Industrial Internet NFV energy-saving power supply control method

Abstract

The invention relates to the technical field of industrial internet energy-saving control, and discloses an industrial internet NFV energy-saving power supply control method. The method comprises the following steps: acquiring energy supply request information in an industrial internet environment, and generating an energy control semantic vector by adopting a context-aware load analysis mechanism; a power supply coordination network architecture is established for the distributed virtual network function entities, nodes are used for representing the entities, edges are used for representing energy interaction or function association, the network form and coordination performance are optimized by means of a real-time updating mechanism, and the power supply coordination network architecture is configured in an industrial cloud distributed processing framework; dynamically determining a regulation and control region matched with energy-saving semantics by applying a multi-level mapping calculation method based on the energy control semantic vector; in the regulation and control area, an alternative regulation and control measure set meeting the energy-saving intention is selected through a matching process of multi-dimensional information integration, optimization and adjustment are implemented, and the selection process fuses energy efficiency relevance, regulation and control measure execution characteristics and environmental context factors.

Description

Technical Field

[0001] This invention relates to the field of energy-saving control technology in the industrial internet, specifically to an energy-saving power supply control method for industrial internet NFV. Background Technology

[0002] With the deepening of industrial digital transformation, the Industrial Internet has become a core infrastructure connecting industrial equipment, data, and services. Through the widespread deployment of sensors, controllers, and network devices, it enables real-time perception and intelligent collaboration throughout the entire industrial production process. Network Functions Virtualization (NFV) technology, as a key support for the architectural innovation of the Industrial Internet, transforms traditional hardware network functions into flexibly deployable virtual network function entities, effectively improving the elastic scalability and resource utilization of industrial networks, and driving the development of industrial control systems towards virtualization and distribution. In the operation of the Industrial Internet, the efficiency of energy supply and control directly affects production costs, equipment lifespan, and overall operational stability. Currently, energy consumption in industrial scenarios continues to expand, the deployment density of various virtual network functional entities is constantly increasing, and the energy demands of different entities vary significantly and dynamically change with production tasks, equipment status, and environmental conditions. However, traditional industrial power supply control methods have many limitations. Existing control strategies mostly rely on preset static rules, lacking in-depth analysis of energy supply request information, making it difficult to accurately capture the dynamic characteristics of load demand and potential energy-saving intentions. For example, when processing energy supply requests, allocation is often simply based on current load data, without fully integrating historical energy consumption statistics and related information from the industrial operation database, leading to a disconnect between energy-saving targets and actual needs. The energy coordination mechanism of distributed virtual network functional entities is inadequate. These entities are physically dispersed, with complex energy interactions and functional relationships among them, but existing network architectures lack the ability to dynamically model and optimize these relationships. The network configuration is usually fixed or only periodically adjusted, unable to respond in real time to changes in entity status and fluctuations in energy demand, resulting in redundant energy consumption during transmission and distribution. Furthermore, most power supply control architectures are not deeply integrated with industrial cloud distributed processing frameworks, making it difficult to leverage the computing power advantages of the cloud to achieve global and refined energy coordination.

[0003] In the stages of determining control areas and selecting measures, existing methods lack precision and adaptability. Control areas are often divided based on fixed physical or logical boundaries, failing to consider the dynamic changes in energy-saving semantics, leading to a mismatch between control measures and actual energy-saving needs in some areas. Furthermore, when selecting control measures, only a single dimension of energy efficiency indicators is often considered, neglecting the execution characteristics of the control measures (such as response speed and implementation cost) and environmental context factors (such as the impact of temperature and humidity on equipment energy consumption). This reduces the applicability of alternative measures and makes it difficult to achieve deep energy savings. Summary of the Invention

[0004] The purpose of this invention is to provide an energy-saving power supply control method for Industrial Internet NFV to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides an energy-saving power supply control method for Industrial Internet NFV, the method comprising: Energy supply request information in the industrial internet environment is obtained. A context-aware load parsing mechanism is adopted. Based on the load demand content, historical energy consumption statistics and industrial operation database contained in the energy supply request information, energy saving intention is identified, and an energy control semantic vector is generated by combining the energy saving intention identification results. For distributed virtual network functional entities in an industrial internet system, a power supply coordination network architecture is established. The power supply coordination network architecture uses nodes to represent virtual network functional entities and edges to represent energy interactions or functional associations between entities. The network structure and coordination performance are continuously optimized through a real-time update mechanism. At the same time, the power supply coordination network architecture is configured within an industrial cloud distributed processing framework. Based on the energy control semantic vector, a multi-level mapping calculation method is applied in the power supply coordination network architecture to dynamically determine the control area that matches the energy-saving semantic. Within the control area, a set of alternative control measures that meet the energy-saving intention identification results are selected and optimized through a matching process that integrates multi-dimensional information. The selection process integrates energy efficiency correlation, control measure execution characteristics, and environmental context factors.

[0006] Preferably, the acquisition of energy supply request information in the industrial internet environment employs a context-aware load parsing mechanism. Based on the load demand content, historical energy consumption statistics, and industrial operation database contained in the energy supply request information, energy-saving intent is identified, and an energy control semantic vector is generated by combining the energy-saving intent identification result, including: The energy supply request information input from the industrial internet environment is input into the neural network load mapping module and transformed into a standardized dimension control semantic space to form load demand embedding features; Collect historical energy consumption statistics of the system, analyze energy consumption fluctuation patterns, abnormal event records and system feedback behavior, construct system energy consumption profile, and convert system energy consumption profile information into energy consumption preference embedding features; Connect to the industrial operation database, identify operational knowledge related to energy supply request information, and extract and embed operational features. The collaborative fusion mechanism is used to merge the embedded features of load demand, energy consumption preference, and operation, and to perform energy-saving intention recognition. An influence weight is assigned to each recognized intention, and an energy control semantic vector is generated based on the weight. The energy control semantic vector covers the current energy-saving intention and reflects the system characteristics and historical operation mode.

