Communication network resource intelligent scheduling method, device, equipment and medium

Abstract

The invention relates to an intelligent scheduling method and device for communication network resources, equipment and a medium. The method comprises the following steps: performing IP packet header analysis on a service data packet to obtain an initial service priority identifier, and performing adjustment in combination with a preset priority rule, an equipment alarm state and a network congestion rate to generate a dynamic service label; based on the dynamic service label, the real-time link state and the energy consumption coefficient, generating a resource allocation strategy and a path scoring weight parameter through a preset deep reinforcement learning network; and calculating a path score based on the resource allocation strategy and the path score weight parameter, generating a path mapping table, generating a port suppression instruction based on the dynamic service label and the resource allocation strategy, and extracting a base station power regulation instruction. According to the method, the dynamic adaptability, the resource utilization rate and the energy consumption management and control capability of communication network resource scheduling are improved by means of dynamically adjusting the service priority, comprehensively modeling a resource allocation strategy and the like, and the transmission guarantee level of high-priority services is enhanced.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, and in particular relates to intelligent scheduling methods, devices, equipment and media for communication network resources. Background Technology

[0002] With the deep integration of 5G communication technology and the Internet of Things (IoT), the types of services in communication networks are becoming increasingly diversified. Different services, such as high-definition video, industrial control, and real-time interaction, have significantly different requirements for network bandwidth, latency, and reliability. Meanwhile, energy consumption management and resource utilization improvement of network equipment such as base stations have become hot topics in the industry. Existing communication network resource scheduling methods mostly adopt a static rule-based allocation model. For example, fixed priorities are pre-set for different service types, and network resources such as time slots and bandwidth are allocated according to priority. While this method is simple to implement, it cannot dynamically adapt to real-time changes in network operating status. For instance, when encountering sudden situations such as equipment alarms or partial network congestion, static priority rules cannot quickly adjust resource allocation strategies, easily leading to the obstruction of high-priority service transmission, while the ineffective occupation of low-priority services also results in resource waste.

[0003] To improve dynamic adaptability, some technologies attempt to introduce priority adjustment mechanisms, but these often rely solely on single indicators such as network congestion rate, failing to incorporate equipment alarm status to form a multi-dimensional decision-making basis. This results in insufficient accuracy in determining dynamic service priorities. Furthermore, in the resource scheduling strategy generation stage, existing communication network resource scheduling technologies often consider link status or service requirements in isolation, lacking comprehensive modeling of service priorities, real-time link quality, and base station energy consumption. This makes it difficult to achieve coordinated management and control of resource utilization and energy consumption optimization. In addition, existing communication network resource scheduling methods often use fixed weights to calculate path scores during path selection, failing to dynamically adjust weight parameters according to service priorities. They also lack dedicated resource protection mechanisms for high-priority services, such as not using port suppression to reduce the resource consumption of low-priority services. Simultaneously, base station power adjustment often relies on manually preset thresholds, failing to link with service scheduling strategies. This results in base station energy consumption not dynamically adapting to service traffic, further exacerbating resource waste and excessive energy consumption. Although some studies in recent years have attempted to introduce artificial intelligence technologies such as reinforcement learning for resource scheduling, existing solutions have not fully explored the correlation between service time-series characteristics and network status, resulting in insufficient targeting of strategy generation. Furthermore, they have not effectively integrated key factors such as base station energy consumption coefficients and dynamic service labels into model inputs, making it difficult to output scheduling strategies that take into account multiple objectives. This fails to meet the current communication network's requirements for high precision, low energy consumption, and high reliability in intelligent scheduling. Summary of the Invention

[0004] Therefore, it is necessary to provide intelligent scheduling methods, devices, equipment and media for communication network resources to address the above-mentioned technical problems, aiming to improve the dynamic adaptability, accuracy and resource utilization of communication network resource scheduling, and enhance the energy consumption control capabilities of communication networks.

[0005] Firstly, this application provides a method for intelligent scheduling of communication network resources, including:

[0006] The input service data packet is parsed by the IP header service type field to obtain the initial service priority identifier. Based on the preset priority rules, device alarm status and network congestion rate, the initial service priority identifier is adjusted to generate dynamic service tags.

[0007] The system acquires real-time link status and energy consumption coefficients. Based on dynamic service tags, real-time link status, and energy consumption coefficients, it processes these data through a pre-defined deep reinforcement learning network to generate resource allocation strategies and path scoring weight parameters.

[0008] The system processes the path scores based on resource allocation strategies and path score weight parameters, calculates path scores, sorts and filters the path scores, generates a path mapping table, and analyzes the dynamic service tags and resource allocation strategies to generate port suppression instructions. It also extracts base station power adjustment instructions from the resource allocation strategies, which are used to perform dynamic adjustment of base station power.

[0009] In one embodiment, the input service data packet is parsed using the IP header service type field to obtain an initial service priority identifier. Based on preset priority rules, device alarm status, and network congestion rate, the initial service priority identifier is adjusted to generate a dynamic service tag, including:

[0010] The first three bits of the Service Type field in the IP header of the service data packet are parsed to obtain the initial service priority identifier;

[0011] Parse the source device information in the IP header of the service data packet to obtain the source device identifier;

[0012] Real-time monitoring of device alarm status and network congestion rate; adjustment of the initial service priority identifier based on preset priority rules to obtain the adjusted priority identifier; wherein the adjusted priority identifier is obtained through the following steps:

[0013] When the device alarm status is detected as an emergency state or the network congestion rate exceeds the preset congestion threshold, the initial service priority identifier is upgraded to the highest priority identifier, resulting in the adjusted priority identifier.

[0014] When no device alarm status is detected as an emergency state and the network congestion rate does not exceed the preset congestion threshold, the initial service priority identifier remains unchanged, and the initial service priority identifier is output as the adjusted priority identifier.

[0015] The system is formatted based on the current system time to generate a timestamp. The adjusted priority identifier, timestamp, and source device identifier are then integrated to generate a dynamic business tag.

[0016] In one embodiment, real-time link status and energy consumption coefficient are obtained. Based on dynamic service tags, real-time link status, and energy consumption coefficient, the data are processed through a preset deep reinforcement learning network to generate resource allocation strategies and path scoring weight parameters, including:

[0017] The network link is detected by a link quality probe to obtain the real-time link status, which includes bandwidth, latency and packet loss rate.

[0018] Obtain the base station load, input the base station load into a preset base station nonlinear energy consumption model for calculation, and output the energy consumption coefficient.

[0019] Dynamic business tags are classified and statistically analyzed according to preset priorities to generate a four-level business traffic matrix;

[0020] The fourth-level business traffic matrix, real-time link status, and energy consumption coefficient are vector-integrated to generate a status input vector.

[0021] The state input vector is processed by a pre-set deep reinforcement learning network to output a four-dimensional action. The pre-set deep reinforcement learning network is a joint network built on a long short-term memory network and a deep deterministic policy gradient algorithm. The four-dimensional action includes four independent control parameters, namely the proportion of dedicated time slots for the first-level service, the path scoring weight parameters shared by the second and third-level services, the load balancing factor for the fourth-level service, and the base station power adjustment command.

[0022] The four-dimensional actions are encapsulated to generate resource allocation strategies, and the path scoring weight parameters shared by the second and third level businesses are extracted from the four-dimensional actions.

[0023] In one embodiment, the process involves processing based on resource allocation strategies and path score weight parameters, calculating path scores, sorting and filtering the path scores, and generating a path mapping table, including:

[0024] Extract bandwidth weight coefficient, latency weight coefficient, and packet loss rate weight coefficient from the path scoring weight parameters;

[0025] Obtain a set of preset available paths, parse each preset available path, and extract the next-hop address and transmission method of each preset available path;

[0026] Link quality is tested for each preset available path to obtain the bandwidth utilization, transmission latency and packet loss rate of each preset available path;

[0027] Based on the bandwidth weighting coefficient, latency weighting coefficient, and packet loss rate weighting coefficient, the bandwidth utilization, transmission latency, and packet loss rate of each preset available path are weighted and calculated to obtain the path score of each preset available path.

