Dynamic bandwidth allocation method and system based on large model, electronic equipment and medium

Abstract

The invention discloses a dynamic bandwidth allocation method and system based on a large model, electronic equipment and a medium, and relates to the technical field of network communication, historical flow data and multi-modal data are modeled and fused by using a pre-training architecture of ChatGLM3, so that fusion features are generated. Based on the fusion features, the system can predict future traffic demands. By defining a dynamic priority adjustment mechanism and combining a flow prediction result, the system realizes multi-service dynamic adjustment and efficient bandwidth allocation; meanwhile, a constructed flow anomaly detection mechanism can monitor flow changes in real time, and the priority of abnormal services is improved; in addition, a global resource optimization strategy ensures comprehensive scheduling of bandwidth resources in the network, and the benefit of bandwidth allocation is measured based on a utility function; the invention provides an intelligent method, so that network resources can be flexibly configured under various service requirements, and the overall service quality and network performance are improved.

Description

Technical Field

[0001] The present invention relates to the field of network communication technology, and in particular to a large model-based dynamic bandwidth allocation method, system, electronic equipment and medium. Background Art

[0002] With the rapid development of information technology and the continuous expansion of network applications, the surge in network traffic has become an indisputable fact. Especially in scenarios with high bandwidth requirements such as video streaming, cloud computing, and online games, users' expectations for network quality are constantly increasing. This change in demand has brought huge challenges to network management and resource allocation. Existing network resource allocation solutions mainly include fixed bandwidth allocation, dynamic allocation based on priority scheduling, and load balancing solutions. Fixed bandwidth allocation cannot adapt to dynamic traffic changes, resulting in network congestion and reduced user experience during peak hours. Although the priority scheduling-based solution can provide priority protection for real-time services such as video calls, its static priority setting lacks intelligence and cannot cope with the challenges brought by burst traffic. At the same time, load balancing technology dynamically adjusts node loads by monitoring traffic. Although it can disperse pressure, it has deficiencies in refined control of bandwidth allocation and traffic prediction, making it difficult to effectively prevent the problem of insufficient bandwidth. Therefore, it is of great significance to develop a method that can intelligently and adaptively allocate network resources. Summary of the invention

[0003] The purpose of the present invention is to provide a large model-based dynamic bandwidth allocation method, system, electronic device and medium to solve the deficiencies in the background technology.

[0004] In order to achieve the above object, the present invention provides the following technical solution: a dynamic bandwidth allocation method based on a large model, comprising: The pre-trained architecture of ChatGLM3 is used to model and fuse historical traffic data and multimodal data to obtain fusion features, and the ChatGLM3 large model is used to predict future traffic based on the fusion features to obtain traffic prediction; Define dynamic priority adjustment mechanism and intelligent bandwidth allocation based on traffic prediction to achieve dynamic adjustment and bandwidth allocation for multiple services; Build a global resource optimization strategy, combine priority adjustment and bandwidth allocation based on traffic prediction, and realize comprehensive scheduling of bandwidth resources in the network.

[0005] In a new embodiment, the steps of using the pre-trained architecture of ChatGLM3 to model and fuse historical traffic data and multimodal data to obtain fusion features, and using the ChatGLM3 large model to predict future traffic based on the fusion features to obtain traffic prediction include: Defining historical traffic datasets for large-scale networks ,in represents the number of time steps, represents the number of devices, and Indicates the device In time Traffic flow at the moment; The pre-trained architecture of ChatGLM3 is used to perform global dependency modeling on the input historical traffic numbers through the self-attention mechanism to automatically capture the correlation in the time series: in, They are historical traffic datasets The query, key and value matrices obtained after linear transformation, is a learnable weight matrix, is the dimension of the key vector, the output of the self-attention mechanism Represents the weighted information of historical traffic data in the time series, which is used to show the trend and dependency of traffic changes; At the same time, ChatGLM3's pre-trained architecture is used to perform multimodal data fusion to obtain a multimodal feature representation vector; Use gating mechanism to fuse the multimodal feature representation vector and the output of self-attention mechanism The fusion features are obtained and the ChatGLM3 large model is used for traffic prediction.

