A primary and secondary fusion complete ring net box energy efficiency optimization system based on cloud edge cooperation

Abstract

The present application relates to the field of edge computing, and particularly to a primary and secondary fusion complete ring network box energy efficiency optimization system based on cloud edge cooperation. The system comprises: an edge device deployment and data acquisition module, which collects operating electrical parameters and device states, and generates a service call request; a service call module, which filters target control parameters under safety constraints based on an energy efficiency optimization service package; a data synchronization module, which unloads data to a cloud platform through a message queue protocol; a cloud service orchestration module, which aggregates data update service packages and distributes them to the edge; and a real-time control module, which converts parameters into instructions to drive device actions. The present application adopts a cloud training and edge reasoning cooperative architecture, solves the problems of lack of adaptability in local control, limited edge resources, and communication time delay in full cloud control, and realizes real-time energy efficiency optimization.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of edge computing, in particular to a primary and secondary fusion complete ring network box energy efficiency optimization system based on cloud edge cooperation. BACKGROUND

[0002] The primary and secondary fusion complete ring network box is an important edge node in the energy internet of things, continuously collecting and generating high-frequency operation data streams. Traditional control logic is deployed on a local controller, and its decision is based on a pre-set simple data threshold, and it performs basic device control operations. This local service application based on a simple threshold lacks flexibility in data processing and instruction exchange.

[0003] However, the large amount of real-time data streams collected by the node has significant time-varying and nonlinear characteristics. This local service application based on a simple threshold lacks the ability to adapt to complex working conditions. When the data fluctuates greatly and the amount of data to be collected is large, this decision-making method is prone to cause frequent sending of instructions or unable to converge to the global optimal system operation efficiency, thereby reducing the system operation performance and device reliability.

[0004] Although a complex distributed application service can improve the accuracy of data decision-making, it faces contradictions in network architecture deployment: on the one hand, the computing power and storage space of edge nodes are limited, and it is difficult to bear the computing load required by such service applications; on the other hand, if a full cloud centralized service architecture is used, the uploading of a large amount of real-time data streams will cause excessive occupation of communication bandwidth, and the wide area network transmission delay and packet loss risk will seriously affect the real-time performance and reliability of distributed network services.

[0005] In summary, the technical problem to be solved by the prior art is: how to overcome the influence of limited computing resources of edge nodes and wide area network communication delay under the demand of large-scale real-time data stream processing, and realize the deployment and real-time execution of data-driven low-delay distributed network service applications.

[0006] Therefore, a primary and secondary fusion complete ring network box energy efficiency optimization system based on cloud edge cooperation is proposed. SUMMARY

[0007] The purpose of the present application is to provide a primary and secondary fusion complete ring network box energy efficiency optimization system based on cloud edge cooperation, which solves the contradiction between insufficient edge computing power and high network transmission delay in an unstable network environment under the demand of large-scale real-time data stream processing through a hierarchical cooperative architecture of cloud model training and edge real-time inference, and realizes real-time and accurate optimization of the energy efficiency of ring network box devices.

[0008] To achieve the above purpose, the present application provides the following technical solutions:

[0009] A primary and secondary fusion integrated ring network box energy efficiency optimization system based on cloud edge collaboration, specifically comprising the following steps:

[0010] Edge device deployment and data acquisition module: deploy edge computing unit in intelligent power distribution terminal; receive real-time power distribution network operating electrical parameters and ring network box device state data, integrate into state data package; and generate corresponding service call request according to state data package;

[0011] Service calling module: based on the service call request, call the energy efficiency optimization service package embedded in the edge computing unit, select a group of data as the target control parameter in the device control parameter space representing the device operating state under the condition of meeting the preset power grid voltage and load safety constraints;

[0012] Data synchronization module: obtain the state data package and format it into a synchronization payload, and unload the synchronization payload to the energy internet of things cloud platform through the MQTT message transmission protocol;

[0013] Cloud service orchestration module: collect the synchronization payload and update the energy efficiency optimization service package as feedback data, and download the updated energy efficiency optimization service package to the edge computing unit through the network according to the preset strategy;

[0014] Real-time control module: encapsulate the target control parameter into a remote device control instruction, and send the remote device control instruction to the controlled device in real time.

[0015] Preferably, the edge computing unit is an intelligent computing terminal integrated with a central processing unit, a memory and a network communication interface, which has a container runtime environment deployed inside, providing isolated computing resources and running base for the energy efficiency optimization service package; the specific way of integrating into a state data package includes: obtaining power distribution network operating electrical parameters including three-phase voltage, three-phase current, active power and reactive power, and ring network box device state data including switch opening and closing position, energy storage state and environmental temperature and humidity, and encapsulating the operating electrical parameters and the ring network box device state data into a data frame of structured data objects, taking the data frame as the state data package; the specific way of generating a corresponding service call request includes: generating an interface call instruction containing a power distribution terminal identifier and the state data package, taking the interface call instruction as the service call request.

[0016] Preferably, the energy efficiency optimization service package is an independent running instance encapsulated as a containerized microservice, which internally integrates an energy efficiency decision model based on deep reinforcement learning and an inference environment for supporting the running of the model; the energy efficiency decision model adopts a deep Q network architecture and characterizes the energy efficiency evaluation index and the power grid voltage and load safety constraint condition by constructing a reward function; wherein the power grid voltage and load safety constraint condition is configured as a boundary penalty term in the reward function, and when the monitored bus voltage deviation rate exceeds the preset threshold, a negative value penalty is applied to the energy efficiency decision model; the energy efficiency evaluation index is configured as a target return item in the reward function, and its specific value is the weighted sum of the line loss reduction rate and the power factor improvement rate per unit power supply period.

