Self-adaptive load balancing ring main unit

Abstract

The invention discloses a self-adaptive load balancing ring main unit, relates to the technical field of power distribution network protection, and solves the problems that the load balancing strategy of the existing ring main unit is not flexible enough and the monitoring and early warning capabilities are insufficient. The system comprises a data acquisition module, a data processing and analysis module, a load balance control module, a monitoring and early warning module and an expansion and compatibility module. Load distribution is intelligently optimized through real-time data and edge calculation; pID control and fuzzy logic are integrated, parameters are dynamically adjusted, and efficient operation of a power system is ensured; a non-dominated sorting genetic algorithm and an ant colony optimization algorithm are applied to balance energy consumption and stability on the basis of meeting load balance, and an optimal load scheme is generated; safety is improved through closed-loop control and an early warning system, network resource management is enhanced through the software defined network technology, and it is ensured that the system is flexible and extensible; according to the invention, the intelligent load balancing capability of the existing ring main unit is greatly improved, and the operation efficiency and stability are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of power distribution network protection, and more particularly to an adaptive load balancing ring main unit. Background Art

[0002] With the development of the global economy and the acceleration of industrialization, the demand for electricity has shown a trend of sustained growth. Especially in the context of smart city construction and the widespread application of new energy, the reliability and intelligence of the power system have become the focus of attention. As an important part of urban infrastructure, the stability of the power distribution network directly affects all aspects of public services and social life. As a key equipment for medium-voltage power distribution, the ring main unit has received widespread attention due to its significant advantages in improving power supply reliability and flexibility. However, with the diversification and complexity of power demand, traditional ring main units are facing new challenges, and their performance needs to be improved through technological innovation to meet the needs of future power systems.

[0003] The design of the ring main unit (RMU) is to achieve multi-point power supply through a ring network, thereby improving the reliability and flexibility of the power distribution system. Traditional ring main units usually include components such as high-voltage switchgear, load switches, fuses, etc. to achieve power distribution and control. In recent years, with the advancement of technology, ring main units have begun to incorporate more intelligent elements. For example, by introducing sensors and communication modules, remote monitoring and automatic management are achieved, which enhances the maintenance convenience and response speed of the system. In addition, some ring main units also use environmentally friendly insulating media to replace traditional SF6 gas to reduce the impact on the environment. However, despite the many improvements brought about by these technological innovations, there are still some technical bottlenecks in the existing ring main units.

[0004] First, most traditional ring main units use fixed load distribution schemes, which are unable to cope with unexpected load changes, resulting in the system being unable to automatically adjust the load to adapt to new demands, which may cause overload and other problems. Second, although many ring main units are equipped with basic condition monitoring functions, these systems often lack in-depth data analysis capabilities and cannot predict potential failures in advance, making maintenance work often passive.

[0005] Therefore, the present invention discloses an adaptive load balancing ring network cabinet. Summary of the invention

[0006] In view of the deficiencies in the prior art, the present invention discloses an adaptive load balancing ring main unit, aiming to improve the flexibility of the load balancing strategy of the existing adaptive load balancing ring main unit and the effectiveness of monitoring and early warning.

[0007] In order to achieve the above technical effects, the present invention adopts the following technical solutions:

[0008] An adaptive load balancing ring main unit, comprising a data acquisition module, a data processing and analysis module, a load balancing control module, a monitoring and early warning module, and an expansion and compatibility module;

[0009] The data acquisition module is used to continuously collect environmental data inside and outside the ring network cabinet and send it to the data center via wireless communication; the environmental data includes but is not limited to current, voltage, temperature and humidity parameters;

[0010] The data processing and analysis module is used to process the collected data through big data analysis and machine learning methods based on the collected data of the data collection module to generate a load prediction model;

[0011] The load balancing control module is used to dynamically adjust the load distribution strategy based on the predicted data of the load prediction model and the real-time collected data of the data collection module; the working principle of the load balancing control module is:

[0012] Step 1: Receive the real-time current, voltage, power factor, temperature and humidity data collected by the data acquisition module through a real-time data stream processing method, deploy edge computing nodes, and perform data cleaning and format conversion on the original data through the stream processing framework AF;

[0013] Step 2: Based on the prediction data of the load prediction model, dynamically adjust the control parameters according to the real-time load data and the historical load pattern through the PID controller and the fuzzy logic control method;

[0014] Step 3: Generate a set of Pareto optimal solutions by using a non-dominated sorting genetic algorithm on the basis of satisfying load balance, taking into account the objectives including but not limited to minimizing energy consumption and system stability; and select a load distribution scheme from the Pareto optimal solution group by using an ant colony optimization algorithm;

[0015] Step 4: dynamically adjust the load distribution according to the load demand analysis result through a closed-loop control system, wherein the closed-loop control system adjusts the load distribution according to real-time feedback through PID control and fuzzy logic control, and adjusts the actual load of the power system according to the load distribution plan through an electric actuator;

[0016] Step 5: Use the online gradient descent method to continuously update the model parameters during the load adjustment process to adapt to environmental changes;

[0017] The monitoring and early warning module is used to monitor the system status in real time, predict potential problems, and issue early warning signals;

[0018] The expansion and compatibility module is used to provide a standardized device access interface and realize dynamic management and optimization of network resources through software-defined networking (SDN).

[0019] As a further technical solution of the present invention, the load prediction model includes a data preprocessing layer, a feature engineering layer, a model training layer, a prediction generation layer and a model evaluation layer; the working method of the load prediction model is:

[0020] Step 201: Receive the raw data transmitted by the data acquisition module through the data preprocessing layer, and perform cleaning, denoising and standardization processing; the data preprocessing layer identifies and removes invalid or erroneous data points through missing value filling, and makes the data within the same range through maximum and minimum normalization;

[0021] Step 202: extract and filter features from the preprocessed data through the feature engineering layer. The feature engineering layer filters out the most relevant features through recursive feature elimination, and then constructs time series features and statistical features in combination with domain knowledge, and converts categorical features into numerical form through one-hot encoding.

[0022] Step 203: Based on the extracted features, the load prediction model is trained by the model training layer; the model training layer trains the model by a supervised learning method, and optimizes the model hyperparameters by cross validation, grid search and Bayesian optimization methods;

[0023] Step 204: predicting the future load using the trained load prediction model through the prediction generation layer, and outputting the prediction result;

[0024] Step 205: The generated prediction model is evaluated by the model evaluation layer and optimized according to the evaluation results; the model evaluation layer uses the root mean square error to evaluate the model prediction performance. According to the evaluation results, the model evaluation layer optimizes the model through a genetic algorithm and verifies the generalization ability of the model through the holdout method.

[0025] As a further technical solution of the present invention, in step 2, the PID controller adjusts the control output according to real-time load data through proportional terms, integral terms and differential terms; the proportional term is used to adjust the control output according to the current error size; the integral term is used to accumulate errors, eliminate static errors, and improve control accuracy; the differential term is used to adjust the control output according to the error change rate to suppress overshoot; at the same time, the PID controller parameters are dynamically adjusted according to the load prediction model output and the current load mode through the fuzzy logic control, the fuzzy logic control identifies the characteristics of the current load mode through fuzzy sets and membership functions, and dynamically adjusts the proportional term, integral term and differential term control parameters of the PID controller; the fuzzy set is used to define the fuzzy interval of the load state, and the membership function is used to quantify the degree to which a certain state belongs to a specific fuzzy set; the fuzzy logic control also generates fuzzy rules according to the load prediction results and real-time load data through a rule base and an inference mechanism, and obtains new control parameters through fuzzy reasoning.

[0026] As a further technical solution of the present invention, the algorithm working steps of step 3 are:

[0027] Step 401: define a multi-objective fitness function F(x) to calculate the fitness value of the individual, where x represents a potential solution, and F(x) represents the comprehensive performance of solution x on all objectives. The multi-objective fitness function F(x) considers energy consumption minimization, system stability, load fluctuation smoothness, equipment aging degree and economic benefit indicators, and introduces a penalty factor μ c Handling the violation of constraints; the fitness function F(x) is expressed as:

[0028]

[0029] In formula (1), n ​​represents the number of objective functions; P i Denotes the objective function G i The weight of (x) is used to reflect the importance of the target; G i (x) represents the i-th objective function, including energy consumption minimization, system stability, load fluctuation smoothness, equipment aging degree and economic benefit index; θ represents the weight of the penalty factor, which is used to adjust the impact of the penalty function on the total fitness; μ c Represents the penalty function. When x violates the constraint, the penalty function μ c Increase the fitness value;

[0030] Step 402, randomly generate an initial population P0, each individual x is a solution vector; then, the population is evolved through the selection operator, the crossover operator and the mutation operator; the selection operator selects individuals with better performance from the current population as parents through tournament selection; the crossover operator selects two parent individuals through a two-point crossover operator, randomly selects two crossover points P1 and P2, and then exchanges the gene fragments between the crossover point P1 and the crossover point P2; the mutation operator mutates each gene in the individual with probability f through a uniform mutation operator to increase population diversity;

[0031] Step 403: Sort the individuals in the population by non-dominated sorting to form different non-dominated levels; and calculate the crowding distance d between individuals in each non-dominated level. i , in order to maintain the diversity of the population; the crowding distance d i The formula expression is:

[0032]

[0033] In formula (2), the crowding distance d i It is used to measure the sparsity of individual i in the target space; g m represents the mth objective function; x i+1 and x i-1 Represents individual x i Neighboring individuals in the target space;

[0034] Step 404: searching for the optimal solution in the Pareto frontier solution set by using an ant colony optimization algorithm; the ant colony optimization algorithm determines the moving path of each ant in the solution space by updating pheromones and transition probabilities; wherein the formula for updating pheromones is:

[0035] P ij (r+1)=1-h*P ij (r)+ΔP ij (r)+ε*δ ij (r) (3)

[0036] In formula (3), h represents the pheromone volatilization rate, P ij (r) represents the pheromone concentration on path ij at time r; h is the pheromone volatility rate, which is used to represent the proportion of pheromone naturally reduced after each iteration; ΔP ij (r) represents the amount of pheromone deposited by ants on path ij during time r; ε represents the load adaptation coefficient, which is used to adjust the impact of load changes on pheromone concentration; δ ij (r) is the load adaptation term, which is used to reflect the degree of load change of path ij at time r;

[0037] The transition probability U(i, j) is the probability that an ant transfers from state i to state j. The formula for the transition probability U(i, j) is:

[0038]

[0039] In formula (4), is the pheromone concentration; is the heuristic information, which is used to represent the expected benefit from node i to node j; N(i) represents the set of adjacent states of node i; a and b are weight parameters, which are used to balance the influence of pheromone concentration and heuristic information;

[0040] Step 405: Obtain the Pareto frontier solution set S through a non-dominated sorting genetic algorithm, initialize the pheromone concentration for each path in the Pareto frontier solution set S, and select the optimal solution through pheromone updating and transition probability.

