Substation data collaborative identification method, device, equipment and storage medium

Abstract

The present application relates to the technical field of substation power distribution, and discloses a method, device, equipment and storage medium for collaborative identification of substation data. The method includes: performing topology identification processing on the real-time voltage and current data of the target substation to obtain a dynamic distribution network topology structure; dynamically adjusting the output characteristics of the distributed energy unit to obtain a power regulation signal; collaboratively optimizing the power regulation signal and the power supply and demand data to obtain a collaborative scheduling strategy; performing health status assessment processing on multi-source heterogeneous data and collaborative scheduling strategies to obtain equipment health status assessment results; dynamically optimizing the dynamic distribution network topology structure and equipment health status assessment results to obtain dynamic protection setting values; optimizing and allocating the dynamic distribution network topology structure, collaborative scheduling strategy and dynamic protection setting values ​​to obtain the power allocation plan corresponding to the target substation. The present application improves the accuracy of collaborative identification of substation data.

Description

Technical Field

[0001] The present application relates to the field of substation power distribution, and in particular to a substation data collaborative identification method, device, equipment and storage medium. Background Art

[0002] Existing substation data analysis and management systems usually adopt a centralized architecture, relying mainly on a single data processing center to collect, analyze, and process data from various devices and sensors. This approach works well when dealing with small-scale systems and can achieve basic monitoring and control functions. However, with the development of smart grids and the widespread access to distributed energy, substation systems are becoming more and more complex, and the amount of data is increasing dramatically. Traditional centralized methods face huge challenges in processing large-scale, heterogeneous, and dynamically changing data.

[0003] The main shortcoming of existing technologies is the lack of an effective data collaborative processing mechanism. First, data from different sources are often processed in isolation, and the correlation and potential value between data cannot be fully explored. Secondly, the system lacks the ability to respond quickly to dynamic changes in network topology, making it difficult to adjust control strategies in a timely manner. Furthermore, traditional methods are difficult to achieve accurate power scheduling and protection configuration when dealing with large-scale distributed energy access. Finally, existing systems usually use fixed analysis models, which are difficult to adapt to complex and changing operating environments, resulting in the accuracy and reliability of analysis results being affected. Summary of the invention

[0004] The present application provides a method, device, equipment and storage medium for collaborative identification of substation data, which are used to improve the accuracy of collaborative identification of substation data, realize accurate modeling, intelligent control and optimized scheduling of complex dynamic systems of substations, and effectively solve the limitations of traditional methods in processing large-scale, dynamic and heterogeneous data.

[0005] In a first aspect, the present application provides a method for collaborative identification of substation data, the method comprising: performing topology identification processing on the real-time voltage and current data of the target substation to obtain a dynamic distribution network topology; dynamically adjusting the output characteristics of the distributed energy units in the dynamic distribution network topology through a virtual inertia control algorithm to obtain a power regulation signal that simulates the characteristics of a synchronous generator; collaboratively optimizing the power regulation signal and pre-acquired power supply and demand data through a multi-agent reinforcement learning algorithm to obtain a collaborative scheduling strategy; performing health status assessment processing on the pre-collected multi-source heterogeneous data of the target substation and the collaborative scheduling strategy through a deep transfer learning algorithm to obtain an equipment health status assessment result; dynamically optimizing the dynamic distribution network topology and the equipment health status assessment result through an adaptive fuzzy logic algorithm to obtain a dynamic protection setting value; optimizing and allocating the dynamic distribution network topology, the collaborative scheduling strategy and the dynamic protection setting value through a distributed optimal power flow algorithm to obtain a power allocation scheme corresponding to the target substation.

[0006] In combination with the first aspect, in a first implementation method of the first aspect of the present application, the real-time voltage and current data of the target substation are topologically identified to obtain a dynamic distribution network topology structure, including: performing data cleaning on the real-time voltage and current data of the target substation to obtain a preprocessed data set; performing time series segmentation processing on the preprocessed data set to obtain multiple time window data subsets; performing feature extraction processing on the multiple time window data subsets through a sliding window algorithm to obtain a time-varying feature vector; performing principal component analysis processing on the time-varying feature vector to obtain a reduced dimension feature set; performing similarity calculation processing on the reduced dimension feature set through an adaptive kernel function algorithm to obtain an inter-node association matrix; performing threshold segmentation processing on the inter-node association matrix to obtain an initial topological connection relationship; performing community detection processing on the initial topological connection relationship through a graph clustering algorithm to obtain a subnet division result; performing boundary node identification processing on the subnet division result to obtain a set of key connection points; performing connection optimization processing on the set of key connection points through a recursive minimum spanning tree algorithm to obtain an optimized topology structure; performing dynamic time warping processing on the optimized topology structure to obtain the dynamic distribution network topology structure.

[0007] In combination with the first aspect, in a second implementation of the first aspect of the present application, the output characteristics of the distributed energy units in the dynamic distribution network topology are dynamically adjusted by a virtual inertia control algorithm to obtain a power adjustment signal that simulates the characteristics of a synchronous generator, including: performing node classification processing on the dynamic distribution network topology to obtain a set of distributed energy units; performing output characteristic parameter extraction processing on each unit in the set of distributed energy units to obtain an initial characteristic parameter matrix; performing parameter optimization processing on the initial characteristic parameter matrix by an adaptive particle swarm optimization algorithm to obtain an optimized characteristic parameter set; performing virtual rotational inertia calculation processing on the optimized characteristic parameter set to obtain a virtual inertia distribution map ; Decompose the grid frequency signal through the wavelet transform algorithm to obtain the multi-scale frequency characteristics; Perform frequency response characteristic matching processing on the multi-scale frequency characteristics and the virtual inertia distribution diagram to obtain the virtual inertia control gain coefficient; Dynamically adjust the virtual inertia control gain coefficient through the fuzzy adaptive PID algorithm to obtain the adaptive control parameter; Perform virtual synchronous machine model mapping processing on the adaptive control parameter to obtain the virtual synchronous machine state variable; Perform feedforward compensation processing on the virtual synchronous machine state variable through the predictive control algorithm to obtain the compensation control quantity; Perform power synthesis processing on the compensation control quantity and the virtual synchronous machine state variable to obtain the power regulation signal simulating the characteristics of the synchronous generator.