[0017] Preferably, the specific way of screening a group of data as target control parameters includes: obtaining the allowable adjustment range of the on-load voltage regulating switch associated with the ring box and the available switching group number of the reactive power compensation switching device, and constructing respective discrete state sets with the minimum adjustment action unit as the step; performing Cartesian product operation on each of the discrete state sets to generate a candidate action library containing all feasible combinations as the device control parameter space; constructing the current state data packet into a state vector and inputting it into the energy efficiency decision model based on deep reinforcement learning; evaluating the value score of each group of parameters in the candidate action library through the energy efficiency decision model, and outputting the target control parameters according to a preset action selection strategy, wherein the target control parameters include reactive power compensation switching instructions and transformer tap position instructions.

[0018] Preferably, the specific way of formatting into a synchronization payload includes: extracting data in the state data packet and encoding it in protobuf serialization data format to generate the synchronization payload; and the specific way of offloading to the energy internet of things cloud platform includes: establishing a message subscription and publishing channel with the energy internet of things cloud platform based on the MQTT message transmission protocol, and pushing the synchronization payload through the channel.

[0019] Preferably, the specific way of collecting synchronization payloads includes: subscribing to the channel of the MQTT message transmission protocol on the energy internet of things cloud platform, receiving and parsing the synchronization payload, and storing the parsed data in the power time series database according to the distribution terminal identifier.

[0020] Preferably, the specific way of updating the energy efficiency optimization service package comprises: monitoring a data aggregation size of the power timing database, and extracting historical operation interaction data containing state-action pairs and performing offline retraining on the deep reinforcement learning-based energy efficiency decision model when the data aggregation size meets a preset retraining threshold; the historical operation interaction data comprises corresponding operation electrical parameters, ring main unit device states, historically executed target control parameters and calculated energy efficiency evaluation indexes at the same time section; and the specific way of issuing to the edge computing unit comprises: checking a version number of the energy efficiency decision model after offline retraining, and issuing an incremental update package of the deep reinforcement learning-based energy efficiency decision model to the edge computing unit through a cloud-edge collaborative container orchestration architecture.

[0021] Preferably, the specific way of packaging the remote device control instruction comprises: converting reactive power compensation device switching instructions and transformer tap position instructions into instruction frames conforming to a Modbus-TCP protocol format; and the specific way of sending to the controlled device comprises: sending the instruction frames to a programmable logic controller in the ring main unit through an industrial Ethernet interface to remotely control terminal devices in a network.

[0022] Compared with the prior art, the present application has the following beneficial effects:

[0023] 1. The present application deploys an edge computing unit in an intelligent power distribution terminal, collects synchronous payloads using a cloud service orchestration module and updates an energy efficiency optimization service package, thereby constructing a cloud-edge collaborative architecture. This scheme offloads high-computing-power training tasks to the cloud, thereby solving the problem of limited edge resources; at the same time, the energy efficiency optimization service package embedded in the edge computing unit is called based on a service call request to perform local inference, thereby avoiding response lag caused by network fluctuations in pure cloud control and ensuring the real-time performance of power distribution network control.

[0024] 2. The energy efficiency decision model in the present application adopts a deep Q network architecture, wherein the power grid voltage and load safety constraint conditions are configured as boundary penalty terms in a reward function, and the energy efficiency evaluation indexes are configured as target return terms in the reward function. This scheme uses boundary penalties to force the model to avoid out-of-limit actions, thereby ensuring voltage safety; at the same time, the weighted sum of line loss reduction rate and power factor improvement rate in a unit power supply period is used to clearly optimize the direction, thereby realizing the collaborative optimization of line loss and power factor in a dynamic load environment.

[0025] 3.The application constructs respective discrete state sets with minimum adjustment action units as steps, and performs a Cartesian product operation on each of the discrete state sets to generate a candidate action library containing all feasible combinations as the device control parameter space.This scheme strictly limits the algorithm search space within the physical range, and prevents device mechanical damage from a mechanism; at the same time, the pre-constructed limited action library reduces the optimization range and reduces the computational complexity of the edge computing unit.

[0026] 4.The application adopts the protobuf serialization data format for encoding to generate the synchronization payload, and establishes a message subscription and publishing channel with the energy internet of things cloud platform based on the MQTT message transmission protocol.Compared with the traditional text format, this scheme significantly compresses the data volume and reduces the protocol overhead, ensuring the transmission stability of data uplink and model update package distribution in the environment of limited communication bandwidth of the power distribution network. BRIEF DESCRIPTION OF DRAWINGS

[0027] Fig. 1 A cloud-edge collaborative primary and secondary fusion complete ring network box energy efficiency optimization system structure diagram is provided for the first embodiment of the application.

[0028] Fig. 2 A cloud-edge collaborative primary and secondary fusion complete ring network box energy efficiency optimization system work flow diagram is provided for the second embodiment of the application. DETAILED DESCRIPTION

[0029] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the application.

[0030] Please refer to Figs. 1-2 The application provides a cloud-edge collaborative primary and secondary fusion complete ring network box energy efficiency optimization system, which comprises:

[0031] An edge device deployment and data acquisition module: the edge computing unit is deployed in the intelligent power distribution terminal; real-time receiving of power grid operation electric parameters and ring network box device state data, integration into a state data packet; and generating a corresponding service call request according to the state data packet;

[0032] A service call module: calling the energy efficiency optimization service package embedded in the edge computing unit based on the service call request, selecting a group of data as the target control parameter in the device control parameter space representing the device operating state under the condition of meeting the preset power grid voltage and load safety constraints;

[0033] The data synchronization module: obtains the state data packet and formats it into a synchronization payload, and unloads the synchronization payload to an energy internet cloud platform through an MQTT message transmission protocol;

[0034] The cloud service orchestration module: collects the synchronization payload, updates an energy efficiency optimization service package as feedback data, and distributes the updated energy efficiency optimization service package to the edge computing unit through a network according to a preset strategy;

[0035] The real-time control module: encapsulates the target control parameter into a remote device control instruction, and sends the remote device control instruction to the controlled device in real time.