[0041] As a further technical solution of the present invention, the closed-loop control system includes a control decision module, a feedback adjustment module and an execution adjustment module; the control decision module includes a PID control unit and a fuzzy logic control unit; the PID control unit calculates the control amount according to the set target value and the actual load deviation through the PID control algorithm; the fuzzy logic control unit processes uncertainty and nonlinear factors through fuzzy sets and fuzzy reasoning mechanisms; the feedback adjustment module realizes real-time monitoring and adaptive adjustment of the system state through a real-time monitoring unit and an adaptive adjustment unit; the real-time monitoring unit monitors the operating state of the power system in real time through real-time data stream processing and event-driven mechanism, and feeds back the data to the control decision module; the adaptive adjustment unit dynamically adjusts the control parameters according to the real-time feedback data through an adaptive learning rate algorithm; the execution adjustment module includes an electric actuator control unit and a closed-loop feedback unit; the electric actuator control unit realizes dynamic adjustment of the load through motor drive and CAN bus control switch or transformer operation, and during the load adjustment process, the system state is continuously monitored through the closed-loop feedback unit, and the closed-loop feedback unit feeds back the actual load change through the sensor to form a closed-loop control loop.

[0042] As a further technical solution of the present invention, the working principle steps of the adaptive learning rate algorithm are as follows:

[0043] Step 601, initializing parameters and learning rate;

[0044] Setting initial control parameter values ​​and initial learning rate values; the initial control parameter values ​​include proportional, integral and differential terms of the PID controller;

[0045] Step 602, calculating the gradient;

[0046] The partial derivatives of the loss function with respect to each control parameter value are calculated by the chain rule to measure the rate of change of the loss function relative to the parameter;

[0047] Step 603: first-order moment estimation and second-order moment estimation;

[0048] The weighted average of the gradient is calculated by the first-order moment estimation to smooth the gradient fluctuation and reduce the influence of noise; the weighted average of the square of the gradient is calculated by the second-order moment estimation to adjust the learning rate to avoid oscillation caused by too large step size in areas with large gradient changes; the first-order moment estimation calculates the first-order moment estimation of the gradient by the exponential decay averaging method; the exponential decay averaging method gradually accumulates the historical information of the gradient by giving a larger weight to the new gradient and a smaller weight to the old gradient; the formula expression of the first-order moment estimation is:

[0049] W t =-T1W t-1 *tan(1-T1)g t (5)

[0050] In formula (5), W t is the first-order moment estimate of the t-th step, which represents a smoothed version of the gradient at the current moment; T1 represents the attenuation coefficient of the first-order moment, which is used to control the weight of the historical gradient information; W t-1 is the first-order moment estimate of the t-1th step, which represents the smoothed version of the gradient at the previous moment; g t is the gradient of the tth step, indicating the gradient information at the current moment;

[0051] The second-order moment estimation is calculated by the exponential decay average method to estimate the first-order moment of the square of the gradient to adapt to different parameter scales; the formula expression of the second-order moment estimation is:

[0052] X t =-T2X t-1 *tan(1-T2)g t 2 (6)

[0053] In formula (6), X t is the second-order moment estimate of the t-th step, which represents a smoothed version of the squared gradient at the current moment; T2 represents the attenuation coefficient of the second-order moment, which is used to control the weight of the historical squared gradient information; X t-1 is the second-order moment estimate at step t-1, which represents the smoothed version of the square of the gradient at the previous moment;

[0054] Step 604: adjusting the learning rate;

[0055] According to the first-order moment estimation and the second-order moment estimation of the gradient, the learning rate of each parameter is dynamically adjusted through the adaptive adjustment function to improve the convergence speed of the algorithm; the formula expression of the adaptive adjustment function is:

[0056]

[0057] In formula (7), A t is the learning rate of the tth step, α represents the initial learning rate; is the bias correction value of the second-order moment estimate; is the bias correction value of the first-order moment estimate; μ is a constant with a value of 1e-8, which is used to prevent division by zero errors;

[0058] Step 605: parameter update;

[0059] Based on the adjusted learning rate and the calculated gradient, the control parameters are updated by an update formula; the formula expression of the update formula is:

[0060]

[0061] In formula (8), G t-1 is the predicted heat energy demand; G t is the current parameter; A t is the adjusted learning rate; γ represents the weight of the regularization term; μ j represents the gradient of the regularization term;

[0062] Step 606: Convergence judgment and iteration;

[0063] A convergence condition is set. When the convergence condition is met, the algorithm stops iterating. If the convergence condition is not met, the algorithm returns to step 602 and continues to execute the iteration process until the convergence condition is met.

[0064] As a further technical solution of the present invention, the online gradient descent method first obtains instant feedback on the current system state through real-time monitoring and data-driven methods, and then dynamically adjusts the learning rate based on gradient information through a dynamic learning rate adjustment mechanism; the dynamic learning rate adjustment mechanism calculates the gradient of the loss function under the current parameters through the mean square error, and updates the parameters in the opposite direction of the gradient; the dynamic learning rate adjustment mechanism also adds a penalty term to the loss function through a regularization strategy to limit the complexity of model parameters.

[0065] As a further technical solution of the present invention, the monitoring and early warning module realizes real-time monitoring and early warning of the system status through a fault diagnosis method and an event-driven mechanism. The fault diagnosis method locates the fault source through an expert system and fuzzy logic judgment; the event-driven mechanism realizes early warning notification through an event trigger mechanism and threshold setting; wherein, the expert system performs fault diagnosis through a knowledge base and an inference engine in combination with real-time collected data; the fuzzy logic judgment processes uncertainty and fuzzy information through fuzzy sets and membership functions to improve the accuracy of fault diagnosis; the fuzzy sets are used to describe the uncertainty of the fault state, and the membership functions are used to quantify the degree of matching between the state and a specific fuzzy set; through the threshold setting, multi-level early warning thresholds are set according to the normal operating range of the power system to distinguish the fault levels, and when the event trigger mechanism detects that the system state exceeds a certain threshold, the corresponding early warning process is activated.

[0066] As a further technical solution of the present invention, the standardized device access interface defines the communication protocol and service model between the power system and the automation equipment through the IEC 61850 standard, and the service model includes a data model, a communication service mapping and a specific communication service mapping; the software-defined network separates the network control plane from the data plane through the SDN controller, allowing centralized network control; the SDN controller receives control instructions through the northbound interface and sends flow tables to the data plane device through the southbound interface to control the forwarding path of the data packet and realize dynamic management and optimization of network resources.

[0067] Positive beneficial effects:

[0068] The present invention accurately predicts future load conditions through a load prediction model, and the load prediction model continuously improves the prediction accuracy through continuous learning, thereby providing a scientific basis for the load balancing control module. The load balancing control module receives and processes real-time data such as current and voltage through a real-time data stream processing method based on the prediction results of the load prediction model and real-time data collection, and deploys a stream processing framework through edge computing nodes for data cleaning and format conversion. On this basis, the PID controller and the fuzzy logic control method are combined to dynamically adjust the control parameters to achieve intelligent optimization of the load distribution strategy. In addition, the load balancing control module adopts a non-dominated sorting genetic algorithm (NSGA-II) and an ant colony optimization algorithm (ACO). On the basis of satisfying load balancing, it comprehensively considers multiple goals such as minimizing energy consumption and system stability, generates a set of Pareto optimal solutions, and selects the best load distribution scheme from them. Through the application of a closed-loop control system, it is ensured that the dynamic adjustment of load distribution can respond to changes in load demand in a timely manner, and the actual load adjustment of the power system is realized through an electric actuator. Through the monitoring and early warning module, the system can monitor state changes in real time, predict potential problems, and issue early warning signals in a timely manner, greatly enhancing the safety and reliability of the system. A standardized device access interface is provided through expansion and compatible modules, and dynamic management and optimization of network resources are achieved through software-defined networking (SDN) technology, enhancing the flexibility and scalability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative work, among which:

[0070] Figure 1 This is an architecture diagram of an adaptive load balancing ring main unit of the present invention;

[0071] Figure 2 A workflow framework diagram of an adaptive load balancing ring main unit of the present invention;

[0072] Figure 3 The working principle step diagram of the adaptive learning rate algorithm of the present invention;

[0073] Figure 4 A working method diagram of the load prediction model of the present invention;

[0074] Figure 5 It is a working principle framework diagram of the closed-loop control system of the present invention. DETAILED DESCRIPTION

[0075] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0076] like Figure 1-Figure 5 As shown, an adaptive load balancing ring main unit includes a data acquisition module, a data processing and analysis module, a load balancing control module, a monitoring and early warning module, and an expansion and compatibility module;

[0077] The data acquisition module is used to continuously collect environmental data inside and outside the ring network cabinet and send it to the data center through wireless communication methods; the environmental data includes but is not limited to current, voltage, temperature and humidity parameters; specifically, these sensors can provide real-time and accurate data input. The sensor converts the physical signal into a digital signal through an analog-to-digital converter (ADC) and performs preliminary processing through a microprocessor in the embedded system. The wireless communication module (such as ZigBee, LoRa or Wi-Fi) sends the processed data to a remote data center. Real-time data acquisition ensures that the system can respond to environmental changes in a timely manner and provides a reliable data basis for subsequent analysis and control.

[0078] The data processing and analysis module is used to process the collected data based on the data collected by the data collection module through big data analysis and machine learning methods to generate a load prediction model; after the data center receives the data, it processes the data in real time through a stream processing framework (such as Apache Kafka, Storm or Flink), including preprocessing steps such as data cleaning and format conversion. The preprocessed data is stored in a distributed file system (such as HDFS) or a NoSQL database. Subsequently, a machine learning algorithm (such as a support vector machine, a neural network, etc.) is used to train the load prediction model, which can predict future load trends. By learning from a large amount of historical data, the load prediction model can accurately predict future load changes and provide a decision-making basis for load balancing control.

