A resource scheduling method and device, electronic equipment and storage medium

Abstract

The present disclosure provides a resource scheduling method and device, electronic equipment and storage medium, relates to the technical field of artificial intelligence, in particular to the technical field of large model, deep learning, network slicing and the like. The method comprises: for at least one service in the current scheduling scene, obtaining multi-source data collected by an edge node and a core network for each service respectively; inputting the multi-source data collected by the edge node and the core network for each service into a corresponding large model for prediction to obtain a first prediction result corresponding to the edge node and a second prediction result corresponding to the core network; based on attribute information of the service, fusing the first prediction result and the second prediction result to obtain a target prediction result corresponding to the service; and generating a resource scheduling scheme according to the target prediction results corresponding to all services, first slice resources at a current moment and network state parameters. Thus, intelligent slice scheduling of center and edge cooperation can be realized, and the individuality and reliability of resource scheduling are improved.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, particularly to the fields of large models, deep learning, and network slicing, and specifically to a resource scheduling method, apparatus, electronic device, and storage medium. Background Technology

[0002] In 5G and future networks, network slicing technology enables customized services for multiple services through logical isolation, becoming a core means to support the low latency and high reliability requirements of vertical industries such as the Industrial Internet and the Internet of Vehicles.

[0003] However, relying on static configuration or simple rules based on thresholds (such as fixed bandwidth allocation and priority queuing) for slice resource scheduling cannot cope with the scheduling requirements of dynamic scenarios such as rapid changes in user behavior over time or space and non-linear fluctuations in business load. It is difficult to simultaneously take into account resource utilization and user experience. Therefore, there is an urgent need for a resource scheduling method that can dynamically perceive and is more personalized. Summary of the Invention

[0004] This disclosure aims to at least partially address one of the technical problems in the related art.

[0005] Therefore, the purpose of this disclosure is to propose a resource scheduling method, device, electronic device and storage medium. By co-deploying a distributed large model in the core network and edge nodes, it is possible to analyze user behavior characteristics, service traffic patterns and network status information in real time, and then dynamically adjust the allocation and priority scheduling of slice resources. While ensuring low latency and high reliability of critical services, it improves the overall utilization rate of slice resources and significantly enhances the intelligence and personalization of network resource allocation.

[0006] According to a first aspect of this disclosure, a resource scheduling method is provided, comprising:

[0007] For at least one service in the current scheduling scenario, acquire multi-source data collected by edge nodes and the core network for each service, wherein the multi-source data includes user behavior data, service traffic data and network status data; The multi-source data collected by the edge nodes and the core network for each service is input into the corresponding large model for prediction, to obtain the first prediction result output by the first large model corresponding to the edge node and the second prediction result output by the second large model corresponding to the core network. Based on the attribute information of the service, the first prediction result and the second prediction result are fused to obtain the target prediction result corresponding to the fused service; Based on the target prediction results for all services, the first slice resources at the current moment, and the network state parameters, a reinforcement learning optimizer is used to generate a resource scheduling scheme for at least one service, wherein the resource scheduling scheme includes the second slice resources corresponding to each service.

[0008] According to a second aspect of this disclosure, a resource scheduling method apparatus is provided, comprising: The data acquisition module is used to acquire multi-source data collected by the edge nodes and the core network for each service in the current scheduling scenario, wherein the multi-source data includes user behavior data, service traffic data and network status data. The prediction module is used to input the multi-source data collected by the edge nodes and the core network for each service into the corresponding large model for prediction, and obtain the first prediction result output by the first large model corresponding to the edge node and the second prediction result output by the second large model corresponding to the core network. The fusion module is used to fuse the first prediction result and the second prediction result based on the attribute information of the service to obtain the fused target prediction result corresponding to the service; The scheduling module is used to generate a resource scheduling scheme for at least one service based on the target prediction results corresponding to all services, the first slice resources at the current time, and network state parameters using a reinforcement learning optimizer, wherein the resource scheduling scheme includes the second slice resources corresponding to each service.

[0009] According to a third aspect of this disclosure, an electronic device is provided, comprising: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the resource scheduling method as described in the first aspect.

[0010] According to a fourth aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions for causing the computer to perform the resource scheduling method as described in the first aspect.

[0011] According to a fifth aspect of this disclosure, a computer program product is provided, including computer instructions that, when executed by a processor, implement the steps of the resource scheduling method as described in the first aspect.

[0012] The resource scheduling method, apparatus, electronic device, and storage medium provided in this disclosure have the following beneficial effects: By collaboratively deploying a distributed large model across the core network and edge nodes, multi-source data such as real-time user behavior characteristics, business traffic patterns, and network status information for each service in concurrent multi-service scenarios are collected and analyzed. Service demands are predicted from both global and local perspectives, and the global and local prediction results are fused together with the service data. Then, a reinforcement learning optimizer is used to dynamically adjust slice resource allocation and priority scheduling. This approach can improve the overall slice resource utilization while ensuring low latency and high reliability for critical services, significantly enhancing the intelligence and personalization of network resource allocation and optimizing user experience.

[0013] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0014] The above and / or additional aspects and advantages of this disclosure will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, which are provided for a better understanding of the present invention and are not intended to limit the scope of this disclosure, wherein: Figure 1 This is a flowchart illustrating a resource scheduling method according to an embodiment of the present disclosure; Figure 2 This is a flowchart illustrating a resource scheduling method according to another embodiment of the present disclosure; Figure 3 This is a flowchart illustrating a resource scheduling method according to another embodiment of the present disclosure; Figure 4 This is a system architecture diagram of a resource scheduling system proposed in this disclosure; Figure 5 This is a schematic diagram of the structure of a resource scheduling method apparatus according to an embodiment of the present disclosure; Figure 6 A block diagram of an exemplary electronic device suitable for implementing embodiments of the present disclosure is shown. Detailed Implementation

[0015] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0016] This disclosure relates to the fields of artificial intelligence technology, such as large models, deep learning, and network slicing.