[0007] Preferably, for the distributed virtual network functional entities in the industrial internet system, a power supply coordination network architecture is established. This architecture uses nodes to represent virtual network functional entities, edges to represent energy interactions or functional associations between entities, and continuously optimizes the network structure and coordination performance through a real-time update mechanism. Furthermore, the power supply coordination network architecture is configured within an industrial cloud distributed processing framework, including: Industrial data processing technology is applied to extract the physical operation characteristics of virtual network functional entities, and each virtual network functional entity is treated as an independent node in the power supply coordination network architecture, and each node is assigned corresponding physical operation characteristics. Edges in the power supply coordination network architecture are created based on the energy interaction or functional association between virtual network functional entities. The initial connection strength of each edge is set to obtain the power supply coordination network architecture. The connection strength is initialized by evaluating the energy interaction frequency between two entities to reflect the degree of association of the edges. Relying on a real-time update mechanism, the system monitors changes in entity status, energy consumption data, and their correlations to ensure the timeliness of the power supply coordination network architecture. When adding new entities or modifying the parameters of existing entities, an incremental update method is adopted, updating only the affected nodes and edges. The connection strength of the edges is dynamically adjusted with reference to system load and energy consumption trends to continuously optimize the network structure and coordination performance. In an industrial cloud environment, a distributed processing framework is selected, and a real-time power coordination network architecture is deployed in a distributed graph storage service within the selected distributed processing framework.

[0008] Preferably, based on the energy control semantic vector, the step of applying a multi-level mapping calculation method in the power supply coordination network architecture to dynamically determine the control region matching the energy-saving semantic includes: The power supply coordination network architecture is processed using graph embedding pre-training technology to generate a node embedding representation for each node. In this process, deep feature encoding is performed on each virtual network functional entity node so that the position of the node in the embedding space simultaneously reflects its own operating attributes and energy interaction relationships with other nodes. The embedding space is composed of all node embedding representations. The energy control semantic vector is mapped to the embedding space using a transformation model, so that the mapped energy control semantic vector is in the same space as the node embedding representation of the power supply coordination network architecture. Calculate the matching degree between the mapped energy control semantic vector and the embedded representation of all nodes, and select the top N nodes with the highest matching degree from the power supply coordination network architecture as the reference nodes to identify the initial control starting point. Based on the neighboring nodes of each baseline node, extend outward to the first level of neighboring nodes. Determine whether the matching degree of these neighboring nodes with the energy control semantic vector exceeds the set threshold. If the matching degree of the neighboring nodes does not exceed the set threshold, the extension is terminated. If the matching degree exceeds the set threshold, add it to the candidate node set and repeat the extension operation for the newly added neighboring nodes to continue exploring the next level of neighboring nodes until the predefined level limit is reached or the cumulative number of related nodes is satisfied. The set of all candidate nodes obtained through multi-level mapping is defined as the control region. The control region is a sub-network of the power supply coordination network architecture and includes network morphology information and complete operational characteristics of the candidate nodes.

[0009] Preferably, within the control area, the selection and optimization of a set of alternative control measures that meet the energy-saving intent identification results through a multi-dimensional information integration matching process includes: For each potentially relevant virtual network functional entity within the control area, extract multi-dimensional feature data; Multi-dimensional information integration technology is applied to fuse feature data from different dimensions. In this process, weighted collaboration, priority allocation, or deep neural networks are used to perform feature fusion to generate integrated feature data for each entity, which is used to represent the multi-dimensional operational semantics of the entity. The correlation between the integrated feature data of each entity and the energy control semantic vector is evaluated, and entities with a correlation higher than a certain threshold are selected as the basis for alternative control measures. The basis for alternative control measures is a set of entities closely related to the energy-saving intention recognition results. Based on correlation, system energy consumption preference, implementation cost and timeliness, the alternative control measures are optimized and ranked to form optimized control instructions and execute them.

[0010] Preferably, the step of inputting the energy supply request information from the industrial internet environment into the neural network load mapping module and transforming it into a standardized dimension control semantic space to form load demand embedding features includes: Preprocess the load demand content in the energy supply request information, remove redundant and noisy data, and extract key load parameters; Key load parameters are input into the neural network load mapping module, and through multi-layer perception and feature transformation operations, they are mapped to a standardized dimension control semantic space to generate load demand embedding features, which are used for subsequent collaborative fusion mechanisms.

[0011] Preferably, the application of industrial data processing technology to extract entity operation features from virtual network functional entities, treating each virtual network functional entity as an independent node in the power supply coordination network architecture, and assigning corresponding entity operation features to each node, including: Obtain real-time operational data of virtual network functional entities from industrial internet systems, including computing resource utilization, energy consumption rate, and task processing latency; Industrial data processing technology is used to analyze operational data and extract entity operational characteristics, including peak load characteristics, steady-state characteristics, and abnormal response characteristics. The operational characteristics of entities are assigned to the corresponding nodes as initial attributes of the power supply coordination network architecture, which are then used to optimize the network structure through subsequent real-time update mechanisms.

[0012] Preferably, the process of using graph embedding pre-training technology to process the power supply coordination network architecture and generate a node embedding representation for each node includes deep feature encoding of each virtual network functional entity node, so that the node's position in the embedding space simultaneously reflects its own operating attributes and energy interaction relationships with other nodes, including: Initialize the graph embedding model of the power supply coordination network architecture and set the pre-training parameters, including the embedding dimension and the number of iterations; Each node is processed by deep feature encoding, and the vector position of the node in the embedding space is calculated by combining the node's own operating attributes and the energy interaction relationship of the edge connections. After pre-training is completed, a set of node embedding representations is generated for subsequent transformation model mapping and matching degree calculation.

[0013] Preferably, the application of multi-dimensional information integration technology fuses feature data from different dimensions. This fusion is performed using weighted collaboration, priority allocation, or deep neural networks to generate integrated feature data for each entity, including: Identify the sources of characteristic data from different dimensions, including real-time energy consumption data, historical operation logs, and environmental sensor information; Multi-dimensional information integration technology is used, weighted collaborative methods are applied to allocate weights, priority allocation methods are used to set the order of feature importance, or deep neural networks automatically learn fusion weights to perform feature fusion; Generate integrated feature data for subsequent correlation evaluation and ranking optimization.