[0036] Embodiment one

[0037] As an embodiment of the present application, referring to Fig. 1 A one-two fusion integrated ring network box energy efficiency optimization system structure diagram based on cloud edge cooperation.

[0038] Further, the edge computing unit is an intelligent computing terminal integrated with a central processor, a memory and a network communication interface, and has a container runtime environment deployed therein, which provides isolated computing resources and a running base for the energy efficiency optimization service package; the specific way of integrating into a state data packet includes: obtaining power distribution network operation electrical parameters including three-phase voltage, three-phase current, active power and reactive power, and ring network box equipment state data including switch opening and closing position, energy storage state and environmental temperature and humidity, and encapsulating the operation electrical parameters and the ring network box equipment state data into a JSON format data frame, and taking the data frame as the state data packet; the specific way of generating a corresponding service call request includes: generating an interface call instruction containing a power distribution terminal identifier and the state data packet, and taking the interface call instruction as the service call request.

[0039] Specifically, the edge computing unit serves as an edge side computing power carrier of the system, adopts an industrial grade embedded architecture on a hardware level, is configured with a central processor supporting floating point operation, such as an industrial grade CPU of ARM architecture or x86 architecture, and a memory for caching real-time data and model weights, including a RAM random access memory and a Flash flash memory; on a software level, the unit runs an embedded operating system and is pre-installed with a container runtime environment, such as Docker Engine or Containerd, so as to be able to load and execute the energy efficiency optimization service package as an independent container instance, realizing decoupling of application logic and underlying hardware. The edge device deployment and data acquisition module establishes a data communication channel with the underlying sensing device, reads large-scale real-time data from the measurement unit, and divides the large-scale real-time data into power distribution network operation electrical parameters representing power load characteristics and ring network box equipment state data representing physical working conditions.

[0040] Specifically, the module utilizes a preset serialization program to map the read heterogeneous data into a standard key-value pair structure and encapsulate it into a JSON format data frame, for example, generating a key-value pair structure containing voltage and switching state, thereby forming a state data packet. Subsequently, the module reads the locally stored unique power distribution terminal identifier, generates an interface call instruction containing the identifier and the state data packet, which explicitly points to the energy efficiency optimization service container interface deployed in the edge computing unit, and triggers the subsequent process as a service call request.

[0041] The present application realizes the decoupling of the underlying data collection logic and the upper energy efficiency optimization algorithm by encapsulating the running electrical parameters and device state data into a unified JSON format data frame and generating an interface call instruction containing the terminal identifier. This technical solution establishes a standardized data interaction format, so that the service call in the edge computing unit no longer depends on specific underlying register addresses, improving the maintainability of the software module and the standardization of data analysis.

[0042] Further, the energy efficiency optimization service package is an independent running instance encapsulated as a containerized microservice, which integrates an energy efficiency decision model based on deep reinforcement learning and an inference environment for supporting the running of the model; the energy efficiency decision model adopts a deep Q network architecture and represents the energy efficiency evaluation index and the power grid voltage and load safety constraint condition by constructing a reward function; wherein the power grid voltage and load safety constraint condition is configured as a boundary penalty term in the reward function, and when the monitored bus voltage deviation rate exceeds the preset threshold, a negative value penalty is applied to the energy efficiency decision model; the energy efficiency evaluation index is configured as a target return term in the reward function, and its specific value is the weighted sum of the line loss reduction rate and the power factor improvement rate per unit power supply period.

[0043] Specifically, the energy efficiency optimization service package is explicitly defined as a containerized microservice independent running instance encapsulated in the storage space of the edge computing unit, for example, a Docker container, which integrates a lightweight inference engine such as TensorFlow Lite to support the running of the energy efficiency decision model. The energy efficiency decision model adopts a deep Q network architecture, which includes an input layer, a fully connected hidden layer with ReLU activation function, and an output layer. The model represents the energy efficiency evaluation index and the safety constraint by constructing a reward function, which consists of a boundary penalty term and a target return term.

[0044] Specifically, the energy efficiency evaluation index refers to the numerical feedback for quantitatively evaluating the degree of improvement of system running efficiency by a single control action. Its calculation logic is as follows: the system first calculates the line loss reduction rate per unit power supply period based on the electrical parameter data before and after the action, denoted as and power factor improvement rate, denoted as The system introduces an adaptive dynamic weight mechanism, which dynamically adjusts the first weight coefficient and the second weight coefficient according to the deviation of the real-time power factor from the evaluation threshold. The and are adaptive dynamic weight coefficients. Their values are dynamically calculated by the system according to the deviation of the real-time power factor from the preset evaluation threshold, and the specific calculation process is as follows: when the power factor is not lower than the evaluation threshold, the economic priority strategy is executed, and a higher proportion of corresponding to the line loss reduction rate is given through a linear function, guiding the model to preferentially reduce loss; when the power factor is lower than the evaluation threshold, the compliance priority strategy is switched to; the compliance priority strategy refers to the strategy adopted by the system to force the model to prioritize correcting the power factor in the reward feedback to ensure that the operation index returns to the standard range. Under this strategy, the system significantly increases corresponding to the power factor improvement rate and correspondingly reduces with the increase of the negative deviation, thereby forcing the model to prioritize correcting the power factor in the reward feedback until the index returns to the standard range. The calculation formula of the reward function is expressed as The parameter in the formula refers to the energy efficiency evaluation index (i.e., the reward value) obtained by the energy efficiency decision-making model in the reinforcement learning training, which represents the pros and cons of the current control action in a numerical form and is directly used as the reward value in the reinforcement learning training.