[0017] Artificial Intelligence (AI) is a new technological science that studies, develops, and applies theories, methods, technologies, and application systems to simulate, extend, and expand human intelligence.

[0018] Large-scale artificial intelligence models (or simply "large models") refer to a class of artificial intelligence models with a large number of parameters built from artificial neural networks. They are typically pre-trained on massive datasets using self-supervised or semi-supervised learning, and then their performance and capabilities are further optimized through fine-tuning based on instructions and human alignment. Large models are characterized by a large number of parameters, large training data sets, and large computational resources, and possess the ability to solve general tasks, follow human instructions, and perform complex reasoning.

[0019] Deep learning learns the inherent patterns and hierarchical representations of sample data. The information gained during this learning process greatly aids in interpreting data such as text, images, and sound. The ultimate goal of deep learning is to enable machines to possess analytical and learning capabilities similar to humans, allowing them to recognize data such as text, images, and sound.

[0020] Network slicing is a technology that builds multiple logically isolated end-to-end virtual networks on a shared network infrastructure, achieving differentiated service guarantees through software-defined networking and network function virtualization. This technology can provide customized logical topologies, quality of service, and security levels for different business types or industry users. Its core value includes reducing private network construction costs, supporting flexible topology customization, and enabling automated slice management.

[0021] The collection, storage, use, processing, transmission, provision, and disclosure of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0022] The resource scheduling method, apparatus, electronic device, and storage medium of this disclosure are described below with reference to the accompanying drawings.

[0023] It should be noted that the execution subject of the resource scheduling method in this embodiment is a resource scheduling method device, which can be implemented by software and / or hardware. The device can be configured in an electronic device, which may include, but is not limited to, a terminal, a server, etc.

[0024] Figure 1 This is a flowchart illustrating a resource scheduling method according to an embodiment of the present disclosure.

[0025] like Figure 1 As shown, the resource scheduling method includes: S101: For at least one service in the current scheduling scenario, obtain multi-source data collected by the edge nodes and the core network for each service.

[0026] Multi-source data can include user behavior data, business traffic data, and network status data.

[0027] In this embodiment of the disclosure, user behavior data may include access time, access location, service type, and terminal characteristics. Service traffic data may include bandwidth requirements, latency sensitivity, and the number of concurrent users. Network status data may include link bandwidth, packet loss rate, latency, and jitter.

[0028] The resource scheduling method proposed in this disclosure can be used for slice resource allocation and scheduling in multi-service concurrent scenarios. For example, in an industrial park where there are high-definition video users and industrial control users, it is known that the current scheduling scenario includes high-definition video services and industrial control services. Therefore, it is necessary to collect data for each service in real time, predict the service demand in the future period, and then comprehensively allocate slice resources for multiple services to ensure the effectiveness and high utilization of resource scheduling in multi-service concurrent scenarios.

[0029] In this embodiment of the disclosure, user behavior data, service traffic data, and network status data can be obtained globally through the Network Functions Virtualization (NFV) interface on the core network side, and user behavior data, service traffic data, and network status data can be obtained locally through lightweight probes on the edge node side.

[0030] In this embodiment, user access data can be acquired at the base station via Radio Resource Control (RRC) signaling. The user access data collected by the base station within the current cell is then transmitted in real-time to the edge node with the closest physical or logical distance. This allows the large-scale model deployed on the edge node to utilize this information (e.g., a sudden influx of high-definition video users at a base station) to quickly identify burst traffic and immediately issue dynamic slice adjustments for that area. All base stations and edge nodes across the network can upload processed (e.g., anonymized or compressed) non-real-time, statistical user behavior data, or feature vectors of user behavior data, to the core network. This enables the large-scale model deployed on the core network to collect access patterns across the entire network, identify global user behavior patterns (e.g., a city-wide surge in video services during lunch breaks), and thereby optimize the allocation of global resources.

[0031] It's important to note that after collecting multi-source data, preprocessing can be performed. Preprocessing steps can include cleaning, normalization, feature extraction, and data packaging. Data cleaning removes outliers and missing values, such as discarding incomplete business logs. Normalization maps data of different dimensions to a unified range. Feature extraction utilizes time series analysis methods to extract user behavior patterns, such as using a sliding window to extract the frequency of business requests over the past 10 minutes, thus obtaining user behavior characteristics from the collected data. Data packaging converts the processed data into tensor format for input into a larger model.

[0032] It should be noted that, in the embodiments of this disclosure, the sampling granularity for core network and edge nodes collecting multi-source data can be set. Thus, during resource scheduling, the sampling granularity can be dynamically adjusted by increasing or decreasing the sampling frequency and / or increasing or decreasing the data dimension. This allows core network and edge nodes to capture sudden changes in business during peak periods, or to reduce system overhead during low-load periods, significantly improving the efficiency of the resource scheduling system, reducing costs, and enhancing adaptability.

[0033] S102: Input the multi-source data collected by the edge nodes and the core network for each service into the corresponding large model for prediction, and obtain the first prediction result output by the first large model corresponding to the edge node and the second prediction result output by the second large model corresponding to the core network.

[0034] The first major model is the edge-lightweight large model, which can be a deep learning Transformer model after knowledge distillation or a hybrid expert model. It has a small parameter size, fast inference speed, and is suitable for bandwidth requirements such as real-time prediction (e.g., 1 to 5 minutes in the future).

[0035] The second major model is the core network model, which can be a multimodal Transformer model. It has a large parameter scale, strong computing power, and is suitable for long-term predictions (such as bandwidth requirements for the next 30 minutes to 1 hour).