[0006] In a new embodiment, the steps of using the pre-trained architecture of ChatGLM3 to perform multimodal data fusion are: Define multimodal data input, respectively , represents the historical data under different modes; Through the pre-trained architecture of ChatGLM3, the feature-level fusion method is used to perform multimodal data fusion, mapping historical data under different modalities into a unified feature space: in, is modal The feature extraction function is is a multimodal feature representation vector.

[0007] In a new embodiment, the steps of using a gating mechanism to fuse the multimodal feature representation vector and the output of the self-attention mechanism to obtain a fused feature, and using the ChatGLM3 large model to perform traffic prediction are: Calculate a gating weight , used to decide and relative importance in the integration process; Gating weight Through and After linear transformation and Sigmoid activation function, we get: in, represents the learnable weight matrix of the gating layer, is the bias term, Represents the Sigmoid activation function and limits the gate weight to the interval [0,1]; Using gated weights make and Perform weighted fusion to obtain fusion features: in, represents element-wise multiplication, when When it is larger, the model will pay more attention to ,when When it is smaller, the model will rely more on ; Based on the fusion features, assuming that the predicted time step is , prediction results of the ChatGLM3 large model It can be expressed as: in, It is the prediction function of the ChatGLM3 model on multimodal data and historical traffic data, and the output It is the traffic forecast.

[0008] In a new embodiment, the steps of defining a dynamic priority adjustment mechanism and intelligent bandwidth allocation based on traffic prediction to achieve dynamic adjustment and bandwidth allocation of multiple services are as follows: Define that there are N businesses in the system and the priority of each business and bandwidth requirements Changes over time; The dynamic adjustment mechanism of priority is realized by the following formula: in, It's time business The priority of It's time Time business The priority of It's business In time The flow demand function at time, It's business In time Traffic information at the moment, is the predicted future traffic demand, is a smoothing coefficient, which is used to balance the weight of current priority and real-time demand and control the speed of priority adjustment. Is the balance factor, used to control the current moment and The relative weight of ; Traffic demand function Used according to Calculate its bandwidth requirement, expressed as: Indicates business In time The traffic proportion at the moment is used to dynamically adjust its priority. When the traffic demand is large, its priority will increase; conversely, when the business traffic is small, its priority will decrease. Similarly, ; When priority The total bandwidth resources are determined by weighted bandwidth allocation method. Dynamically allocated to different services, expressed as: in, yes Time allocated to business bandwidth, yes Total bandwidth resources at the moment, It's business The priority of It is the sum of all business priorities; Formula (8) shows that the bandwidth allocation obtained by each service is proportional to its priority. The higher the priority of the service, the more bandwidth resources it obtains. Build a traffic anomaly detection mechanism to detect traffic anomaly services and determine whether the priority of traffic anomaly services needs to be adjusted.

[0009] In a new embodiment, the steps of constructing a traffic anomaly detection mechanism for detecting traffic anomaly services and determining whether the priority of traffic anomaly services needs to be adjusted are: Collect the traffic change rate of each business and the standard deviation of flow rate variation ,in for: And assuming there is time points, the standard deviation for: in It represents the average value of the rate of change; when When the system considers that abnormal traffic occurs, Threshold parameters for controlling anomaly detection; When a traffic abnormality service occurs, the system increases the priority weight of the traffic abnormality service until the traffic of the traffic abnormality service returns to normal.

[0010] In a new embodiment, the steps of constructing a global resource optimization strategy, combining priority adjustment and bandwidth allocation based on traffic prediction, and realizing comprehensive scheduling of bandwidth resources in the network are: The bandwidth resources in the network are divided into a finite number of clusters, and a utility function is defined to measure the benefits of bandwidth allocation: Based on formula (8), the global optimization goal of bandwidth resources in the network is: in, Is a cluster The weight is used to further reflect the business importance of the cluster.