[0008] In combination with the first aspect, in a third implementation of the first aspect of the present application, the power regulation signal and the pre-acquired power supply and demand data are collaboratively optimized through a multi-agent reinforcement learning algorithm to obtain a collaborative scheduling strategy, including: performing time series alignment processing on the power regulation signal and the pre-acquired power supply and demand data to obtain a synchronized data set; performing multi-scale analysis processing on the synchronized data set through a wavelet decomposition algorithm to obtain a feature subset; performing principal component analysis processing on the feature subset to obtain a reduced-dimensional feature vector; performing agent division processing on the reduced-dimensional feature vector through a clustering algorithm to obtain a multi-agent system structure; and The system structure is subjected to state space modeling processing to obtain the state transfer function of each agent; the state transfer function of each agent is subjected to action strategy learning processing through the deep deterministic policy gradient algorithm to obtain the initial strategy set; the initial strategy set is subjected to multi-agent collaboration mechanism design processing to obtain the inter-agent communication protocol; the inter-agent communication protocol is subjected to strategy optimization processing through the distributed dominant actor critic algorithm to obtain the optimized strategy set; the optimized strategy set is subjected to reward function design processing to obtain the global value evaluation index; the global value evaluation index is subjected to strategy comprehensive processing through the hierarchical reinforcement learning algorithm to obtain the collaborative scheduling strategy.

[0009] In combination with the first aspect, in a fourth implementation of the first aspect of the present application, the health status assessment is performed on the multi-source heterogeneous data of the target substation collected in advance and the collaborative scheduling strategy through a deep transfer learning algorithm to obtain an equipment health status assessment result, including: performing outlier detection and interpolation processing on the multi-source heterogeneous data of the target substation collected in advance to obtain a corrected data set; performing high-order feature extraction processing on the corrected data set through a tensor decomposition algorithm to obtain a multi-dimensional feature tensor; performing spatiotemporal correlation analysis processing on the multi-dimensional feature tensor and the collaborative scheduling strategy to obtain a spatiotemporal feature matrix; and performing a variational inference algorithm on the spatiotemporal feature tensor to obtain a multi-dimensional feature tensor. The probability distribution estimation processing is performed on the feature matrix to obtain the latent variable distribution; the migration mapping processing is performed on the latent variable distribution to obtain the cross-domain feature representation; the topological structure learning processing is performed on the cross-domain feature representation through the graph attention network algorithm to obtain the device association graph; the spectral clustering processing is performed on the device association graph to obtain the device status group; the domain invariant feature extraction processing is performed on the device status group through the adversarial training algorithm to obtain the robustness feature set; the multi-task learning processing is performed on the robustness feature set to obtain the multidimensional health index; the uncertainty quantification processing is performed on the multidimensional health index through the Bayesian inference algorithm to obtain the device health status assessment result.

[0010] In combination with the first aspect, in a fifth implementation of the first aspect of the present application, the dynamic distribution network topology and the equipment health status assessment results are dynamically optimized by an adaptive fuzzy logic algorithm to obtain a dynamic protection setting value, including: performing sensitivity analysis on the dynamic distribution network topology to obtain a key node set; performing reliability quantification on the equipment health status assessment results to obtain an equipment reliability index; performing comprehensive evaluation on the key node set and the equipment reliability index by a fuzzy c-means clustering algorithm to obtain a fuzzy rule base; performing adaptive adjustment on the fuzzy rule base to obtain a dynamic fuzzy inference system; performing member function optimization on the dynamic fuzzy inference system by a genetic algorithm to obtain an optimized fuzzy set; performing TS fuzzy model construction on the optimized fuzzy set to obtain a nonlinear system model; performing parameter identification on the nonlinear system model by a particle swarm optimization algorithm to obtain a model parameter set; performing sensitivity analysis on the model parameter set to obtain a key parameter ranking; performing adaptive adjustment on the key parameter ranking by a fuzzy neural network algorithm to obtain a dynamic protection threshold; and performing coordinated optimization on the dynamic protection threshold to obtain the dynamic protection setting value.

[0011] In combination with the first aspect, in a sixth implementation method of the first aspect of the present application, the dynamic distribution network topology, the collaborative dispatching strategy and the dynamic protection setting value are optimized and allocated by a distributed optimal power flow algorithm to obtain a power allocation scheme corresponding to the target substation, including: performing graph segmentation processing on the dynamic distribution network topology to obtain segmentation graph data; performing boundary condition setting processing on the segmentation graph data by a Lagrangian relaxation algorithm to obtain distributed optimization constraints; performing time series decomposition processing on the collaborative dispatching strategy to obtain a multi-time scale dispatching target; performing objective function construction processing on the multi-time scale dispatching target and the distributed optimization constraints by an alternating direction multiplier method to obtain a distribution The invention provides an optimization problem model; the dynamic protection setting value is transformed into a safety constraint to obtain a power flow calculation safety boundary; the distributed optimization problem model is iteratively solved by a distributed primal-dual algorithm to obtain an initial power allocation scheme; the initial power allocation scheme is subjected to node voltage sensitivity analysis to obtain a voltage control strategy; the voltage control strategy is locally optimized by a distributed Newton method to obtain an optimized voltage distribution; the optimized voltage distribution and the initial power allocation scheme are coordinated and matched to obtain a balanced power allocation scheme; the balanced power allocation scheme is globally converged by a distributed consistency algorithm to obtain the power allocation scheme corresponding to the target substation.