[0028] The path scores of each preset available path are sorted from high to low, and the top preset number of available paths are selected from the sorting results to obtain the selected paths.

[0029] Assign corresponding path identifiers to the selected paths, establish a mapping relationship between each path identifier and its corresponding next-hop address and transmission mode, and generate a path mapping table.

[0030] In one embodiment, analysis is performed based on dynamic service tags and resource allocation strategies to generate port suppression instructions, including:

[0031] The priority level of dynamic service tags is identified, and the scope of preset low-priority services is defined by combining the priority division rules corresponding to the four-level service traffic matrix, and the port identifier association rules corresponding to the preset low-priority services are obtained; the preset low-priority services are the fourth-level services in the four-level service traffic matrix.

[0032] The dynamic business tags are parsed and processed in real time to detect the priority level in the dynamic business tags and obtain the detection results.

[0033] When the detection result indicates that the highest priority tag exists in the dynamic business tags, the fourth-level business load balancing factor is extracted from the resource allocation strategy.

[0034] The port suppression duration is calculated based on the fourth-level service load balancing factor and the preset base duration.

[0035] Based on the port identifier association rules, the target port identifier corresponding to the preset low-priority service is determined, and the target port identifier is bound to the port suppression duration to generate a port suppression command.

[0036] In one embodiment, the preset deep reinforcement learning network is trained through the following steps:

[0037] Obtain historical network status datasets, which include historical service label sequences, historical link status, and historical base station energy consumption data.

[0038] Construct a simulated network environment model, which integrates a preset base station nonlinear energy consumption model, historical network topology, and service transmission rules.

[0039] Temporal feature extraction processing of historical network state datasets is performed using a long short-term memory network to obtain time slot association feature vectors;

[0040] Construct a reward function with the goal of maximizing network resource utilization and minimizing base station energy consumption;

[0041] The time-slot-related feature vectors are processed by a deep deterministic policy gradient algorithm to make policy decisions and output initial action instructions.

[0042] The initial action command is input into the simulated network environment model to conduct service scheduling simulation test, and the network resource utilization, actual energy consumption of the base station and base station load during the test are obtained after the test.

[0043] The base station load during the test is input into a preset base station nonlinear energy consumption model for calculation to obtain the predicted energy consumption value of the base station.

[0044] The reward value of the reward function is calculated based on the network resource utilization rate and actual energy consumption of the base station after testing. The network parameters of the joint network are adjusted according to the reward value. The deviation between the actual energy consumption of the base station and the predicted energy consumption of the base station is calculated. The deviation is used as a negative feedback term to update the network weight of the joint network and generate the optimized joint network.

[0045] The optimized joint network is used as the network basis for a new round of training. Iterative parameter optimization is performed until the reward value of the reward function meets the preset reward threshold, thus obtaining the preset deep reinforcement learning network after training.

[0046] In one embodiment, the port suppression duration is calculated using the following formula:

[0047]

[0048] in, Port suppression duration, To preset the base duration, For load balancing factor weights, This is the fourth-level business load balancing factor. This is the maximum transmit power of the base station. This represents the current transmit power of the base station. This is a correction factor for the number of high-priority services. The number of highest priority service tags in real time. This represents the total number of real-time business tags. This represents the number of business priority levels.

[0049] Secondly, this application also provides an intelligent scheduling device for communication network resources, including:

[0050] The service priority dynamic adjustment module is used to parse the IP header service type field of the input service data packet to obtain the initial service priority identifier. Based on the preset priority rules, device alarm status and network congestion rate, the initial service priority identifier is adjusted to generate dynamic service tags.

[0051] The resource allocation and strategy generation module is used to obtain real-time link status and energy consumption coefficient. Based on dynamic service tags, real-time link status and energy consumption coefficient, it is processed through a preset deep reinforcement learning network to generate resource allocation strategies and path score weight parameters.

[0052] The path mapping and instruction generation module is used to process the path score based on the resource allocation strategy and path score weight parameters, calculate the path score, sort and filter the path scores, generate a path mapping table, and analyze the dynamic service tags and resource allocation strategy to generate port suppression instructions. It also extracts the base station power adjustment instructions from the resource allocation strategy, which are used to perform dynamic adjustment of the base station power.

[0053] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the first aspect.

[0054] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the first aspect.

[0055] The aforementioned intelligent scheduling method, device, equipment, and medium for communication network resources firstly parses input service data packets to generate initial priority identifiers. Combined with preset priority rules, device alarm status, and network congestion rate adjustments, this solves the problem of traditional static priority failing to adapt to sudden network conditions, improving the dynamic adaptability and accuracy of service priority determination. Secondly, based on dynamic service tags, real-time link status, and energy consumption coefficients, a resource allocation strategy and path scoring weight parameters are generated through preset deep reinforcement learning networks. This achieves comprehensive modeling of multi-dimensional network states, improving the global coordination and scientific decision-making of resource scheduling strategies. Furthermore, path scores are calculated based on resource allocation strategies and path scoring weight parameters, and a path mapping table is generated, improving the rationality of path selection and the reliability of service transmission. Finally, port suppression instructions are generated by combining dynamic service tags, and base station power adjustment instructions are extracted to execute dynamic power adjustments. This not only ensures dedicated resources for high-priority services but also solves the problem of excessive energy consumption from static base station power control, improving resource utilization and energy consumption management. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 A flowchart of an intelligent scheduling method for communication network resources provided as an exemplary embodiment of the present invention;

[0058] Figure 2 A flowchart of a method for generating port suppression instructions is provided as an exemplary embodiment of the present invention;

[0059] Figure 3 A schematic diagram of the structure of an intelligent scheduling device for communication network resources provided as an exemplary embodiment of the present invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0061] In one embodiment, such as Figure 1 As shown, an intelligent scheduling method for communication network resources is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and is implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0062] S101: Parse the IP header service type field of the input service data packet to obtain the initial service priority identifier. Based on the preset priority rules, device alarm status and network congestion rate, adjust the initial service priority identifier to generate a dynamic service label.

[0063] Specifically, the process of parsing the IP header Type of Service (TOS) field of the input service data packets can be based on the Type of Service (TOS) or Differential Code Point of Service (DSCP) fields defined in the IP protocol. These fields reflect the priority information of the service data packets. However, relying solely on the initial priority identifier cannot meet the resource scheduling requirements in a dynamic network environment. Therefore, the initial service priority identifier can also be adjusted based on preset priority rules, device alarm status, and network congestion rate. The preset priority rules are set according to the default priority of service types (such as high-definition video, industrial control, real-time interaction, etc.), but these rules need to be dynamically adjusted according to the real-time status in the actual network environment. For example, when a device in the network experiences an alarm, it may affect the transmission quality of some services. Therefore, the priority of related services can be reduced to avoid resource waste and ensure the transmission of other normal services. At the same time, the network congestion rate is also an important adjustment basis. When local network congestion occurs, the transmission of high-priority services may be affected. Therefore, service priorities can be dynamically adjusted according to the degree of congestion to ensure that critical services can receive priority resource allocation. In this way, the generated dynamic service tags can more accurately reflect the actual priority of the service in the current network environment, providing a more reliable basis for subsequent resource allocation and path selection.

[0064] S102: Obtain real-time link status and energy consumption coefficient. Based on dynamic service labels, real-time link status and energy consumption coefficient, process them through a preset deep reinforcement learning network to generate resource allocation strategies and path score weight parameters.