[0011] The present invention also provides a large model-based dynamic bandwidth allocation system, comprising: Traffic prediction model module: Use the pre-trained architecture of ChatGLM3 to model and fuse historical traffic data and multimodal data to obtain fusion features, and use the ChatGLM3 large model to predict future traffic based on the fusion features to obtain traffic prediction; Dynamic priority adjustment module: connected with the traffic prediction model module, defines dynamic priority adjustment mechanism and intelligent bandwidth allocation based on traffic prediction, and realizes dynamic adjustment and bandwidth allocation of multiple services; Global resource optimization module: connected with the dynamic priority adjustment module, builds a global resource optimization strategy, combines priority adjustment and bandwidth allocation based on traffic prediction, and realizes comprehensive scheduling of bandwidth resources in the network.

[0012] In the above technical solution, the technical effects and advantages provided by the present invention are: 1. This invention uses a pre-trained architecture to model and fuse historical traffic data and multimodal data input, significantly improving the accuracy of traffic prediction. Through the self-attention mechanism, this method can automatically capture the correlation and trend in the time series, providing a sufficient information basis for the model. Based on the use of historical and real-time data, ChatGLM3 can understand the complex patterns in the data and make accurate predictions of future traffic changes; 2. The present invention ensures efficient use of network resources by constructing a dynamic priority adjustment mechanism. The system automatically adjusts the priority of each service according to the traffic demand at a specific time. When the traffic demand of a service increases, its priority also increases to ensure that the service can obtain the required bandwidth. Conversely, when the traffic demand decreases, the priority will also decrease. Such an adjustment can effectively avoid waste or shortage of resources under static strategies. By introducing smoothing coefficients and weight factors, the system can flexibly control the speed and amplitude of priority adjustment, thereby improving management accuracy. This mechanism ensures that key services can still be given priority when the network is highly loaded, enhancing the service quality and user experience of the entire network, and making it more adaptable in a changing environment; 3. The present invention constructs an intelligent bandwidth allocation method, makes full use of the results of traffic prediction and dynamic priority adjustment, and adopts a weighted method to allocate bandwidth to different services. High-priority services obtain more bandwidth in the allocation, while low-priority services obtain relatively less bandwidth, ensuring that network resources are reasonably used. In addition, the system can monitor and analyze the traffic changes of each service in real time, and effectively identify any abnormal traffic service by calculating the traffic change rate and standard deviation of each service, and increase the priority through the anomaly detection mechanism, so that it can obtain the necessary bandwidth support and ensure the stable operation of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0014] Figure 1 is a flow chart of the method of the present invention; Figure 2 It is a system block diagram of the present invention. DETAILED DESCRIPTION