[0036] In this embodiment of the disclosure, local user behavior data, edge node service traffic, and network status data can be input into a first large model to obtain a first prediction result output by the first large model, namely, short-term service load prediction and local user demand prediction. Network-wide user behavior data, service traffic data, and network status data can be input into a second large model to obtain a second prediction result output by the second large model, namely, global service load prediction and slice resource demand prediction.

[0037] In this embodiment of the disclosure, after the multi-source data collected by the edge nodes is preprocessed and input into the first large model instance for local prediction, some of the preprocessed data from the edge nodes can be uploaded to the second large model for global prediction. Some of the data can be generated through a data sensitivity filtering mechanism, such as high-dimensional feature vectors after edge preprocessing, local prediction results and errors, and global key statistical indicators, so as to improve the global prediction accuracy of the core network while minimizing the network backhaul load.

[0038] To reduce bandwidth consumption during data backhaul, the system does not upload all raw packets collected at the edge. Instead, it uploads standardized features extracted by the preprocessing module, such as user behavior feature vectors and service traffic trend features after normalization and feature extraction. This allows the second model to utilize these features for macro-level modeling across the entire network without processing massive amounts of raw, underlying data. The local prediction results generated by the first model, as well as the actual monitored prediction errors, can also serve as reference data for core network predictions. This data is used for the fusion calculation of core network and edge prediction results and can trigger knowledge distillation or parameter synchronization mechanisms to maintain model consistency. Because the core network needs to maintain awareness of the overall network status, it also needs to aggregate core statistical information from each edge node, such as the total number of access users in the edge region, the average bandwidth utilization of each slice, and the average latency indicators of key services (such as industrial control). In this embodiment, the first model is responsible for local prediction, and the second model is responsible for global optimization. The two models maintain consistency through parameter synchronization and knowledge distillation, ensuring that the models can run correctly and reliably in a distributed environment and avoiding errors or performance degradation caused by inconsistent data or states.

[0039] In this embodiment, in order to deploy large models at edge nodes, knowledge distillation and lightweighting of the large models deployed in the core network are required to ensure consistency between the models deployed at edge nodes and in the core network. The large model in the core network can be used as the teacher model, and the large model at the edge can be used as the student model. The distillation loss function is defined as shown in the following formula (1). Through knowledge distillation, the large model at the edge can significantly reduce computational complexity while maintaining high prediction accuracy.

[0040] (1) In formula (1), It is the cross-entropy loss, used to measure the difference between the student model's prediction and the true label. It is the Kullback-Leibler divergence, used to measure the difference between the student model output distribution and the teacher model output distribution. It is the output distribution of the teacher model. It is the student model output distribution. It is a weighting coefficient used to balance the two parts of the loss.

[0041] In this embodiment of the disclosure, the large model deployed on the edge node can also be compressed and optimized by lightweight processing methods such as parameter pruning to remove redundant parameters, reducing the amount of computation, compressing floating-point parameters into low-bit representation to reduce storage requirements, and activating only some sub-models to improve inference efficiency through a hybrid expert mechanism.

[0042] It should be noted that, in addition to using the core network big model as the teacher model during the training phase to periodically perform knowledge distillation on the edge big model to ensure long-term consistency, the core network big model can also periodically issue lightweight parameter update packages during the runtime phase. This allows the edge big model to maintain short-term consistency through incremental updates. Thus, through a two-layer mechanism of offline distillation and online synchronization, the lightweight and efficient nature of the edge model is guaranteed, while avoiding the computational overhead caused by real-time distillation.

[0043] S103: Based on the attribute information of the business, the first prediction result and the second prediction result are fused to obtain the target prediction result corresponding to the fused business.

[0044] The attribute information of a business can include the business name (i.e., identifier), application scenario, and dynamic characteristics such as volatility intensity.

[0045] In this embodiment, since the prediction results of the core network big model and the edge big model differ, it is necessary to fuse the first prediction result and the second prediction result. The fusion formula can be shown in the following formula (2): (2) In formula (2), This is the first prediction result. This is the second prediction result. It is the fusion weight, which takes a value in the range of 0 to 1.

[0046] In this embodiment, the fusion weights can be adaptively adjusted. When business scenarios change drastically, edge prediction becomes more reliable. As the value increases, the core prediction becomes more reliable when the business scenario is relatively stable. The value decreases. Alternatively, the fusion weights can be calculated using the weight adjustment formula based on the historical prediction errors of the large models in the core network and the large models in the edge nodes. Or, in cold start or new business scenarios, due to the lack of historical prediction errors, a scenario-priority rule and a fixed initial weight initialization strategy can be used to determine the fusion weights, and then, as feedback data accumulates, gradually switch to adaptive weight calculation based on prediction errors.

[0047] For example, in a certain urban area, the number of video users increases rapidly during the lunch break, requiring prediction of the service load for the next 10 minutes to adjust slice resources in advance. Edge nodes collect data showing 500 current video users, an average bandwidth requirement of 2Mbps, and a total requirement of 1000Mbps. The core network collects data showing 50,000 video users across the entire network, an average bandwidth requirement of 2Mbps, and a total requirement of 10Gbps. The data collected by the edge nodes can be input into the first large model for local prediction, resulting in a prediction that the number of users will increase by 50% to 750 in the next 5 minutes, with a predicted bandwidth requirement of 1500Mbps. The data collected by the core network can be input into the second large model for global prediction, resulting in a prediction that the number of users across the entire network will increase by 20% to 60,000 in the next 30 minutes, with a predicted bandwidth requirement of 12Gbps. Based on historical video service prediction data, the standard deviation of the edge prediction error is 0.1, and the standard deviation of the core prediction error is 0.3. Using the weight adjustment formula, the fusion weight is determined to be 0.75. The fused target prediction result is... If the bandwidth requirement is 1425 Mbps, the system predicts that the bandwidth requirement for video services in the next 5 minutes will be 1425 Mbps, and instructs the scheduling module to allocate more bandwidth to video service slices in advance accordingly.