[0065] Specifically, real-time link status can include key parameters such as link bandwidth utilization, latency, and packet loss rate, which reflect the current transmission performance of the network link. Energy consumption coefficient is related to the energy consumption level of network devices (such as base stations), reflecting the energy consumption of the devices under different operating conditions. By comprehensively considering dynamic service labels, real-time link status, and energy consumption coefficient, the allocation requirements of network resources can be more comprehensively evaluated. Based on this information, a pre-set deep reinforcement learning network can be used to process the data and generate resource allocation strategies and path scoring weight parameters. Deep reinforcement learning is a machine learning model that can learn the optimal resource allocation strategy based on multi-dimensional input data such as service priority, link status, and energy consumption coefficient. Through continuous interaction with the environment, the network can adjust its strategy based on feedback reward signals (such as resource utilization, energy consumption level, and service transmission quality), thereby achieving dynamic optimization and outputting resource allocation strategies and path scoring weight parameters. The resource allocation strategy can explicitly specify the resource allocation ratio for different services on different links and devices, while the path scoring weight parameters can be used for quantitative evaluation of subsequent path selection.

[0066] S103: Based on the resource allocation strategy and path score weight parameters, the path score is calculated, sorted and filtered, a path mapping table is generated, and based on the dynamic service label and resource allocation strategy, a port suppression instruction is generated. The base station power adjustment instruction is extracted from the resource allocation strategy and used to perform dynamic adjustment of the base station power.

[0067] Specifically, a score can be calculated for each path based on link state parameters such as bandwidth, latency, and packet loss rate, combined with service priority and energy consumption coefficient. A higher path score indicates that the path is more suitable for transmitting a specific service under the current network conditions. The path scores can then be sorted and filtered to generate a path mapping table. This table records the priority ranking of different services on different paths, providing a clear basis for path selection in actual service transmission. This process ensures that high-priority services preferentially select the optimal path for transmission, thereby improving the reliability and efficiency of service transmission.

[0068] Furthermore, analysis based on dynamic service tags and resource allocation strategies can generate port suppression instructions. By suppressing the transmission rate of low-priority services or temporarily interrupting their transmission, the network resource consumption of low-priority services can be reduced, freeing up more resources for high-priority services. In addition, base station power adjustment instructions can be extracted from the resource allocation strategy. These instructions are used to dynamically adjust base station power. For example, when service traffic is low, the base station power can be appropriately reduced to save energy, while when service traffic is high, the base station power can be appropriately increased to meet service transmission demands. In this way, base station power can be dynamically adjusted according to actual service needs, avoiding the problem of base station energy consumption being disconnected from service traffic in traditional methods, further improving resource utilization and energy consumption control capabilities.

[0069] The above method firstly achieves accurate determination and dynamic adaptation of service priorities by parsing the IP headers of service data packets and dynamically adjusting their priorities, thus enhancing scheduling flexibility. Secondly, based on dynamic service tags and real-time network status, it utilizes deep reinforcement learning to generate resource allocation strategies and path scoring weights, overcoming the shortcomings of traditional methods that isolate link or service requirements in resource allocation, thereby improving the comprehensiveness and accuracy of resource scheduling. Finally, by using path scoring sorting and filtering and generating port suppression instructions, it solves the problem of low-priority services ineffectively occupying resources, enhancing the guarantee capability of high-priority services. Furthermore, it can dynamically adjust base station power in conjunction with base station power adjustment instructions, avoiding the problem of base station energy consumption being disconnected from service traffic in traditional methods, thus improving resource utilization and energy consumption management capabilities.

[0070] In one embodiment, the IP header service type field of the input service data packet is parsed to obtain an initial service priority identifier. Based on preset priority rules, device alarm status, and network congestion rate, the initial service priority identifier is adjusted to generate a dynamic service tag, including:

[0071] The first three bits of the Service Type field in the IP header of the service data packet are parsed to obtain the initial service priority identifier;

[0072] Parse the source device information in the IP header of the service data packet to obtain the source device identifier;

[0073] Real-time monitoring of device alarm status and network congestion rate; adjustment of the initial service priority identifier based on preset priority rules to obtain the adjusted priority identifier; wherein the adjusted priority identifier is obtained through the following steps:

[0074] When the device alarm status is detected as an emergency state or the network congestion rate exceeds the preset congestion threshold, the initial service priority identifier is upgraded to the highest priority identifier, resulting in the adjusted priority identifier.

[0075] When no device alarm status is detected as an emergency state and the network congestion rate does not exceed the preset congestion threshold, the initial service priority identifier remains unchanged, and the initial service priority identifier is output as the adjusted priority identifier.

[0076] The system is formatted based on the current system time to generate a timestamp. The adjusted priority identifier, timestamp, and source device identifier are then integrated to generate a dynamic business tag.

[0077] Specifically, based on the standardized definition of the Type of Service (TOS) field in the IP packet header of the TCP / IP protocol stack, the first three bits of this field are priority identifiers, specifically used to carry the basic service quality level requirements of the service. Its value ranges from 0 to 7 (binary 000 to 111), with different values ​​corresponding to different default transmission priorities. Therefore, this field can be extracted using a protocol parsing module deployed on the network access gateway. This module's parsing process strictly follows the IP protocol format specified in RFC 791. It first locates the start position of the IP packet header using the frame synchronization module, then locates the Type of Service field using the field offset, extracts its first three bits, and obtains the initial service priority identifier through binary-to-decimal logical operations. For example, when the extracted bit is 110, the corresponding initial priority identifier is 6. This parsing method can directly obtain the service's native priority requirements, providing a basis for subsequent adjustments and avoiding priority misjudgments due to parsing errors.

[0078] Subsequently, source device information can be parsed from the IP header of the service data packets. The underlying principle is that the source IP address or source MAC address field in the IP header possesses unique device identification characteristics. Since MAC addresses have globally unique hardware encoding attributes and can be directly associated with specific terminal devices, this embodiment uses the source MAC address as the basis for extracting the source device identifier. For example, after completing priority bit parsing, the protocol parsing module continues to locate the source MAC address field using the field offset. After extracting the 48-bit binary MAC address data, it generates a standard format source device identifier through hexadecimal conversion and colon-separated formatting, for example, converting 001B44113A5B to 00:1B:44:11:3A:5B. During this process, the module can also incorporate MAC address validity verification logic. By verifying the validity of the first three vendor identifier segments, it eliminates abnormal data packets corresponding to invalid MAC addresses, ensuring the reliability of the source device identifier.

[0079] Specifically, the alarm status of devices can be monitored through an alarm collection module deployed in the network management system. This module uses Simple Network Management Protocol (SNMP) to establish communication connections with various network devices (such as base stations, switches, and routers), periodically collecting alarm status information. Alarm status is divided into three levels: emergency, important, and general. Emergency status indicates a core function failure (such as a base station RF unit failure), important status indicates a non-core function failure (such as port rate degradation), and general status indicates a warning message (such as slightly high device temperature). After collection, the alarm level can be converted into a 0-2 numerical identifier through status coding conversion, where 0 represents emergency, 1 represents important, and 2 represents general. Network congestion rate monitoring can be achieved through a network traffic analysis module. This module collects link-layer data packets through mirrored ports and can calculate the congestion rate based on a sliding time window algorithm. The calculation formula is as follows:

[0080]

[0081] Where R is the network congestion rate. This represents the number of data packets dropped within the sliding time window. This represents the total number of received data packets within the sliding time window. The length of the sliding time window can be configured according to service latency requirements. For example, a 100ms window can be configured for real-time interactive services. Within the window, the number of dropped data packets is counted by checking the sequence number continuity of the data packets and verifying the validity of the TTL field. Total number of received data packets The congestion threshold is directly calculated using a counter. Preset congestion thresholds can be set based on the packet loss tolerance of different services; for example, a 10% threshold can be set for industrial control services, and a 20% threshold for high-definition video services.

[0082] Specifically, the preset priority rules are stored in the rule engine of the scheduling controller. The core logic is a dual-trigger adjustment mechanism based on alarm status and congestion rate. When a device alarm status of emergency (numerical identifier 0) or a network congestion rate R exceeds the preset congestion threshold is detected, the rule engine can trigger a priority upgrade process, directly setting the initial service priority identifier to 7 (highest priority), thus obtaining the adjusted priority identifier. When no device alarm status of emergency (numerical identifier 1 or 2) is detected and the network congestion rate R does not exceed the preset congestion threshold, the rule engine can trigger a priority maintenance process, directly outputting the initial service priority identifier as the adjusted priority identifier.