[0015] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0016] Example 1, please refer to Figure 1 As shown, the large model-based dynamic bandwidth allocation method described in this embodiment includes: S1. Use the pre-trained architecture of ChatGLM3 to model and fuse historical traffic data and multimodal data to obtain fusion features, and use the ChatGLM3 large model to predict future traffic based on the fusion features to obtain traffic prediction; S2. Define dynamic priority adjustment mechanism and intelligent bandwidth allocation based on traffic prediction to achieve dynamic adjustment and bandwidth allocation for multiple services; S3. Build a global resource optimization strategy, combine priority adjustment and bandwidth allocation based on traffic prediction, and realize comprehensive scheduling of bandwidth resources in the network; As described in the above steps S1-S3, with the rapid development of information technology and the continuous expansion of network applications, the surge in network traffic has become an indisputable fact. Especially in scenarios with high bandwidth requirements such as video streaming, cloud computing, and online games, users' expectations for network quality are constantly increasing. This change in demand poses huge challenges to network management and resource allocation. Existing network resource allocation schemes mainly include fixed bandwidth allocation, dynamic allocation based on priority scheduling, and load balancing schemes. Fixed bandwidth allocation cannot adapt to dynamic traffic changes, resulting in peak network congestion and reduced user experience. Although the priority scheduling-based scheme can provide priority protection for high-real-time services such as video calls, its static priority setting lacks intelligence and cannot cope with the challenges brought by burst traffic. At the same time, load balancing technology dynamically adjusts node loads by monitoring traffic. Although it can disperse pressure, it has deficiencies in refined control of bandwidth allocation and traffic prediction, and it is difficult to effectively prevent the problem of insufficient bandwidth. The present invention models and fuses historical traffic data and multimodal data inputs through a pre-training architecture, significantly improving the accuracy of traffic prediction. Through the self-attention mechanism, the method can automatically capture the correlation and trend in the time series, providing a sufficient information basis for the model. Based on the use of historical and real-time data, ChatGLM3 can understand the complex patterns in the data and make accurate predictions about future traffic changes. By building a dynamic priority adjustment mechanism, it ensures efficient use of network resources. The system automatically adjusts the priority of each service based on its traffic demand at a specific time. When the traffic demand of a service increases, its priority also increases to ensure that the service can obtain the required bandwidth. Conversely, when traffic demand decreases, the priority will also be reduced. Such adjustments can effectively avoid waste or shortage of resources under static strategies. By introducing smoothing coefficients and weight factors, the system can flexibly control the speed and amplitude of priority adjustment, thereby improving management accuracy. This mechanism ensures that key services can still be given priority when the network is under high load, enhancing the service quality and user experience of the entire network, making it more adaptable in a changing environment; at the same time, by constructing an intelligent bandwidth allocation method, making full use of the results of traffic prediction and dynamic priority adjustment, and allocating bandwidth to different services in a weighted manner, high-priority services get more bandwidth in the allocation, while low-priority services get relatively less, ensuring that network resources are used reasonably. In addition, the system can monitor and analyze the traffic changes of each service in real time, and effectively identify any abnormal traffic service by calculating the traffic change rate and standard deviation of each service, and increase its priority through the anomaly detection mechanism, so that it can get the necessary bandwidth support and ensure the stable operation of the system.

[0017] In one embodiment, the step S1 of using the ChatGLM3 pre-trained architecture to model and fuse historical traffic data and multimodal data to obtain fusion features, and using the ChatGLM3 large model to predict future traffic based on the fusion features to obtain traffic prediction includes: S11. Define historical traffic datasets for large-scale networks ,in represents the number of time steps, represents the number of devices, and Indicates the device In time Traffic flow at the moment; S12. Use the pre-trained architecture of ChatGLM3 to perform global dependency modeling on the input historical traffic numbers through the self-attention mechanism to automatically capture the correlation in the time series: S13, among which, They are historical traffic datasets The query, key and value matrices obtained after linear transformation, is a learnable weight matrix, is the dimension of the key vector, the output of the self-attention mechanism Represents the weighted information of historical traffic data in the time series, which is used to show the trend and dependency of traffic changes; S14. At the same time, ChatGLM3's pre-trained architecture is used to perform multimodal data fusion to obtain a multimodal feature representation vector; S15. Use the gating mechanism to fuse the multimodal feature representation vector and the output of the self-attention mechanism Get fusion features and use ChatGLM3 large model for traffic prediction; As described in the above steps S11-S15, there will be a large amount of historical traffic data in the network. Define the historical traffic data set, use the ChatGLM3 pre-training architecture to perform global dependency modeling on the input historical traffic data through the self-attention mechanism to obtain ,The historical traffic data X is linearly transformed by dynamically calculating the query (Q), key (K) and value (V) matrices. These matrices are combined with the learnable weight matrix to generate attention scores to measure the importance of traffic at different time points at the current moment. The output of the attention mechanism can be regarded as a weighted representation of historical traffic data, which effectively extracts information from the time series, enabling the model to accurately identify the trends and dependencies of traffic changes.

[0018] In one embodiment, the step S14 of using the pre-trained architecture of ChatGLM3 to perform multimodal data fusion includes: S141. Define multimodal data input, which are , Represents historical data in different modes; S142. Through the pre-training architecture of ChatGLM3, the feature-level fusion method is used to perform multimodal data fusion, and the historical data under different modalities are mapped to a unified feature space: S143, among which, is modal The feature extraction function is is a multimodal feature representation vector; As described in the above steps S141-S143, in addition to the network traffic time series data, information such as the device status in the network, user behavior patterns, and environmental changes also affect the fluctuation of network traffic. Assuming that the system has multimodal input, and defining the historical traffic data set of each modality, the pre-trained architecture of ChatGLM3 is used to perform multimodal data fusion using the feature-level fusion method, and the historical traffic data under different modalities are mapped to a unified feature space. For example, Represented as user behavior history data, Represented as environmental historical data, the feature representation of user behavior history data and environmental history data is extracted through the pre-trained architecture of ChatGLM3 and the extracted feature representation is mapped into a unified feature space.