[0048] S104: Based on the target prediction results for all services, the first slice resources at the current moment, and the network state parameters, use a reinforcement learning optimizer to generate a resource scheduling scheme for at least one service.

[0049] The resource scheduling scheme includes a second slice of resources corresponding to each service.

[0050] The slice resources may include, but are not limited to: communication resources (such as bandwidth, spectrum blocks, physical resource blocks (PRBs)), computing resources (such as processor (Central Processing Unit) cycles, virtual machine / container quotas), storage resources, and network service quality (QoS) parameters (such as scheduling priority, latency thresholds), etc. The first slice resource represents the resource allocation of each slice before scheduling, and the second slice resource represents the resource allocation of each slice after scheduling.

[0051] Among them, reinforcement learning optimizers can use algorithms such as Proximal Policy Optimization (PPO) to update network parameters in order to maximize cumulative rewards.

[0052] In this embodiment of the disclosure, an optimal slice resource allocation scheme can be generated in a dynamic network environment based on user demand predicted by a large model and real-time network status.

[0053] In this embodiment of the disclosure, the target prediction result obtained from the large model can be combined with the real-time network state to form a state space as shown in the following formula (3): (3) In formula (3), This represents the target prediction results for all business operations. This represents all currently allocated first slice resources. These represent network status parameters, such as latency, packet loss rate, and bandwidth utilization.

[0054] Then, the reinforcement learning optimizer receives the state space S, determines and outputs the action space A through the optimizer's policy network, and converts the action space A into a scheduling scheme for slice resource allocation. The action space A is shown in the following formula (4): (4) In formula (4), Indicates the first The bandwidth adjustment amount for each slice, N, is the same as the number of services that need to be scheduled in the current scheduling scenario. It is the first The scheduling priority of each slice in the resource scheduling scheme.

[0055] It should be noted that the scheduling of slice resources can simultaneously consider three aspects: resource utilization, user experience quality, and slice switching overhead. This ensures that the generated slice resource allocation scheme can minimize bandwidth waste, guarantee low latency and high reliability for critical services, and avoid signaling overhead and performance jitter caused by frequent switching. Therefore, this disclosure designs a multi-objective optimization objective function. After an action is executed, the system can calculate the reward brought by that action, and this reward value is calculated from the objective function. Through continuous training iterations, the optimizer learns to select actions that maximize the objective function (i.e., the reward), thereby ensuring that the generated resource scheduling scheme is optimal.

[0056] In this embodiment, by collaboratively deploying a distributed large model in the core network and edge nodes, multi-source data such as real-time user behavior characteristics, service traffic patterns, and network status information for each service under multi-service concurrent scenarios are collected and analyzed. Service demands are predicted from both global and local perspectives, and the global and local prediction results are fused together with the service data. Then, a reinforcement learning optimizer is used to dynamically adjust slice resource allocation and priority scheduling. This can improve the overall slice resource utilization while ensuring low latency and high reliability of critical services, significantly enhance the intelligence and personalization of network resource allocation, and optimize user experience.

[0057] The following example, using a scenario of concurrent video and industrial control during peak hours, illustrates the process of generating a resource scheduling scheme. Based on the large model's prediction of multi-source data, the prediction result is that the number of video users will increase by 100% in the next 5 minutes, while the number of industrial control users will remain stable. The real-time network status is that the current video slice bandwidth utilization is 60%, and the industrial control slice latency is 4ms. It can be determined that in the state space S, With increasing demand for video and stable industrial demand, For video slicing, 100Mbps; for industrial slicing, 20Mbps. Assuming a video utilization rate of 60% and an industrial latency of 4ms, the reinforcement learning optimizer makes scheduling decisions based on the state space S. In the action space A, the bandwidth of the video slice is increased by 50Mbps, while the bandwidth of the industrial slice remains unchanged. The scheduling priority of the industrial slice is higher than that of the video slice. Therefore, the generated scheduling scheme includes adjusting the bandwidth of the video slice to 150Mbps, keeping the industrial slice at 20Mbps, and maintaining the highest priority.

[0058] It should be noted that, in some embodiments of this disclosure, when fusing the first prediction result and the second prediction result, the fusion weight can be determined in various ways based on historical prediction errors or business characteristics, so as to ensure the reliability of the fusion of the central prediction result and the edge prediction result and improve the accuracy of the prediction result.

[0059] Figure 2 This is a schematic flowchart of a resource scheduling method proposed in another embodiment of this disclosure.

[0060] like Figure 2 As shown, the resource scheduling method includes: S201: For at least one service in the current scheduling scenario, obtain multi-source data collected by the edge nodes and the core network for each service.

[0061] S202: The multi-source data collected by the edge nodes and the core network for each service is input into the corresponding large model for prediction, and the first prediction result output by the first large model corresponding to the edge node and the second prediction result output by the second large model corresponding to the core network are obtained.

[0062] For a detailed description of steps S201 and S202 above, please refer to other embodiments of this disclosure, which will not be repeated here.

[0063] S203: Based on the business identifier in the attribute information, obtain the first standard deviation of the first largest model's historical prediction of the business, and the second standard deviation of the second largest model's historical prediction of the business.

[0064] In this embodiment of the disclosure, the error generated when the first large model predicts the business with the identified business in the historical prediction process can be determined according to the business identifier. Then, the first standard deviation is calculated from the error of multiple predictions. Similarly, the second standard deviation of the historical prediction of the business by the second large model can be obtained.

[0065] S204: The first weight is calculated based on the first standard deviation and the second standard deviation.

[0066] In this embodiment of the disclosure, the first weight can be calculated by the weight adjustment formula shown in the following formula (5).