[0083] Furthermore, the system can be formatted based on the current system time to generate a timestamp. This process integrates the adjusted priority identifier, timestamp, and source device identifier, adding timing attributes to the tag to generate a dynamic service tag. This facilitates differentiation of the service reception order during subsequent scheduling, avoiding scheduling conflicts caused by inconsistent service arrival times. The timestamp can be generated using the clock module built into the scheduling controller, which employs a GPS-synchronized clock. The integration process is achieved through string concatenation, with the concatenation order being "source device identifier | adjusted priority identifier | timestamp," using a vertical bar as a separator to generate the final dynamic service tag. In this tag structure, the choice of separator avoids special characters in the device identifier and timestamp, ensuring that subsequent parsing allows for quick extraction of each field's information through simple string splitting. This provides complete information, including service source, real-time priority, and reception timing, to support the generation of subsequent resource allocation strategies.

[0084] In one embodiment, real-time link status and energy consumption coefficients are obtained, and based on dynamic service tags, real-time link status, and energy consumption coefficients, they are processed through a preset deep reinforcement learning network to generate resource allocation strategies and path scoring weight parameters, including:

[0085] The network link is detected by a link quality probe to obtain the real-time link status, which includes bandwidth, latency and packet loss rate.

[0086] Obtain the base station load, input the base station load into a preset base station nonlinear energy consumption model for calculation, and output the energy consumption coefficient.

[0087] Dynamic business tags are classified and statistically analyzed according to preset priorities to generate a four-level business traffic matrix;

[0088] The fourth-level business traffic matrix, real-time link status, and energy consumption coefficient are vector-integrated to generate a status input vector.

[0089] The state input vector is processed by a pre-set deep reinforcement learning network to output a four-dimensional action. The pre-set deep reinforcement learning network is a joint network built on a long short-term memory network and a deep deterministic policy gradient algorithm. The four-dimensional action includes four independent control parameters, namely the proportion of dedicated time slots for the first-level service, the path scoring weight parameters shared by the second and third-level services, the load balancing factor for the fourth-level service, and the base station power adjustment command.

[0090] The four-dimensional actions are encapsulated to generate resource allocation strategies, and the path scoring weight parameters shared by the second and third level businesses are extracted from the four-dimensional actions.

[0091] Specifically, link quality probes can employ an embedded hardware architecture, deployed at key link nodes such as core network gateways, aggregation layer switches, and base station access nodes. Each probe is directly connected to the link interface via fiber optic cable. Real-time link status parameters can then be acquired using a differentiated acquisition mechanism. Bandwidth parameters can be obtained by periodically reading the ifInOctets and ifOutOctets fields in the link interface's MIB library using Simple Network Management Protocol (SNMP), calculating the number of bytes transmitted per unit time, and comparing this to the link's rated bandwidth to obtain real-time bandwidth utilization. Latency parameters can be obtained using the Internet Control Message Protocol (ICMP) echo request mechanism. The probe sends fixed-size echo request packets to the peer node, recording the difference between the request sending time and the response receiving time. The average value of multiple continuously collected data packets is taken as the link latency. Packet loss rate parameters can be implemented using a data packet sequence number marking method. The probe appends consecutive sequence numbers when sending detection data packets. Upon receiving the packets, the peer node sends back sequence number confirmation information. The packet loss rate is calculated by statistically analyzing the ratio of the number of packets with missing sequence numbers to the total number of packets sent. During the data collection process, each probe can synchronize its time using the Network Time Protocol (NTP) to ensure consistency of the time base for latency calculation. The collection cycle can be dynamically adjusted according to the business type. After outliers are removed by the built-in filtering module of the probe, the collected data can be uploaded to the scheduling center via the Transmission Control Protocol (TCP).

[0092] Specifically, base station load can be obtained through a resource management unit deployed on the base station side. This unit communicates with the base station baseband processing unit (BBU) via an internal bus to collect core load indicators of the base station in real time, including the number of current access users, downlink time slot resource utilization, number of active carriers, and uplink interference level. The collected data is normalized to form a comprehensive load value L in the 0-1 range. The preset base station nonlinear energy consumption model is constructed based on the base station hardware characteristics and uses a quadratic polynomial to fit the nonlinear relationship between base station load and actual power consumption. The model expression is:

[0093]

[0094] Where P represents the real-time power consumption of the base station, L represents the comprehensive load value of the base station, and α, β, and γ are inherent coefficients of the base station hardware. α is a non-linear power consumption coefficient, reflecting the accelerated change in power consumption as the load increases; β is a linear power consumption coefficient, reflecting the linear correlation between load and power consumption; and γ is a static power consumption coefficient, representing the base station's base power consumption when unloaded. These coefficients can be determined through base station factory testing. During testing, multiple sets of load-power consumption data are generated by adjusting the simulated load, and the least squares method is used for fitting calculations. The energy consumption coefficient is the ratio of the base station's real-time power consumption to its rated full-load power consumption, ranging from 0 to 1. A larger value indicates lower current energy efficiency of the base station, requiring resource scheduling to reduce the load and optimize energy consumption.

[0095] Specifically, dynamic service tags can be categorized and statistically analyzed according to preset priorities to generate a four-level service traffic matrix. This transforms discrete service data into structured traffic features, providing a macroscopic representation of service requirements for deep learning models. The preset priority rules map priority identifiers in dynamic service tags to four levels: highest priority, second-highest priority, medium priority, and lowest priority. These mappings are stored in the classification rule base of the scheduling center. Subsequently, the traffic analysis unit of the scheduling center performs classification and statistics using a sliding time window mechanism with a window length of 1 second. All dynamic service tags received within the window are prioritized, and the number of data packets and total bytes for each priority level are counted to generate the four-level service traffic matrix. This matrix is ​​a 4-row, 2-column two-dimensional data structure. Each row corresponds to a priority level, the first column shows the number of data packets for that level, and the second column shows the total number of bytes for that level. The data in the matrix is ​​updated in real time to ensure it reflects the latest service traffic distribution characteristics.

[0096] Specifically, based on the feature engineering units of the scheduling center, a combination of standardized processing and dimensional concatenation can be used to vectorize the four-level service traffic matrix, real-time link status, and energy consumption coefficient. For example, firstly, the number of data packets and bytes in the four-level service traffic matrix are linearly normalized, mapping the values ​​to the 0-1 interval, resulting in an 8-dimensional normalized vector (4 levels × 2 indicators). The bandwidth utilization, latency, and packet loss rate in the real-time link status are normalized, with latency normalized by dividing by a preset maximum allowable latency, resulting in a 3-dimensional normalized vector. The energy consumption coefficient itself is already a 0-1 value, so it can be directly used as a 1-dimensional vector. Then, the 8-dimensional, 3-dimensional, and 1-dimensional vectors are concatenated in the order of service traffic-link status-energy consumption to generate a 12-dimensional state input vector. The order and scale of the data in each dimension of the vector are calibrated through offline testing to ensure input consistency during model training and inference.

[0097] Specifically, the pre-defined deep reinforcement learning network is constructed using a Long Short-Term Memory (LSTM) network and a Deep Deterministic Policy Gradient (DDPG) algorithm, and can be deployed on edge computing nodes with GPU acceleration capabilities. The LSTM module extracts temporal features from the state input vector, thus addressing the problem that traditional models cannot capture dynamic changes in network state. Its input can be the state input vector for the current and four past time windows (a total of 5 × 12 = 60 dimensions). The network structure can contain two hidden layers: the first hidden layer has 64 nodes, and the second has 32, outputting a 32-dimensional temporal feature vector through a tanh activation function. The DDPG algorithm consists of an Actor network and a Critic network. The Actor network receives the temporal feature vector output from the LSTM, maps it to a four-dimensional action vector through a fully connected layer, and uses a sigmoid activation function in the output layer to constrain the parameters of each dimension to the 0-1 range. The Critic network receives the temporal feature vector and the action vector output from the Actor network, calculates the value score of the action, and uses this score to guide network parameter updates.