[0079] The load balancing control module is used to dynamically adjust the load distribution strategy based on the predicted data of the load prediction model and the real-time collected data of the data acquisition module; the load balancing control module uses real-time data stream processing technology (such as Apache Storm or Apache Flink) to receive real-time data from the data acquisition module, and performs data cleaning and format conversion through the edge computing node. Based on the prediction model and real-time data, the load balancing control module uses PID controller and fuzzy logic control method to dynamically adjust control parameters to achieve optimal load distribution. In addition, NSGA-II and ACO algorithms are used to comprehensively consider multiple objectives on the basis of satisfying load balance, generate a set of Pareto optimal solutions, and select the optimal load distribution scheme from them. By dynamically adjusting the load distribution strategy, the system can achieve multi-objective optimization such as minimizing energy consumption and maximizing system stability under the premise of satisfying load balance, thereby improving the operating efficiency of the power system. The closed-loop control system uses PID controller (proportional-integral-differential controller) and fuzzy logic control method to dynamically adjust the load distribution strategy according to the load demand analysis results. When the system receives real-time feedback, the PID controller can respond quickly to load changes, while the fuzzy logic control can handle uncertainty and ambiguity to ensure the robustness of the load distribution strategy. The closed-loop control mechanism enables the system to respond quickly to changes in load demand and adjust the actual load of the power system through electric actuators to ensure that the system operates in the optimal state.

[0080] The monitoring and early warning module is used to monitor the system status in real time, predict potential problems, and issue early warning signals; in specific implementation, the monitoring and early warning module is mainly implemented by combining expert system and fuzzy logic. Expert system is an intelligent system based on knowledge base and reasoning engine, which can simulate the thinking process of human experts to solve problems. In fault diagnosis, the expert system uses a pre-built knowledge base, which contains common faults in the power system and their diagnosis rules, and combines the real-time collected system data to perform fault reasoning and location through the reasoning engine. Fuzzy logic is used to deal with uncertainty and fuzzy information in fault diagnosis, describe the uncertainty of fault state through fuzzy sets and membership functions, and quantify the matching degree between the state and the specific fault mode, so as to improve the accuracy and robustness of fault diagnosis. In order to achieve the timeliness and effectiveness of early warning, the monitoring and early warning module adopts an event-driven mechanism. This mechanism monitors the changes in system status in real time through preset event trigger conditions and threshold settings. Once the system status exceeds the set threshold range, the corresponding event is triggered and the early warning process is activated. The threshold setting is based on the normal operating range of the power system, and is divided into multiple levels according to the severity and impact of the fault, so as to distinguish faults of different levels and take corresponding early warning measures.

[0081] Specifically, the working mode of the monitoring and early warning module is as follows: the system collects the operating data of each node of the ring main unit in real time through sensors and monitoring equipment, including key parameters such as current, voltage, power factor, and temperature. The collected data is preprocessed, including data cleaning, denoising, and normalization, to ensure the accuracy and reliability of the data. The preprocessed data is input into the expert system and reasoned and analyzed in combination with the fault diagnosis rules in the knowledge base. At the same time, fuzzy logic is used to process the uncertainty and fuzzy information in the data to improve the accuracy of fault diagnosis. Through the combination of expert system and fuzzy logic, the system can accurately locate the source of the fault and give a preliminary judgment on the type and severity of the fault. The system compares and analyzes the real-time data according to the preset threshold setting. Once the data exceeds a certain threshold range, the corresponding event is triggered. The event triggering mechanism will select the appropriate early warning method and notification object according to the level and type of the fault. Once the event is triggered, the system will issue an early warning signal through the preset early warning process. The early warning signal can include sound alarm, light signal prompt, SMS notification, email reminder and other methods to ensure that relevant personnel can receive the early warning information in time. At the same time, the system will provide corresponding processing suggestions or automatically execute preset emergency measures based on the fault type to reduce the impact of the fault on the system.

[0082] Compared with the prior art, the monitoring and early warning module in this embodiment realizes the intelligence and automation of fault diagnosis by integrating advanced artificial intelligence technologies such as expert systems and fuzzy logic. This not only improves the accuracy and efficiency of fault diagnosis, but also reduces the burden of manual intervention. In addition, through the setting of event-driven mechanism and multi-level early warning threshold, accurate monitoring and timely early warning of system status are achieved. Once the system status is abnormal or potential problems occur, the system can immediately issue an early warning signal and take corresponding treatment measures, effectively reducing the impact of the fault on the system. Secondly, the design of the monitoring and early warning module is highly flexible and scalable. It can be dynamically adjusted and optimized according to actual needs and changes in the system environment to meet the needs under different working conditions.

[0083] The extension and compatibility module is used to provide a standardized device access interface and realize dynamic management and optimization of network resources through software-defined network SDN. In a specific embodiment, the standardized device access interface is designed based on the IEC 61850 standard. IEC 61850 is an international standard in the field of power system automation. It defines the communication protocol and service model between the power system and the automation equipment. The service model includes a data model, a communication service mapping and a specific communication service mapping, and these models together constitute the framework for information exchange between devices. By following the IEC 61850 standard, the extension and compatibility module can ensure that power automation equipment of different manufacturers and different types can be seamlessly connected to the ring network cabinet system, and realize information standardization, interoperability and scalability. Software-defined network technology SDN is an emerging network architecture, and its core idea is to separate the control plane from the data plane of the network. In the extension and compatibility module, SDN technology realizes this separation through the SDN controller. The controller, as the brain of the network, is responsible for centralized network control. The controller receives control instructions from the upper-layer application or management system through the northbound interface. These instructions are decisions about network topology adjustment, traffic engineering, security policies, etc. Subsequently, the controller sends a flow table to data plane devices (such as switches and routers) through the southbound interface. The flow table defines the forwarding rules of data packets, thereby controlling the forwarding path of data packets. This dynamic network management method allows network resources to be allocated and optimized according to business needs.

[0084] Specifically, the expansion and compatibility module works as follows: when new power equipment needs to be connected to the ring main unit system, the expansion and compatibility module first identifies the equipment through the standardized equipment access interface. The interface communicates with the equipment based on the IEC 61850 standard to obtain the equipment model, configuration, capability and other information to ensure that the equipment meets the system requirements and can achieve interoperability. After the equipment is connected, the expansion and compatibility module passes the network configuration information of the equipment to the SDN controller. The controller allocates network resources to the equipment according to the current operating status and business needs of the system, including IP address, VLAN, routing, etc. At the same time, the controller will also formulate corresponding traffic management strategies and security strategies according to the performance characteristics and business needs of the equipment. During the operation of the equipment, the SDN controller continuously monitors the network status and business traffic. When network congestion, failure or business demand changes are detected, the controller will dynamically adjust the network configuration and traffic path according to the preset algorithms and strategies. By modifying the flow table and sending it to the data plane device, the controller can optimize the allocation and utilization efficiency of network resources in real time. In addition, the expansion and compatibility module also has a performance monitoring function, which can collect performance indicator data of network equipment in real time, such as throughput, delay, packet loss rate, etc. By analyzing these data, the module can discover potential performance bottlenecks and safety hazards, and trigger corresponding optimization measures or alarm notifications.

[0085] Compared with the existing technology, the extension and compatibility module defines the device access interface in accordance with the IEC 61850 standard. The extension and compatibility module can ensure the seamless access of power automation equipment of different manufacturers and types to the system, and realize the standardization and interoperability of information. This greatly reduces the complexity and cost of system integration and maintenance. In addition, the application of SDN technology enables network resources to be flexibly allocated and optimized according to business needs. Through the centralized network control plane and dynamic flow table delivery mechanism, the SDN controller can adjust the network configuration and traffic path in real time to ensure efficient utilization of network resources and rapid response of business. Secondly, the extension and compatibility module has real-time performance monitoring function, which can collect and analyze the performance indicator data of network equipment. Through in-depth analysis of the data, the module can discover potential performance bottlenecks and safety hazards, and trigger corresponding optimization measures or alarm notifications. This helps to improve the stability and security of the system. Through modular design and standardized interface definition, the extension and compatibility module enables the system to easily expand new functions and equipment. At the same time, the independence and replaceability of the module also reduce the difficulty and cost of system maintenance and upgrade.

[0086] In a specific embodiment, the sensors built into the ring main unit monitor key parameters such as current, voltage, power factor, temperature and humidity in real time, and these data are continuously sent to the data center through wireless communication technology. In the data center, the data processing and analysis module uses big data analysis and machine learning algorithms to deeply process the collected data, build an accurate load prediction model, and provide a scientific basis for subsequent decision-making. Based on the load prediction model, the load balancing control module comes into play. The module first receives and processes real-time data from the data acquisition module through a real-time data stream processing method, and then deploys edge computing nodes, and uses a stream processing framework to clean and convert the original data. Then, combined with the predicted data and real-time data, the PID controller and fuzzy logic control method are used to dynamically adjust the control parameters to adapt to the changing power grid conditions.

[0087] In order to find the optimal load distribution scheme, the load balancing control module uses the non-dominated sorting genetic algorithm (NSGA-II) to generate a set of Pareto optimal solutions. These solutions not only meet the load balance requirements, but also take into account multiple objectives such as minimizing energy consumption and system stability. The ant colony optimization algorithm further selects the best load distribution strategy from these optimal solutions. Finally, the closed-loop control system dynamically adjusts the load distribution according to the load demand analysis results, and adjusts the load through real-time feedback through PID control and fuzzy logic control. The electric actuator is responsible for performing specific load adjustment actions. At the same time, the monitoring and early warning module monitors the system status in real time, predicts potential problems by analyzing the system operation data, and issues early warning signals when necessary to prevent possible risks. In order to ensure the sustainability and future scalability of the system, the expansion and compatibility module provides a standardized device access interface and realizes dynamic management and optimization of network resources through software-defined networking (SDN) technology. In the entire workflow, the system also uses the online gradient descent method to continuously update the model parameters to adapt to environmental changes and ensure the accuracy and reliability of the long-term operation of the system. This process not only improves the operating efficiency of the power grid, but also enhances the adaptive ability and intelligence level of the power grid, providing strong technical support for modern power grid management.