[0012] In a second aspect, the present application provides a substation data collaborative identification device, the substation data collaborative identification device comprising:

[0013] A processing module is used to perform topology identification processing on the real-time voltage and current data of the target substation to obtain a dynamic distribution network topology structure;

[0014] A regulation module, used for dynamically regulating the output characteristics of the distributed energy units in the dynamic distribution network topology structure through a virtual inertia control algorithm to obtain a power regulation signal simulating the characteristics of a synchronous generator;

[0015] A collaborative module, used to perform collaborative optimization processing on the power regulation signal and the pre-acquired power supply and demand data through a multi-agent reinforcement learning algorithm to obtain a collaborative scheduling strategy;

[0016] An evaluation module is used to perform health status evaluation processing on the pre-collected multi-source heterogeneous data of the target substation and the collaborative scheduling strategy through a deep transfer learning algorithm to obtain a health status evaluation result of the equipment;

[0017] An optimization module, used for dynamically optimizing the dynamic distribution network topology and the equipment health status evaluation results through an adaptive fuzzy logic algorithm to obtain a dynamic protection setting value;

[0018] The allocation module is used to optimize and allocate the dynamic distribution network topology, the coordinated scheduling strategy and the dynamic protection setting value through a distributed optimal power flow algorithm to obtain a power allocation plan corresponding to the target substation.

[0019] The third aspect of the present application provides a substation data collaborative identification device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the substation data collaborative identification device executes the above-mentioned substation data collaborative identification method.

[0020] A fourth aspect of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the above-mentioned substation data collaborative identification method.

[0021] In the technical solution provided by the present application, the real-time voltage and current data are processed for topological identification through the adaptive graph structure analysis algorithm, which can quickly and accurately capture the dynamic changes of the distribution network topology structure, greatly improving the system's response speed and adaptability to changes in network structure. Secondly, the introduction of the virtual inertia control algorithm solves the problem of insufficient system inertia caused by the access of distributed energy resources. By dynamically adjusting the output characteristics of distributed energy units, the effective simulation of the characteristics of synchronous generators is achieved, which significantly enhances the stability and reliability of the system. Furthermore, the application of multi-agent reinforcement learning algorithms enables the system to autonomously learn the optimal collaborative scheduling strategy in a complex and changing environment, greatly improving the flexibility and efficiency of energy scheduling, effectively balancing the supply and demand relationship, and reducing operating costs. In addition, the introduction of deep transfer learning algorithms solves the adaptability problem of traditional health status assessment methods in the face of data scarcity or new equipment. Through transfer learning technology, the system can use limited data to quickly build an accurate health status assessment model, improving the efficiency of equipment management and the accuracy of predictive maintenance. At the same time, the application of adaptive fuzzy logic algorithms enables the system to adjust the protection setting value in real time according to the dynamically changing network topology and equipment health status, greatly improving the safety and reliability of the system. Finally, the introduction of the distributed optimal power flow algorithm solves the complex optimization problems in large-scale distribution networks. Through distributed computing, the solution efficiency is significantly improved. At the same time, network constraints, equipment limitations and safety boundaries are taken into account to achieve more accurate and reliable power distribution. This comprehensive collaborative identification method not only improves the operating efficiency and reliability of the substation, but also enhances the intelligence level and adaptive ability of the system. By integrating a variety of advanced algorithms, this method can better cope with the complexity and dynamics of the power grid. The substation data collaborative identification method proposed in the present invention has achieved remarkable results in improving system performance, reducing operating costs, and enhancing reliability and safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0023] Figure 1 A schematic diagram of an embodiment of a method for collaborative identification of substation data in an embodiment of the present application;

[0024] Figure 2 It is a schematic diagram of an embodiment of a substation data collaborative identification device in an embodiment of the present application. DETAILED DESCRIPTION

[0025] Embodiments of the present application provide a method, apparatus, device and storage medium for collaborative identification of substation data. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0026] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 , an embodiment of the substation data collaborative identification method in the embodiment of the present application includes:

[0027] Step S101, performing topology identification processing on the real-time voltage and current data of the target substation to obtain a dynamic distribution network topology structure;

[0028] Step S102, dynamically adjusting the output characteristics of the distributed energy units in the dynamic distribution network topology structure by using a virtual inertia control algorithm to obtain a power adjustment signal simulating the characteristics of a synchronous generator;

[0029] Step S103: performing collaborative optimization processing on the power regulation signal and the pre-acquired power supply and demand data through a multi-agent reinforcement learning algorithm to obtain a collaborative scheduling strategy;

[0030] Step S104: Perform health status assessment on the pre-collected multi-source heterogeneous data and collaborative dispatching strategy of the target substation through a deep transfer learning algorithm to obtain a health status assessment result of the equipment;

[0031] Step S105, dynamically optimizing the dynamic distribution network topology and equipment health status assessment results through an adaptive fuzzy logic algorithm to obtain a dynamic protection setting value;

[0032] Step S106: Optimize and allocate the dynamic distribution network topology, coordinated dispatch strategy and dynamic protection setting value through a distributed optimal power flow algorithm to obtain a power allocation plan corresponding to the target substation.

[0033] It is understandable that the execution subject of the present application may be a substation data collaborative identification device, or a terminal or a server, which is not specifically limited here. The present application embodiment is described by taking a server as the execution subject as an example.

[0034] Specifically, the real-time voltage and current data of the target substation are topologically identified and processed to obtain the dynamic distribution network topology. This process uses an adaptive graph structure analysis algorithm to perform data cleaning, time series segmentation, feature extraction and other processing on the voltage and current data, and finally constructs a dynamic topology that reflects the current state of the distribution network. Then, the virtual inertia control algorithm is used to dynamically adjust the output characteristics of the distributed energy units in the dynamic distribution network topology to obtain a power regulation signal that simulates the characteristics of the synchronous generator. The virtual inertia control algorithm generates a power regulation signal that can simulate the characteristics of the traditional synchronous generator through the steps of parameter optimization of the output characteristics of the distributed energy unit, virtual moment of inertia calculation, frequency response characteristic matching, etc., thereby enhancing the system stability. The multi-agent reinforcement learning algorithm is used to coordinately optimize the power regulation signal and the pre-acquired power supply and demand data to obtain a coordinated scheduling strategy. The algorithm regards each distributed energy unit as an agent, constructs a multi-agent system structure through the alignment of time series data, multi-scale analysis, feature dimensionality reduction and other processing, and learns the optimal collaborative scheduling strategy through methods such as deep deterministic policy gradient algorithm and distributed dominant actor critic algorithm.