[0067] (5) In formula (5), The first standard deviation, The second standard deviation is given. The weight adjustment formula shown in formula (5) ensures that the model with smaller error has a larger weight in the fusion.

[0068] S205: The target prediction result is calculated from the first weight, the first prediction result, and the second prediction result.

[0069] For example, for video services, the first prediction result is 1500Mbps, the second prediction result is 1200Mbps, and historical predictions for video services show that the standard deviation of the edge prediction error is 0.1, and the standard deviation of the core prediction error is 0.3. Using the weight adjustment formula (5), the first weight can be calculated as follows: If the value is 0.75, then the fused target prediction result is... =1425 Mbps.

[0070] S206: Based on the target prediction results for all services, the first slice resources at the current moment, and network state parameters, use a reinforcement learning optimizer to generate a resource scheduling scheme for at least one service.

[0071] For a detailed description of S204 to S206, please refer to the above embodiments, which will not be repeated here.

[0072] In this embodiment, the volatility and uncertainty of the prediction results of the central and edge nodes are quantified by the standard deviation of the historical predictions of the core network and edge nodes. Then, the standard deviation is used to determine the weight of the fusion of the prediction results of the core network and the prediction results of the edge nodes, which can effectively reduce the prediction error and improve the overall prediction accuracy.

[0073] It should be noted that, in this disclosure, the second weight corresponding to the business can also be determined based on the volatility characteristics of the business in the attribute information, and then the target prediction result can be calculated from the second weight, the first prediction result, and the second prediction result.

[0074] In this embodiment of the disclosure, edge prediction results can be prioritized when the business scenario is highly volatile (such as video services), while core network prediction results can be prioritized when the business scenario is stable (such as industrial control). Alternatively, when the volatility of the business cannot be determined, a default fusion weight of 0.5 can be used, that is, edge and core prediction results are fused with equal weight.

[0075] In this embodiment, a highly skewed weighting approach is preferred. This can involve considering only one factor. For example, when business scenarios are highly volatile (such as a sudden large-scale live event), the core network's long-term predictions often underestimate instantaneous peaks due to smoothing effects. In this case, the second weight can be set to 1, meaning that only edge prediction results can respond to demands most quickly. Alternatively, for system stability, a global perspective is usually not completely abandoned. Therefore, a mapping relationship between volatility quantification levels and weights can be defined. The higher the volatility level, the higher the second weight (but it should be less than 1). This mapping relationship can be preset based on experience.

[0076] In this embodiment, by determining the weight of the fusion between the core network prediction results and the edge node prediction results based on the volatility of the service, it is possible to dynamically adapt to changes in the service, more accurately reflect the actual changes in the service, improve prediction accuracy, and further improve the reliability of resource scheduling.

[0077] In this embodiment, the resource scheduling system has a multi-level cold start protection mechanism. First, it attempts to match scenario priority rules with business features to obtain more targeted experience weights. If business features are missing, it retreats to a fixed initial weight strategy for equal weight fusion. Then, based on the real-time error data returned after monitoring and feedback of sliced ​​resources, the system can seamlessly switch to a fast adaptive update mode based on the standard deviation of prediction error, realizing a smooth transition from experience-driven to data-driven.

[0078] In this disclosure, after generating a resource scheduling scheme, the usage of each slice's resources after resource scheduling is completed according to the scheme can be monitored in real time, and the results can be fed back to the large model and optimizer to form a closed-loop optimization. This constructs a closed-loop optimization mechanism that combines fast and slow feedback, which not only corrects the generation strategy of the scheduling scheme, but also enables the self-evolution of the large model.

[0079] Figure 3 This is a schematic flowchart of a resource scheduling method proposed in another embodiment of this disclosure.

[0080] like Figure 3 As shown, the resource scheduling method includes: S301: For at least one service in the current scheduling scenario, obtain multi-source data collected by the edge nodes and the core network for each service.

[0081] S302: Input the multi-source data collected by the edge nodes and the core network for each service into the corresponding large model for prediction, and obtain the first prediction result output by the first large model corresponding to the edge node and the second prediction result output by the second large model corresponding to the core network.

[0082] S303: Based on the attribute information of the business, the first prediction result and the second prediction result are merged to obtain the target prediction result corresponding to the business after fusion.

[0083] S304: Based on the target prediction results of all services, the first slice resources at the current moment, and the network state parameters, use a reinforcement learning optimizer to generate a resource scheduling scheme for at least one service.

[0084] The descriptions of steps S301 to S304 above can be found in the above embodiments, and will not be repeated here.

[0085] S305: Distribute the resource scheduling plan to network devices, monitor the usage of each second slice resource, and determine multiple indicator values ​​associated with the second slice resource.

[0086] The monitoring metrics may include, but are not limited to, bandwidth utilization, latency, packet loss rate, and slice isolation. Bandwidth utilization is the ratio of the actual bandwidth used by a slice to the allocated bandwidth; latency is the end-to-end transmission latency; packet loss rate is the ratio of the number of lost packets to the total number of data packets; and slice isolation is the degree of interference between different slices.

[0087] In this embodiment of the application, an alarm can be triggered when the latency of a certain slice exceeds a threshold, a resource reclamation mechanism can be triggered when the bandwidth utilization rate is lower than a certain threshold (such as 30%), or an emergency resource reallocation can be triggered when the packet loss rate exceeds a certain threshold (such as 5%).

[0088] It should be noted that after obtaining the monitoring results of the second slice of resources rescheduled according to the resource scheduling scheme, this disclosure can directly feed back to the scheduling module containing the optimizer through a feedback mechanism that combines fast and slow feedback, so as to correct the scheduling strategy, and feed back to the large model as training samples for online fine-tuning of model parameters.

[0089] S306: Based on the index values ​​associated with all second slice resources, the resource scheduling scheme, and the preset objective function, calculate the reward value for this resource scheduling by the reinforcement learning optimizer.