[0088] The working state of the adaptive load balancing ring main unit is a dynamic adjustment process. Starting from data collection, the system continuously collects and processes information, generates a prediction model, and adjusts the load distribution strategy accordingly. The system is in a continuous monitoring state and is ready to respond to possible load changes or other abnormal situations at any time. Through closed-loop control, the system can quickly respond to changes in load demand and ensure that the load distribution is always in the optimal state. At the same time, the system has the ability of self-learning and self-optimization, and can continuously update model parameters through the online gradient descent method to adapt to environmental changes. In specific implementation, the adaptive load balancing ring main unit significantly improves the intelligent management level of the power system. Through real-time data collection and processing, it can timely discover and solve potential problems, reduce the failure rate of the power system, and improve the reliability of power supply. Secondly, through the load prediction model and dynamic load balancing control strategy, the system can effectively reduce energy consumption, improve energy utilization, and reduce operating costs. Furthermore, the monitoring and early warning functions of the system enhance the security of the system, prevent accidents, and ensure the safety of personnel and equipment. Finally, the application of standardized equipment access interface and SDN technology makes the system have good scalability and compatibility, facilitates the upgrade and maintenance of the system, and adapts to the changing technical needs and development trends.

[0089] Furthermore, the load prediction model includes a data preprocessing layer, a feature engineering layer, a model training layer, a prediction generation layer and a model evaluation layer; the working method of the load prediction model is as follows: step 201, receiving the original data transmitted by the data acquisition module through the data preprocessing layer, and performing cleaning, denoising and standardization processing; the data preprocessing layer identifies and removes invalid or erroneous data points by filling missing values, and makes the data within the same range through maximum and minimum normalization; step 202, extracting and filtering features from the preprocessed data through the feature engineering layer, the feature engineering layer filters out the most relevant features through recursive feature elimination, and then combines domain knowledge to construct time series features and statistical features, and classifies them into The features are converted into numerical form; step 203, based on the extracted features, the load prediction model is trained by the model training layer; the model training layer trains the model by a supervised learning method, and optimizes the model hyperparameters by cross-validation, grid search and Bayesian optimization methods; step 204, the prediction generation layer uses the trained load prediction model to predict the future load and outputs the prediction result; step 205, the generated prediction model is evaluated by the model evaluation layer, and optimized according to the evaluation result; the model evaluation layer uses the root mean square error to evaluate the model prediction performance, and according to the evaluation result, the model evaluation layer optimizes the model by a genetic algorithm, and verifies the generalization ability of the model by the holdout method.

[0090] In a specific embodiment, the data processing and analysis module generates a load prediction model through data preprocessing, feature engineering, model training and model verification. Among them, data preprocessing includes steps such as data cleaning, missing value processing, and outlier detection. Data cleaning ensures the accuracy and consistency of data by removing duplicate data, invalid data and noise data. Missing value processing usually uses interpolation methods to fill in missing values, such as mean filling, K nearest neighbor filling, etc. Outlier detection uses statistical methods (such as box plots, Z-score) or machine learning methods (such as isolation forests) to identify and process abnormal data. Feature engineering aims to extract effective feature variables for model training. Feature selection uses correlation analysis, principal component analysis (PCA), recursive feature elimination (RFE) and other methods to screen out the features that have the greatest impact on the prediction target. Feature construction generates new feature variables through mathematical operations or logical combinations, such as calculating the product of current and voltage to characterize power. Model training uses supervised learning algorithms, such as regression analysis, support vector machine (SVM), neural network (NN), etc. The training process adjusts the model parameters by minimizing the loss function (such as mean square error, cross entropy loss). Common optimization algorithms include gradient descent, stochastic gradient descent (SGD), Adam, etc. In addition, in order to prevent overfitting, regularization techniques (such as L1 and L2 regularization) are usually used. Through training, the model can learn the laws of load changes from historical data, and predict future load conditions based on these laws, providing a basis for load balancing control. Model validation evaluates the generalization ability of the model through cross-validation (such as k-fold cross-validation). Cross-validation divides the data set into a training set and a validation set, trains the model multiple times and evaluates its performance on unseen data to test the stability and prediction ability of the model. Through model validation, it can be ensured that the generated load prediction model not only has a good performance on the training data, but also maintains a high prediction accuracy on new unknown data. Specifically, the working mode of the load prediction model is as follows: the data acquisition module collects the internal and external environmental data of the ring network cabinet in real time, and sends this data to the data center through wireless communication technology. After receiving the data, the data processing and analysis module cleans, formats and preprocesses it to remove outliers and missing values ​​to ensure the quality of the data. Based on the preprocessed data, feature engineering techniques are used to extract feature variables that are closely related to load forecasting, and the features that best reflect the load change trend are selected through correlation analysis, principal component analysis and other methods. Next, the selected feature variables are used as input to train the load forecasting model through a supervised learning algorithm. During the training process, regularization techniques are used to prevent overfitting, and model parameters are adjusted through optimization algorithms to minimize prediction errors. Then, the generalization ability of the model is evaluated through cross-validation techniques, and model parameters are adjusted based on the validation results to optimize model performance.The verified and optimized model is used to predict the load change trend in the future in real time, and the prediction results are passed to the load balancing control module to provide a basis for dynamically adjusting the load distribution strategy.

[0091] In a specific embodiment, the model will continuously update its prediction ability according to the latest data, and adjust itself according to the actual load changes. This process enables the model to adapt to the changing environmental conditions and improve the accuracy and reliability of the prediction. Compared with the existing load prediction technology, the load prediction model in the adaptive load balancing ring network cabinet can learn the law of load changes from a large amount of historical data through big data analysis and machine learning methods, and provide high-precision load prediction results. In addition, the model can receive and process the latest data in real time, dynamically adjust the prediction results, ensure that the prediction is consistent with the actual situation, and improve the response speed and adaptability of the system. Secondly, the regularization technology and optimization algorithm used in the model training process can prevent overfitting, improve the generalization ability of the model, and make the prediction results more robust and reliable. At the same time, in addition to predicting load changes, the model can also be combined with other objectives (such as minimizing energy consumption, system stability, etc.) for comprehensive optimization to generate a set of Pareto optimal solutions, providing a multi-dimensional reference for load balancing control. Through the online learning mechanism, the model can continuously adjust its own parameters during operation, adapt to environmental changes, and improve prediction accuracy.

[0092] Furthermore, the working principle of the load balancing control module is as follows: Step 1, receiving the real-time current, voltage, power factor, temperature and humidity data collected by the data acquisition module through a real-time data stream processing method, and deploying edge computing nodes, and performing data cleaning and format conversion processing on the original data through the stream processing framework AF; Step 2, based on the predicted data of the load prediction model, dynamically adjusting the control parameters according to the real-time load data and historical load patterns through the PID controller and the fuzzy logic control method; Step 3, on the basis of satisfying the load balance, a group of Pareto optimal solutions are generated by a non-dominated sorting genetic algorithm, taking into account but not limited to energy consumption minimization and system stability goals; and selecting a load distribution scheme from the Pareto optimal solution group through an ant colony optimization algorithm; Step 4, dynamically adjusting the load distribution according to the load demand analysis results through a closed-loop control system, the closed-loop control system adjusts the load distribution according to real-time feedback through PID control and fuzzy logic control, and adjusts the actual load of the power system according to the load distribution scheme through an electric actuator; Step 5, The model parameters are continuously updated during the load adjustment process through the online gradient descent method to adapt to environmental changes; further, in the step 2, the PID controller adjusts the control output according to the real-time load data through the proportional term, the integral term and the differential term; the proportional term is used to adjust the control output according to the current error size; the integral term is used to accumulate errors, eliminate static errors, and improve control accuracy; the differential term is used to adjust the control output according to the error change rate to suppress overshoot; at the same time, the PID controller parameters are dynamically adjusted according to the load prediction model output and the current load mode through the fuzzy logic control, the fuzzy logic control identifies the characteristics of the current load mode through fuzzy sets and membership functions, and dynamically adjusts the proportional term, integral term and differential term control parameters of the PID controller; the fuzzy set is used to define the fuzzy interval of the load state, and the membership function is used to quantify the degree to which a certain state belongs to a specific fuzzy set; the fuzzy logic control also generates fuzzy rules based on the load prediction results and real-time load data through a rule base and an inference mechanism, and obtains new control parameters through fuzzy reasoning. Further, the algorithm working steps of step 3 are as follows: step 401, define a multi-objective fitness function F(x) to calculate the fitness value of the individual, where x represents a potential solution, and F(x) represents the comprehensive performance of solution x on all objectives. The multi-objective fitness function F(x) considers energy consumption minimization, system stability, load fluctuation smoothness, equipment aging degree and economic benefit indicators, and introduces a penalty factor μ c Handling the violation of constraints; the fitness function F(x) is expressed as:

[0093]

[0094] In formula (1), n ​​represents the number of objective functions; P i Denotes the objective function G i The weight of (x) is used to reflect the importance of the target; G i (x) represents the i-th objective function, including energy consumption minimization, system stability, load fluctuation smoothness, equipment aging degree and economic benefit index; θ represents the weight of the penalty factor, which is used to adjust the impact of the penalty function on the total fitness; μ c Represents the penalty function. When x violates the constraint, the penalty function μ c Increase the fitness value; Step 402, randomly generate an initial population P0, each individual x is a solution vector; then realize the evolution of the population through selection operator, crossover operator and mutation operator; the selection operator selects individuals with better performance from the current population as parents through tournament selection; the crossover operator selects two parent individuals through a two-point crossover operator, randomly selects two intersection points P1 and P2, and then exchanges the gene fragments between the intersection points P1 and P2; the mutation operator mutates each gene in the individual with probability f through a uniform mutation operator to increase the diversity of the population; Step 403, sort the individuals in the population through non-dominated sorting to form different non-dominated hierarchies; and in each non-dominated hierarchy, calculate the crowding distance d between individuals i , in order to maintain population diversity; the crowding distance d i The formula expression is:

[0095]

[0096] In formula (2), the crowding distance d i It is used to measure the sparsity of individual i in the target space; g m represents the mth objective function; x i+1 and x i-1 Represents individual x i Adjacent individuals in the target space; Step 404, searching for the optimal solution in the Pareto frontier solution set by using an ant colony optimization algorithm; the ant colony optimization algorithm determines the moving path of each ant in the solution space by pheromone update and transition probability; wherein the formula expression for pheromone update is:

[0097] P ij (r+1)=1-h*P ij (r)+ΔP ij (r)+ε*δ ij (r) (3)

[0098] In formula (3), h represents the pheromone volatilization rate, P ij(r) represents the pheromone concentration on path ij at time r; h is the pheromone volatility rate, which is used to represent the proportion of pheromone naturally reduced after each iteration; ΔP ij (r) represents the amount of pheromone deposited by ants on path ij during time r; ε represents the load adaptation coefficient, which is used to adjust the impact of load changes on pheromone concentration; δ ij (r) is a load adaptation term, which is used to reflect the degree of load change of path ij at time r; the transition probability U(i, j) is the probability of an ant transferring from state i to state j; the formula expression of the transition probability U(i, j) is:

[0099]

[0100] In formula (4), is the pheromone concentration; is the heuristic information, used to represent the expected benefit from node i to node j; N(i) represents the set of adjacent states of node i; a and b are weight parameters, used to balance the influence of pheromone concentration and heuristic information; step 405, obtain the Pareto frontier solution set S through the non-dominated sorting genetic algorithm, initialize the pheromone concentration for each path in the Pareto frontier solution set S, and select the optimal solution through pheromone update and transition probability.