[0035] In order to evaluate the health status of equipment, the method uses a deep transfer learning algorithm to evaluate the health status of the multi-source heterogeneous data and collaborative dispatching strategy of the pre-collected target substation, and obtains the equipment health status evaluation results. The deep transfer learning algorithm realizes the knowledge transfer from the source domain to the target domain by processing the outlier detection, high-order feature extraction, and spatiotemporal correlation analysis of multi-source heterogeneous data, effectively solves the problem of data scarcity, and improves the accuracy of health status evaluation. The dynamic distribution network topology and equipment health status evaluation results are dynamically optimized using an adaptive fuzzy logic algorithm to obtain the dynamic protection setting value. The algorithm realizes the dynamic optimization of the protection setting value through the steps of sensitivity analysis of key nodes, quantification of equipment reliability, and construction of fuzzy rule base, thereby improving the safety and reliability of the system.

[0036] The distributed optimal power flow algorithm is used to optimize the dynamic distribution network topology, collaborative dispatching strategy and dynamic protection setting value, and obtain the power allocation scheme corresponding to the target substation. The distributed optimal power flow algorithm constructs a distributed optimization problem model through graph segmentation, boundary condition setting, multi-time scale dispatching target decomposition, and obtains the final power allocation scheme through a distributed solution algorithm. For example, a substation is connected to multiple distributed photovoltaic power generation units and wind power generation units. Through the real-time collected voltage and current data, the adaptive graph structure analysis algorithm identifies the dynamic topology of the current distribution network, including 10 key nodes and 15 connecting lines. The virtual inertia control algorithm dynamically adjusts these distributed energy units and calculates the virtual inertia distribution map, where the virtual inertia value of the photovoltaic unit is 0.5 seconds and that of the wind power unit is 0.8 seconds. The multi-agent reinforcement learning algorithm generates a 24-hour collaborative dispatching strategy based on power regulation signals and power supply and demand data, which takes into account load fluctuations and output prediction of renewable energy. The deep transfer learning algorithm processed multi-source heterogeneous data including equipment operating parameters and environmental data, and evaluated that the health index of the transformer was 0.92 and the reliability index of the main switchgear was 0.95. The adaptive fuzzy logic algorithm dynamically adjusted the protection setting value based on this information, adjusting the overcurrent protection setting value of a key line from 500A to 480A. Finally, the distributed optimal power flow algorithm comprehensively considered all factors and generated an optimized power distribution plan, which increased the utilization rate of renewable energy to 85% while ensuring the safe and stable operation of the system. Through this collaborative identification method, the substation has achieved more efficient and reliable operation.

[0037] In the embodiment of the present application, the real-time voltage and current data are processed for topological identification by an adaptive graph structure analysis algorithm, which can quickly and accurately capture the dynamic changes of the distribution network topology structure, greatly improving the system's response speed and adaptability to changes in network structure. Secondly, the introduction of the virtual inertia control algorithm solves the problem of insufficient system inertia caused by the access of distributed energy resources. By dynamically adjusting the output characteristics of distributed energy units, the effective simulation of the characteristics of synchronous generators is achieved, which significantly enhances the stability and reliability of the system. Furthermore, the application of the multi-agent reinforcement learning algorithm enables the system to autonomously learn the optimal collaborative scheduling strategy in a complex and changing environment, greatly improving the flexibility and efficiency of energy scheduling, effectively balancing the supply and demand relationship, and reducing operating costs. In addition, the introduction of the deep transfer learning algorithm solves the adaptability problem of traditional health status assessment methods in the face of data scarcity or new equipment. Through the transfer learning technology, the system can use limited data to quickly build an accurate health status assessment model, which improves the efficiency of equipment management and the accuracy of predictive maintenance. At the same time, the application of the adaptive fuzzy logic algorithm enables the system to adjust the protection setting value in real time according to the dynamically changing network topology and equipment health status, greatly improving the safety and reliability of the system. Finally, the introduction of the distributed optimal power flow algorithm solves the complex optimization problems in large-scale distribution networks. Through distributed computing, the solution efficiency is significantly improved. At the same time, network constraints, equipment limitations and safety boundaries are taken into account to achieve more accurate and reliable power distribution. This comprehensive collaborative identification method not only improves the operating efficiency and reliability of the substation, but also enhances the intelligence level and adaptive ability of the system. By integrating a variety of advanced algorithms, this method can better cope with the complexity and dynamics of the power grid. The substation data collaborative identification method proposed in the present invention has achieved remarkable results in improving system performance, reducing operating costs, and enhancing reliability and safety.

[0038] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0039] (1) Perform data cleaning on the real-time voltage and current data of the target substation to obtain a preprocessed data set;

[0040] (2) Perform time series segmentation processing on the preprocessed data set to obtain multiple time window data subsets;

[0041] (3) Perform feature extraction on multiple time window data subsets through a sliding window algorithm to obtain time-varying feature vectors;

[0042] (4) Perform principal component analysis on the time-varying feature vector to obtain a reduced-dimensional feature set;

[0043] (5) The similarity calculation of the reduced-dimensional feature set is performed through the adaptive kernel function algorithm to obtain the node association matrix;

[0044] (6) Perform threshold segmentation on the node association matrix to obtain the initial topological connection relationship;

[0045] (7) Perform community detection on the initial topological connection relationship through the graph clustering algorithm to obtain the subnet division result;

[0046] (8) Perform boundary node identification processing on the subnet division results to obtain a set of key connection points;

[0047] (9) The key connection point set is optimized through the recursive minimum spanning tree algorithm to obtain the optimized topological structure;

[0048] (10) The optimized topology is dynamically time-warped to obtain the dynamic distribution network topology.