[0014] Preferably, the process of forming and executing the optimized control command includes: The optimized and ranked alternative control measures are then converted into a sequence of specific control instructions. The control command sequence is sent to the target virtual network functional entity through the execution interface of the industrial internet system; Monitor energy consumption changes and system response after command execution, collect feedback data to update historical energy consumption statistics and optimize subsequent power supply coordination network architecture.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This industrial internet NFV energy-saving power supply control method demonstrates significant advantages in the precision, dynamism, and coordination of energy control through multi-stage collaborative design. In the energy-saving intent identification stage, a context-aware load parsing mechanism is employed to deeply integrate load demand content from energy supply request information with historical energy consumption statistics and industrial operation databases. This enables the extraction of potential energy-saving needs from complex industrial scenario data. This mechanism overcomes the limitations of traditional static rules, moving beyond a superficial analysis of the current load. Instead, it combines historical operating patterns and industrial production logic, making the generated energy control semantic vector more aligned with the energy optimization direction in actual production, providing a practical basis for subsequent precise regulation. A power supply coordination network architecture was established for distributed virtual network functional entities. Through the concrete modeling of nodes and edges, the energy interactions and functional relationships between entities were clearly presented. The introduction of a real-time update mechanism allows the network structure to dynamically adjust according to changes in entity states and fluctuations in energy demand, avoiding the problem of lagging energy coordination in traditional fixed architectures. By configuring this architecture within an industrial cloud distributed processing framework and leveraging the powerful computing power and data processing capabilities of the cloud, global coordination and optimization of large-scale distributed entities were achieved. This makes the flow of energy between entities more rational and reduces energy waste caused by information silos or poor coordination. By applying a multi-level mapping calculation method based on energy control semantic vectors to determine the control area, a precise match between energy-saving targets and control ranges is achieved. The multi-level mapping process refines the correspondence between semantic vectors and physical and functional regions layer by layer, avoiding control deviations caused by traditional fixed boundary divisions. This dynamically determined control area can be adjusted in real time according to changes in energy-saving intentions, ensuring that energy control resources are concentrated on the areas most in need of optimization, thus improving the targeting and effectiveness of energy control. The multi-dimensional information integration and matching process within the control area takes into account energy efficiency correlation, the implementation characteristics of control measures, and environmental context factors, breaking through the limitations of single-indicator evaluation. When selecting alternative control measures, it not only focuses on the energy-saving potential of the measures themselves but also considers their feasibility in the current industrial environment, their impact on production processes, and their adaptability to environmental conditions. This comprehensive evaluation approach makes the selected set of control measures more practically valuable, enabling the achievement of energy-saving goals while ensuring the stability and continuity of industrial production. Through optimization and adjustment, these measures can be more smoothly integrated into the operational processes of the Industrial Internet, forming a complete closed loop from intent recognition to measure implementation. This drives the transformation of energy control in the Industrial Internet of Things (NFV) environment from extensive management to refined and intelligent management, improving the overall energy utilization level. Attached Figure Description

[0016] Figure 1 This is a timing diagram of the Industrial Internet NFV energy-saving power supply control method described in this invention; Figure 2 A flowchart for generating energy control semantic vectors; Figure 3 Flowchart for establishing a power supply coordination network architecture; Figure 4 A flowchart for dynamically determining the control area. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 This invention provides an energy-saving power supply control method for Industrial Internet NFV, the method comprising: Energy supply request information is received from the industrial internet environment, and an energy-saving intent identification mechanism is performed using a context-aware load parsing mechanism. This mechanism parses load demand content, historical energy consumption statistics, and industrial operation databases to identify energy-saving intents. The energy-saving intent identification results are integrated with system characteristics and historical operation patterns to generate an energy control semantic vector. For distributed virtual network functional entities, a power supply coordination network architecture is constructed: each virtual network functional entity is represented as a node, and edges represent energy interactions or functional associations between entities. This architecture continuously optimizes network morphology and coordination performance through a real-time update mechanism and is deployed in an industrial cloud distributed processing framework. Based on the energy control semantic vector, a multi-level mapping calculation method is applied in the power supply coordination network architecture. This method dynamically determines the control area that matches the energy-saving semantics. Within the control area, a multi-dimensional information integration matching process is executed: integrating feature data to evaluate correlation and screening candidate control measures; ranking based on correlation, system energy consumption preferences, implementation costs, and timeliness to form an optimized control measure set and implement adjustments. The entire process ensures dynamic matching between energy-saving control and the system.

[0019] Example 1: See Figure 2 The energy supply request information input from the Industrial Internet environment undergoes a preprocessing process. Redundant and noisy data is filtered out using a rule-based filtering algorithm, retaining key load parameters. These key load parameters cover the type of computing resource requirement, task execution time window constraints, peak power consumption threshold, and load fluctuation cycle characteristics. The preprocessed parameters are then input to a neural network load mapping module. This module consists of a three-layer fully connected network structure. The input layer receives the parameter vector, the hidden layer performs a nonlinear transformation using the ReLU activation function, and the output layer maps the result to a standardized dimension control semantic space through normalization. During the mapping process, the weight matrix is ​​dynamically adjusted based on the training data, and the number of hidden layer nodes is configured according to the parameter complexity. The output generates a fixed-dimensional load demand embedding feature vector, where each dimension of the vector represents the distribution characteristics of resource demand in the semantic space. This feature serves as one of the input elements for subsequent multi-source data fusion.

[0020] The system's historical energy consumption statistics database is accessed synchronously, and the energy consumption data stream is processed by the parsing engine. The parsing process identifies energy consumption fluctuation patterns, including periodic fluctuations, triggering conditions for sudden events, and steady-state operating ranges. Abnormal event records are extracted using timestamp alignment to determine the abnormality type label and duration. System feedback behavior is analyzed by extracting response delay records and compensation operation records from the logs. The parsing results are input into the feature construction layer to generate a system energy consumption profile data model. This model is processed through an embedding transformation layer to learn potential energy consumption preference patterns and outputs an energy consumption preference embedding feature vector. The dimension of this vector is consistent with the load demand embedding feature.

[0021] The industrial operations database is connected via a semantic query interface, with query conditions based on load type and resource identifiers in the energy supply request information. Relevant operational knowledge entities are retrieved, including equipment collaborative operation procedures, historical energy efficiency optimization plans, and fault handling contingency plans. The knowledge extraction process uses a graph neural network to traverse the operational knowledge graph nodes, extracting node attributes and associated edge features to generate an operational feature embedding vector. This vector integrates the contextual semantics of the operational scenario.

[0022] A collaborative fusion mechanism is implemented using load demand embedding features, energy consumption preference embedding features, and operation feature embedding features as inputs. This mechanism consists of a feature concatenation layer, an intent recognition layer, and a weight allocation layer. The feature concatenation layer concatenates the three vectors along their dimensions; the intent recognition layer uses convolutional kernels to scan the concatenated vectors, detect energy-saving intent feature patterns, and output intent category labels and confidence levels; the weight allocation layer dynamically calculates the influence weight of each intent based on intent confidence and historical operation frequency. Weight values ​​are adjusted based on a preset rule base; for example, high-confidence intents correspond to high weights, and frequent operation intents correspond to diminishing weights. The intent labels and weight values ​​are integrated to generate an energy control semantic vector, whose dimensions include energy-saving intent encoding, weight coefficients, and context identifiers. This vector represents the core semantic guidance of current energy-saving regulation.