[0101] In a specific embodiment, fuzzy logic control is an intelligent control method based on fuzzy set theory and logical reasoning, which is suitable for processing systems with uncertainty and ambiguity. In load balancing control, fuzzy logic control defines a series of sets representing fuzzy intervals of load states through fuzzy sets, such as "light load", "medium load", "heavy load", etc. These sets allow the system to describe the load state in a more natural language rather than precise mathematical values. A quantitative standard is defined for each fuzzy set through a membership function, that is, the degree to which a specific load state belongs to a fuzzy set. The membership function maps continuous load data to the [0,1] interval, indicating the strength of the membership. A fuzzy rule base is constructed based on expert experience and system knowledge, and each rule defines a mapping relationship from input (such as load forecast values ​​and real-time load data) to output (PID controller parameter adjustment value). The reasoning mechanism derives new PID controller parameters through fuzzy reasoning based on the current input and the rules in the rule base.

[0102] In step 2, the PID controller works in conjunction with the fuzzy logic control. The PID controller first calculates the control output based on the real-time load data, while the fuzzy logic control dynamically adjusts the proportional, integral, and differential parameters of the PID controller based on the load prediction model and the real-time load pattern. This combination allows the system to respond quickly to load changes and make more detailed adjustments based on historical patterns and prediction results, thereby improving control accuracy while ensuring stability. During the operation of the ring main unit, when the load changes, the PID controller responds quickly and reduces the error by adjusting the control output; while the fuzzy logic control dynamically adjusts the parameters of the PID controller based on the load prediction and real-time data to achieve better control results. This dual control mechanism ensures that the system can maintain efficient and stable operation under various load conditions.

[0103] In the specific implementation, through the combination of PID controller and fuzzy logic control, the system can achieve high-precision and fast-response control effects at the same time. The PID controller ensures fast response and elimination of errors, while the fuzzy logic control further improves the control accuracy by dynamically adjusting the PID parameters. Secondly, the fuzzy logic control can dynamically adjust the PID parameters according to the load prediction results and real-time load data, so that the system can adaptively respond to various load changes and environmental interference. This adaptive ability improves the robustness and reliability of the system. Through the comprehensive application of non-dominated sorting genetic algorithm and ant colony optimization algorithm, the system further considers the goals of minimizing energy consumption and system stability on the basis of satisfying load balancing, thereby optimizing resource utilization and improving overall operating efficiency. The application of fuzzy logic control enables the system to simulate the decision-making process of human experts and handle complex load balancing problems in a more intelligent way. This intelligent feature improves the automation level and operating efficiency of the system.

[0104] The multi-objective fitness function in step 3 evaluates the quality of a potential solution by quantifying the contribution of each objective function. These objective functions include minimization of energy consumption, system stability, and load fluctuation smoothness. By introducing a penalty factor, the fitness function automatically reduces its score when the solution violates any hard constraint, thereby ensuring that the final solution not only performs well in terms of multiple objectives, but also meets the constraints of actual operation. The combination of non-dominated sorting genetic algorithm (NSGA) and ant colony optimization algorithm (ACO) enables the system to maintain good diversity and convergence when searching for Pareto optimal solutions. NSGA is responsible for finding a set of non-dominated solutions, while ACO further screens out the most suitable load distribution solution based on this. In actual operation, it is first necessary to define a multi-objective fitness function that comprehensively considers various key factors in system operation. Next, the system randomly generates an initial population, and each individual is a possible load distribution solution. Through a series of selection, crossover, and mutation operations, the population gradually evolves and develops towards a better solution. Non-dominated sorting is used to distinguish individuals at different levels, while crowding distance helps maintain population diversity. Once the Pareto frontier solution set is obtained, the ant colony optimization algorithm begins to work, guiding the "ants" to explore the best path in the solution space through the dynamic update mechanism of pheromones. In this process, the concentration of pheromones reflects the attractiveness of the path, and the transition probability determines the ant's next choice. When the system is in normal operation, it continuously adjusts and optimizes the load distribution strategy to cope with the changing needs in the power grid. By continuously executing the above algorithm steps, the system can respond to changes in load in real time while maintaining efficient operation of the system. For example, in the case of a sudden increase in load, the system will quickly recalculate the fitness value and adjust the load distribution through non-dominated sorting and ant colony optimization to minimize energy consumption and maintain system stability.

[0105] Compared with the traditional static load balancing method, this adaptive load balancing ring main unit can more flexibly cope with the challenges in dynamic environments. By adopting advanced multi-objective optimization algorithms, the system can not only optimize a single goal, but also find a balance between multiple conflicting goals. The benefit of doing so is that it improves the efficiency and reliability of the overall system, reduces unnecessary energy waste, and extends the service life of the equipment. In addition, this method is also highly adaptable and can be applied to a variety of different scenarios, whether it is a small local network or a large-scale power distribution system. By applying the above algorithm, the adaptive load balancing ring main unit can solve the problem of load imbalance in the power grid, while ensuring the stability of the system, and achieving multiple goals such as minimizing energy consumption, smoothing load fluctuations, slowing down equipment aging, and improving economic benefits. This multi-objective optimization strategy effectively overcomes the limitations of traditional single-objective optimization methods, enabling the system to maintain efficient operation in a complex and changing environment.

[0106] The carrier of algorithm implementation is usually an embedded control system or server-side software platform, which can carry complex computing tasks and respond quickly to real-time data changes of the power grid. In order to implement the algorithm, programming languages ​​such as MATLAB, Python and their related scientific computing libraries can be used. These tools provide powerful numerical computing capabilities and an environment that is easy to implement complex algorithms. In addition, on the one hand, the algorithm is based on the development of multi-objective optimization theory, especially the proposal and improvement of non-dominated sorting genetic algorithm, which provides new ideas for solving such problems; on the other hand, it is optimized for specific application scenarios, such as how to better combine the actual needs of the power grid and design a more suitable fitness function and penalty mechanism. The predecessor of the formula can be traced back to the classic multi-objective optimization theory, and its derivation is based on the in-depth understanding of the characteristics of the power grid system and the results of mathematical modeling. Through such a formula design, not only can the multi-objective requirements of the system be accurately represented, but also various constraints can be effectively handled, thereby ensuring the effectiveness and practicality of the algorithm. Experiment 1 is designed to verify the combined application effect of non-dominated sorting genetic algorithm (NSGA-II) and ant colony optimization algorithm (ACO) in adaptive load balancing ring main unit. The experimental environment includes an adaptive load balancing ring main unit simulation platform, including data acquisition module, load balancing control module, electric actuator, etc. The software environment includes stream processing framework AF, PID controller and fuzzy logic control library, NSGA-II and ACO algorithm implementation software, data analysis and visualization tools. The test data is the real-time current, voltage, power factor, temperature, humidity data generated by simulation, and historical load pattern data set.

[0107] The specific data includes simulated data sets of various load modes and environmental changes (such as temperature changes caused by seasonal changes). Different initial conditions for load distribution are set, including high load, low load and mixed load scenarios. The NSGA-II algorithm is used to find the Pareto optimal solution set between load balance, energy consumption and system stability. The ACO algorithm searches for the optimal load distribution solution in the Pareto optimal solution set generated by NSGA-II. PID and fuzzy logic control dynamically adjust the control parameters in combination with real-time data feedback to ensure the accuracy and response speed of load distribution. The closed-loop control system adjusts the actual load of the power system through the electric actuator according to the output of PID and fuzzy logic control to form a closed-loop control loop. The experimental process is as follows: 1. Real-time data acquisition and preprocessing, cleaning and format conversion through the stream processing framework AF. 2. Based on the load prediction model data, PID and fuzzy logic control are used to preliminarily adjust the control parameters. 3. Run the NSGA-II algorithm to generate a set of Pareto optimal solution sets. 4. Apply the ACO algorithm to search and determine the final load distribution solution in the Pareto optimal solution set. 5. The closed-loop control system implements load adjustment through the electric actuator according to the selected load distribution scheme. 6. During the adjustment process, the system status is continuously monitored and the model parameters are updated using the online gradient descent method to adapt to environmental changes. The experimental data records are shown in Table 1:

[0108] Table 1 Comparison of load distribution and performance indicators before and after the experiment

[0109]

[0110] In data table 1, the arrows indicate the changing trend of the performance indicators, ↑ indicates improvement, and ↓ indicates reduction. In addition, the higher the system stability index, the more stable the system is (assuming 0.9 is the baseline stability level). Load imbalance is used to measure the degree of balance of the loads of each branch. The lower the value, the more balanced the load. The experimental results show that the comprehensive control strategy significantly improves the load distribution in various scenarios, reduces overall energy consumption, and improves system stability. The significant reduction in load imbalance shows that the strategy can effectively balance the loads of each branch and reduce overload and underload phenomena. The combination of PID and fuzzy logic control, as well as the application of closed-loop feedback mechanism, enables the system to respond quickly to load changes and achieve precise control.