[0098] Specifically, the four-dimensional actions output by the pre-defined deep reinforcement learning network all have clear scheduling significance. The proportion of dedicated time slots for the first-level service is the percentage of dedicated time slot resources allocated to the highest-priority service, ensuring the transmission timeliness of high-priority services. The path scoring weight parameters shared by the second and third-level services are used for weighted indexing during subsequent path selection, achieving path optimization for medium-priority services. The fourth-level service load balancing factor is used to adjust the load distribution of the lowest-priority service, avoiding local congestion. The base station power adjustment command can be a relative adjustment ratio of the base station's transmit power, used for dynamic control of base station energy consumption. Furthermore, the four adjustment parameters of the four-dimensional actions can be associated with the corresponding service level and timestamp using JSON data format to generate a resource allocation strategy. This strategy includes executable information such as time slot allocation schemes, power adjustment commands, and load balancing rules, and is distributed to the base station and switches for execution via a base station interface protocol (such as the S1AP protocol). Furthermore, parameters shared by the second and third level services can be directly extracted from the four-dimensional action to obtain the path scoring weight parameters. These parameters are three-dimensional vectors, corresponding to the scoring weights of bandwidth, latency, and packet loss rate, respectively. After extraction, they can be stored in the parameter cache area of ​​the path selection module, providing a direct basis for subsequent path scoring calculations.

[0099] In one embodiment, a pre-defined deep reinforcement learning network is trained through the following steps:

[0100] Obtain historical network status datasets, which include historical service label sequences, historical link status, and historical base station energy consumption data.

[0101] A simulated network environment model is constructed, which integrates a preset base station nonlinear energy consumption model, historical network topology, and service transmission rules. The historical network state dataset is processed by extracting time-series features through a long short-term memory network to obtain time slot correlation feature vectors.

[0102] Construct a reward function with the goal of maximizing network resource utilization and minimizing base station energy consumption;

[0103] The time slot-related feature vectors are processed by a deep deterministic policy gradient algorithm to make policy decisions and output initial action instructions. The initial action instructions are then input into a simulated network environment model for service scheduling simulation testing to obtain the network resource utilization, actual base station energy consumption, and base station load during the test.

[0104] The base station load during the test is input into a preset base station nonlinear energy consumption model for calculation to obtain the predicted energy consumption value of the base station.

[0105] The reward value of the reward function is calculated based on the network resource utilization rate and actual energy consumption of the base station after testing. The network parameters of the joint network are adjusted according to the reward value. The deviation between the actual energy consumption of the base station and the predicted energy consumption of the base station is calculated. The deviation is used as a negative feedback term to update the network weight of the joint network and generate the optimized joint network.

[0106] The optimized joint network is used as the network basis for a new round of training. Iterative parameter optimization is performed until the reward value of the reward function meets the preset reward threshold, thus obtaining the preset deep reinforcement learning network after training.

[0107] Specifically, historical service tag sequences can be obtained from the service log storage unit of the gateway device. This unit archives dynamic service tags for all service data packets in timestamp order. The sequence length is configured according to training requirements, and each sequence contains service tag data within a preset time span, covering the distribution characteristics of services with different priorities. Historical link status can be taken from the historical collection database of the link quality probe, containing bandwidth utilization, latency, and packet loss rate data within the corresponding time span, aligned with the service tag sequences by timestamp. Historical base station energy consumption data can be taken from the energy consumption monitoring unit logs on the base station side, containing data such as power consumption and load rate of the base station at the corresponding time node. After data acquisition, outliers in the link status and energy consumption data can be removed using the 3σ criterion, and missing data can be filled in using linear interpolation. Finally, the data is organized into a structured dataset according to the structure of timestamp-service tag-link status-energy consumption.

[0108] Specifically, by constructing a simulated network environment model, the operational characteristics of a real network can be simulated, providing a repeatable and controllable interactive environment for network training and avoiding the risk of service interruption caused by training directly on a real network. This model can adopt a discrete event simulation architecture, deployed on a training server with a multi-core processor, and all three core elements integrated are consistent with those of a real network. The preset base station nonlinear energy consumption model directly reuses the quadratic polynomial model used in online scheduling, ensuring consistency in energy consumption calculations. The historical network topology can be exported from the network management system's topology database, including the node distribution and link connections of the core network, aggregation layer, and access layer. The model configures attributes such as rated bandwidth and transmission delay limits for each link according to real parameters, while the service transmission rules can be defined based on the TCP / IP protocol stack, including the forwarding mechanism of service data packets, priority scheduling logic, and port suppression execution logic, which are completely consistent with the service processing flow of a real network.

[0109] Specifically, due to the strong temporal dependence between network state and service traffic, traditional deep learning models cannot effectively capture this temporal correlation. Therefore, Long Short-Term Memory (LSTM) networks can be used to extract temporal features from historical network state datasets. For example, the historical dataset is first divided into sample sequences according to fixed time windows. Each sample sequence contains network state data for multiple consecutive time steps (e.g., 5), and each time step has a 12-dimensional data dimension (consistent with the state input vector dimension of online scheduling: 8-dimensional service traffic + 3-dimensional link state + 1-dimensional energy consumption). This network can extract the temporal dimension information of the sample sequences through forward propagation, eliminate redundant information, and retain key temporal correlations (such as the relationship between peak service traffic and link congestion). Finally, it outputs a 32-dimensional time-slot correlation feature vector, which serves as the core feature input for subsequent policy decisions.

[0110] Subsequently, by constructing a reward function, a clear optimization guide can be provided for network training, enabling the network to achieve synergistic optimization between the two objectives of resource utilization and energy consumption control. The reward function can be in the form of a weighted sum, comprehensively reflecting the achievement of the two optimization objectives. The specific expression is as follows:

[0111]

[0112] Where R is the reward value, ranging from 0 to 1. A larger value indicates that the action instructions output by the network are better. This is the resource utilization weighting coefficient, with a value ranging from 0 to 1. It can be adjusted according to business needs. For example, for industrial control scenarios, λ can be set to 0.6 to prioritize resource utilization. Network resource utilization rate is the ratio of the total amount of time slots, bandwidth, and other resources actually occupied by services to the total available network resources. E is the normalized value of base station energy consumption, which is obtained by the ratio of the actual energy consumption of the base station to the rated full-load energy consumption (i.e., the energy consumption coefficient in online scheduling). By using the form 1-E in this function, the contribution of this term is made larger when the energy consumption is lower.

[0113] Specifically, the process of making policy decisions based on slot-related feature vectors using the Deep Deterministic Policy Gradient (DPRQ) algorithm is the core of transforming temporal features into specific scheduling actions. This algorithm is applicable to decision-making problems in continuous action spaces. The DPRQ algorithm comprises two core modules: an Actor network and a Critic network. The Actor network is the policy execution module, receiving a 32-dimensional slot-related feature vector. It performs feature mapping through two fully connected layers (64 nodes in the first layer and 32 nodes in the second). The output layer uses a sigmoid activation function to constrain the output value to the 0-1 range, obtaining a four-dimensional initial action instruction (consistent with the four-dimensional actions in online scheduling). The Critic network is the value evaluation module, receiving the slot-related feature vector and the initial action instruction output by the Actor network. It calculates the value score of the action through fully connected layers, which is used as the basis for subsequent network parameter adjustments.

