utilization.

[0037] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A fiber optic access traffic scheduling method based on edge computing and AI, characterized in that, Includes the following steps: The service traffic, queue status and link status data of each user-side fiber access device are obtained at the edge computing node of the fiber access network to form a time series dataset. The time series data is preprocessed and the input feature vector sequence is constructed by arranging the identifiers of the user-side fiber optic access devices in chronological order. An improved spatiotemporal traffic prediction model, including graph adaptive convolutional structure and temporal bi-branch dilated convolutional structure, is deployed on edge computing nodes, and the model parameters are initialized. The feature vector sequence is input into the graph adaptive convolution structure. Based on the topology of the optical fiber access network, a graph Laplacian matrix is ​​constructed and spectral domain convolution and graph convolution operations are performed to obtain spatial correlation features. Spatial correlation features are input into a temporal dual-branch dilated convolutional structure. The temporal features output from the two branches are concatenated along the channel dimension and input into the prediction layer to obtain traffic prediction results and congestion risk indicators. Based on traffic prediction results and congestion risk indicators, the target bandwidth requirements of each user-side fiber optic access device are calculated at the edge computing node, and corresponding dynamic bandwidth allocation parameters are generated. Based on dynamic bandwidth allocation parameters, the available bandwidth of each user-side fiber optic access device and the priority and service order of each service queue are determined to form a fiber optic access traffic scheduling scheme.

2. The fiber optic access traffic scheduling method based on edge computing and AI according to claim 1, characterized in that, The formation of the time series dataset specifically includes: At the edge computing node of the fiber optic access network, a unique device identifier is assigned to each of the user-side fiber optic access devices, a device list is generated, the sampling start time and sampling period are configured in the edge computing node, a data communication channel with each user-side fiber optic access device is established, and a periodic data acquisition task is created. At each sampling trigger moment, the edge computing node sends data acquisition instructions to each user-side fiber optic access device in sequence according to the device list, and acquires the service traffic data, queue status data and link status data in the current sampling period, and records the corresponding device identifier and sampling timestamp for each data, forming the original sampling data set; After completing continuous sampling, the edge computing node reads the original sampling data set of each sampling time from the cache area, organizes and processes it according to the order of sampling time to obtain the system status record of the current sampling time, and then adds multiple system status records into the same time series data structure in chronological order to form a time series dataset.

3. The fiber optic access traffic scheduling method based on edge computing and AI according to claim 1, characterized in that, The construction of the input feature vector sequence specifically includes: Outlier removal and missing data imputation are performed sequentially on the feature data of each user-side fiber optic access device in the time series dataset to obtain time series feature data that is continuous in time and has no missing features. The average level and fluctuation of the entire time series are calculated according to the feature type of the time series feature data. The feature values ​​of each sampling time and each user-side fiber optic access device are numerically normalized. At the same time, at each sampling time, the normalized features of all devices are arranged in the order of device identification to form a feature matrix using the user-side fiber optic access device identifier as an index. The feature matrices are then added to the same time series structure in the order of sampling time to construct the input feature vector sequence.

4. The fiber optic access traffic scheduling method based on edge computing and AI according to claim 1, characterized in that, The deployment process of the improved spatiotemporal flow prediction model specifically includes: An improved spatiotemporal traffic prediction model based on artificial intelligence deep learning is created in the edge computing nodes of the fiber optic access network. The input feature vector sequence is registered as the data input source. The functional module regions of graph adaptive convolution structure, temporal double-branch dilated convolution structure and prediction output layer are pre-divided, and computing and storage resources are allocated to each functional module to form the basic framework of the model. Using the basic model framework as the configuration object, the graph adaptive convolutional structure within the model instance is configured in detail in the edge computing node. The configuration results are written into the topology configuration unit of the graph adaptive convolutional structure, setting the number of layers, input channels, and output channels of the graph adaptive convolutional structure, and registering the output interface information of the graph adaptive convolutional structure in the model instance. Using the output interface of the graph adaptive convolutional structure as the input, the temporal bi-branch dilated convolutional structure is configured in the edge computing node. The input of the temporal bi-branch dilated convolutional structure is bound to the output interface of the graph adaptive convolutional structure. The number of convolutional layers, time window, and dilation sampling interval of the long-period branch and the short-period branch are set respectively. The concatenation order of the two branch outputs in the channel dimension is determined, and the concatenated output interface is associated with the input of the prediction output layer. Initialization operations are performed on the weight parameters, bias parameters, and normalization parameters of the temporal bi-branch dilated convolutional structure and the prediction output layer.