[0090] In this embodiment of the disclosure, the preset objective function can be represented by the following formula (6): (6) In formula (6), Represents the overall utility function. To improve the utilization rate of slice resources, For the sake of user experience quality, For slice switching overhead, , and This is a weighting coefficient used to balance the importance of the three factors, and can be set based on experience.

[0091] In this embodiment of the disclosure, The ratio of the actual bandwidth used to the allocated bandwidth for all slices is calculated using the following formula (7): (7) In formula (7), For the first The actual bandwidth used by each slice For the first The bandwidth allocated to each slice, where N is the number of slices. The higher the value, the less resource waste there is.

[0092] In this embodiment of the disclosure, Taking into account latency, packet loss rate, and bandwidth satisfaction, the calculation formula is shown in formula (8) below: (8) In formula (8), This is the actual delay. For the maximum allowable delay, For packet loss rate, To allocate bandwidth in practice Request bandwidth for the user. , and These are the weighting coefficients. A higher value indicates a better user experience.

[0093] In this embodiment of the disclosure, Defined as the additional overhead caused by switching all slices, the calculation formula is shown in the following formula (9): (9) In formula (9), For the first Whether a slice has been switched can be determined by a value of 0 or 1, where 0 indicates no switch and 1 indicates a switch. Additional overhead incurred during handover (such as signaling overhead and latency jitter). The lower the value, the better.

[0094] In this embodiment of the disclosure, whether a slice is switched depends not only on whether the bandwidth increases or decreases, but also on whether the slice's status or configuration changes. Configuration changes may be due to changes in key slice parameters (such as Service Level Agreement (SLA) level, isolation level, or large-scale resource block remapping).

[0095] In this embodiment of the disclosure, the objective function, by introducing a slice switching overhead term, can establish a balance between the sensitivity of dynamic resource adaptation and the stability of the network signaling system.

[0096] In this embodiment of the disclosure, the reward function for calculating the reward value is consistent with the objective function, and the reward function can be expressed as follows: (10) (10) In formula (10), , and This is the reward weight. The higher the value of R, the better the scheduling scheme.

[0097] It should be noted that in formulas (6) and (10), the weights corresponding to the same term are logically consistent, meaning the reward function must be completely consistent with the objective function in direction. If the weight ratios of the two (i.e., the importance ranking of each indicator) are inconsistent, the reinforcement learning optimizer will optimize in a direction that deviates from the business objective. and b and c and All are corresponding relationships. , and It can be against , and The values ​​are normalized to ensure that the fluctuation range of the reward value is within the stable convergence interval of the reinforcement learning algorithm (such as PPO). That is, the reward function weights of the reinforcement learning optimizer and the utility weights of the objective function of the business layer maintain a linear mapping relationship, ensuring that the evolution direction of the optimization decision always conforms to the principle of maximizing the comprehensive benefits of network computing power and communication resources.

[0098] S307: Update the reinforcement learning optimizer based on the reward value.

[0099] Taking the PPO algorithm as an example, this describes the process of updating the reinforcement learning optimizer based on the reward value. First, the optimizer's policy network and value network are initialized. Then, the state at time t before scheduling is sampled in the environment. The action performed at time t Based on the reward value after the action scheduling and the new state after scheduling Then the action can be calculated and measured. The advantage function relative to the average level is shown in the following formula (11). Discount factor: (11) Then, the policy network can be updated using the following formula (12), the value network can be updated using the following formula (13), and iterative training can be performed based on the updated network until convergence, thus completing the update of the reinforcement learning optimizer.

[0100] (12) In formula (12), The ratio between the old and new strategies. Used to limit the update magnitude and ensure training stability.

[0101] (13) In formula (13), the accuracy of value estimation is improved by minimizing the mean square error.

[0102] In this embodiment of the disclosure, in order to ensure the convergence of the reinforcement learning optimizer in a dynamic environment, the following strategies can be adopted: pre-training the optimizer model using historical data in an offline environment, requiring only small-scale online fine-tuning during runtime; storing historical scheduling experience to avoid training instability due to drastic environmental fluctuations; and reverting to a rule-based safe scheduling strategy when the optimizer strategy fails to converge, ensuring that critical business operations are not affected.

[0103] In this embodiment, by monitoring the resource usage of the scheduling scheme and combining it with a multi-objective optimization function, the strategy of the reinforcement learning optimizer in generating subsequent scheduling schemes is corrected, so that the scheduling scheme can take into account resource utilization, user experience and slice switching overhead, significantly improving the stability and reliability of the real-time generated scheduling scheme.

[0104] In some possible embodiments, after determining multiple index values ​​associated with the second slice resource, the sampling frequency of multi-source data of the edge node and the core network for the services corresponding to the second slice resource can be adjusted according to the multiple index values ​​associated with the second slice resource.

[0105] For example, during peak hours, the sampling frequency can be increased to the millisecond level to capture sudden changes in business operations; or, during low-load periods, the sampling frequency can be reduced to the second level to reduce system overhead.

[0106] In this embodiment, by dynamically adjusting the collection granularity based on the monitoring results of resource usage, the system can dynamically adapt to business changes and further improve the accuracy of resource scheduling.

[0107] In some possible embodiments, after determining multiple indicator values ​​associated with the second slice resource, the contribution of each structural unit in the first major model to the business prediction corresponding to the second slice resource can be determined based on the multiple indicator values.

[0108] Then, based on the contribution, the first major model is pruned, and the pruned major model is determined as the target major model for subsequent business prediction.

[0109] In this embodiment of the disclosure, in order to ensure the real-time performance of edge nodes, a model acceleration mechanism is introduced at the edge. Through hardware acceleration, model pruning, and hierarchical inference, the large edge model can complete prediction within a millisecond timescale, meeting the needs of low-latency services such as industrial control.