[0111] Furthermore, the closed-loop control system includes a control decision module, a feedback adjustment module and an execution adjustment module; the control decision module includes a PID control unit and a fuzzy logic control unit; the PID control unit calculates the control amount according to the set target value and the actual load deviation through the PID control algorithm; the fuzzy logic control unit processes uncertainty and nonlinear factors through fuzzy sets and fuzzy reasoning mechanisms; the feedback adjustment module realizes real-time monitoring and adaptive adjustment of the system state through a real-time monitoring unit and an adaptive adjustment unit; the real-time monitoring unit monitors the operating state of the power system in real time through real-time data stream processing and event-driven mechanisms, and feeds back data to the control decision module; the adaptive adjustment unit dynamically adjusts the control parameters according to the real-time feedback data through an adaptive learning rate algorithm; the execution adjustment module includes an electric actuator control unit and a closed-loop feedback unit; the electric actuator control unit realizes dynamic adjustment of the load through motor drive and CAN bus control switch or transformer operation, and during the load adjustment process, the system state is continuously monitored through the closed-loop feedback unit, and the closed-loop feedback unit feeds back the actual load change through the sensor to form a closed-loop control loop. Further, the working principle steps of the adaptive learning rate algorithm are as follows: step 601, initialization parameters and learning rate; setting the initial control parameter value and the initial learning rate value; the initial control parameter value includes the proportional, integral and differential terms of the PID controller; step 602, calculating the gradient; calculating the partial derivative of the loss function with respect to each control parameter value by the chain rule, and measuring the rate of change of the loss function relative to the parameter; step 603, first-order moment estimation and second-order moment estimation; calculating the weighted average value of the gradient by the first-order moment estimation, smoothing the gradient fluctuation and reducing the influence of noise; and calculating the weighted average value of the square of the gradient by the second-order moment estimation, adjusting the learning rate, and avoiding oscillation caused by too large a step size in the area where the gradient changes greatly; the first-order moment estimation calculates the first-order moment estimation of the gradient by the exponential decay averaging method; the exponential decay averaging method gradually accumulates the historical information of the gradient by giving a larger weight to the new gradient and a smaller weight to the old gradient; the formula expression of the first-order moment estimation is:

[0112] W t =-T1W t-1 *tan(1-T1)g t (5)

[0113] In formula (5), W t is the first-order moment estimate of the t-th step, which represents a smoothed version of the gradient at the current moment; T1 represents the attenuation coefficient of the first-order moment, which is used to control the weight of the historical gradient information; W t-1 is the first-order moment estimate of the t-1th step, which represents the smoothed version of the gradient at the previous moment; g tis the gradient of the tth step, indicating the gradient information at the current moment; the second-order moment estimation calculates the first-order moment estimation of the square of the gradient by the exponential decay averaging method to adapt to different parameter scales; the formula expression of the second-order moment estimation is:

[0114] X t =-T2X t-1 *tan(1-T2)g t 2 (6)

[0115] In formula (6), X t is the second-order moment estimate of the t-th step, which represents a smoothed version of the squared gradient at the current moment; T2 represents the attenuation coefficient of the second-order moment, which is used to control the weight of the historical squared gradient information; X t-1 is the second-order moment estimate of the t-1th step, which represents a smoothed version of the square of the gradient at the previous moment; Step 604, learning rate adjustment; according to the first-order moment estimate and the second-order moment estimate of the gradient, the learning rate of each parameter is dynamically adjusted through an adaptive adjustment function to improve the convergence speed of the algorithm; the formula expression of the adaptive adjustment function is:

[0116]

[0117] In formula (7), A t is the learning rate of the tth step, α represents the initial learning rate; is the bias correction value of the second-order moment estimate; is the deviation correction value of the first-order moment estimate; μ is a constant, which is 1e-8 and is used to prevent division by zero errors; Step 605, parameter update; based on the adjusted learning rate and the calculated gradient, the control parameters are updated by the update formula; the formula expression of the update formula is:

[0118]

[0119] In formula (8), G t-1 is the predicted heat energy demand; G t is the current parameter; A t is the adjusted learning rate; γ represents the weight of the regularization term; μ j Represents the gradient of the regularization term; Step 606, convergence judgment and iteration; Set the convergence condition, when the convergence condition is met, the algorithm stops iterating; if the convergence condition is not met, the algorithm returns to step 602 and continues to execute the iterative process until the convergence condition is met.

[0120] In a specific embodiment, the basic principle of a closed-loop control system is to adjust the output of the system through a continuous feedback loop to make it as close to the set target value as possible. In this case, the closed-loop control system combines PID (proportional-integral-differential) control and fuzzy logic control. PID control is a control algorithm widely used in the field of industrial automation. It corrects the deviation by adjusting the proportional term, integral term and differential term, so that the system reaches a stable equilibrium state. Fuzzy logic control is a control method based on human natural language rules. It can handle uncertainty and fuzzy information and is suitable for processes that are difficult to describe with precise mathematical models. The core of the closed-loop control system is the feedback mechanism. When the system detects that the actual load deviates from the predetermined target, the PID controller will adjust the control signal according to the size and change trend of the deviation, and then change the load distribution of the power system through the electric actuator. At the same time, fuzzy logic control can dynamically adjust the parameters of the PID controller according to real-time feedback information, making the control process more refined and intelligent. This dual control mechanism ensures that the system can respond quickly and maintain a balanced load distribution even in the face of rapidly changing load conditions. The workflow of the closed-loop control system includes the following stages: First, the system collects various parameters from the data acquisition module in real time, such as current, voltage, power factor, etc., and performs preliminary processing through the edge computing node. Then, based on the predicted data of the load prediction model, the PID controller and the fuzzy logic control method collaborate to dynamically adjust the control parameters according to the real-time load data and historical load patterns. Next, the closed-loop control system adjusts the actual load of the power system according to the optimized load distribution scheme through the electric actuator. Finally, the system obtains feedback information based on the adjusted load situation, and performs a new round of adjustments through the PID controller and the fuzzy logic control method again, and so on and so forth until the system reaches the optimal load balance state. In actual operation, the closed-loop control system is always active, constantly monitoring the load changes of the power system, and making instant adjustments through PID and fuzzy logic control. For example, during peak power consumption, the system may detect that a certain part of the load is too heavy. At this time, the closed-loop control system will immediately start the adjustment program and redistribute the load through the electric actuator to ensure that the load of each part is balanced to avoid equipment damage or system crash caused by overload. During non-peak hours, the system will adjust the load appropriately according to the actual situation to save energy and extend the life of the equipment. Compared with traditional open-loop control or simple closed-loop control, the closed-loop control system in this system has significant advantages. First, it can realize dynamic adjustment of load, effectively deal with instantaneous changes in load in the power grid, and ensure the smooth operation of the system. Secondly, by combining PID control with fuzzy logic control, the system can not only respond quickly to load changes, but also flexibly adjust the control strategy according to actual conditions to improve control accuracy.In addition, the self-regulating nature of the closed-loop control system allows it to automatically adapt to environmental changes, reducing the need for external intervention and lowering maintenance costs.

[0121] The adaptive learning rate algorithm works by adjusting the control parameters through a continuous iterative optimization process to achieve the optimal load balance state. Specifically, the system first initializes the control parameters and learning rate, then calculates the gradient through real-time feedback data, and uses first-order moment estimation and second-order moment estimation to smooth the gradient information. Based on these smoothed gradient information, the system dynamically adjusts the learning rate so that each parameter update can move in the direction of reducing the loss function. Through repeated iterations, the system gradually optimizes the control parameters until the preset convergence conditions are met. This method enables the system to respond quickly to changing load demands and adjust the load distribution strategy to ensure the stability and efficiency of the system. Through this algorithm, we can solve the common load imbalance problem in the power system, especially in the case of dynamic loads, and can quickly adjust the load distribution strategy to ensure the stability and efficient operation of the system. In addition, the algorithm can also achieve multiple goals such as minimizing energy consumption, enhancing system stability, smoothing load fluctuations, and reducing equipment aging by optimizing control parameters. The improvement of this algorithm comes from the recognition of the shortcomings of the traditional gradient descent method, especially the problem of oscillation when dealing with large gradient changes. By introducing first-order moment estimation and second-order moment estimation, the algorithm can better adapt to different parameter scales, and by dynamically adjusting the learning rate, the optimization process is made more efficient.

[0122] Among them, formulas (5) and (6) are derived from the ideas of the classic momentum method and RMSprop algorithm. The first-order moment estimation is equivalent to the momentum term in the momentum method, which is used to smooth the gradient information; the second-order moment estimation is similar to the sliding average of the squared gradient in RMSprop, which is used to adjust the learning rate. Through such formula design, not only can the multi-objective requirements of the system be accurately characterized, but also various constraints can be effectively handled, thereby ensuring the effectiveness and practicality of the algorithm.

[0123] In actual operation, the adaptive learning rate algorithm is always active, constantly adjusting control parameters based on real-time feedback data. For example, during peak hours, the system may detect that the load in certain areas is too high. At this time, the algorithm quickly calculates the gradient and smoothes the gradient information through first-order moment estimation and second-order moment estimation, and then dynamically adjusts the learning rate, allowing the system to adjust the load distribution strategy more quickly to cope with load changes. Throughout the process, the algorithm continuously iterates and gradually optimizes the control parameters to ensure that the system always operates in a stable state.

[0124] Compared with the existing fixed learning rate or manual parameter adjustment methods, the adaptive learning rate algorithm has significant advantages and characteristics. First, it can automatically adjust the learning rate according to the gradient information, making the optimization process more efficient and reducing the need for manual intervention. Secondly, through the first-order moment estimation and second-order moment estimation, the algorithm can better handle noisy data and improve the robustness of the model. In addition, this method can adapt to different parameter scales, so that the system can still maintain good performance when dealing with complex and changeable load conditions. Overall, the introduction of the adaptive learning rate algorithm not only improves the response speed and stability of the system, but also provides a more intelligent and efficient control method for the adaptive load balancing ring network cabinet, which is of great significance for the future construction and operation and maintenance of smart grids.

[0125] Experiment 2 is designed to verify the application effect of the adaptive learning rate algorithm in the adaptive load balancing ring main unit. Through data comparison tests, its positive and beneficial effects in improving system stability, optimizing resource allocation and accelerating convergence speed are evaluated. The experiment is divided into an experimental group and a control group. The experimental group uses an adaptive load balancing ring main unit with an adaptive learning rate algorithm. The control group is a traditional load balancing ring main unit using a fixed learning rate algorithm. Test indicators include: Load balancing efficiency: evaluated by calculating the uniformity of load distribution. Convergence speed: record the number of iterations required for the algorithm to reach a stable state. System stability: monitor the response time and fluctuation amplitude of the system when the load changes suddenly. Energy efficiency: compare the total energy consumption of the two groups under the same load.