[0023] After the optimization and prioritization of control measures are completed, the candidate control measures are converted into a sequence of specific control instructions. The conversion process employs an instruction compilation engine: high-priority measures are mapped to corresponding instruction templates; for example, computing resource scaling instructions use a Kubernetes container control template, and energy consumption optimization instructions use a Modbus protocol register configuration template. Instruction parameters are automatically filled based on entity characteristics. The protocol adaptation layer encapsulates the instructions into industrial communication messages, including OPCUA data packets and MQTT message formats. The instruction sequence is issued through the industrial internet system execution interface, with the target address pointing to the list of virtual network function entity IPs within the control area. Instruction execution covers dynamic allocation of computing resources, adjustment of energy supply curves, and reset of task scheduling strategies.

[0024] After the execution interface is started, the monitoring agent collects the target entity's operating status in real time. Energy consumption change data is obtained through the smart meter interface, recording current values ​​and power factors; system response data is captured by the agent program, including CPU utilization curves, task queue backlog, and network latency indicators. Feedback data is transmitted to the analysis engine via message middleware, and compared with preset thresholds to generate an execution performance report. This report is input into the historical energy consumption statistics database update process: new energy consumption data is entered into the database using a time series compression algorithm; system response data is marked with operation timestamps and stored in the log analysis pool. The database update triggers a power supply coordination network architecture optimization task: a new energy consumption pattern is added to the graph structure reconfiguration algorithm, automatically correcting the connection strength between nodes; response latency data drives the entity's running feature update module, adjusting node attribute values. The data closed-loop process maintains consistency between the system profile and the actual state.

[0025] Example 2: See Figure 3 The Industrial Internet system periodically acquires real-time operational data streams of virtual network functional entities through a data acquisition agent. Data sources include resource monitoring interfaces, energy consumption meters, and task scheduling logs. Acquired data entries include percentage values ​​of computing resource utilization, wattage of energy consumption rate, and millisecond-level timestamps of task processing latency. This data is transmitted to the feature processing engine in time-series format, where a sliding window mechanism is used to segment the raw data stream into fixed time periods.

[0026] The data analysis process utilizes the functions of industrial data processing technology components. The peak load feature extraction module receives the computing resource occupancy sequence and identifies periods of sudden increases in occupancy using a dynamic threshold detection algorithm. The set threshold is automatically adjusted based on historical peak distribution; feature recording is triggered when the duration of a sudden increase exceeds a set lower limit. The steady-state feature analysis unit processes the energy consumption rate data stream, applies a moving average algorithm to filter out instantaneous fluctuations, and calculates the mean power consumption and fluctuation range as steady-state features within a time interval where the variance value remains below a preset stability threshold. The anomaly response feature extractor receives the task processing delay sequence and uses an isolated forest anomaly detection model to identify abnormal delay events; snapshots of the entity states associated with the events are captured, including instantaneous resource occupancy values ​​and concurrent task volume. The analysis results output a set of entity operation features, with each feature labeled with time attributes and duration metadata.

[0027] In the node initialization process of the power supply coordination network architecture, each virtual network functional entity is assigned a unique node identifier. The node attribute database receives the entity's operational characteristic set and stores it in key-value pairs: peak load characteristics are recorded as time-indexed arrays, steady-state characteristics are saved as floating-point triples, and abnormal response characteristics are stored as a list of tagged events. The attribute binding service maps the characteristic data to the corresponding node identifiers, generating an initial node configuration list.

[0028] The edge creation process in the architecture is based on entity interaction logs and functional dependency topology graphs. Energy interaction logs record energy transfer events between entities, including source entity identifier, target entity identifier, transfer amount, and timestamp; functional association logs record service call chain data, recording the caller and called party identifiers. An interaction frequency counter groups log events by entity pair and calculates the cumulative number of interaction events per unit time as the initial connection strength baseline. Edge type attributes are labeled according to the interaction data category: energy transfer edges are marked with energy flow arrows, and functional association edges are marked with service dependency levels. The initial connection strength of all edges is written to the edge relationship table in the format of a four-tuple sequence of [source node, target node, strength value, type label].

[0029] The real-time update mechanism consists of event listeners and incremental updaters. Listeners subscribe to entity state change events, including new entity registration requests, entity configuration parameter modification notifications, and energy consumption data anomaly alarms. When a new entity is registered, the node generation service automatically assigns a node identifier, and the attribute extraction module synchronously collects its initial runtime data; the incremental updater only adds the node and its related edges to the relationship graph, avoiding full graph reconstruction. When existing entity parameters are modified, the associated node attribute database performs a partial update: resource quota adjustments trigger peak load characteristic recalculation, and energy consumption model changes guide steady-state characteristic updates. Energy consumption change data streams are input to the network optimizer in real time. This component compares the similarity of energy consumption trends between adjacent nodes: if the consistency rate of fluctuation directions between two nodes exceeds a set percentage within three consecutive sampling periods, the strength of their connection edge is automatically increased; if the fluctuation divergence continues, the connection strength is gradually weakened. Edge strength adjustment uses a smooth transition algorithm to avoid numerical jumps.

[0030] The network morphology optimization service periodically performs an overall coordination assessment. The assessment process loads the node attribute table and edge relationship table, and calculates the energy efficiency coordination index of the node cluster: this index is calculated based on the weighted average of the state differences between adjacent nodes according to the strength of the connecting edges. When the coordination index falls below a preset lower limit, a topology reconfiguration suggestion is triggered: after confirmation by the system administrator, the strength of low-cooperation edges is reset or selectively removed to maintain a reasonable sparsity in the network architecture.

[0031] The industrial cloud environment selects framework components with distributed computing capabilities. Distributed processing framework configuration parameters are set as follows: the number of worker nodes dynamically scales according to entity size, with a fixed minimum number of nodes to retain; the task scheduling strategy adopts a latency-sensitive mode. During the deployment of the power supply coordination network architecture, the graph data segmenter divides nodes into different storage shards based on entity geographical location. Each shard is deployed in a graph database instance, with node attributes compressed using a columnar storage format and edge relationships stored using an adjacency list structure. The graph query interface supports parallel retrieval: baseline queries are routed to the corresponding shard via consistent hashing, and cross-shard traversal operations have their results merged by the coordinating node. The data persistence service performs periodic graph structure snapshot storage, with snapshots containing complete node attribute versions and edge relationship versions. The runtime memory caching module retains the latest network architecture copy, and the cache invalidation strategy uses a write-through mechanism—all real-time update operations synchronously update the cache and persistence layer.