[0049] Specifically, outliers are first removed, missing data are filled, and noise is eliminated to obtain a reliable preprocessed data set. Subsequently, the preprocessed data set is segmented into time series to divide the continuous data stream into multiple time window data subsets, each of which represents the substation operation status within a specific time period. The sliding window algorithm is used to extract features from multiple time window data subsets. The sliding window algorithm extracts statistical features (such as mean, variance, peak, etc.) in each window by sliding a fixed-size window on the data sequence, thereby obtaining a time-varying feature vector that reflects the characteristics of the data over time. In order to reduce the data dimension and retain key information, the time-varying feature vector is processed by principal component analysis. Principal component analysis calculates the covariance matrix of the feature vector and selects the principal components with the largest contribution rate to obtain a reduced-dimensional feature set.

[0050] The adaptive kernel function algorithm is used to calculate the similarity of the reduced dimension feature set. The adaptive kernel function algorithm automatically adjusts the kernel function parameters according to the data distribution characteristics, calculates the similarity between the feature vectors, and finally obtains the association matrix that reflects the strength of the relationship between the nodes. The node association matrix is ​​threshold segmented, and the node pairs with similarity higher than the set threshold are regarded as connections, thereby obtaining the initial topological connection relationship. In order to further optimize the network structure, the initial topological connection relationship is processed by community detection through the graph clustering algorithm. The graph clustering algorithm uses the Laplace matrix eigenvalues ​​of the graph to divide the network into multiple tightly connected subnetworks to obtain the subnetwork partitioning results. Subsequently, the subnetwork partitioning results are processed for boundary node identification to find the key nodes connecting different subnetworks and obtain the key connection point set. The recursive minimum spanning tree algorithm is used to optimize the connection of the key connection point set. The recursive minimum spanning tree algorithm continuously constructs and optimizes the minimum spanning tree to minimize the total connection cost while ensuring network connectivity, thereby obtaining the optimized topological structure. Finally, the optimized topological structure is processed by dynamic time warping, and the static topological structure is combined with time information to obtain the dynamic distribution network topology that changes with time.

[0051] For example, a substation has 100 monitoring points, collecting voltage and current data once a second. First, the 24-hour raw data is cleaned to remove about 0.5% of outliers and noise data. Then the cleaned data is segmented into 15-minute time windows to obtain 96 time window data subsets. Through the sliding window algorithm, 10 statistical features are extracted for each subset to form a 960-dimensional time-varying feature vector. The principal component analysis reduces the feature vector to 50 dimensions, retaining 95% of the information. The adaptive kernel function algorithm calculates a 100x100 node association matrix, sets the similarity threshold to 0.8, and obtains the initial topological connection relationship. The graph clustering algorithm divides the network into 5 subnets and identifies 15 key connection points. After optimization by the recursive minimum spanning tree algorithm, an optimized topological structure with 100 nodes and 120 edges is obtained. Finally, through dynamic time warping, a dynamic model reflecting the changes in the topological structure of the distribution network within 24 hours is obtained.

[0052] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0053] (1) Classify the nodes of the dynamic distribution network topology to obtain a set of distributed energy units;

[0054] (2) Extract the output characteristic parameters of each unit in the distributed energy unit set to obtain an initial characteristic parameter matrix;

[0055] (3) The initial characteristic parameter matrix is ​​optimized by using the adaptive particle swarm optimization algorithm to obtain the optimized characteristic parameter set;

[0056] (4) Calculate the virtual moment of inertia of the optimized characteristic parameter set to obtain a virtual moment of inertia distribution diagram;

[0057] (5) Decompose the power grid frequency signal through the wavelet transform algorithm to obtain multi-scale frequency characteristics;

[0058] (6) Perform frequency response characteristic matching processing on the multi-scale frequency characteristics and the virtual inertia distribution diagram to obtain the virtual inertia control gain coefficient;

[0059] (7) The virtual inertia control gain coefficient is dynamically adjusted through the fuzzy adaptive PID algorithm to obtain the adaptive control parameters;

[0060] (8) Mapping the adaptive control parameters to the virtual synchronous machine model to obtain the virtual synchronous machine state variables;

[0061] (9) Perform feedforward compensation on the virtual synchronous machine state variables through a predictive control algorithm to obtain a compensation control quantity;

[0062] (10) The compensation control quantity and the virtual synchronous machine state variable are subjected to power synthesis processing to obtain a power regulation signal simulating the characteristics of the synchronous generator.

[0063] Specifically, the nodes of the dynamic distribution network topology are first classified to identify distributed energy units, such as photovoltaic power generation, wind power generation, etc., to form a distributed energy unit set. Subsequently, the output characteristic parameters of each unit in the set are extracted, including rated power, maximum output power, response time, etc., to obtain the initial characteristic parameter matrix. The adaptive particle swarm optimization algorithm is used to optimize the parameters of the initial characteristic parameter matrix. The algorithm simulates the behavior of the particle swarm, continuously adjusts the particle position and speed, and searches for the optimal solution. In this process, the algorithm will adaptively adjust the inertia weight, learning factor and other parameters according to the objective function, and finally obtain the optimized characteristic parameter set.

[0064] The virtual moment of inertia is calculated and processed for the optimized characteristic parameter set to obtain the virtual inertia distribution diagram. The virtual moment of inertia is a key parameter for simulating the rotor inertia of synchronous generators. By calculating the virtual inertia value of each distributed energy unit, the virtual inertia distribution diagram of the entire network is formed. The grid frequency signal is decomposed and processed by the wavelet transform algorithm. The wavelet transform can analyze the signal in both the time and frequency domains, decompose the grid frequency signal into sub-signals of different scales, and obtain multi-scale frequency characteristics. These characteristics reflect the changing characteristics of the grid frequency at different time scales.