[0110] In this embodiment of the disclosure, the marginal contribution of each layer of the model to the prediction of different grade slices (such as industrial grade and civilian grade) can be back-calibrated by monitoring the slice performance indicators returned by the feedback. Based on the contribution, the first model is structured and pruned, and only the parameters that contribute the most to the prediction results are retained, thus realizing a differentiated parameter retention strategy.

[0111] In this embodiment, under extremely low latency scenarios, edge nodes can operate only a shallow network to quickly provide prediction results, which are then corrected by the core network's large model. Edge nodes can also dynamically disable feature extraction channels related to inactive slice services based on user access behavior collected by the current base station, thereby achieving millisecond-level inference without retraining the model.

[0112] It should be noted that by associating and storing the resource scheduling scheme with the data obtained from monitoring after scheduling, the stored data can be used as training samples for a large model when a certain amount of data is reached or the interval between the last model training meets a threshold, thus achieving self-evolution.

[0113] In this embodiment, by using the data obtained from monitoring after the scheduling scheme is executed, the large model can be pruned and trained, which can dynamically optimize the accuracy of the prediction results of the large model and make the prediction efficiency higher.

[0114] This disclosure also proposes a resource scheduling system that can execute the resource scheduling method according to the above embodiments. The system can include four modules: a data acquisition and preprocessing module, a distributed large model inference module, an intelligent slice scheduling and optimization module, and a slice resource monitoring and feedback module. The system architecture diagram of the resource scheduling system is shown below. Figure 4 As shown.

[0115] Depend on Figure 4 It is evident that there is a close interactive relationship among the four modules, and the entire invocation process follows a logical chain of data-driven, predictive analysis, intelligent scheduling, and closed-loop feedback. The data acquisition and preprocessing module provides standardized data input, the distributed large-scale model inference module performs predictive analysis on the input data, the intelligent slice scheduling and optimization module generates a scheduling scheme based on the prediction results, and the slice resource monitoring and feedback module monitors the usage of each slice resource after the scheduling scheme is executed, feeding the monitoring results back to the intelligent slice scheduling and optimization module and the distributed large-scale model inference module to correct the scheduling strategy, train the large-scale model, and form a closed-loop optimization. The system can continuously optimize slice resource allocation in a dynamic environment, improving utilization and ensuring user experience.

[0116] Figure 5 This is a schematic diagram of the structure of a resource scheduling method apparatus proposed in an embodiment of this disclosure.

[0117] like Figure 5 As shown, the resource scheduling method apparatus 50 includes: The acquisition module 501 is used to acquire multi-source data collected by the edge node and the core network for each service in the current scheduling scenario, wherein the multi-source data includes user behavior data, service traffic data and network status data. Prediction module 502 is used to input the multi-source data collected by the edge nodes and the core network for each service into the corresponding large model for prediction, and obtain the first prediction result output by the first large model corresponding to the edge node and the second prediction result output by the second large model corresponding to the core network. The fusion module 503 is used to fuse the first prediction result and the second prediction result based on the attribute information of the business to obtain the fused target prediction result corresponding to the business. The scheduling module 504 is used to generate a resource scheduling scheme for at least one service based on the target prediction results of all services, the first slice resources at the current time, and network state parameters using a reinforcement learning optimizer. The resource scheduling scheme includes the second slice resources corresponding to each service.

[0118] In some possible embodiments, the fusion module 503 may specifically be used for: Based on the business identifier in the attribute information, obtain the first standard deviation of the first model's historical prediction of the business, and the second standard deviation of the second model's historical prediction of the business. The first weight is calculated based on the first standard deviation and the second standard deviation; The target prediction result is calculated from the first weight, the first prediction result, and the second prediction result.

[0119] In some possible embodiments, the fusion module 503 may specifically be used for: Based on the volatility characteristics of the business in the attribute information, determine the second weight corresponding to the business; The target prediction result is calculated from the second weight, the first prediction result, and the second prediction result.

[0120] In some possible embodiments, the resource scheduling method apparatus 50 further includes an update module, which can be specifically used for: The resource scheduling plan is distributed to network devices to monitor the usage of each second slice resource and determine multiple indicator values ​​associated with the second slice resource. Based on the index values ​​associated with all second slice resources, the resource scheduling scheme, and the preset objective function, calculate the reward value of the reinforcement learning optimizer for this resource scheduling. Update the reinforcement learning optimizer based on the reward value.

[0121] In some possible embodiments, the resource scheduling method apparatus 50 further includes an update module, which can be specifically used for: Based on the multiple indicator values ​​associated with the second slice resource, adjust the sampling frequency of multi-source data for the services corresponding to the second slice resource by the edge nodes and the core network.

[0122] In some possible embodiments, the resource scheduling method apparatus 50 further includes an update module, which can be specifically used for: Based on multiple indicator values, the contribution of each structural unit in the first major model to the business prediction corresponding to the second slice resource is determined. Based on contribution, the first major model is pruned, and the pruned major model is determined as the target major model for subsequent business prediction.

[0123] It should be noted that the foregoing explanation of the resource scheduling method also applies to the resource scheduling method apparatus of this embodiment, and will not be repeated here.

[0124] In this embodiment, by collaboratively deploying a distributed large model in the core network and edge nodes, multi-source data such as real-time user behavior characteristics, service traffic patterns, and network status information for each service under multi-service concurrent scenarios are collected and analyzed. Service demands are predicted from both global and local perspectives, and the global and local prediction results are fused together with the service data. Then, a reinforcement learning optimizer is used to dynamically adjust slice resource allocation and priority scheduling. This can improve the overall slice resource utilization while ensuring low latency and high reliability of critical services, significantly enhance the intelligence and personalization of network resource allocation, and optimize user experience.