[0032] The network architecture version management component maintains the update history. Each graph structure adjustment generates a change log: a list of new node IDs and their timestamps; an attribute change log comparing modified node IDs with their old / new values; and an edge change log showing the difference matrix before and after strength adjustments. Historical versions support rollback operations by time point for system failure recovery scenarios.

[0033] Example 3: See Figure 4 The preprocessing stage of the power supply coordination network architecture initiates the graph embedding pre-training process. The pre-training parameter configuration interface sets the embedding dimension to a fixed value, determined based on the total dimension of node attribute features in the network architecture; the iteration count parameter is adaptively adjusted according to the network size, with the base value set as a logarithmic function of the number of nodes. The graph embedding model initialization process loads the adjacency matrix and node feature matrix of the network architecture. The adjacency matrix element values ​​are normalized using edge connection strength, and the node feature matrix contains standardized values ​​of entity operational features. The model is built based on a graph neural network framework, including three layers of graph convolutional modules and one layer of pooling modules. Each graph convolutional operation aggregates the feature information of adjacent nodes, achieving information transfer between nodes through a feature transformation matrix. The pooling module outputs a fixed-dimensional node embedding representation, the calculation process of which is as follows:

[0034] in Indicates the first Layer nodes Embedded vector, It is a node The neighborhood group, It is a node and The strength of the connection between them and These are nodes and The degree, For the first The trainable weight matrix of the layer, This represents a non-linear activation function. This formula ensures that the generated node embeddings simultaneously encode their own attributes and local topological features.

[0035] Energy control semantic vectors are mapped to the embedding space through a transformation model. This model employs a two-layer fully connected neural network, with the input layer dimension matching the original semantic vector dimension and the output layer dimension matching the node embedding space dimension. The mapping process preserves the semantic distribution characteristics of the vectors: high-weighted intent features retain significant numerical differences after transformation. The transformed vectors are then matched with the node embedding representations for similarity. A modified cosine similarity metric is used to calculate the similarity, introducing a moderating factor to balance the influence of node degree. The matching results generate a ranked list, with the top N nodes selected as baseline nodes. The value of N is dynamically adjusted based on the network density.

[0036] The control area extension algorithm iteratively expands from the baseline node. The first-level neighbor node retrieval utilizes the adjacency query interface of the graph database to obtain a set of directly connected nodes. The matching degree evaluation module calculates the correlation between these nodes and the energy control semantic vector, integrating node feature similarity and edge connection strength as evaluation metrics. The threshold judgment condition is set to a dynamic value, fluctuating and adjusting according to the current network load status: when the system load is higher than the average level, the threshold is increased accordingly to narrow the control range. Eligible neighbor nodes are added to the candidate set, and their hop distance from the baseline node is marked. The extension process continues until the termination condition is met: the maximum hop limit is set to a preset value based on the network diameter; the cumulative node count limit is calculated based on the control resource budget. The terminated candidate node set constitutes the complete control area, and the boundary nodes of the area are specially marked for subsequent processing.

[0037] The entity feature extraction process within the control area initiates multi-dimensional data acquisition. The real-time energy consumption monitoring interface provides instantaneous values ​​of current, voltage, and power factor; the historical operation log database retrieves recent operation records and configuration change history; and the environmental sensor network returns measured values ​​of physical quantities such as temperature, humidity, and vibration intensity. The feature integration engine receives this heterogeneous data and performs time alignment and unit standardization preprocessing. In the weighted collaborative method, feature importance weights are calculated using the entropy method: for energy consumption data, periods with large fluctuations are assigned higher weights; for environmental data, measurement points deviating significantly from standard values ​​receive additional weights. The priority allocation scheme sets the priority level of operation history features, assigning higher priority to configuration parameters from recently successfully performed operations. The deep neural network fusion branch employs an attention mechanism to automatically learn the interaction relationships between features of different dimensions, and the attention weight matrix is ​​optimized through backpropagation during the training phase.

[0038] The correlation assessment module processes the integrated feature data. The assessment algorithm calculates a multi-dimensional distance metric between the feature vector and the energy control semantic vector, including Euclidean distance and angular distance components. When converting the distance value into a correlation score, an adjustment coefficient is introduced to reflect the current state of the system: when energy-saving needs are urgent, the coefficient of energy-related features is increased; when system stability is prioritized, the coefficient of operational history features is increased. The threshold value setting adopts a dynamic quantile method, taking the upper quartile of the correlation score distribution as the screening criterion. Entities that pass the threshold enter the set of candidate control measures, and each element in the set is accompanied by a complete correlation analysis report.

[0039] The optimization process loads a set of alternative control measures. Key ranking factors include four dimensions: a correlation score reflecting semantic matching; system energy consumption preference parameters derived from historical optimal configuration templates; measure implementation cost estimation based on a resource occupancy prediction model; and a timeliness index calculating the expected delay time for measure effectiveness. A multi-objective optimization algorithm generates a Pareto front solution set, and the decision engine selects the final ranking scheme based on the real-time system strategy: in energy efficiency-first mode, the weights of correlation and energy consumption preference are increased; in stability-first mode, the weights of implementation cost and timeliness are increased. When the ranking results are converted into a control instruction sequence, high-priority measures are mapped to immediately executed instructions, medium-priority measures generate conditionally triggered instructions, and low-priority measures are converted into advisory prompts. The encoding format of the instruction sequence is compatible with industry protocol standards, including OPCUA node path specifications and Modbus function code definitions.

[0040] The morphological analysis service for the sub-networks in the control area runs concurrently. The network analyzer calculates the sub-network's topological indicators: average path length reflects the efficiency of information transmission between nodes; clustering coefficient measures local connection density; degree distribution statistics identify key hub nodes. These indicators are input into the network optimization model to generate structural adjustment suggestions: for node groups with high clustering coefficients, it is recommended to merge them into logical units for unified control; for node pairs with long path connections, it is suggested to add direct edges to improve coordination efficiency. The suggested solutions are presented through the management interface, and after confirmation, an incremental update process for the network architecture is triggered. The sub-network operational characteristic data is archived into a case knowledge base as a reference for subsequent similar control scenarios. The knowledge base uses a graph structure to store cases, facilitating similarity retrieval and analogical reasoning applications.