[0065] The frequency response characteristics of the multi-scale frequency characteristics and the virtual inertia distribution diagram are matched to obtain the virtual inertia control gain coefficient. This step determines the most suitable control gain coefficient by analyzing the corresponding relationship between the frequency characteristics and the virtual inertia to achieve good frequency response characteristics. The virtual inertia control gain coefficient is dynamically adjusted by the fuzzy adaptive PID algorithm. The fuzzy adaptive PID algorithm combines the advantages of fuzzy logic and PID control, and can automatically adjust the PID parameters according to the system state to obtain adaptive control parameters. The adaptive control parameters are mapped to the virtual synchronous machine model to obtain the virtual synchronous machine state variables. The virtual synchronous machine model is a control strategy that simulates the characteristics of the synchronous generator. By mapping the adaptive control parameters to the virtual synchronous machine model, the state variables reflecting the operating state of the virtual synchronous machine are obtained.

[0066] The virtual synchronous machine state variables are fed-forward compensated by the predictive control algorithm. The predictive control algorithm predicts the future state based on the system model and optimizes the control input to obtain the compensation control quantity. This feed-forward compensation can respond to system changes in advance and improve control performance. The compensation control quantity and the virtual synchronous machine state variables are processed by power synthesis to obtain a power regulation signal that simulates the characteristics of the synchronous generator. This power regulation signal combines the results of virtual inertial control and predictive control, and can make the distributed energy unit exhibit characteristics similar to those of a synchronous generator.

[0067] For example, a substation is connected to 10 distributed energy units, including 5 photovoltaic power generation units and 5 wind power generation units. Through node classification processing, these 10 units are identified and their output characteristic parameters are extracted, such as the maximum output power of the photovoltaic unit is 2MW, the response time is 0.1 seconds, the maximum output power of the wind power unit is 3MW, the response time is 0.3 seconds, etc., forming a 10x5 initial characteristic parameter matrix. The adaptive particle swarm optimization algorithm is iterated 1000 times to obtain the optimized characteristic parameter set. The virtual moment of inertia calculation process obtains the virtual inertia distribution diagram, in which the virtual inertia value of the photovoltaic unit is 0.5 seconds on average and that of the wind power unit is 0.8 seconds on average. The wavelet transform algorithm decomposes the collected 24-hour grid frequency signal into frequency characteristics of 5 scales. By matching the frequency response characteristics, the initial virtual inertia control gain coefficient matrix is ​​obtained. The fuzzy adaptive PID algorithm dynamically adjusts the PID parameters according to the real-time operation data, such as adjusting the proportional gain from 1.2 to 1.5 and the integral time from 0.5 seconds to 0.4 seconds. The virtual synchronous machine model maps the state variables that reflect the operating status of each unit, such as rotor angular velocity, phase angle, etc. The predictive control algorithm calculates the compensation control quantity based on the load forecast for the next 5 minutes. Finally, through power synthesis processing, the power regulation signal of each distributed energy unit is obtained, so that it can respond to system frequency changes like a synchronous generator, improving the stability of the distribution network.

[0068] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0069] (1) Perform time series alignment processing on the power regulation signal and the pre-acquired power supply and demand data to obtain a synchronized data set;

[0070] (2) Perform multi-scale analysis on the synchronized data set using a wavelet decomposition algorithm to obtain feature subsets;

[0071] (3) Perform principal component analysis on the feature subset to obtain a reduced-dimensional feature vector;

[0072] (4) Using a clustering algorithm to divide the reduced-dimensional feature vector into agents, we obtain the multi-agent system structure.

[0073] (5) Perform state space modeling on the multi-agent system structure to obtain the state transfer function of each agent;

[0074] (6) Perform action strategy learning on the state transfer function of each agent through a deep deterministic policy gradient algorithm to obtain an initial strategy set;

[0075] (7) Design a multi-agent collaboration mechanism for the initial strategy set to obtain an inter-agent communication protocol;

[0076] (8) The distributed dominant actor-critic algorithm is used to optimize the inter-agent communication protocol to obtain an optimized strategy set;

[0077] (9) Design a reward function for the optimization strategy set to obtain a global value evaluation index;

[0078] (10) The global value evaluation indicators are strategically processed through a hierarchical reinforcement learning algorithm to obtain a collaborative scheduling strategy.

[0079] Specifically, the power regulation signal and the pre-acquired power supply and demand data are first aligned in time series. This step unifies the data of different time scales into the same time series through interpolation or resampling techniques to obtain a synchronized data set. Subsequently, the synchronized data set is subjected to multi-scale analysis by a wavelet decomposition algorithm. The wavelet decomposition algorithm uses wavelet functions of different scales to decompose the original signal into multiple frequency subbands, thereby obtaining a feature subset that reflects the characteristics of the data at different time scales. The feature subset is subjected to principal component analysis to obtain a reduced-dimensional feature vector. Principal component analysis calculates the eigenvalues ​​and eigenvectors of the covariance matrix and selects several principal components with the largest contribution rate to achieve data dimensionality reduction while retaining key information. Then, the reduced-dimensional feature vector is subjected to agent division by a clustering algorithm. Clustering algorithms, such as K-means or hierarchical clustering, group data points according to the similarity of feature vectors, and each group represents an agent, thereby obtaining a multi-agent system structure.

[0080] The state space modeling of the multi-agent system structure is carried out to obtain the state transition function of each agent. The state space model describes the relationship between the current state, action and next state of the agent, laying the foundation for subsequent reinforcement learning. The action strategy learning process of the state transition function of each agent is carried out through the deep deterministic policy gradient algorithm. This algorithm combines deep learning and deterministic policy gradient methods, approximates the value function and policy function through neural networks, and learns the initial policy set. The multi-agent collaboration mechanism design process is carried out on the initial policy set to obtain the inter-agent communication protocol. This step defines how the agents exchange information and coordinate actions, including message passing format, collaboration strategy, etc. The inter-agent communication protocol is optimized through the distributed advantage actor-critic algorithm. In the multi-agent environment, each agent simultaneously learns an actor network (for selecting actions) and a critic network (for evaluating action value), improves the coordination and efficiency of the strategy through distributed learning, and obtains the optimized strategy set.