[0114] Specifically, the training server can send initial action commands to the simulated network environment model via a network interface. The model executes corresponding scheduling operations according to the commands, namely, allocating time slot resources according to the proportion of dedicated time slots, performing path selection according to path scoring weight parameters, adjusting the distribution of low-priority services according to load balancing factors, and adjusting base station transmit power according to power adjustment commands. During the simulation test, the model can collect and record key indicators in real time. Among them, network resource utilization is calculated by statistically analyzing resource occupancy during the simulation period, the actual energy consumption of the base station is output by the model's built-in energy consumption calculation module (reusing the preset base station nonlinear energy consumption model), and the base station load during the test is collected in real time by the model's resource management module (including comprehensive indicators such as the number of access users and time slot occupancy rate). After the test, the model can package the three types of data and feed them back to the training server as the basis for subsequent parameter optimization. Furthermore, the specific calculation process of inputting the base station load during the test into the preset base station nonlinear energy consumption model to calculate the predicted energy consumption value of the base station is consistent with the energy consumption coefficient calculation logic in online scheduling. That is, the comprehensive value L of the base station load during the test is input into the quadratic polynomial model to obtain the predicted energy consumption value of the base station.

[0115] Specifically, to improve network performance, we can first assess the network resource utilization rate based on simulation test feedback. The reward value R is calculated by substituting the actual energy consumption E of the base station into the reward function. The parameters of the joint network are then adjusted based on the reward value. The Critic network calculates the mean squared error loss based on the reward value and its predicted value score, and updates the network weights using gradient descent to improve the accuracy of value assessment. The Actor network updates its policy parameters based on the value gradient of the Critic network, enabling subsequent actions to obtain higher reward values. The deviation between the actual and predicted energy consumption of the base station can then be calculated using relative error. This deviation is used as a negative feedback term to update the weights of the joint network of the Long Short-Term Memory network and the Deep Deterministic Policy Gradient Algorithm through backpropagation, reducing energy consumption prediction bias and improving the accuracy of energy consumption control. Parameter updates can employ mini-batch stochastic gradient descent, with the batch size determined by the training server's memory configuration. The learning rate uses an adaptive adjustment mechanism to avoid gradient explosion or vanishing during training. The optimized joint network is used as the basis for iterative optimization in each iteration. During each iteration, the training server randomly extracts new sample sequences from the historical dataset, repeatedly executing the complete process of temporal feature extraction, policy decision-making, simulation testing, and parameter optimization. The preset reward threshold is determined through offline testing and can be the value of the reward function under ideal scheduling conditions (e.g., 0.8). When the average reward value of multiple consecutive training batches (e.g., 100 batches) stably exceeds this threshold, it indicates that the network has learned a stable optimization strategy, and the training process terminates. After training, the parameters of the joint network (including the weights and biases of each layer of the Long Short-Term Memory network and the Deep Deterministic Policy Gradient Algorithm) can be saved as a model file and deployed to edge computing nodes with online scheduling to provide decision support for real-time scheduling.

[0116] In one embodiment, the process involves processing based on resource allocation strategies and path score weight parameters, calculating path scores, sorting and filtering the path scores, and generating a path mapping table, including:

[0117] Extract bandwidth weight coefficient, latency weight coefficient, and packet loss rate weight coefficient from the path scoring weight parameters;

[0118] Obtain a set of preset available paths, parse each preset available path, and extract the next-hop address and transmission method of each preset available path;

[0119] Link quality is tested for each preset available path to obtain the bandwidth utilization, transmission latency and packet loss rate of each preset available path;

[0120] Based on the bandwidth weighting coefficient, latency weighting coefficient, and packet loss rate weighting coefficient, the bandwidth utilization, transmission latency, and packet loss rate of each preset available path are weighted and calculated to obtain the path score of each preset available path.

[0121] The path scores of each preset available path are sorted from high to low, and the top preset number of available paths are selected from the sorting results to obtain the selected paths.

[0122] Assign corresponding path identifiers to the selected paths, establish a mapping relationship between each path identifier and its corresponding next-hop address and transmission mode, and generate a path mapping table.

[0123] Specifically, the path scoring weight parameters can be stored as a 3D vector in the parameter cache unit of the scheduling center. The dimensions of the vector are arranged in a fixed order of bandwidth-latency-packet loss rate, and the storage format uses binary floating-point to ensure accuracy. When reading this vector, the three coefficients can be located using vector indices, and are respectively labeled as follows: , , And the three coefficients satisfy the normalization constraint conditions. + + =1. The preset available path set can be pre-configured in the path management database based on network topology planning results. The set records complete link information for each path, including the origin node identifier, destination node identifier, sequence of nodes along the path, and link type. For example, by retrieving all available paths matching the current service source and destination devices from the database using a structured query language, a preset available path set can be formed. Subsequently, a routing protocol parsing engine can be used to parse this set. This engine supports mainstream routing protocols such as Open Shortest Path First (OSPF) and Border Gateway Protocol (BGP), and can extract the next-hop address (presented in IPv4 address format) from the routing entries of the parsed path, and determine the transmission mode based on the link layer protocol type. Transmission modes can include Transmission Control Protocol (TCP), User Datagram Protocol (UDP), and Synchronous Digital Hierarchy (SDH), etc. After parsing, the next-hop address of each path is associated with and stored with the transmission mode.

[0124] The link quality testing process for each preset available path can be executed by a distributed link quality testing unit. This unit is directly connected to the starting node of each path and employs a testing mechanism matched to the path's transmission method. For example, for TCP / UDP paths, an active probing method is used, sending fixed-length probe data packets and receiving feedback to collect metrics. For leased line paths such as SDH, a passive monitoring method can be used, capturing service data packets transmitted in the link for metric statistics. Bandwidth utilization is obtained by calculating the ratio of the actual amount of data transmitted per unit time to the path's rated bandwidth. Transmission latency is calculated by recording the round-trip time of probe data packets from sending to receiving and taking half of that value. Packet loss rate is calculated by calculating the percentage difference between the total number of probe data packets sent and the number of successfully received packets. During the testing process, the testing cycle for each path is related to the service priority; paths corresponding to higher priority services have shorter testing cycles, ensuring that the testing data reflects path performance fluctuations in real time.

[0125] Specifically, a path score can be obtained by weighting link quality indicators based on three weighting coefficients. Before calculation, the link quality indicators can be normalized to eliminate the influence of differences in the units of measurement of different indicators. Bandwidth utilization can be directly expressed as the original ratio (range 0-1), transmission delay can be normalized by dividing by the preset maximum allowable delay to obtain a normalized delay value (range 0-1), and packet loss rate can be expressed as the original ratio (range 0-1). The formula for calculating the path score is:

[0126]

[0127] Where S is the path score, ranging from 0 to 1, with a larger value indicating better path adaptability; ω1 is the bandwidth weight coefficient, reflecting the influence of bandwidth performance on path selection; B is the path's bandwidth utilization rate; ω2 is the latency weight coefficient, reflecting the influence of latency performance on path selection; T is the path's normalized latency value; ω3 is the packet loss rate weight coefficient, reflecting the influence of packet loss performance on path selection; and L is the path's packet loss rate. Since lower latency and packet loss rate indicate better path performance, the formula uses the forms (1-T) and (1-L), thus ensuring a positive correlation between the indicators and the score, consistent with the scoring logic. Furthermore, the path scores can be sorted and filtered to obtain the selected paths, ensuring that services are allocated the best-performing transmission resources. The sorting process can be implemented by the path sorting module using a fast sorting algorithm, with the path score as the sorting benchmark and the sorting direction from high to low, outputting a path list sorted in descending order of path score. By extracting the first preset number of paths from the sorted list, the selected paths can be obtained.

[0128] Based on the transmission reliability requirements of the business, the preset number can be determined after verifying the success rate of transmission under different numbers of services through offline testing. For example, for critical services such as industrial control, the preset number is set to 3 to ensure path redundancy, while for ordinary data services, the preset number is set to 1 to simplify resource consumption. After filtering, the paths can also be load balanced. If there is a path with excessive load (load rate exceeding the preset threshold) among the filtered paths, it is replaced with a subsequent qualified path in the sorted list to ensure that the filtered paths have both performance advantages and reasonable load.