[0041] Example 4: Energy Supply Request Information Preprocessing Process. Raw data packets are collected from industrial internet edge devices. The packet format verification module checks data integrity and discards abnormal packets that fail verification. Valid packets enter the content parsing stage; the parsing rule base contains protocol adaptation templates for different device manufacturers. Taking an energy request packet from a CNC machine tool as an example, the parser identifies key fields in the packet: the resource request type is marked as "urgent processing task," the load intensity value is 85%, and the duration is marked as 120 minutes. The noise filtering component eliminates interference data according to preset rules: removing periodic fluctuation records generated by heartbeat signals and filtering instantaneous peak values ​​generated by sensor false alarms. The extracted key load parameters are stored in a structured storage area, with parameter items stored in association with metadata. See Table 1 for the preprocessed parameters.

[0042] Table 1: Parameters after preprocessing.

[0043] Equipment identification Resource types Load strength Duration Peak power consumption Fluctuation cycle CNC-2032 Emergency processing 85% 120min 4.2kW 15min PLC-8816 Routine testing 32% 240min 1.8kW 30min ROB-4495 Precision assembly 67% 90min 3.5kW 20min The neural network load mapping module receives standardized parameter data. The number of input layer nodes matches the parameter dimension; in this example, 6 input nodes correspond to the 6 parameters in the table. The hidden layers employ a dimensionality reduction design: the first hidden layer has 128 nodes, and the second hidden layer is compressed to 64 nodes. A variant of LeakyReLU is chosen as the activation function to avoid the loss of negative value information. During training, historical load data is used to fine-tune the network weights, and the loss function calculates the deviation between the predicted load and the actual energy consumption. The control semantic space for the standardized dimension is defined as a 32-dimensional vector space, and the output layer uses the Tanh activation function to constrain the values ​​to the [-1, 1] interval. Taking CNC-2032 equipment data as an example, the generated load demand embedding features exhibit a specific distribution pattern: high load intensity corresponds to the right region of the semantic space, while long-duration features are concentrated in the upper part of the space.

[0044] The Industrial Internet system collects operational data of virtual network functional entities through monitoring agents. The data collection frequency is set to once every 10 seconds, and the raw data stream undergoes aggregation processing within a time window. Computational resource utilization statistics track the CPU and memory usage of container instances. Monitoring records for a specific virtual machine instance show that between 15:00 and 15:15, CPU utilization fluctuated between 78% and 92%, while memory utilization remained stable at 64%. The energy consumption metering module collects power data, with power readings varying between 3.8 and 4.1 kW during the same period. Task processing latency data comes from the scheduler logs, which show an average task queue latency of 23 ms and a maximum peak latency of 156 ms.

[0045] The entity operation feature extraction process employs multi-algorithm collaborative analysis. A peak load detector scans the CPU utilization data stream, identifying a sustained high load segment between 15:08 and 15:12, marking it as a peak event. A steady-state analyzer processes power data, determining that steady-state operation has begun when the fluctuation range is less than 5% for five consecutive sampling points, recording an average power consumption of 3.95kW. An anomaly detection model analyzes latency data, triggering anomaly marking when the value exceeds three times the standard deviation, and associating it with snapshots of the CPU utilization at 91% and memory utilization at 66% at that time. Feature data is organized as structured records. Taking the VM-2056 virtual machine instance as an example, its feature records include: peak load period markers, steady-state power consumption values, anomaly event timestamps, and associated resource snapshots.

[0046] The node attribute binding service in the power supply coordination network architecture performs feature mapping. Each virtual network functional entity is assigned a unique node ID, and the node attribute table stores feature data. The attribute table adopts a hierarchical structure: the base layer stores raw monitoring data, the feature layer stores analysis results, and the derived layer records calculated metrics. Taking the VM-2056 node as an example, its attribute records include: a 15-minute monitoring dataset from the base layer, three peak event markers from the feature layer, and the average steady-state power consumption value from the derived layer. After node initialization, the architecture service generates a network-wide node status overview, visually displaying the load distribution and energy consumption characteristics of each node.

[0047] Real-time data updates trigger an incremental feature recalculation mechanism. When new monitoring data arrives, only the feature data of the affected nodes is recalculated. For example, a sudden increase in the CPU utilization of a container instance triggers the peak detector to rescan the data of the last 5 minutes, and the updated feature tags are synchronized to the corresponding node attributes. The relationships between nodes are communicated through the event bus. Relevant nodes receive feature change notifications and adjust their locally cached state data. The architecture maintenance service periodically verifies the timeliness of node features. Feature data that has expired triggers a re-collection command to ensure that the feature data in the network architecture remains synchronized with the actual operating state. The feature version management component records each update operation and supports backtracking to historical feature states by time point for use in anomaly analysis and fault diagnosis scenarios.

[0048] Example 5: The initial parameter configuration of the graph embedding model adopts an adaptive strategy. The embedding dimension value is dynamically determined based on the total number of node attributes in the power supply coordination network architecture, and the base dimension is set as an integer multiple of the node feature dimension. The upper limit of the number of iterations is set as a floating threshold based on the network size, and the initial value is the logarithm of the number of nodes rounded up. The model construction adopts a three-layer graph convolutional neural network architecture, with the output dimension of each layer decreasing in a geometric sequence. During training, the feature encoding module processes both the inherent attributes and structural relationships of nodes. The node's own operational attributes are normalized by the input layer and then used to generate the initial embedding through a fully connected network; the connection relationship between adjacent nodes is analyzed by the edge attribute processor, which calculates the weighted average of the features of adjacent nodes. Attribute embedding and relation embedding are fused in the intermediate layer, and the fusion operation uses a channel concatenation method. The final layer outputs the coordinate vector of the node in the embedding space, and each dimension of the vector implies the functional characteristics and cooperation mode of the node. The entire encoding process ensures that nodes with similar physical distances have similar directional distributions in the embedding space. After encoding, the node embedding representation is stored in an in-memory database, and the vector format uses a single-precision floating-point array for storage.