[0045] Specifically, the system reads the grid operation standard, and when the monitored bus voltage deviation rate exceeds the preset threshold, such as exceeding the range of plus or minus 7%, triggers the boundary penalty term, applies a negative value penalty to the energy efficiency decision-making model, such as giving negative 100 points, forcing the algorithm to avoid out-of-limit actions. At the same time, to avoid the mechanical life consumption of transformer tap adjustment and reactive power compensation switching devices due to frequent operation, the system also configures an action cost penalty mechanism: when the control instruction output by the model causes the transformer tap position or capacitor group state to change, the system will introduce a negative cost related to the action type and frequency into the reward value at the current time step. The negative cost value will be multiplied by a preset action penalty coefficient An additional penalty term is applied to the model. The purpose is to quantify the mechanical wear cost of the equipment as negative feedback to the energy efficiency decision-making model, guiding it to learn a strategy that minimizes the frequency of control actions while meeting energy efficiency targets. Simultaneously, the energy efficiency evaluation index is constructed using a tiered penalty mechanism. The reduction in line loss is configured as the main reward term, reflecting operational economic benefits; the power factor is configured as a soft constraint term, imposing a non-linear deduction penalty when it falls below the assessment standard, and not imposing an additional reward when it exceeds the standard, thereby avoiding optimization redundancy caused by physical quantity coupling. Furthermore, the system uses a Min-Max normalization method based on historical extreme values ​​to map the reward value to a fixed interval, eliminating the interference of the moving target effect on Q-value estimation.

[0046] As a preferred implementation, the system introduces a dynamic normalization mechanism when constructing the reward function. This mechanism maintains the moving average and variance of the line loss reduction rate and power factor improvement rate in real time, and uses the Z-Score normalization method to map the two indicators with different dimensions to the same numerical range, followed by weighted summation. This technical solution solves the gradient update bias problem caused by the difference in the magnitude of the indicators in multi-objective optimization, improves the stability of model training convergence, and ensures the balance of the two optimization objectives in weight allocation.

[0047] This invention employs containerized microservices to deploy the energy efficiency decision-making model, providing an independent operating environment and isolated computing resources, thus avoiding resource contention during concurrent application deployments. Simultaneously, by constructing a reward function that includes boundary penalty terms and target reward terms, it achieves a numerical representation of physical safety constraints and energy efficiency optimization goals. The boundary penalty term ensures that the model learns voltage safety boundaries during training, preventing outputs from exceeding limits; the target reward term, through weighted calculation, clarifies the optimization direction, ensuring that the system performs line loss reduction and power factor improvement while meeting safety constraints.

[0048] Furthermore, the specific method for selecting a set of data as the target control parameters is based on the construction of a discretized action space and the inference mechanism of a deep neural network. First, during the initialization phase, the system defines the operating boundaries of the physical equipment and reads the adjustment range of the mechanical tap changer of the on-load tap changer, for example, an integer sequence from -8 to +8, totaling 17 tap positions; simultaneously, it reads the number of physical groups of the reactive power compensation switching device, for example, an integer sequence from 0 to 4, totaling 5 switching states. The system discretizes the above physical boundaries based on the minimum adjustable accuracy as the preset sampling step size, generating a set of tap position parameters and a set of switching levels. Subsequently, the system performs a Cartesian product operation, performing full permutations and combinations of the set of tap position parameters and the set of switching levels to construct a two-dimensional matrix containing all physically feasible combinations. This matrix is ​​defined as the preset equipment control parameter space, i.e., the candidate action library, and its dimension is determined by the cardinality product of the two sets.

[0049] Specifically, in the real-time inference stage, the system collects the current operating electrical parameters and equipment state data through sensors. The state vector is constructed as follows: the system selects the effective values of three-phase voltage, line active power and reactive power at the current time and past historical time points as the time series characteristics part, and encodes the current tap position and the number of capacitor switching groups as discrete characteristics part, and splices all the above features to form a one-dimensional state feature vector. The data processing module performs maximum and minimum normalization processing on the above data to eliminate dimensional differences. The state feature vector is input into the pre-trained energy efficiency decision model, which adopts a deep Q network architecture. The deep neural network layer inside the model performs feature extraction and nonlinear mapping on the input vector, and outputs a value score vector consistent in dimension with the candidate action library. Each element in the vector represents the expected cumulative energy efficiency return of executing the corresponding action in the current state. The system applies a pre-set action selection strategy, such as a greedy strategy or a maximization strategy, to lock the action index corresponding to the maximum value score in the value score vector, and maps the index back to the specific combination in the candidate action library, thereby determining the target control parameters including the reactive power compensation device switching instruction and the transformer tap position instruction.

[0050] Specifically, the deep Q network architecture adopts a Dueling DQN improved architecture. It contains an input layer, a shared feature extraction layer, a value function branch and an advantage function branch. The number of input layer neurons is consistent with the dimension of the state feature vector; the shared feature extraction layer is composed of 3 fully connected layers, each containing 128 neurons, and ReLU activation function is used between layers; the value function branch is used to output the value score of the current state, and the advantage function branch is used to output the advantage value of each action relative to the average action. The outputs of the two branches are aggregated in the output layer to calculate the Q value of each action. The number of output layer nodes is consistent with the cardinality of the candidate action library, which is used to output the Q value of each action.

[0051] The present application defines the physical boundaries and completeness of the algorithm search space by constructing a discrete state set with the minimum adjustment action unit as the step size and generating a candidate action library using Cartesian product operation. Compared with continuous space search, the pre-constructed finite discrete action library effectively reduces the computational complexity of the edge side inference process, and mechanically excludes non-physical feasible solutions, ensuring that the control instructions meet the adjustment step size requirements of the hardware equipment.