[0125] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0126] Figure 6 A schematic block diagram of an example electronic device 600 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0127] like Figure 6 As shown, device 600 includes a computing unit 601, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 602 or a computer program loaded into random access memory (RAM) 603 from storage unit 608. RAM 603 may also store various programs and data required for the operation of device 600. The computing unit 601, ROM 602, and RAM 603 are interconnected via bus 604. Input / output (I / O) interface 605 is also connected to bus 604.

[0128] Multiple components in device 600 are connected to I / O interface 605, including: input unit 606, such as keyboard, mouse, etc.; output unit 607, such as various types of monitors, speakers, etc.; storage unit 608, such as disk, optical disk, etc.; and communication unit 609, such as network card, modem, wireless transceiver, etc. Communication unit 609 allows device 600 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0129] The computing unit 601 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 601 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 601 performs the various methods and processes described above, such as resource scheduling methods. For example, in some embodiments, the resource scheduling method may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 608. In some embodiments, part or all of the computer program may be loaded and / or installed on device 600 via ROM 602 and / or communication unit 609. When the computer program is loaded into RAM 603 and executed by the computing unit 601, one or more steps of the resource scheduling method described above may be performed. Alternatively, in other embodiments, the computing unit 601 may be configured to perform resource scheduling methods by any other suitable means (e.g., by means of firmware).

[0130] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0131] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0132] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0133] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0134] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), the Internet, and blockchain networks.

[0135] Computer systems can include clients and servers. Clients and servers are generally geographically separated and typically interact via communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. A server can be a cloud server, also known as a cloud computing server or cloud host, a hosting product within the cloud computing service system that addresses the shortcomings of traditional physical hosts and VPS (Virtual Private Server) services, such as high management difficulty and weak business scalability. Servers can also be servers for distributed systems or servers incorporating blockchain technology.

[0136] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0137] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this disclosure, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified. In the description of this disclosure, the words "if" and "suppose" as used may be interpreted as "when," "when," "in response to determination," or "in the circumstances."

[0138] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A resource scheduling method, characterized in that, include: For at least one service in the current scheduling scenario, acquire multi-source data collected by edge nodes and the core network for each service, wherein the multi-source data includes user behavior data, service traffic data and network status data; The multi-source data collected by the edge nodes and the core network for each service is input into the corresponding large model for prediction, to obtain the first prediction result output by the first large model corresponding to the edge node and the second prediction result output by the second large model corresponding to the core network. Based on the attribute information of the service, the first prediction result and the second prediction result are fused to obtain the target prediction result corresponding to the fused service; Based on the target prediction results for all services, the first slice resources at the current moment, and the network state parameters, a reinforcement learning optimizer is used to generate a resource scheduling scheme for at least one service, wherein the resource scheduling scheme includes the second slice resources corresponding to each service.

2. The method as described in claim 1, characterized in that, The step of fusing the first prediction result and the second prediction result based on the attribute information of the service to obtain the fused target prediction result corresponding to the service includes: Based on the identifier of the business in the attribute information, obtain the first standard deviation of the first largest model's historical prediction of the business, and the second standard deviation of the second largest model's historical prediction of the business; The first weight is calculated based on the first standard deviation and the second standard deviation; The target prediction result is calculated from the first weight, the first prediction result, and the second prediction result.

3. The method as described in claim 2, characterized in that, The step of fusing the first prediction result and the second prediction result based on the attribute information of the service to obtain the fused target prediction result corresponding to the service includes: Based on the volatility characteristics of the service in the attribute information, determine the second weight corresponding to the service; The target prediction result is calculated from the second weight, the first prediction result, and the second prediction result.

4. The method according to any one of claims 1-3, characterized in that, After determining the resource scheduling scheme for the at least one service using a reinforcement learning optimizer, the method further includes: The resource scheduling scheme is distributed to network devices to monitor the usage of each second slice resource and determine multiple indicator values ​​associated with the second slice resource. Based on the index values ​​associated with all second slice resources, the resource scheduling scheme, and the preset objective function, the reward value of the reinforcement learning optimizer for this resource scheduling is calculated. The reinforcement learning optimizer is updated based on the reward value.

5. The method as described in claim 4, characterized in that, After determining the multiple index values ​​associated with the second slice resource, the method further includes: Based on multiple indicator values ​​associated with the second slice resource, the sampling frequency of multi-source data for the services corresponding to the second slice resource by the edge node and the core network is adjusted.

6. The method as described in claim 4, characterized in that, After determining the multiple index values ​​associated with the second slice resource, the method further includes: Based on the multiple indicator values, the contribution of each structural unit in the first large model to the business prediction corresponding to the second slice resource is determined. Based on the contribution, the first large model is pruned, and the pruned large model is determined as the target large model for subsequent prediction of the business.

7. A resource scheduling method apparatus, characterized in that, include: The data acquisition module is used to acquire multi-source data collected by the edge nodes and the core network for each service in the current scheduling scenario, wherein the multi-source data includes user behavior data, service traffic data and network status data. The prediction module is used to input the multi-source data collected by the edge nodes and the core network for each service into the corresponding large model for prediction, and obtain the first prediction result output by the first large model corresponding to the edge node and the second prediction result output by the second large model corresponding to the core network. The fusion module is used to fuse the first prediction result and the second prediction result based on the attribute information of the service to obtain the fused target prediction result corresponding to the service; The scheduling module is used to generate a resource scheduling scheme for at least one service based on the target prediction results corresponding to all services, the first slice resources at the current time, and network state parameters using a reinforcement learning optimizer, wherein the resource scheduling scheme includes the second slice resources corresponding to each service.

8. An electronic device, comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the resource scheduling method according to any one of claims 1-6.

9. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, in, The computer instructions are used to cause the computer to execute the resource scheduling method according to any one of claims 1-6.

10. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the steps of the resource scheduling method according to any one of claims 1-6.

Images