5. The fiber optic access traffic scheduling method based on edge computing and AI according to claim 1, characterized in that, The formation of the spatial correlation features specifically includes: In the graph adaptive convolutional structure, the input feature vector sequence is used as input, and each user-side fiber optic access device in the device list is used as a graph node. Based on the physical topology of the fiber optic access network, it is determined whether there is a physical connection between any two devices. The connected device pairs are registered as graph edge information. The number of connections for each device is counted to generate the corresponding node degree information. Inside the graph adaptive convolutional structure, the initial graph Laplacian matrix configuration data is generated based on the node degree information and the graph edge information. The initial graph Laplacian matrix configuration data is loaded into the spectral domain convolution processing unit of the graph adaptive convolution structure. The spectral domain convolution processing is performed sequentially on the feature vector sequence corresponding to each sampling time. The spectral range corresponding to the initial graph Laplacian matrix is ​​divided into different frequency bands of low frequency, mid frequency and high frequency. A set of filter weight parameters is pre-configured for each frequency band. The different frequency bands are weighted and combined to obtain spectral feature data. The spectral features of each frequency band are concatenated in the feature dimension to generate a comprehensive graph spectral feature record. In the edge computing node, based on the comprehensive graph spectral feature record, the graph adjacency matrix inside the graph adaptive convolution structure is dynamically updated. For each pair of user-side fiber access devices that are registered as having a physical connection relationship, the corresponding feature vector of the device in the comprehensive graph spectral feature is extracted, and the similarity index is obtained according to the preset similarity calculation rules. At the same time, the current similarity index and the preset smoothing coefficient are combined in a weighted manner to generate a new edge weight value. The new edge weight value is written into the graph adjacency matrix configuration unit to complete the update of the edge weights of all connected devices and form an adaptive graph adjacency matrix. The adaptive graph structure and the feature vector sequence at the corresponding sampling time are input into the graph convolution operation unit of the graph adaptive convolution structure. According to the preset number of graph convolution layers and multi-order neighbor aggregation rules, the node features and multi-order neighbor node features are weighted, aggregated and linearly transformed in each layer to obtain spatial correlation features.

6. The fiber optic access traffic scheduling method based on edge computing and AI according to claim 1, characterized in that, The traffic prediction results and congestion risk indicators are obtained specifically through the following methods: Spatial correlation features are sequentially written into the input buffer of the temporal dual-branch dilated convolution structure according to the sampling time, forming a temporally ordered sequence of time windows, which are then associated with the input ports of the long-period branch and the short-period branch. In the long-period branch, for each time window, the periodic index of the historical traffic change of each user-side fiber access device is calculated. The hole factor value used by the current convolutional layer is selected from the preset different hole factor candidate sets. The multi-layer one-dimensional causal hole convolutional layers are enabled in the order of increasing selected hole factors. Layer-by-layer convolution operation is performed on the spatial correlation features at different sampling times within the same time window to grow the periodic time feature sequence. In the short-period branch, for each time window, the local fluctuation index of traffic and queue status of each user-side fiber access device is calculated in a smaller time neighborhood. Based on the local fluctuation index, the hole factor and convolution kernel used in the current convolution layer are selected from the preset combination of small hole factor and convolution kernel. Multi-layer one-dimensional causal hole convolution operation is performed on the spatial correlation features of the current sampling time and a small number of adjacent sampling times. In each layer, channel gating weights are generated based on the local fluctuation index and the convolution output is scaled by channel to obtain the short-period time feature sequence. At each sampling time, the long-period time feature sequence and the short-period time feature sequence are concatenated in channel order, and the channel weights are calculated according to the preset channel weight generation rules. The concatenated multi-scale time feature matrix is ​​input into the prediction layer, and the corresponding traffic prediction results and congestion risk indicators are output.

7. The fiber optic access traffic scheduling method based on edge computing and AI according to claim 1, characterized in that, The process of generating the dynamic bandwidth allocation parameters specifically includes: The traffic prediction results and congestion risk indicators are written into the prediction data cache according to the user-side fiber access equipment identifier and sampling time to form a prediction data set. Using the predicted data set as input, the traffic prediction value of each user-side fiber access device in the target scheduling period is read one by one as the basic demand, and the corresponding congestion risk index is read as the demand increment factor. The basic demand is incrementally calculated to form the initial bandwidth demand value, and the initial bandwidth demand value is written into the temporary bandwidth demand list. Read the allocable total bandwidth from the temporary list of bandwidth requirements, perform total bandwidth constraint processing on all initial bandwidth requirements, scale the initial requirement values ​​of each device, and form dynamic bandwidth allocation parameters.

8. The fiber optic access traffic scheduling method based on edge computing and AI according to claim 1, characterized in that, The formation of the fiber optic access traffic scheduling scheme specifically includes: The dynamic bandwidth allocation parameters are used to construct a scheduling parameter sequence with the scheduling period as the time index; Based on the scheduling parameter sequence, the current total bandwidth is read one time slot at a time within the scheduling period. The available bandwidth is calculated according to the dynamic bandwidth allocation parameters. Combined with the service queue status and service level, the service priority and service order of each service queue in the current time slot are sorted to form a time slot bandwidth allocation list and time slot scheduling information. Based on the time slot bandwidth allocation list and time slot scheduling information, the available bandwidth of all time slots within the scheduling period is combined with the service queue order to form a fiber access traffic scheduling scheme, which is then distributed to the central office optical line terminal equipment for execution through the edge computing node.

Images