[0019] In one embodiment, the gating mechanism is used to fuse the multimodal feature representation vector and the output of the self-attention mechanism. Step S15 of obtaining the fusion features and using the ChatGLM3 large model to perform traffic prediction includes: S151. Calculate a gating weight , used to decide and relative importance in the integration process; S152, gate weight Through and After linear transformation and Sigmoid activation function, we get: S153, among which, represents the learnable weight matrix of the gating layer, is the bias term, Represents the Sigmoid activation function and limits the gate weight to the interval [0,1]; S154, using gated weights make and Perform weighted fusion to obtain fusion features: S155, among which, represents element-wise multiplication, when When it is larger, the model will pay more attention to ,when When it is smaller, the model will rely more on ; S156. Based on the fusion features, assuming that the predicted time step is , prediction results of the ChatGLM3 large model It can be expressed as: S157, among which, It is the prediction function of the ChatGLM3 model on multimodal data and historical traffic data, and the output It is the traffic forecast; As described in the above steps S151-S157, in order to determine the relative importance of different modal features in the feature fusion process, a gating weight is first calculated. This weight will be used to adjust the contribution of each modal input in the fusion process, and the multimodal feature representation vector and the historical traffic processed by the attention mechanism are weighted and fused using the gating weight. By calculating the gating weight, the ChatGLM3 large model can autonomously determine the relative importance of various input features in prediction under dynamic changes. For example, in some cases, user behavior characteristics may significantly affect traffic prediction. At this time, when the value of the gating weight is high, the contribution of user behavior characteristics will be amplified. At the same time, the combination of linear transformation and Sigmoid activation enables the gating weight to have the ability of self-learning and can be continuously adjusted according to the traffic pattern observed during the forward propagation process.

[0020] In one embodiment, the step S2 of defining a dynamic priority adjustment mechanism and intelligent bandwidth allocation based on traffic prediction to achieve dynamic adjustment and bandwidth allocation of multiple services includes: S21. Define that there are N businesses in the system and the priority of each business and bandwidth requirements Changes over time; S22, the dynamic adjustment mechanism of priority is realized by the following formula: S23, among which, It's time business The priority of It's time Time business The priority of It's business In time The flow demand function at time, It's business In time Traffic information at the moment, is the predicted future traffic demand, is a smoothing coefficient, which is used to balance the weight of current priority and real-time demand and control the speed of priority adjustment. Is the balance factor, used to control the current moment and The relative weight of ; S24. Flow demand function Used according to Calculate its bandwidth requirement, the expression is: S25, indicating business In time The traffic proportion at the moment is used to dynamically adjust its priority. When the traffic demand is large, its priority will increase; conversely, when the business traffic is small, its priority will decrease. Similarly, ; S26, when priority The total bandwidth resources are determined by weighted bandwidth allocation method. Dynamically allocated to different services, expressed as: S27, among which, yes Time allocated to business bandwidth, yes Total bandwidth resources at the moment, It's business The priority of It is the sum of all business priorities; S28, Formula (8) shows that the bandwidth allocation obtained by each service is proportional to its priority. The higher the priority of the service, the more bandwidth resources it obtains; S29, constructing a traffic anomaly detection mechanism to detect traffic anomaly services and determine whether the priority of traffic anomaly services needs to be adjusted; As described in the above steps S21-S29, it is defined that there are N services in the system, and the priority and bandwidth demand of each service change over time. A dynamic adjustment formula for service priority is defined. The dynamic adjustment formula is based on the traffic demand function combined with a smoothing coefficient and a weight factor to achieve real-time priority adjustment. The traffic demand function is used to calculate the bandwidth demand of the service according to its current traffic situation, and the specific form of the function can be selected according to the actual situation of the network application. The present invention uses a linear proportional function, which can represent the traffic proportion of a certain service at the current moment and is used to dynamically adjust its priority. When the traffic demand of a certain service is large, its priority will increase accordingly; conversely, when the traffic of a certain service is small, the priority will be appropriately reduced. Through this traffic proportion adjustment method, it can be adaptively guaranteed that high-demand services obtain more bandwidth resources. When the priority of the service is determined, a bandwidth allocation formula (8) is defined based on the priority. The formula shows that the bandwidth allocation obtained by each service is proportional to its priority. The higher the priority, the more bandwidth the service obtains. Through this weighted allocation strategy, the system can dynamically adjust the bandwidth allocation according to the real-time priority to ensure that key services are given priority when there is high traffic demand.