[0081] The reward function is designed for the set of optimization strategies to obtain the global value evaluation index. The reward function defines the goals of the system, such as minimizing costs, maximizing renewable energy utilization, etc. The global value evaluation index comprehensively considers the trade-offs of multiple goals. Finally, the global value evaluation index is subjected to strategy comprehensive processing through the hierarchical reinforcement learning algorithm. Hierarchical reinforcement learning decomposes complex decision-making problems into multiple levels. The high-level layer formulates macro strategies, and the low-level layer performs specific actions. The final collaborative scheduling strategy is obtained through interactive learning between levels.

[0082] For example, a substation manages 10 distributed energy units, including photovoltaic, wind power and energy storage systems. First, the power regulation signals of these units with a sampling interval of 5 minutes are time-aligned with the power supply and demand forecast data with a sampling interval of 15 minutes, and unified into a synchronized data set with a sampling interval of 5 minutes. Through the wavelet decomposition algorithm, the 24-hour synchronized data set is decomposed into feature subsets of 4 scales, reflecting the change characteristics of 15 minutes, 1 hour, 4 hours and 12 hours respectively. The principal component analysis reduces the original 100-dimensional features to 20 dimensions, retaining 95% of the information. The K-means clustering algorithm divides the 10 energy units into 3 intelligent groups: power generation group, energy storage group and load group.

[0083] State space modeling defines state variables (such as current output power, remaining capacity) and action variables (such as power adjustment) for each agent. The deep deterministic policy gradient algorithm learns the initial strategy of each agent through a 3-layer neural network (20 nodes in the input layer, 50 nodes in the hidden layer, and the output layer is the action dimension). The multi-agent collaboration mechanism designs a communication protocol based on a shared experience pool, and the agents exchange state information every 5 minutes. The distributed advantage actor-critic algorithm optimizes the collaboration strategy in 1000 iterations, increasing the initial average reward from 100 to 150.

[0084] The reward function design comprehensively considers the three goals of cost minimization (weight 0.4), maximization of renewable energy utilization (weight 0.3) and load balancing (weight 0.3). The hierarchical reinforcement learning algorithm designs a two-layer structure: the top layer formulates the overall allocation strategy every hour, and the bottom layer performs specific power regulation every 5 minutes. After 10,000 rounds of training, the obtained collaborative scheduling strategy achieved 95% renewable energy utilization in simulation tests, while controlling load fluctuations within ±3%.

[0085] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0086] (1) Perform outlier detection and interpolation processing on the multi-source heterogeneous data of the target substation collected in advance to obtain a corrected data set;

[0087] (2) Using the tensor decomposition algorithm to extract high-order features from the corrected data set to obtain a multi-dimensional feature tensor;

[0088] (3) Perform spatiotemporal correlation analysis on the multi-dimensional feature tensor and the collaborative scheduling strategy to obtain the spatiotemporal feature matrix;

[0089] (4) Using the variational inference algorithm, the probability distribution of the spatiotemporal feature matrix is ​​estimated to obtain the distribution of latent variables;

[0090] (5) Perform migration mapping on the latent variable distribution to obtain cross-domain feature representation;

[0091] (6) The topological structure of cross-domain feature representation is learned through the graph attention network algorithm to obtain the device association graph;

[0092] (7) Perform spectral clustering on the equipment association graph to obtain equipment status groups;

[0093] (8) Using the adversarial training algorithm, domain-invariant feature extraction is performed on the device state group to obtain a robust feature set;

[0094] (9) Perform multi-task learning on the robust feature set to obtain multi-dimensional health indicators;

[0095] (10) The uncertainty of multidimensional health indicators is quantified through the Bayesian reasoning algorithm to obtain the equipment health status assessment results.

[0096] Specifically, outlier detection and interpolation processing are first performed on the pre-collected multi-source heterogeneous data of the target substation. Outlier detection uses statistical methods or machine learning models to identify outliers, while interpolation processing fills missing values ​​through time series interpolation or similarity-based methods to finally obtain a corrected data set. Subsequently, the corrected data set is subjected to high-order feature extraction processing by a tensor decomposition algorithm. Tensor decomposition algorithms, such as Tucker decomposition or CP decomposition, decompose multidimensional data into core tensors and factor matrices to obtain a multidimensional feature tensor that reflects the high-order structure of the data. The multidimensional feature tensor and the collaborative scheduling strategy are subjected to spatiotemporal correlation analysis processing. This step obtains a spatiotemporal feature matrix by calculating the correlation in the time dimension and the spatial dimension, fusing the feature tensor and the scheduling strategy information. Then, the spatiotemporal feature matrix is ​​subjected to probability distribution estimation processing by a variational inference algorithm. Variational inference obtains the probability distribution of latent variables by optimizing the variational lower bound to approximate the complex posterior distribution, which provides a basis for subsequent transfer learning.

[0097] The latent variable distribution is transferred and mapped to obtain cross-domain feature representation. Transfer mapping aims to transfer the knowledge of the source domain to the target domain, and realize the cross-domain application of knowledge by minimizing the distribution difference or learning domain-invariant features. The topological structure of the cross-domain feature representation is learned by the graph attention network algorithm. The graph attention network uses the attention mechanism to capture the important relationship between nodes and learns the device association graph that reflects the mutual influence between devices. The device association graph is spectrally clustered to obtain device state groups. Spectral clustering uses the Laplacian matrix eigenvalues ​​of the graph to divide the devices into different state groups for subsequent analysis and processing. The device state groups are subjected to domain-invariant feature extraction by the adversarial training algorithm. Adversarial training learns domain-invariant feature representations through the game between the generator and the discriminator, thereby obtaining a robust feature set.

[0098] Multi-task learning is performed on the robustness feature set to obtain multi-dimensional health indicators. Multi-task learning simultaneously optimizes multiple related tasks, such as equipment remaining service life prediction and fault type identification, so as to obtain multi-dimensional health indicators that fully reflect the equipment status. Finally, the uncertainty of multi-dimensional health indicators is quantified through the Bayesian reasoning algorithm. Bayesian reasoning combines prior knowledge and observation data to calculate the posterior distribution, thereby obtaining equipment health status assessment results containing uncertainty information.