[0052] Further, the specific manner of screening a group of data as a target control parameter includes: obtaining an allowable adjustment range of a load voltage regulating switch associated with the ring main unit and a number of available switching groups of a reactive power compensation switching device, constructing respective discrete state sets with a minimum adjustment action unit as a step; performing a Cartesian product operation on each of the discrete state sets to generate a candidate action library containing all feasible combinations as the device control parameter space; constructing a state vector from the current state data packet and inputting it into the deep reinforcement learning-based energy efficiency decision model; evaluating the value score of each group of parameters in the candidate action library through the energy efficiency decision model, and outputting the target control parameter according to a preset action selection strategy, wherein the target control parameter includes a reactive power compensation switching instruction and a transformer tap position instruction.

[0053] Specifically, the system first obtains the allowable adjustment range of the load voltage regulating switch, for example, negative 8 to positive 8, and the number of available switching groups of the reactive power compensation switching device, for example, 0 to 4 groups, in the initialization phase. The system constructs respective discrete state sets with the minimum adjustment action unit of the device as a step. The Cartesian product operation is used to construct a complete candidate action library, and the calculation formula is expressed as: . Wherein, represents the discrete state set of the voltage regulating switch, represents the discrete state set of the reactive power compensation device, and the ordered pair represents a group of joint control instructions. The system performs the Cartesian product operation on the voltage regulating switch state set and the reactive power compensation switching set, thereby generating a candidate action library containing all physically feasible combinations, which constitutes the device control parameter space.

[0054] Specifically, in the real-time reasoning phase, the system constructs a normalized state vector from the current state data packet and inputs it into the deep Q network architecture. The model outputs a value score vector consistent with the dimension of the candidate action library, which represents the expected return of each action. According to the preset action selection strategy, such as the greedy strategy, the system locks the index with the maximum value from the score vector and maps it back to the specific combination in the candidate action library, thereby determining the target control parameter containing the reactive power compensation switching instruction and the transformer tap position instruction.

[0055] As a preferred embodiment, the offline retraining process adopts a priority experience replay mechanism. The system calculates the sample priority according to the time series difference error in the historical interaction data, and gives higher sampling probability to high error samples. In addition, before generating the incremental update package, the cloud performs INT8 training and post-quantization processing on the new model. This technical solution improves the model iteration efficiency by focusing on high error samples, and at the same time, through the quantization technology, it compresses the model volume, reduces the transmission load of the cloud-edge collaboration channel and the storage occupation of the edge side.

[0056] The application adopts the Protobuf binary serialization format in combination with the MQTT message transmission protocol for data transmission, reduces the data load volume and bandwidth occupation of network communication, and adapts to the unstable network environment on the edge side of the power distribution network. The cloud end stores the power time series database by analyzing and storing the synchronization payload, establishes a structured historical working condition data index, provides a data basis for subsequent model training and state backtracking, and supports efficient writing and querying of massive time series data.

[0057] Further, the specific way of formatting the synchronization payload includes extracting data in the state data packet and encoding in the protobuf serialization data format to generate the synchronization payload; and the specific way of unloading to the energy internet of things cloud platform includes establishing a message subscription and publishing channel with the energy internet of things cloud platform based on the MQTT message transmission protocol, and pushing the synchronization payload through the channel.

[0058] Specifically, the data synchronization module performs efficient data serialization and transmission. The system defines the data structure according to the ProtocolBuffers protocol standard, extracts data in the state data packet and encodes in the format to generate a compact binary byte stream as a synchronization payload, so as to reduce the data volume.

[0059] Specifically, the module constructs a lightweight communication link based on the MQTT message transmission protocol, connects to the cloud platform as a client and establishes a message subscription and publishing channel. For uplink data, the module encapsulates the binary synchronization payload into an MQTT publishing message and pushes it to the cloud platform through the channel; and subscribes to the downlink control topic to receive the reverse control or model update instruction from the cloud end.

[0060] The application collects synchronization payloads by subscribing to the MQTT message transmission protocol channel on the cloud end, and uses the power time series database for persistent storage, thereby constructing a structured data foundation indexed by the power distribution terminal identifier; the storage mode adapts to the writing characteristics of high-frequency sampling data of the power system, supports efficient retrieval and backtracking of massive historical operating conditions, and provides a complete data basis for data-driven training of subsequent energy efficiency decision-making models.

[0061] Further, the specific manner of updating the energy efficiency optimization service package comprises: monitoring a data aggregation size of the power time sequence database, and extracting historical running interaction data containing state-action pairs and performing offline retraining on the deep reinforcement learning-based energy efficiency decision model when the data aggregation size meets a preset retraining threshold; the historical running interaction data comprises corresponding running electrical parameters, ring main unit device states, historically executed target control parameters and calculated energy efficiency evaluation indexes at the same time section; and the specific manner of issuing to the edge computing unit comprises: checking a version number of the energy efficiency decision model after offline retraining, and issuing an incremental update package of the deep reinforcement learning-based energy efficiency decision model to the edge computing unit through a cloud-edge collaborative container orchestration architecture.