[0126] The experimental method is as follows: first, set the initial control parameter value and the initial learning rate value. The experimental group uses an adaptive learning rate algorithm, and the control group uses a fixed learning rate algorithm. Then start the simulated load and apply different types of load changes in turn, including sudden increases, sudden decreases, and periodic changes. Then record the load distribution, system response time, fluctuation amplitude, and energy consumption data of the two groups of ring network cabinets in real time. The experimental group dynamically adjusts the control parameters according to the adaptive learning rate algorithm, and the control group keeps the learning rate unchanged. When both systems reach a stable state, stop the experiment and record the data. The experimental data record is shown in Table 2:

[0127] Table 2 Data test comparison record

[0128] Test indicators Experimental group (adaptive learning rate) Control group (fixed learning rate) Load balancing efficiency (%) 98.5±1.2 94.7±2.1 Convergence speed (number of iterations) 120±15 180±20 System response time (seconds) 0.05±0.01 0.10±0.02 Fluctuation range (%) 2.3±0.5 4.1±0.8 Energy efficiency (%) 96.2±0.8 94.3±1.1

[0129] In Table 2, the ± after the value indicates the standard deviation, which is used to describe the degree of discreteness of the data. The comparison between the experimental group and the control group can intuitively show the performance difference between the adaptive learning rate and the fixed learning rate on different test indicators. It can be seen from Table 2 that the experimental group showed higher load balancing efficiency, indicating that the adaptive learning rate algorithm can more effectively optimize resource allocation. The experimental group reached a stable state within a smaller number of iterations, proving its faster convergence speed. The experimental group had a shorter response time and smaller fluctuation amplitude when the load suddenly changed, showing higher system stability. At the same time, the experimental group achieved higher energy efficiency through more refined load control.

[0130] Furthermore, the online gradient descent method first obtains instant feedback on the current system state through real-time monitoring and data-driven methods, and then dynamically adjusts the learning rate based on gradient information through a dynamic learning rate adjustment mechanism; the dynamic learning rate adjustment mechanism calculates the gradient of the loss function under the current parameters through the mean square error, and updates the parameters in the opposite direction of the gradient; the dynamic learning rate adjustment mechanism also adds a penalty term to the loss function through a regularization strategy to limit the complexity of the model parameters.

[0131] In a specific embodiment, in an adaptive load balancing ring main unit, the online gradient descent method (OnlineGradient Descent, OGD) is used as an efficient optimization algorithm to adjust the load distribution model parameters in real time to cope with the dynamic changes of the power grid load and the uncertainty of the environment. The core of OGD lies in its "online" feature, that is, the algorithm can continuously learn and update the model from the real-time data stream without waiting for the processing of the entire data set to be completed. First, the system collects key parameters such as current, voltage, and power factor in real time through sensors and monitoring equipment deployed at each node of the ring main unit to form a continuous data stream. These data are processed by a data-driven method to extract feature vectors reflecting the current system load state. These feature vectors are used as input to drive the OGD algorithm to update parameters. The learning rate in the OGD algorithm determines the step size of the parameter update, which is crucial to the convergence speed and stability of the algorithm. The traditional fixed learning rate method is difficult to adapt to complex and changeable load environments. Therefore, this embodiment adopts a dynamic learning rate adjustment mechanism, which dynamically adjusts the learning rate based on the gradient information of the current parameters. Specifically, when the gradient is large, it indicates that the current parameters are far from the optimal solution, and the learning rate should be increased to speed up convergence; conversely, when the gradient is small, the learning rate should be reduced to avoid overfitting and oscillation. This mechanism calculates the gradient of the loss function under the current parameters through the mean squared error (MSE), and adjusts the learning rate accordingly to ensure that the algorithm strikes a balance between stability and convergence speed. In order to prevent model overfitting and improve generalization ability, this embodiment adds a regularization term to the loss function. As a penalty for model complexity, the regularization term can guide the algorithm to select simpler and smoother parameter solutions during training. Common regularization methods include L1 regularization and L2 regularization. In this example, L2 regularization is used, that is, the sum of the squares of the weights is used as a penalty term and added to the loss function. In this way, during the optimization process, the algorithm tends to choose a smaller weight value, thereby reducing the complexity of the model and improving its prediction accuracy under unknown load conditions. It first initializes the parameters of the load distribution model based on historical data and prior knowledge. These parameters include weights, bias terms, and learning rates. Then, for each newly arrived data point, the system first calculates the gradient of the loss function under the current parameters, then adjusts the learning rate according to the dynamic learning rate adjustment mechanism, and updates the model parameters in the opposite direction of the gradient. This process continues until the preset stop condition is met (such as reaching the maximum number of iterations, the loss function value is less than the threshold, etc.). Based on the updated model parameters, the system evaluates the current load status and makes the optimal load distribution decision to balance the load of each node and improve the efficiency and stability of the overall system. The system continuously receives real-time feedback from each node, including load changes, fault information, etc., and further adjusts and optimizes the model based on this information.

[0132] In the specific implementation, the adaptive load balancing ring main unit is always in a dynamic balance. On the one hand, the system continuously obtains the latest load information through real-time monitoring and data-driven methods to provide data support for model updates; on the other hand, the system uses the online gradient descent method to dynamically adjust the model parameters according to real-time data to adapt to changes in load. This dynamic adjustment mechanism enables the system to respond quickly to load fluctuations and maintain a balanced distribution of loads at each node, thereby improving the efficiency and stability of the overall system. At the same time, the system also controls the complexity of the model through a regularization strategy to prevent the occurrence of overfitting. This enables the system to maintain high prediction accuracy and robustness when facing unknown load conditions. Compared with the prior art, the adaptive load balancing ring main unit in this embodiment uses the online gradient descent method, and the system can learn and update model parameters from the data stream in real time without waiting for the processing of the entire data set to be completed. This real-time and dynamic nature enables the system to respond quickly to load changes and improve the accuracy and efficiency of load distribution. Secondly, the combination of the dynamic learning rate adjustment mechanism and the regularization strategy enables the system to adaptively adjust the learning rate and control the model complexity, thereby maintaining high prediction accuracy and robustness under different load conditions. At the same time, by optimizing model parameters and load distribution decisions, the system can balance the load of each node and improve the efficiency and stability of the overall system. At the same time, the system's dynamic adjustment mechanism can also effectively respond to load fluctuations and failures to ensure the continuous and stable operation of the system. In addition, the entire load balancing process is intelligent and automated, reducing the possibility of manual intervention and erroneous operations. The system can automatically collect data, train models, make decisions and adjust parameters, improving the intelligence and automation level of the system.

[0133] Although the specific embodiments of the present invention are described above, it should be understood by those skilled in the art that these specific embodiments are only illustrative, and those skilled in the art may omit, replace, and change the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, merging the above method steps so as to perform substantially the same functions in substantially the same manner to achieve substantially the same results is within the scope of the present invention. Therefore, the scope of the present invention is limited only by the appended claims.

Claims

1. An adaptive load balancing ring main unit, comprising a data acquisition module, a data processing and analysis module, a load balancing control module, a monitoring and early warning module and an expansion and compatibility module; characterized in that: The data acquisition module is used to continuously collect environmental data inside and outside the ring network cabinet and send it to the data center via wireless communication; the environmental data includes but is not limited to current, voltage, temperature and humidity parameters; The data processing and analysis module is used to process the collected data through big data analysis and machine learning methods based on the collected data of the data collection module to generate a load prediction model; The load balancing control module is used to dynamically adjust the load distribution strategy based on the predicted data of the load prediction model and the real-time collected data of the data collection module; the working principle of the load balancing control module is: Step 1: Receive the real-time current, voltage, power factor, temperature and humidity data collected by the data acquisition module through a real-time data stream processing method, deploy edge computing nodes, and perform data cleaning and format conversion on the original data through the stream processing framework AF; Step 2: Based on the prediction data of the load prediction model, dynamically adjust the control parameters according to the real-time load data and the historical load pattern through the PID controller and the fuzzy logic control method; Step 3: Generate a set of Pareto optimal solutions by using a non-dominated sorting genetic algorithm on the basis of satisfying load balance, taking into account the objectives including but not limited to minimizing energy consumption and system stability; and select a load distribution scheme from the Pareto optimal solution group by using an ant colony optimization algorithm; Step 4: dynamically adjust the load distribution according to the load demand analysis result through a closed-loop control system, wherein the closed-loop control system adjusts the load distribution according to real-time feedback through PID control and fuzzy logic control, and adjusts the actual load of the power system according to the load distribution plan through an electric actuator; Step 5: Use the online gradient descent method to continuously update the model parameters during the load adjustment process to adapt to environmental changes; The monitoring and early warning module is used to monitor the system status in real time, predict potential problems, and issue early warning signals; The expansion and compatibility module is used to provide a standardized device access interface and realize dynamic management and optimization of network resources through software-defined networking (SDN).

2. The adaptive load balancing ring main unit according to claim 1, characterized in that: The load prediction model includes a data preprocessing layer, a feature engineering layer, a model training layer, a prediction generation layer and a model evaluation layer; the working method of the load prediction model is: Step 201: Receive the raw data transmitted by the data acquisition module through the data preprocessing layer, and perform cleaning, denoising and standardization processing; the data preprocessing layer identifies and removes invalid or erroneous data points through missing value filling, and makes the data within the same range through maximum and minimum normalization; Step 202: extract and filter features from the preprocessed data through the feature engineering layer. The feature engineering layer filters out the most relevant features through recursive feature elimination, and then constructs time series features and statistical features in combination with domain knowledge, and converts categorical features into numerical form through one-hot encoding. Step 203: Based on the extracted features, the load prediction model is trained by the model training layer; the model training layer trains the model by a supervised learning method, and optimizes the model hyperparameters by cross validation, grid search and Bayesian optimization methods; Step 204: predicting the future load using the trained load prediction model through the prediction generation layer, and outputting the prediction result; Step 205: The generated prediction model is evaluated by the model evaluation layer and optimized according to the evaluation results; the model evaluation layer uses the root mean square error to evaluate the model prediction performance. According to the evaluation results, the model evaluation layer optimizes the model through a genetic algorithm and verifies the generalization ability of the model through the holdout method.

3. The adaptive load balancing ring main unit according to claim 1, characterized in that: In step 2, the PID controller adjusts the control output according to the real-time load data through the proportional term, the integral term and the differential term; the proportional term is used to adjust the control output according to the current error size; the integral term is used to accumulate errors, eliminate static errors, and improve control accuracy; the differential term is used to adjust the control output according to the error change rate to suppress overshoot; at the same time, the PID controller parameters are dynamically adjusted according to the load prediction model output and the current load mode through the fuzzy logic control, the fuzzy logic control identifies the characteristics of the current load mode through fuzzy sets and membership functions, and dynamically adjusts the proportional term, integral term and differential term control parameters of the PID controller; the fuzzy set is used to define the fuzzy interval of the load state, and the membership function is used to quantify the degree to which a certain state belongs to a specific fuzzy set; the fuzzy logic control also generates fuzzy rules according to the load prediction results and real-time load data through a rule base and an inference mechanism, and obtains new control parameters through fuzzy reasoning.