[0129] By assigning path identifiers to selected paths and establishing mapping relationships to generate a path mapping table, the path matching efficiency during service transmission can be significantly improved. The path identifiers can be generated using a "path type-serial number" encoding rule. The path type is identified by a single letter (e.g., T represents a TCP path, U represents a UDP path), and the serial number is a three-digit number, ensuring the uniqueness of the identifier. For example, the first path identifier for TCP is T001. By associating the identifier of each selected path with its corresponding next-hop address and transmission mode, a key-value pair structure can be formed (the identifier is the key, and the next-hop address and transmission mode are the values). Furthermore, the path mapping table can be stored in a two-dimensional table format, containing three columns: path identifier, next-hop address, and transmission mode. This table is stored in the high-speed cache of the routing controller, supporting fast lookups by path identifier. During subsequent service transmission, the routing controller can determine the priority by parsing the dynamic service tag of the service, and then match and call the corresponding path information in the path mapping table based on the priority, achieving fast routing and forwarding of the service.

[0130] In one embodiment, such as Figure 2 As shown, based on dynamic service tags and resource allocation strategies, analysis is performed to generate port suppression commands, including:

[0131] S201: Identify the priority level of dynamic service tags, combine the priority division rules corresponding to the four-level service traffic matrix, define the scope of preset low-priority services, and obtain the port identifier association rules corresponding to the preset low-priority services; the preset low-priority services are the fourth-level services in the four-level service traffic matrix.

[0132] S202: Perform real-time parsing and processing of dynamic business tags, detect the priority level in the dynamic business tags, and obtain the detection results;

[0133] S203: When the detection result indicates that there is a tag with the highest priority level among the dynamic service tags, extract the fourth-level service load balancing factor from the resource allocation strategy;

[0134] S204: The port suppression duration is calculated based on the fourth-level service load balancing factor and the preset base duration.

[0135] S205: Based on the port identifier association rules, determine the target port identifier corresponding to the preset low-priority service, bind the target port identifier with the port suppression duration, and generate a port suppression command.

[0136] Specifically, the service priority parsing module can identify the priority level of dynamic service tags. This module can read the priority field in the dynamic service tag and match it with the priority division rules of the four-level service traffic matrix pre-stored in the priority rule base of the scheduling center. This priority division rule can map service priority identifiers to four levels, with the fourth level corresponding to the lowest priority service, thereby defining the scope of preset low-priority services. Port identifier association rules are pre-configured in the port management database. They are built based on the binding relationship between services and access ports in the network topology, which can clearly define the set of access ports corresponding to the fourth-level service, and also include the execution logic of port suppression (such as rate limiting ratio).

[0137] Furthermore, the dynamic service tags can be parsed in real time through the service flow parsing unit. This unit is directly connected to the gateway's service forwarding module, parses the priority field in the dynamic service tags according to the order of data packet arrival, counts the existence status of each priority service within the current time window, and generates detection results. Since the highest priority service usually corresponds to high timeliness requirements (such as industrial control commands), it is necessary to reduce the resource consumption of low priority services to avoid interference with the transmission of high priority services. Therefore, when the detection result shows the presence of the highest priority tag, the subsequent port suppression process can be triggered.

[0138] Since the Level 4 service load balancing factor is a core parameter reflecting the uniformity of Level 4 service load distribution and directly determines the intensity of port suppression, it can be extracted from the resource allocation strategy when the detection result shows the presence of the highest priority tag among the dynamic service tags. The Level 4 service load balancing factor can be stored in the parameter field of the resource allocation strategy and extracted by the load factor reading module. This module directly accesses the storage area of ​​the resource allocation strategy through memory, enabling it to locate and read the factor's value, ensuring the real-time and accurate acquisition of the parameter.

[0139] Subsequently, based on the fourth-level service load balancing factor and the preset base duration, the suppression duration can be accurately adapted through comprehensive calculation of multi-dimensional parameters, avoiding over-suppression or under-suppression, thus obtaining the port suppression duration. This calculation is performed by the duration calculation unit, using the following formula:

[0140]

[0141] in, Port suppression duration, which is the length of time that a port corresponding to a low-priority service must be suppressed; The preset base duration is the reference time for port suppression. It is pre-configured in the parameter library of the duration calculation unit and is determined based on the average transmission cycle of the service. The load balancing factor weight is a pre-configured coefficient used to balance the impact of low-priority service load and base station power status on suppression duration. This is the fourth-level business load balancing factor, which is a parameter extracted from the resource allocation strategy. The maximum transmit power of the base station is determined by the base station hardware specifications, stored in the configuration library of the base station resource management unit, and read through the interface; The current transmit power of the base station is collected by the base station real-time power consumption monitoring unit and synchronized to the dispatch center. This is a correction coefficient for the number of high-priority services, pre-configured in the duration calculation unit, used to dynamically adjust the suppression duration based on the number of high-priority services. The number of highest priority business tags in real time is calculated by the business statistics module within the current time window. The total number of business tags in real time is calculated synchronously by the business statistics module, which counts all business tags within the current time window. The number of service priority levels corresponds to the priority level of the four-level service traffic matrix in this embodiment. During calculation, the above parameters can be obtained first, and then the weighted sum of load and power items, the correction term for the number of high-priority services, and finally multiplied by the preset base duration to obtain the port suppression duration.

[0142] Specifically, the port matching module can invoke port identifier association rules to match the corresponding active port as the target port identifier based on the fourth-level service. Then, the target port identifier and port suppression duration are encapsulated into a standardized command format containing the command type, target port, and suppression duration. This command can be sent to the corresponding access switch via the network management protocol, and the switch's port control module will execute port suppression operations (such as limiting the port's packet forwarding rate). The duration is the calculated port suppression duration, thus reserving sufficient transmission resources for high-priority services.

[0143] Based on the same inventive concept, this application also provides a communication network resource intelligent scheduling device for implementing the above-mentioned intelligent scheduling method for communication network resources. The solution provided by this device is similar to the implementation described in the above method; therefore, the specific limitations in one or more embodiments of the intelligent scheduling device for communication network resources provided below can be found in the limitations of the intelligent scheduling method for communication network resources described above, and will not be repeated here.

[0144] In one exemplary embodiment, such as Figure 3 As shown, a communication network resource intelligent scheduling device 300 is provided, comprising:

[0145] The service priority dynamic adjustment module 301 is used to parse the IP header service type field of the input service data packet to obtain the initial service priority identifier, and adjust the initial service priority identifier based on the preset priority rules, device alarm status and network congestion rate to generate dynamic service tags.

[0146] The resource allocation and strategy generation module 302 is used to obtain real-time link status and energy consumption coefficient, and based on dynamic service tags, real-time link status and energy consumption coefficient, it is processed by a preset deep reinforcement learning network to generate resource allocation strategy and path score weight parameters.

[0147] The path mapping and instruction generation module 303 is used to process the path score based on the resource allocation strategy and path score weight parameters, calculate the path score, sort and filter the path scores, generate a path mapping table, and analyze the dynamic service tags and resource allocation strategy to generate port suppression instructions. It also extracts the base station power adjustment instructions from the resource allocation strategy, and uses the base station power adjustment instructions to perform dynamic adjustment of the base station power.

[0148] In one exemplary embodiment, the present invention also provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the intelligent scheduling method for communication network resources of this application. A multi-core processor is preferred to improve the parallel processing capability of the system. The memory provides sufficient temporary storage space to support program execution and data processing. The memory capacity should be large enough to accommodate large amounts of data and computational tasks.

[0149] In one exemplary embodiment, the present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the intelligent scheduling method for communication network resources of the present application. The computer-readable storage medium may include: a read-only memory, a random access memory (RAM), a solid-state drive (SSD), or an optical disc, etc.