[0062] Specifically, the cloud end constructs a data-driven model self-evolution closed loop. The system monitors the data aggregation size of the power time sequence database, and triggers iteration when the number of new samples meets the preset retraining threshold, for example, reaches 100,000 samples. The system extracts historical running interaction data containing state-action pairs, which is formatted as experience replay tuples, and explicitly includes running electrical parameters, device states, historical execution actions and energy efficiency evaluation indexes. The cloud end uses these data to perform offline retraining on the deep Q network architecture to update the weights. After training is completed, the system checks the model version number, and if the version number is updated, the cloud end first locks the latest model weight parameters generated in this training, and reads the old model weight parameters issued to the edge side in the last version. Using the parameter difference strategy, the weight values of the two versions of the model at the corresponding network layer positions are subjected to element-level subtraction operation, thereby obtaining the weight change amount. The weight change amount is packaged as an incremental update package, and is issued to the edge computing unit through a cloud-edge collaborative container orchestration architecture, such as a KubeEdge architecture. After receiving, the edge computing unit parses the weight change amount, and the energy efficiency optimization service container in the edge computing unit is preloaded with a model hot update agent process. After receiving the weight incremental update package (.diff file) issued by the cloud end, the agent process reads the weight matrix of the current running model, performs matrix addition operation, and reloads the updated weight into the inference engine memory, thereby completing online synchronization of the model parameters without restarting the container.

[0063] As a preferred embodiment, the formatting and pushing process combines a dead zone compression strategy based on the rotating door algorithm. The edge computing unit locally maintains a dynamic dead zone threshold, and calculates the change rate of the running electrical parameters in real time; only when the data change amplitude exceeds the dead zone threshold, Protobuf serialization and MQTT pushing are triggered; otherwise, the data is cached locally. This technical solution realizes an on-demand transmission mode, eliminates steady-state redundant data while retaining key transient characteristics, and saves communication traffic and cloud database storage space.

[0064] The application performs offline retraining by using historical interaction data containing state-action pairs, builds a data-driven model iteration closed loop, and enables the algorithm to adapt to the time-varying nature of the load characteristics of the power distribution network. At the same time, the incremental update package of the model is issued under the cloud-edge collaborative architecture, reducing the data transmission amount in the model deployment process, and realizing the version update and maintenance of the edge side inference model under the premise of occupying less network resources.

[0065] Further, the specific way of packaging the remote device control instruction includes: converting the reactive power compensation device switching instruction and the transformer tap position instruction into an instruction frame conforming to the Modbus-TCP protocol format; and the specific way of sending to the controlled device includes: sending the instruction frame to the programmable logic controller in the ring network box through an industrial Ethernet interface to remotely control the terminal device over the network.

[0066] Specifically, the real-time control module adopts a hierarchical asynchronous control strategy. The edge computing unit issues an optimization instruction in advance of a control cycle based on short-term load forecasting, offsetting the time delay accumulation of network transmission and device action. At the same time, millisecond-level safety interlocking logic is implanted in the programmable logic controller. When receiving the edge side instruction, the controller first checks the current bus voltage instantaneous value. If the action may cause instantaneous overrun, it will refuse to execute and trigger local protection, thereby ensuring the physical safety of the system under the non-real-time communication link. The real-time control module executes protocol conversion and physical driving according to the pre-set mapping table. The real-time control module converts the target control parameters into an instruction frame conforming to the Modbus-TCP protocol format, which contains the target register address and function code. Then, the module sends the instruction frame to the programmable logic controller (PLC) in the ring network box through an industrial Ethernet interface. The controller parses the instruction and drives the contactor to attract or the servo motor to rotate, thereby physically adjusting the capacitor bank and transformer tap, realizing remote network control and energy efficiency closed-loop optimization of the terminal device.

[0067] The application converts the target control parameters into an instruction frame conforming to the Modbus-TCP protocol standard, realizing the interoperability between algorithm decision results and industrial control devices. The programmable logic controller is driven through an industrial Ethernet interface, ensuring that the digital control strategy can be correctly parsed and executed by the existing underlying hardware, and realizing remote network control of the reactive power compensation device and the transformer tap.

[0068] Embodiment Two

[0069] This embodiment shows an application method of a cloud-edge collaborative primary and secondary fusion integrated ring network box energy efficiency optimization system provided by the application in a high-tech industrial park power distribution network with distributed photovoltaic access; refer to Fig. 2, the specific steps are as follows:

[0070] Further, the photovoltaic grid-connected point and the low-voltage side bus of the ring network box are accessed in real time through the intelligent instrument interface to read real-time data with a sampling period of 100 milliseconds; the real-time data is divided into two categories, one is the operating electrical parameters representing the photovoltaic reverse power flow and the load characteristics, specifically including four-quadrant active and reactive power that can reflect bidirectional energy flow, three-phase voltage amplitude of the grid-connected point, and total harmonic distortion rate of the current introduced by the photovoltaic inverter; the other is the ring network box equipment state data representing the physical working condition, specifically including the current gear of the on-load voltage regulating transformer, the switching code of the reactive power compensation capacitor bank, and the cabinet temperature and humidity environment value; using the preset energy internet protocol template, the above multi-source heterogeneous data is mapped into a unified key-value pair structure and encapsulated into a JSON format data frame to form a state data packet containing photovoltaic output characteristics; then the unique identifier of the high-tech zone power distribution terminal stored locally is read to generate an interface call instruction containing the identifier and the state data packet, the instruction explicitly points to the "source network load storage coordination optimization" business interface in the local containerized microservice, and is used as a service call request to trigger the subsequent process.