[0021] In one embodiment, the step S29 of constructing a traffic anomaly detection mechanism for detecting traffic anomaly services and determining whether the priority of traffic anomaly services needs to be adjusted includes: S291. Collect the traffic change rate of each service and the standard deviation of flow rate variation ,in for: S292, and assuming that time points, the standard deviation for: S293, among which It represents the average value of the rate of change; S294, when When the system considers that abnormal traffic occurs, Threshold parameters for controlling anomaly detection; S295. When an abnormal traffic service occurs, the system increases the priority weight of the abnormal traffic service until the traffic of the abnormal traffic service returns to normal; As described in the above steps S291-S295, in order to ensure the reliability of the dynamic priority adjustment mechanism, a traffic anomaly detection mechanism is introduced. When a sudden anomaly occurs in the traffic of a certain service, such as a sudden and substantial increase or decrease, the system will trigger the anomaly detection mechanism, and by analyzing historical traffic data and current traffic changes, determine whether it is necessary to immediately adjust the priority of the service, or to perform a larger-scale bandwidth reallocation. By calculating the traffic change rate of each service and the standard deviation of the traffic change, and setting a judgment threshold, once a traffic anomaly is detected in a certain service, the system will temporarily increase the priority of the abnormal service according to its actual needs, which can be achieved by increasing the priority weight of the service. When resources are allocated, the service can obtain more bandwidth. If such situations are not responded to in a timely manner, the service quality of the abnormal service may be significantly reduced or even interrupted. Increasing the priority can help the service quickly obtain resources in case of emergencies and meet emergency needs. After temporarily increasing the priority of the abnormal service, the system will reallocate global bandwidth resources. The new bandwidth allocation priority order is usually: critical service > abnormal service > normal service, which means that during the existence of the abnormal service, its priority will be temporarily higher than that of the normal service, but still lower than that of the critical service. When the traffic of the abnormal service returns to normal, the system will stop the dynamic adjustment of its priority and allocate bandwidth according to the normal priority again.

[0022] In one embodiment, the step S3 of constructing a global resource optimization strategy, combining priority adjustment and bandwidth allocation based on traffic prediction, to achieve comprehensive scheduling of bandwidth resources in the network includes: S31. Divide the bandwidth resources in the network into a finite number of clusters and define a utility function to measure the benefits of bandwidth allocation: S32. Based on formula (8), the global optimization target of bandwidth resources in the network is: S33, among which, Is a cluster The weight is used to further reflect the business importance of the cluster; As described in the above steps S31-S33, in order to further improve the utilization efficiency of the overall network resources, a global resource optimization strategy is proposed. Combined with the priority update and bandwidth allocation based on traffic prediction, the comprehensive scheduling of bandwidth resources in the network is realized. The global optimization strategy divides the bandwidth resources of each node in the network into multiple clusters through the cluster scheduling method, and allocates bandwidth using the multi-objective optimization algorithm. Through the bandwidth allocation formula (8), it is ensured that the high-priority cluster obtains more bandwidth resources to meet the bandwidth requirements of important services. Therefore, a utility function is defined, and a logarithmic utility function is used to measure the benefits of bandwidth allocation to reflect the marginal diminishing effect. This utility function ensures the efficient utilization of bandwidth resources while avoiding excessive concentration of allocation to a certain cluster, which is conducive to fairness. The global optimization goal of bandwidth allocation is defined (10). By optimizing this objective function, we maximize the total utility under the constraint of total bandwidth resources and achieve the best resource utilization.