[0150] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for intelligent scheduling of communication network resources, characterized in that, The method includes: The input service data packet is parsed by the IP header service type field to obtain the initial service priority identifier. Based on the preset priority rules, device alarm status and network congestion rate, the initial service priority identifier is adjusted to generate a dynamic service tag. The real-time link status and energy consumption coefficient are obtained. Based on the dynamic service tag, the real-time link status and the energy consumption coefficient, the data are processed through a preset deep reinforcement learning network to generate resource allocation strategies and path score weight parameters. The path score is calculated based on the resource allocation strategy and the path score weight parameters. The path scores are sorted and filtered to generate a path mapping table. Based on the dynamic service tags and the resource allocation strategy, a port suppression instruction is generated. The base station power adjustment instruction is extracted from the resource allocation strategy and is used to perform dynamic adjustment of the base station power.

2. The method according to claim 1, characterized in that, The process involves parsing the IP header service type field of the input service data packet to obtain an initial service priority identifier. Based on preset priority rules, device alarm status, and network congestion rate, the initial service priority identifier is adjusted to generate a dynamic service tag, including: The first three bits of the service type field in the IP header of the service data packet are parsed to obtain the initial service priority identifier; The source device information of the IP header of the service data packet is parsed to obtain the source device identifier; The device alarm status and network congestion rate are monitored in real time, and the initial service priority identifier is adjusted based on the preset priority rules to obtain the adjusted priority identifier; wherein the adjusted priority identifier is obtained through the following steps: When the device alarm status is detected as an emergency state or the network congestion rate exceeds a preset congestion threshold, the initial service priority identifier is upgraded to the highest priority identifier, resulting in the adjusted priority identifier. When no emergency state is detected in the device alarm and the network congestion rate does not exceed the preset congestion threshold, the initial service priority identifier remains unchanged, and the initial service priority identifier is output as the adjusted priority identifier. The system time is formatted to generate a timestamp. The adjusted priority identifier, the timestamp, and the source device identifier are then integrated to generate the dynamic service tag.

3. The method according to claim 1, characterized in that, The process of obtaining real-time link status and energy consumption coefficient, based on the dynamic service tag, the real-time link status, and the energy consumption coefficient, is processed through a preset deep reinforcement learning network to generate resource allocation strategies and path scoring weight parameters, including: The network link is detected by a link quality probe to obtain the real-time link status, which includes bandwidth, latency and packet loss rate. Obtain the base station load, input the base station load into a preset base station nonlinear energy consumption model for calculation, and output the energy consumption coefficient. The dynamic service tags are classified and statistically analyzed according to a preset priority to generate a four-level service traffic matrix; The four-level service traffic matrix, the real-time link status, and the energy consumption coefficient are vector-integrated to generate a status input vector. The state input vector is processed by the preset deep reinforcement learning network to output a four-dimensional action. The preset deep reinforcement learning network is a joint network constructed based on a long short-term memory network and a deep deterministic policy gradient algorithm. The four-dimensional action includes four independent control parameters, namely the proportion of dedicated time slots for the first-level service, the path scoring weight parameters shared by the second and third-level services, the load balancing factor for the fourth-level service, and the base station power adjustment command. The four-dimensional actions are encapsulated to generate the resource allocation strategy, and the path scoring weight parameters shared by the second-level and third-level services are extracted from the four-dimensional actions.

4. The method according to claim 1, characterized in that, The process of processing based on the resource allocation strategy and the path score weight parameters to calculate the path score, sorting and filtering the path scores, and generating a path mapping table includes: Extract the bandwidth weight coefficient, latency weight coefficient, and packet loss rate weight coefficient from the path scoring weight parameters; Obtain a set of preset available paths, parse each preset available path, and extract the next-hop address and transmission method of each preset available path; Perform link quality testing on each of the preset available paths to obtain the bandwidth utilization, transmission latency, and packet loss rate of each preset available path; Based on the bandwidth weighting coefficient, the latency weighting coefficient, and the packet loss rate weighting coefficient, the bandwidth utilization, transmission latency, and packet loss rate corresponding to each preset available path are weighted and calculated to obtain the path score of each preset available path. The path scores of each of the preset available paths are sorted from high to low, and the top preset number of preset available paths are selected from the sorting results to obtain the selected paths. Assign corresponding path identifiers to the selected paths, establish a mapping relationship between each path identifier and the corresponding next-hop address and transmission mode, and generate the path mapping table.

5. The method according to claim 3, characterized in that, The analysis based on the dynamic service tag and the resource allocation strategy generates a port suppression command, including: Priority levels are identified for the dynamic service tags. Based on the priority division rules corresponding to the four-level service traffic matrix, the scope of preset low-priority services is defined, and the port identifier association rules corresponding to the preset low-priority services are obtained. The preset low-priority services are the fourth-level services in the four-level service traffic matrix. The dynamic service tags are parsed in real time to detect the priority level in the dynamic service tags and obtain the detection results. When the detection result indicates that there is a tag with the highest priority level among the dynamic service tags, the fourth-level service load balancing factor is extracted from the resource allocation strategy; The port suppression duration is calculated based on the fourth-level service load balancing factor and the preset base duration. Based on the port identifier association rule, the target port identifier corresponding to the preset low-priority service is determined, and the target port identifier is bound to the port suppression duration to generate the port suppression instruction.

6. The method according to claim 3, characterized in that, The preset deep reinforcement learning network is trained through the following steps: Obtain a historical network status dataset, which includes historical service label sequences, historical link status, and historical base station energy consumption data; A simulated network environment model is constructed, which integrates the preset base station nonlinear energy consumption model, historical network topology, and service transmission rules. The historical network state dataset is processed by the Long Short-Term Memory network to extract time-series features, resulting in a time-slot-related feature vector. Construct a reward function that aims to maximize network resource utilization and minimize base station energy consumption. The time slot-related feature vector is processed by the deep deterministic policy gradient algorithm to make policy decisions and output initial action instructions. The initial action command is input into the simulated network environment model to perform service scheduling simulation test, and the network resource utilization rate, actual energy consumption of the base station and base station load during the test are obtained after the test. The base station load during the test is input into the preset base station nonlinear energy consumption model for calculation to obtain the predicted energy consumption value of the base station. The reward value of the reward function is calculated based on the network resource utilization rate after the test and the actual energy consumption of the base station. The network parameters of the joint network are adjusted according to the reward value. The deviation between the actual energy consumption of the base station and the predicted energy consumption of the base station is calculated. The deviation is used as a negative feedback term to update the network weight of the joint network and generate an optimized joint network. The optimized joint network is used as the network basis for a new round of training. Iterative parameter optimization is performed until the reward value of the reward function meets the preset reward threshold, thus obtaining the trained preset deep reinforcement learning network.

7. The method according to claim 5, characterized in that, The port suppression duration is calculated using the following formula: in, The port suppression duration, The preset base duration, For load balancing factor weights, This refers to the fourth-level service load balancing factor. This is the maximum transmit power of the base station. This represents the current transmit power of the base station. This is a correction factor for the number of high-priority services. The number of highest priority service tags in real time. This represents the total number of real-time business tags. This represents the number of business priority levels.

8. A communication network resource intelligent scheduling device, characterized in that, The device includes: The service priority dynamic adjustment module is used to parse the IP header service type field of the input service data packet to obtain the initial service priority identifier, and adjust the initial service priority identifier based on the preset priority rules, device alarm status and network congestion rate to generate dynamic service tags. The resource allocation and strategy generation module is used to obtain real-time link status and energy consumption coefficient. Based on the dynamic service tag, the real-time link status and the energy consumption coefficient, it is processed by a preset deep reinforcement learning network to generate resource allocation strategy and path score weight parameters. The path mapping and instruction generation module is used to process the resource allocation strategy and the path score weight parameters, calculate the path score, sort and filter the path scores, generate a path mapping table, analyze the dynamic service tag and the resource allocation strategy, generate a port suppression instruction, extract the base station power adjustment instruction from the resource allocation strategy, and the base station power adjustment instruction is used to perform dynamic adjustment of base station power.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images