[0071] Further, a deep reinforcement learning energy efficiency decision model based on physical information constraints deployed in containerized microservices is adopted. In view of the strong randomness of photovoltaic output and high risk of voltage overrun, the bus voltage deviation rate, photovoltaic instantaneous penetration rate and line load rate are taken as the state input of the model, and the switching combination of the reactive power compensation device and the lifting instruction of the transformer tap are taken as the action output. In the algorithm logic, a reward function is constructed to represent the constraints and objectives: the voltage allowable deviation range specified in the grid national standard is configured as the boundary penalty term in the reward function; when the photovoltaic back-feeding during high-illumination period causes the predicted voltage to possibly break through the upper limit, a high negative penalty is applied to the model, forcing the policy to preferentially adjust the transformer tap or remove the capacitor. In addition, the energy efficiency evaluation index is configured as the target return item in the reward function, and its value is obtained by calculating the weighted sum of the line loss reduction and the power factor improvement per unit of power supply, wherein the setting of the weight coefficient tends to reduce the line loss caused by the reverse flow of the tidal current, while taking into account the power factor assessment; the model also integrates a physical information neural network architecture to construct a composite loss function containing a reinforcement learning main loss term and a physical constraint regularization term. The main loss term is calculated based on the time difference error between the value score predicted by the deep Q network and the target value score; the physical constraint regularization term is determined by calculating the deviation between the voltage prediction value implied by the network output and the theoretical voltage value derived from Kirchhoff's voltage law. The two are weighted and summed by introducing a preset weight penalty coefficient, thereby embedding the law of conservation of physics into the gradient descent process, so that the neural network can automatically converge to the feasible region that meets the principles of circuit physics while optimizing the energy efficiency strategy.

[0072] Further, in the initialization phase, the adjustable physical boundary of the ring network box is defined, the adjustment range of the on-load voltage regulating switch and the number of capacitor groups of the reactive power compensation switching device are read, and the discrete state set of each is constructed with the smallest adjustment action unit as the step, and the candidate action library containing all physical feasible combinations is constructed through Cartesian product operation, that is, the device control parameter space. In the real-time reasoning phase, the current grid operating electrical parameters containing a high proportion of photovoltaic access are collected, and after normalization processing, the state vector is input into the energy efficiency decision model; the value score (Q value) of each group of parameters in the candidate action library is calculated through the deep Q network architecture, which represents the expected return of executing the action under the current photovoltaic output level to smooth voltage fluctuations and reduce line loss; the greedy strategy or the maximum strategy is applied to lock the group of data with the maximum score to determine the target control parameters, for example, during the period of high photovoltaic output, the transformer tap is instructed to lower the gear and part of the capacitor is removed to pre-compensate for the voltage lifting effect.

[0073] Further, according to the Protocol Buffers protocol standard, a description file is written to define the structure Schema of the high-tech park power data, and the floating-point data such as bidirectional power and voltage harmonics and the enumerated data such as switch state are allocated unique Tag numbers, and the state data packet is mapped and compressed into a binary byte stream form of synchronous payload. In the data offloading stage, a communication link with the energy Internet of Things cloud platform is built based on the MQTT message transmission protocol, and the system is used as a client to publish data, and a dead zone compression transmission strategy based on a rotating door algorithm is adopted: the system maintains a dynamic dead zone threshold associated with the volatility of photovoltaic power, and only when the change rate of the collected data exceeds the threshold due to sudden changes in light (such as cloud cover) does the serialization and message pushing trigger; otherwise, the stable data is cached locally, and the cloud platform subsequently restores the full curve by interpolation, which greatly reduces the network overhead caused by continuous uploading while capturing the transient characteristics of photovoltaic power.

[0074] Further, relying on the KubeEdge open source edge computing framework to deploy the cloud service orchestration module, the cloud as the control plane and the edge side ring network box terminal as the computing node, the energy efficiency optimization service package is encapsulated as a Docker container image for unified orchestration. In the data aggregation link, the cloud subscribes to all park terminals through the MQTT wildcard, receives and analyzes the synchronous payload, extracts the terminal identifier and time series data, and stores them in the power time series database, thereby accumulating massive photovoltaic output curve and load characteristic data to provide a data foundation for analyzing the impact of photovoltaic power on the power grid.

[0075] Further, the power time series database is monitored, and when the newly accessed photovoltaic fluctuation data (such as new samples generated by seasonal light changes) reaches a preset retraining threshold, a data-driven self-evolution closed-loop mechanism is constructed, an automatic model iteration pipeline is triggered, historical operation interaction data containing state-action pairs are extracted, and offline retraining is performed on the energy efficiency decision model; the historical data explicitly contains operation electrical parameters, device states, historically executed target control parameters and energy efficiency evaluation indexes in the same time section. The update process adopts a knowledge distillation strategy based on a deep Q network. In this embodiment, the core component in the DQN architecture, the deep neural network used to fit the action value function, is designed as an asymmetric teacher-student structure to solve the problem of limited inference resources on the edge side. Specifically, a teacher model based on an improved deep residual network (1D-ResNet) is maintained on the cloud, which modifies the first layer convolution kernel from a 7x7 two-dimensional convolution to a one-dimensional time series convolution with a kernel size of 7 to adapt to the power time series data with an input tensor shape of (time step, feature dimension). The student model on the edge runs based on a lightweight fully connected network (3 layers of Hidden Layer). The teacher model is trained on the global data until convergence, then a KL divergence term between the soft labels output by the teacher model and the prediction probability distribution output by the student model is introduced into the loss function, and a distillation temperature coefficient is set

[0076] Further, according to the pre-set PLC address mapping table, the target control parameters output by the model containing reactive compensation switching instructions and transformer tap instructions are packaged into Modbus-TCP instruction frames, and sent to the programmable logic controller in the ring network box through the industrial Ethernet interface. After the PLC is parsed, the contactor is driven to act or the servo motor is rotated through the DO output, and the transformer gear adjustment or capacitor switching is physically executed. This process realizes the closed loop from cloud AI decision to physical device execution, ensures that the power grid is in an optimal operating state through preventive adjustment before the voltage exceeds the limit caused by the dramatic change of photovoltaic output, and completes the value evaluation and execution of the discrete action space by the deep Q network architecture.

[0077] The application trains and pushes the cloud edge collaborative architecture by deploying an edge computing unit in the intelligent power distribution terminal and combining a cloud service arrangement module to realize iterative updating of the energy efficiency optimization service package. Compared with full cloud centralized control, the local reasoning based on the local service package effectively reduces the influence of wide area network communication delay and link fluctuation on control real-time performance, and overcomes the physical bottleneck that the embedded device on the edge side cannot bear complex model training tasks due to limited computing power and storage resources, ensuring adaptive updating and accurate execution of the energy efficiency optimization strategy.