[0023] Example 2, please refer to Figure 2 As shown, the large model-based dynamic bandwidth allocation system described in this embodiment includes: Traffic prediction model module: Use the pre-trained architecture of ChatGLM3 to model and fuse historical traffic data and multimodal data to obtain fusion features, and use the ChatGLM3 large model to predict future traffic based on the fusion features to obtain traffic prediction; Dynamic priority adjustment module: connected with the traffic prediction model module, defines dynamic priority adjustment mechanism and intelligent bandwidth allocation based on traffic prediction, and realizes dynamic adjustment and bandwidth allocation of multiple services; Global resource optimization module: connected with the dynamic priority adjustment module, builds a global resource optimization strategy, combines priority adjustment and bandwidth allocation based on traffic prediction, and realizes comprehensive scheduling of bandwidth resources in the network.

[0024] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A dynamic bandwidth allocation method based on a large model, characterized in that: The pre-trained architecture of ChatGLM3 is used to model and fuse historical traffic data and multimodal data to obtain fusion features, and the ChatGLM3 large model is used to predict future traffic based on the fusion features to obtain traffic prediction; Define dynamic priority adjustment mechanism and intelligent bandwidth allocation based on traffic prediction to achieve dynamic adjustment and bandwidth allocation for multiple services; Build a global resource optimization strategy, combine priority adjustment and bandwidth allocation based on traffic prediction, and realize comprehensive scheduling of bandwidth resources in the network.

2. The method for dynamic bandwidth allocation based on a large model according to claim 1, characterized in that: The steps of using the ChatGLM3 pre-training architecture to model and fuse historical traffic data and multimodal data to obtain fusion features, and using the ChatGLM3 large model to predict future traffic based on the fusion features to obtain traffic prediction are as follows: Defining historical traffic datasets for large-scale networks ,in represents the number of time steps, represents the number of devices, and Indicates the device In time Traffic flow at the moment; The pre-trained architecture of ChatGLM3 is used to perform global dependency modeling on the input historical traffic numbers through the self-attention mechanism to automatically capture the correlation in the time series: in, They are historical traffic datasets The query, key and value matrices obtained after linear transformation, is a learnable weight matrix, is the dimension of the key vector, the output of the self-attention mechanism Represents the weighted information of historical traffic data in the time series, used to show the trend and dependency of traffic changes; At the same time, ChatGLM3's pre-trained architecture is used to perform multimodal data fusion to obtain a multimodal feature representation vector; Use gating mechanism to fuse the multimodal feature representation vector and the output of self-attention mechanism The fusion features are obtained and the ChatGLM3 large model is used for traffic prediction.

3. The method for dynamic bandwidth allocation based on a large model according to claim 2, characterized in that: The steps of using the ChatGLM3 pre-training architecture to perform multimodal data fusion are: Define multimodal data input, respectively , represents the historical data under different modes; Through the pre-trained architecture of ChatGLM3, the feature-level fusion method is used to perform multimodal data fusion, mapping historical data under different modalities into a unified feature space: in, is modal The feature extraction function is is a multimodal feature representation vector.

4. The method for dynamic bandwidth allocation based on a large model according to claim 3, characterized in that: The gating mechanism is used to fuse the multimodal feature representation vector and the output of the self-attention mechanism The steps to obtain fusion features and use the ChatGLM3 model for traffic prediction are as follows: Calculate a gating weight , used to decide and relative importance in the integration process; Gating weight Through and After linear transformation and Sigmoid activation function, we get: in, represents the learnable weight matrix of the gating layer, is the bias term, Represents the Sigmoid activation function and limits the gate weight to the interval [0,1]; Using gated weights make and Perform weighted fusion to obtain fusion features: in, Represents element-wise multiplication, when When it is larger, the model will pay more attention to ,when When it is smaller, the model will rely more on ; Based on the fusion features, assuming that the predicted time step is , prediction results of the ChatGLM3 large model It can be expressed as: in, It is the prediction function of the ChatGLM3 model on multimodal data and historical traffic data, and the output It is the traffic forecast.