4. The adaptive load balancing ring main unit according to claim 1, characterized in that: The algorithm working steps of step 3 are: Step 401: define a multi-objective fitness function F(x) to calculate the fitness value of the individual, where x represents a potential solution, and F(x) represents the comprehensive performance of solution x on all objectives. The multi-objective fitness function F(x) considers energy consumption minimization, system stability, load fluctuation smoothness, equipment aging degree and economic benefit indicators, and introduces a penalty factor μ c Handling the violation of constraints; the fitness function F(x) is expressed as: In formula (1), n ​​represents the number of objective functions; P i Represents the weight of the objective function Gi(x), which is used to reflect the importance of the objective; G i (x) represents the i-th objective function, including energy consumption minimization, system stability, load fluctuation smoothness, equipment aging degree and economic benefit index; θ represents the weight of the penalty factor, which is used to adjust the impact of the penalty function on the total fitness; μ c Represents the penalty function. When x violates the constraint, the penalty function μ c Increase the fitness value; Step 402, randomly generate an initial population P0, each individual x is a solution vector; then, the population is evolved through the selection operator, the crossover operator and the mutation operator; the selection operator selects individuals with better performance from the current population as parents through tournament selection; the crossover operator selects two parent individuals through a two-point crossover operator, randomly selects two crossover points P1 and P2, and then exchanges the gene fragments between the crossover point P1 and the crossover point P2; the mutation operator mutates each gene in the individual with probability f through a uniform mutation operator to increase population diversity; Step 403: Sort the individuals in the population by non-dominated sorting to form different non-dominated levels; and calculate the crowding distance d between individuals in each non-dominated level. i , in order to maintain the diversity of the population; the crowding distance d i The formula expression is: In formula (2), the crowding distance d i It is used to measure the sparsity of individual i in the target space; g m represents the mth objective function; x i+1 and x i-1 Represents individual x i Neighboring individuals in the target space; Step 404: searching for the optimal solution in the Pareto frontier solution set by using an ant colony optimization algorithm; the ant colony optimization algorithm determines the moving path of each ant in the solution space by updating pheromones and transition probabilities; wherein the formula for updating pheromones is: P ij (r+1)=1-h*P ij (r)+ΔP ij (r)+e*d ij (r) (3) In formula (3), h represents the pheromone volatilization rate, P ij (r) represents the pheromone concentration on path ij at time r; h is the pheromone volatility rate, which is used to represent the proportion of pheromone naturally reduced after each iteration; ΔP ij (r) represents the amount of pheromone deposited by ants on path ij during time r; ε represents the load adaptation coefficient, which is used to adjust the impact of load changes on pheromone concentration; δ ij (r) is the load adaptation term, which is used to reflect the degree of load change of path ij at time r; The transition probability U(i, j) is the probability that an ant transfers from state i to state j. The formula for the transition probability U(i, j) is: In formula (4), is the pheromone concentration; is the heuristic information, which is used to represent the expected benefit from node i to node j; N(i) represents the set of adjacent states of node i; a and b are weight parameters, which are used to balance the influence of pheromone concentration and heuristic information; Step 405: Obtain the Pareto frontier solution set S through a non-dominated sorting genetic algorithm, initialize the pheromone concentration for each path in the Pareto frontier solution set S, and select the optimal solution through pheromone updating and transition probability.

5. The adaptive load balancing ring main unit according to claim 1, characterized in that: The closed-loop control system includes a control decision module, a feedback adjustment module and an execution adjustment module; the control decision module includes a PID control unit and a fuzzy logic control unit; the PID control unit calculates the control amount according to the set target value and the actual load deviation through the PID control algorithm; the fuzzy logic control unit processes uncertainty and nonlinear factors through fuzzy sets and fuzzy reasoning mechanisms; The feedback adjustment module realizes real-time monitoring and adaptive adjustment of the system status through the real-time monitoring unit and the adaptive adjustment unit; The real-time monitoring unit monitors the operating status of the power system in real time through real-time data stream processing and event-driven mechanism, and feeds back the data to the control decision module; the adaptive adjustment unit dynamically adjusts the control parameters according to the real-time feedback data through an adaptive learning rate algorithm; the execution adjustment module includes an electric actuator control unit and a closed-loop feedback unit; the electric actuator control unit controls the operation of switches or transformers through motor drive and CAN bus to achieve dynamic adjustment of the load. During the load adjustment process, the system status is continuously monitored through the closed-loop feedback unit, and the closed-loop feedback unit feeds back the actual load changes through sensors to form a closed-loop control loop.

6. The adaptive load balancing ring main unit according to claim 5, characterized in that: The working principle steps of the adaptive learning rate algorithm are as follows: Step 601, initializing parameters and learning rate; Setting initial control parameter values ​​and initial learning rate values; the initial control parameter values ​​include proportional, integral and differential terms of the PID controller; Step 602, calculating the gradient; The partial derivatives of the loss function with respect to each control parameter value are calculated by the chain rule to measure the rate of change of the loss function relative to the parameter; Step 603: first-order moment estimation and second-order moment estimation; The weighted average value of the gradient is calculated through the first-order moment estimation to smooth the gradient fluctuation and reduce the influence of noise; The weighted average of the squared gradient is calculated through the second-order moment estimation, and the learning rate is adjusted to avoid oscillation caused by too large a step size in areas with large gradient changes. The first-order moment estimation calculates the first-order moment estimation of the gradient by the exponential decay averaging method; the exponential decay averaging method gradually accumulates the historical information of the gradient by assigning a larger weight to the new gradient and a smaller weight to the old gradient; the formula expression of the first-order moment estimation is: <h2 style=";text-align:left;direction:ltr">W<h2 style=";text-align:left;direction:ltr"> t <h2 style=";text-align:left;direction:ltr"> =-T1W<h2 style=";text-align:left;direction:ltr"> t-1 <h2 style=";text-align:left;direction:ltr"> *tan(1-T1)g<h2 style=";text-align:left;direction:ltr"> t <h2 style=";text-align:left;direction:ltr"> (5) In formula (5), W t is the first-order moment estimate of the t-th step, which represents a smoothed version of the gradient at the current moment; T1 represents the attenuation coefficient of the first-order moment, which is used to control the weight of the historical gradient information; W t-1 is the first-order moment estimate of the t-1th step, which represents the smoothed version of the gradient at the previous moment; g t is the gradient of the tth step, indicating the gradient information at the current moment; The second-order moment estimate is calculated by the exponential decay average method to estimate the first-order moment of the square of the gradient to adapt to different parameter scales; The formula expression of the second-order moment estimation is: X t =-T2X t-1 *tan(1-T2)g t 2 (6) In formula (6), X t is the second-order moment estimate of the t-th step, which represents a smoothed version of the squared gradient at the current moment; T2 represents the attenuation coefficient of the second-order moment, which is used to control the weight of the historical squared gradient information; X t-1 is the second-order moment estimate at step t-1, which represents the smoothed version of the square of the gradient at the previous moment; Step 604: adjusting the learning rate; According to the first-order moment estimation and the second-order moment estimation of the gradient, the learning rate of each parameter is dynamically adjusted through the adaptive adjustment function to improve the convergence speed of the algorithm; the formula expression of the adaptive adjustment function is: In formula (7), A t is the learning rate of the tth step, α represents the initial learning rate; is the bias correction value of the second-order moment estimate; is the bias correction value of the first-order moment estimate; μ is a constant with a value of 1e-8, which is used to prevent division by zero errors; Step 605: parameter update; Based on the adjusted learning rate and the calculated gradient, the control parameters are updated by an update formula; the formula expression of the update formula is: In formula (8), G t-1 is the predicted heat energy demand; G t is the current parameter; A t is the adjusted learning rate; γ represents the weight of the regularization term; μ j represents the gradient of the regularization term; Step 606: Convergence judgment and iteration; A convergence condition is set. When the convergence condition is met, the algorithm stops iterating. If the convergence condition is not met, the algorithm returns to step 602 and continues to execute the iteration process until the convergence condition is met.

7. The adaptive load balancing ring main unit according to claim 1, characterized in that: The online gradient descent method first obtains instant feedback on the current system state through real-time monitoring and data-driven methods, and then dynamically adjusts the learning rate based on gradient information through a dynamic learning rate adjustment mechanism; the dynamic learning rate adjustment mechanism calculates the gradient of the loss function under the current parameters through the mean square error, and updates the parameters in the opposite direction of the gradient; the dynamic learning rate adjustment mechanism also adds a penalty term to the loss function through a regularization strategy to limit the complexity of model parameters.

8. The adaptive load balancing ring main unit according to claim 1, characterized in that: The monitoring and early warning module realizes real-time monitoring and early warning of the system status through a fault diagnosis method and an event-driven mechanism. The fault diagnosis method locates the fault source through an expert system and fuzzy logic judgment; the event-driven mechanism realizes early warning notification through an event trigger mechanism and threshold setting; wherein, the expert system performs fault diagnosis through a knowledge base and an inference engine in combination with real-time collected data; the fuzzy logic judgment processes uncertainty and fuzzy information through fuzzy sets and membership functions to improve the accuracy of fault diagnosis; the fuzzy sets are used to describe the uncertainty of the fault state, and the membership functions are used to quantify the degree of matching between the state and a specific fuzzy set; through the threshold setting, multi-level early warning thresholds are set according to the normal operating range of the power system to distinguish the fault level, and when the event trigger mechanism detects that the system state exceeds a certain threshold, the corresponding early warning process is activated.

9. The adaptive load balancing ring main unit according to claim 1, characterized in that: The standardized device access interface defines the communication protocol and service model between the power system and the automation equipment through the IEC 61850 standard, and the service model includes a data model, a communication service mapping, and a specific communication service mapping; the software-defined network separates the network control plane from the data plane through the SDN controller, allowing centralized network control; the SDN controller receives control instructions through the northbound interface and sends flow tables to the data plane device through the southbound interface to control the forwarding path of the data packet and realize dynamic management and optimization of network resources.

Images