[0049] A mapping index table is established to identify the sources of multi-dimensional feature data. Real-time energy consumption data streams are assigned data channel numbers, historical operation logs are labeled with version tags, and environmental sensor information is marked with geographic coordinates. The index table records metadata such as the data source type, sampling frequency, and numerical range. The data acquisition module configures access interfaces based on the index table: current and voltage data are read from the power quality analyzer, operation logs are retrieved through a database query interface, and temperature and humidity data are subscribed to from the environmental monitoring system. The acquisition timing alignment mechanism is based on hardware clock synchronization, aligning the timestamps of each data source with nanosecond-level precision. The numerical conversion module unifies the units of measurement: current values ​​are converted to amperes, temperature readings to degrees Celsius, and operation record parameters are processed to be dimensionless.

[0050] The feature fusion process offers three modes: weighted collaboration, priority allocation, and neural network. In the weighted collaboration mode, the feature weight allocation algorithm calculates the information entropy index: the weight of environmental temperature data is determined based on the temperature fluctuation variance, and its weight decreases under stable environmental conditions; the weight of operation logs decays exponentially with the time interval between the event and the current event. The priority allocation mode loads a rule engine, with preset rules including: when energy consumption data fluctuates abnormally, the priority of real-time energy consumption features is increased to the highest level; when historical operation conflicts are detected, log features are automatically downgraded. The deep neural network fusion mode adopts an encoder-decoder structure. In the encoding stage, features of each dimension are independently input into branch networks, with the branch structure being a three-layer convolutional neural network to extract local feature patterns. The outputs of each branch are fused at the bottleneck layer, and the fusion layer applies a multi-head attention mechanism to learn the feature interaction relationships. The attention scoring network generates the fusion weights for each feature, and the decoder reconstructs the output to integrate the feature data. Training samples use historical control cases, and the loss function calculates the error between the fused features and the actual control effect.

[0051] The processing strategy for boundary nodes is set independently during the extension process. When extending to a network partition boundary node, the retrieval logic automatically queries related nodes across subgraphs. The boundary relationship analyzer detects the external connections of boundary nodes, including only strongly correlated external nodes (connection strength above 30% of the average). A new node attribute converter maps external node attributes to the current subgraph feature space, using a linear transformation method to maintain feature distribution consistency. After the extension terminates, a final candidate node set is generated, containing the original nodes and the mapped boundary nodes.

[0052] The network architecture version updates include a conflict resolution mechanism. When real-time optimization updates conflict with periodic refactoring suggestions, the state coordinator compares their timestamps. The latest operation automatically gains priority, but the conflict event is recorded simultaneously. The conflict arbitrator analyzes the scope of the operation's impact: local node updates can immediately override global operations, while global structural adjustments must be delayed until the maintenance window. The version snapshot storage service records the operation sequence and supports version comparison. The administrator interface displays change conflict location markers, and specific changes can be forcibly overridden after manual confirmation.

[0053] The decision factor configuration interface for multi-objective optimization ranking supports dynamic weight settings. Interface controls include a correlation baseline value adjustment slider, an energy consumption preference template selector, an implementation cost estimation parameter input box, and a timeliness window setting. The decision engine reloads the weight matrix after receiving the configuration parameters. The ranking generation process employs a constraint-solving algorithm: first, an initial correlation set is generated; then, implementation cost constraints are applied to filter out schemes exceeding the limit; finally, the ranking tiers are output based on timeliness. High-tier measures trigger immediate execution, mid-tier measures enter a timed execution queue, and low-tier measures are marked as observation suggestions. An execution status tracker records the effective time of each measure; measures exceeding the preset effective window are automatically canceled and fed back to the strategy optimization module.

[0054] Graph structure storage optimization employs a compressed adjacency list format. Node and edge indexes are established using hash mapping tables, and attribute data is stored hierarchically based on access frequency. Attributes of hot nodes reside in an in-memory cache, while data for inactive nodes is transferred to a columnar storage partition. The query service prioritizes searching the cache layer, triggering a disk prefetch mechanism when a cache miss occurs. An access pattern analyzer learns query frequency characteristics and dynamically adjusts the cache replacement strategy. Frequently accessed subgraph structures generate materialized views, which are maintained using incremental refresh methods, with a full reconstruction performed only when network structure changes exceed 5%. The data archiving system partitions storage volumes weekly; expired volume data is anonymized and transferred to a cold backup repository, freeing up main storage space while retaining historical traceability.

[0055] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0056] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An industrial internet NFV energy saving power supply control method, characterized in that, The method comprises: Obtaining energy supply request information in an industrial internet environment, using a context-aware load analysis mechanism, and performing energy-saving intention identification according to load demand content, historical energy consumption statistical data and an industrial operation database contained in the energy supply request information, and generating an energy control semantic vector in combination with the energy-saving intention identification result; For the distributed virtual network function entities in the industrial internet system, a power supply coordination network architecture is established, wherein the power supply coordination network architecture uses nodes to represent virtual network function entities and edges to represent energy interaction or functional association between entities, and continuously optimizes network morphology and coordination performance by means of a real-time updating mechanism, and at the same time, the power supply coordination network architecture is configured in an industrial cloud distributed processing framework; Based on the energy control semantic vector, a multi-level mapping calculation method is applied in the power supply coordination network architecture to dynamically determine a regulation and control area matched with the energy-saving semantic; In the regulation and control area, a multi-dimensional information integration matching process is used to select a set of alternative regulation and control measures that meet the energy-saving intention identification result and implement optimization and adjustment, wherein the selection process integrates energy efficiency correlation, regulation and control measure execution characteristics and environmental context factors. 2.The industrial internet NFV energy-saving power supply control method of claim 1, wherein, The method comprises: The energy supply request information input by the industrial internet environment is input into a neural network load mapping module and converted into a standardized dimensional control semantic space to form load demand embedding features; Historical energy consumption statistical data of the system are collected, energy consumption fluctuation patterns, abnormal event records and system feedback behaviors are analyzed, a system energy consumption portrait is constructed, and system energy consumption portrait information is converted into energy consumption preference embedding features; An industrial operation database is connected to identify operation knowledge related to the energy supply request information and extract operation feature embedding features; A collaborative fusion mechanism is used to combine the load demand embedding features, the energy consumption preference embedding features and the operation feature embedding features, perform energy-saving intention identification, assign an influence weight to each identified intention, and generate an energy control semantic vector based on the weight, wherein the energy control semantic vector covers the current energy-saving intention and embodies system characteristics and historical operation patterns. 3.The industrial internet NFV energy-saving power supply control method of claim 1, wherein, The method comprises: Industrial data processing technology is applied to extract entity running features of the virtual network function entities, each virtual network function entity is taken as an independent node in the power supply coordination network architecture, and each node is given corresponding entity running features; An edge in the power supply coordination network architecture is created according to the energy interaction or functional association between virtual network function entities, and the initial connection strength of each edge is set, so as to obtain the power supply coordination network architecture, wherein the connection strength is initialized by evaluating the energy interaction frequency of two entities to reflect the association degree of the edge; Relying on a real-time updating mechanism, the changes in entity state, energy consumption data and their associated relationships are monitored to ensure the instantaneity of the power supply coordination network architecture, wherein when a new entity is added or an existing entity parameter is modified, an incremental updating method is used to update only the affected nodes and edges, and the connection strength of the edges is dynamically adjusted with reference to the system load and energy consumption trend to continuously optimize the network form and coordination performance; In an industrial cloud environment, a distributed processing framework is selected, and the real-time power supply coordination network architecture is deployed in the distributed graph storage service of the selected distributed processing framework. 4.The industrial internet NFV energy-saving power supply control method of claim 1, wherein, The multi-level mapping calculation method is applied in the power supply coordination network architecture based on the energy control semantic vector to dynamically determine a regulation and control region matching the energy saving semantics, including: The power supply coordination network architecture is processed by using a graph embedding pre-training technology to generate a node embedding representation of each node, wherein each virtual network function entity node is deep feature coded, so that the position of the node in the embedding space reflects both its own running attributes and the energy interaction relationship with other nodes, and the embedding space is composed of all node embedding representations; A conversion model is used to map the energy control semantic vector into the embedding space, so that the mapped energy control semantic vector and the node embedding representation of the power supply coordination network architecture are in the same space; The matching degree between the mapped energy control semantic vector and all node embedding representations is calculated, and the top N nodes with the highest matching degree are selected from the power supply coordination network architecture as reference nodes to identify the initial regulation and control starting point; Based on the adjacent nodes of each reference node, the first level of adjacent nodes is extended outward, and it is judged whether the matching degree of these adjacent nodes with the energy control semantic vector exceeds a set threshold, if the matching degree of the adjacent nodes does not exceed the set threshold, the extension is terminated, if the matching degree exceeds the set threshold, the adjacent nodes are added to a candidate node set, and the new adjacent nodes are repeatedly extended to continue to explore the next level of adjacent nodes until a predefined level limit is reached or a cumulative number of related nodes condition is met; All candidate nodes obtained by multi-level mapping extension are defined as a regulation and control region, which is a subnetwork of the power supply coordination network architecture and contains network form information and complete running features of the candidate nodes.