5. The method for dynamic bandwidth allocation based on a large model according to claim 1, characterized in that: The steps of defining a dynamic priority adjustment mechanism and intelligent bandwidth allocation based on traffic prediction to achieve dynamic adjustment and bandwidth allocation of multiple services are as follows: Define that there are N businesses in the system and the priority of each business and bandwidth requirements Changes over time; The dynamic adjustment mechanism of priority is realized by the following formula: in, It's time business The priority of It's time Time business The priority of It's business In time The flow demand function at time, It's business In time Traffic information at the moment, is the predicted future traffic demand, is a smoothing coefficient, which is used to balance the weight of current priority and real-time demand and control the speed of priority adjustment. is the balance factor, used to control the current moment and The relative weight of ; Traffic demand function Used according to Calculate its bandwidth requirement, expressed as: Indicates business In time The traffic proportion at the moment is used to dynamically adjust its priority. When the traffic demand is large, its priority will increase; conversely, when the business traffic is small, its priority will decrease. Similarly, ; When priority The total bandwidth resources are determined by weighted bandwidth allocation method. Dynamic allocation to different services, expressed as: in, yes Time allocated to business bandwidth, yes Total bandwidth resources at the moment, It's business The priority of It is the sum of all business priorities; Formula (8) shows that the bandwidth allocation obtained by each service is proportional to its priority. The higher the priority of the service, the more bandwidth resources it obtains. Build a traffic anomaly detection mechanism to detect traffic anomaly services and determine whether the priority of traffic anomaly services needs to be adjusted.

6. The method for dynamic bandwidth allocation based on a large model according to claim 5, characterized in that: The steps of constructing a traffic anomaly detection mechanism for detecting traffic anomaly services and determining whether the priority of traffic anomaly services needs to be adjusted are as follows: Collect the traffic change rate of each business and the standard deviation of flow rate variation ,in for: And assuming there is time points, the standard deviation for: in It represents the average value of the rate of change; when When the system considers that abnormal traffic occurs, Threshold parameters for controlling anomaly detection; When a traffic abnormality service occurs, the system increases the priority weight of the traffic abnormality service until the traffic of the traffic abnormality service returns to normal.

7. The method for dynamic bandwidth allocation based on a large model according to claim 1, characterized in that: The steps of constructing a global resource optimization strategy, combining priority adjustment and bandwidth allocation based on traffic prediction, and realizing comprehensive scheduling of bandwidth resources in the network are as follows: The bandwidth resources in the network are divided into a finite number of clusters, and a utility function is defined to measure the benefits of bandwidth allocation: Based on formula (8), the global optimization goal of bandwidth resources in the network is: in, Is a cluster The weight is used to further reflect the business importance of the cluster.

8. A large model-based dynamic bandwidth allocation system, used to implement a large model-based dynamic bandwidth allocation method according to any one of claims 1 to 7, characterized in that: Traffic prediction model module: Use the pre-trained architecture of ChatGLM3 to model and fuse historical traffic data and multimodal data to obtain fusion features, and use the ChatGLM3 large model to predict future traffic based on the fusion features to obtain traffic prediction; Dynamic priority adjustment module: connected with the traffic prediction model module, defines dynamic priority adjustment mechanism and intelligent bandwidth allocation based on traffic prediction, and realizes dynamic adjustment and bandwidth allocation of multiple services; Global resource optimization module: connected with the dynamic priority adjustment module, builds a global resource optimization strategy, combines priority adjustment and bandwidth allocation based on traffic prediction, and realizes comprehensive scheduling of bandwidth resources in the network.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the computer program is executed by a processor, the steps of a large model-based dynamic bandwidth allocation method described in any one of claims 1 to 7 are implemented.

10. A computer readable medium having a computer program stored thereon, characterized in that When the computer program is executed by a processor, the steps of a large model-based dynamic bandwidth allocation method according to any one of claims 1 to 7 are implemented.

Images