[0078] Although embodiments of the present application have been shown and described, it is to be understood that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the present application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A cloud-edge collaborative primary and secondary fusion integrated ring network box energy efficiency optimization system, characterized in that, Comprise: Edge device deployment and data acquisition module: deploy edge computing unit in intelligent power distribution terminal; Real-time receive power grid operation electric parameters and ring main unit equipment state data, integrate into state data package; and generate corresponding service call request according to state data package; Service call module: call energy efficiency optimization service package embedded in edge computing unit based on the service call request, The energy efficiency optimization service package is an independent running instance encapsulated as a containerized microservice, which integrates an energy efficiency decision model based on deep reinforcement learning and an inference environment for supporting the model running inside; The energy efficiency decision model adopts deep Q network architecture, and the reward function is constructed to represent the energy efficiency evaluation index and the power grid voltage and load safety constraint condition; wherein, the power grid voltage and load safety constraint condition is configured as a boundary penalty term in the reward function, and when the monitored bus voltage deviation rate exceeds the preset threshold, a negative value penalty is applied to the energy efficiency decision model; The energy efficiency evaluation index is configured as the target return item in the reward function, and its specific value is the weighted sum of line loss reduction rate and power factor improvement rate per unit power supply period; Under the condition of meeting the preset power grid voltage and load safety constraint condition, a group of data is selected as the target control parameter in the device control parameter space representing the device operating state; The specific way of selecting a group of data as the target control parameter includes: obtaining the allowable adjustment range of the on-load voltage regulating switch associated with the ring main unit and the available switching group number of the reactive power compensation switching device, and constructing the respective discrete state set with the minimum adjustment action unit as the step; Perform Cartesian product operation on each of the discrete state sets to generate a candidate action library containing all feasible combinations as the device control parameter space; The current state data package is constructed into a state vector and input into the energy efficiency decision model based on deep reinforcement learning; Evaluate the value score of each group of parameters in the candidate action library through the energy efficiency decision model, and output the target control parameter according to the preset action selection strategy, which contains reactive power compensation switching instruction and transformer tap position instruction Data synchronization module: obtain the state data package and format it into a synchronization payload, and unload the synchronization payload to the energy internet of things cloud platform through the MQTT message transmission protocol; Cloud service orchestration module: collect the synchronization payload and update the energy efficiency optimization service package as feedback data, and download the updated energy efficiency optimization service package to the edge computing unit through the network according to the preset strategy; Real-time control module: encapsulate the target control parameter into a remote device control instruction, and send the remote device control instruction to the controlled device in real time.

2. The cloud-edge collaborative primary and secondary fusion ring main unit energy efficiency optimization system according to claim 1, characterized in that, The edge computing unit is an intelligent computing terminal integrating a central processor, a memory and a network communication interface, and has a container runtime environment arranged therein to provide an isolated computing resource and a running base for the energy efficiency optimization service package; the specific manner of integrating into the state data package includes: acquiring power grid operation electrical parameters including three-phase voltage, three-phase current, active power and reactive power, and ring network box equipment state data including switch opening and closing position, energy storage state and environmental temperature and humidity, and encapsulating the operation electrical parameters and the ring network box equipment state data into a data frame of a structured data object, and taking the data frame as the state data package; the specific manner of generating a corresponding service call request includes: generating an interface call instruction containing a power distribution terminal identifier and the state data package, and taking the interface call instruction as the service call request. 3.The cloud-edge collaborative primary and secondary fusion ring main unit energy efficiency optimization system according to claim 1, characterized in that, The specific manner of formatting into a synchronization payload includes: extracting data in the state data package and encoding in a protobuf serialization data format to generate the synchronization payload; the specific manner of unloading to an energy internet of things cloud platform includes: establishing a message subscription and publishing channel with the energy internet of things cloud platform based on an MQTT message transmission protocol, and pushing the synchronization payload through the channel.

4. The cloud-edge collaborative primary and secondary fusion ring main unit energy efficiency optimization system according to claim 1, characterized in that, The specific manner of collecting synchronization payloads includes: subscribing to the channel of the MQTT message transmission protocol on the energy internet of things cloud platform, receiving and analyzing the synchronization payloads, and storing the analyzed data into a power time series database according to power distribution terminal identifiers.

5. The cloud-edge collaborative primary and secondary fusion ring main unit energy efficiency optimization system according to claim 1, characterized in that, The specific manner of updating the energy efficiency optimization service package includes: monitoring the data aggregation size of the power time series database, and in the case that the data aggregation size meets a preset retraining threshold, extracting historical running interaction data containing state-action pairs, and performing offline retraining on the energy efficiency decision model based on deep reinforcement learning; the historical running interaction data includes corresponding operation electrical parameters, ring network box equipment states, historical executed target control parameters and calculated energy efficiency evaluation indexes at the same time section; the specific manner of issuing to the edge computing unit includes: checking the version number of the energy efficiency decision model after offline retraining, and issuing an incremental update package of the energy efficiency decision model based on deep reinforcement learning to the edge computing unit through a cloud-edge collaborative container orchestration architecture. 6.The cloud-edge collaborative primary and secondary fusion ring main unit (RMU) energy efficiency optimization system of claim 1, wherein, The specific manner of encapsulating into a remote device control instruction includes: converting reactive power compensation device switching instructions and transformer tap position instructions into instruction frames conforming to the Modbus-TCP protocol format; the specific manner of sending to the controlled device includes: sending the instruction frames to the programmable logic controller in the ring network box through an industrial Ethernet interface to remotely control the terminal device over the network.

Images