5. The industrial internet NFV energy saving power supply control method of claim 1, wherein, In the regulation and control region, a matching process of multi-dimensional information integration is used to select a candidate regulation and control measure set meeting the energy saving intention recognition result and implement optimization and adjustment, including: Multi-dimensional feature data of each possible related virtual network function entity in the regulation and control region is extracted; Multi-dimensional information integration technology is applied to fuse feature data of different dimensions, wherein weighted cooperation, priority allocation or deep neural network is used to perform feature fusion to generate integrated feature data of each entity for representing multi-dimensional operation semantics of the entity. The association degree between the integrated feature data of each entity and the energy control semantic vector is evaluated, and entities with an association degree higher than a certain threshold are screened as a candidate regulation measure basis, which is an entity set closely related to the energy-saving intention recognition result; The candidate regulation measure basis is optimized and sorted according to the association degree, system energy consumption preference, measure implementation cost and timeliness, an optimized regulation instruction is formed and executed.

6. The industrial internet NFV energy saving power supply control method of claim 2, wherein, The energy supply request information input into the industrial internet environment is input into the neural network load mapping module and converted into a standardized dimension control semantic space to form load demand embedding features, including: The load demand content in the energy supply request information is preprocessed to remove redundant noise data and extract key load parameters; The key load parameters are input into the neural network load mapping module, which is mapped to the standardized dimension control semantic space through multi-layer perception and feature transformation operations to generate load demand embedding features, which are used for subsequent collaborative fusion mechanisms.

7. The industrial internet NFV energy saving power supply control method of claim 3, wherein, The entity running feature extraction is performed on the virtual network function entity by using the industrial data processing technology, each virtual network function entity is taken as an independent node in the power supply coordination network architecture, and the corresponding entity running feature is assigned to each node, including: Real-time running data of the virtual network function entity is obtained from the industrial internet system, including the computing resource occupancy rate, energy consumption rate and task processing delay; The industrial data processing technology is used to analyze the running data and extract the entity running features, including the peak load feature, steady state feature and abnormal response feature; The entity running features are assigned to the corresponding nodes as the initial attributes of the power supply coordination network architecture, which are used for subsequent real-time update mechanism to optimize the network form. 8.The industrial internet NFV energy saving power supply control method of claim 4, wherein, The graph embedding pre-training technology is used to process the power supply coordination network architecture to generate the node embedding representation of each node, wherein the deep feature coding is performed on each virtual network function entity node, so that the position of the node in the embedding space reflects the running attribute of the node itself and the energy interaction relationship with other nodes, including: The graph embedding model of the power supply coordination network architecture is initialized, and the pre-training parameters are set, including the embedding dimension and the number of iterations; Each node is processed by deep feature coding, combining the running attribute of the node itself and the energy interaction relationship of the edge connection to calculate the vector position of the node in the embedding space; After pre-training, the node embedding representation set is generated, which is used for subsequent conversion model mapping and matching degree calculation. 9.The industrial internet NFV energy saving power supply control method of claim 5, wherein, The multi-dimensional information integration technology is applied to fuse feature data of different dimensions, wherein the feature fusion is performed by using weighted collaboration, priority allocation or deep neural network to generate integrated feature data of each entity, including: The sources of feature data of different dimensions are identified, including real-time energy consumption data, historical operation logs and environmental sensor information; The multi-dimensional information integration technology is used to allocate weights by using the weighted collaboration method, set the importance order of features by using the priority allocation method, or automatically learn the fusion weights by using the deep neural network to perform feature fusion; The integrated feature data is generated and used for subsequent association degree evaluation and optimization sorting. 10.The industrial internet NFV energy-saving power supply control method of claim 1, wherein, The optimized regulation instruction is formed and executed, including: Convert the optimized and sequenced candidate regulation measure basis into a specific control instruction sequence; Through the execution interface of the industrial internet system, the control instruction sequence is issued to the target virtual network function entity; Monitor the energy consumption changes and system responses after the execution of the instructions, collect feedback data, and use it to update the historical energy consumption statistical data and optimize the subsequent power coordination network architecture.

Images