[0099] For example, a substation has 50 main devices, including transformers, circuit breakers, and busbars. First, the operating data of these devices within one year is preprocessed to identify about 2% of outliers and correct them through time series interpolation methods. The tensor decomposition algorithm decomposes the corrected data (50 devices × 365 days × 24 hours × 10 parameters) into a core tensor and three factor matrices, extracting 100-dimensional high-order features. The spatiotemporal correlation analysis combines these features with the collaborative scheduling strategy to obtain a 50×100 spatiotemporal feature matrix.

[0100] The variational inference algorithm estimates the probability distribution of each device state, such as the transformer oil temperature follows a normal distribution with a mean of 65°C and a standard deviation of 5°C. Migration mapping migrates these distributions from normal operating conditions to different load conditions to obtain cross-domain feature representations. The graph attention network learns the associations between devices, such as finding that the state of circuit breaker 1 is highly correlated with transformer A (attention weight 0.8). Spectral clustering divides 50 devices into 5 state groups, such as "normal operation group" and "pending maintenance group".

[0101] Adversarial training extracts 20-dimensional domain-invariant features through 10,000 iterations of the generator and discriminator. Multi-task learning simultaneously predicts the remaining service life and failure probability of the equipment, and obtains a two-dimensional health indicator reflecting these two aspects. Finally, Bayesian reasoning gives the health status assessment results of each device, such as the remaining service life of transformer A is 5.2±0.5 years (95% confidence interval), and the probability of failure in the next month is 0.02±0.005. These results provide reliable data support for preventive maintenance and asset management of substations.

[0102] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0103] (1) Perform sensitivity analysis on the dynamic distribution network topology to obtain the key node set;

[0104] (2) Perform reliability quantification on the equipment health status assessment results to obtain equipment reliability indicators;

[0105] (3) Comprehensively evaluate the key node set and equipment reliability indicators through the fuzzy c-means clustering algorithm to obtain the fuzzy rule base;

[0106] (4) Adaptively adjust the fuzzy rule base to obtain a dynamic fuzzy reasoning system;

[0107] (5) The membership function of the dynamic fuzzy inference system is optimized by genetic algorithm to obtain the optimized fuzzy set;

[0108] (6) Perform TS fuzzy model construction on the optimized fuzzy set to obtain a nonlinear system model;

[0109] (7) Perform parameter identification on the nonlinear system model through the particle swarm optimization algorithm to obtain the model parameter set;

[0110] (8) Perform sensitivity analysis on the model parameter set to obtain the ranking of key parameters;

[0111] (9) Adaptively adjust the key parameter ranking through the fuzzy neural network algorithm to obtain the dynamic protection threshold;

[0112] (10) The dynamic protection threshold is coordinated and optimized to obtain the dynamic protection setting value.

[0113] Specifically, the topology of the dynamic distribution network is subjected to sensitivity analysis. By perturbing the parameters of each node and observing the system response, the sensitivity analysis calculates the degree of influence of each node on the system stability, thereby obtaining a set of key nodes. Subsequently, the reliability of the equipment health status assessment results is quantified, and the qualitative health status is converted into a quantitative reliability index, such as quantifying the "good" state as a reliability value of 0.9. The key node set and equipment reliability index are comprehensively evaluated by the fuzzy c-means clustering algorithm. Fuzzy c-means clustering allows data points to belong to multiple clusters to varying degrees. By iteratively optimizing the cluster center and membership matrix, a fuzzy rule base is finally obtained. The fuzzy rule base is adaptively adjusted, and the weights and parameters of the rules are dynamically adjusted according to the real-time operation data to form a dynamic fuzzy reasoning system.

[0114] The dynamic fuzzy inference system is optimized by genetic algorithm. Genetic algorithm simulates natural selection and genetic process, optimizes the shape and parameters of membership function through crossover, mutation and other operations, and obtains optimized fuzzy set. The optimized fuzzy set is constructed by TS fuzzy model. TS fuzzy model divides the input space into multiple fuzzy regions, each region is described by a linear equation, so as to obtain a system model that can describe complex nonlinear relationships. The nonlinear system model is identified by particle swarm optimization algorithm. Particle swarm optimization algorithm simulates group intelligence, continuously adjusts particle position and speed, searches for the optimal parameter combination, and obtains the model parameter set. The model parameter set is subjected to sensitivity analysis, and the influence of each parameter on the model output is calculated, so as to obtain the key parameter ranking. The key parameter ranking is adaptively adjusted by fuzzy neural network algorithm. Fuzzy neural network combines the reasoning ability of fuzzy logic and the learning ability of neural network, and can dynamically adjust parameters according to input data to obtain dynamic protection threshold. Finally, the dynamic protection threshold is coordinated and optimized, considering the mutual influence between equipment and protection coordination requirements, and the final dynamic protection setting value is obtained.

[0115] For example, a substation contains 50 nodes. Through sensitivity analysis, 10 key nodes are identified, among which node 5 has the highest sensitivity of 0.85. The equipment health status assessment results show that the reliability index of the main transformer is 0.92 and that of the circuit breaker is 0.88. Fuzzy c-means clustering divides these data into 5 fuzzy clusters and generates 20 initial fuzzy rules. After dynamic adjustment, the weight of a rule in the rule base, such as "IF the voltage at node 5 is low AND the reliability of the main transformer is high THEN slightly increase the protection threshold", is adjusted from 0.6 to 0.8. After 100 generations of evolution, the genetic algorithm optimizes the parameters of the member function, such as adjusting the membership function of "low voltage" from a triangle to a Gaussian function. The TS fuzzy model constructs 5 fuzzy rules, each corresponding to a linear equation. The particle swarm optimization algorithm determines the coefficients of these linear equations through 1000 iterations. For example, the equation of rule 1 is , sensitivity analysis found that (the voltage at node 5) is the most critical parameter, with a sensitivity of 0.8, is the load rate of the main transformer. Based on real-time data, the fuzzy neural network dynamically adjusts the voltage protection threshold of node 5 from 0.9 per unit to 0.92 per unit. Finally, considering the coordination of adjacent circuit breakers, the final dynamic protection setting value is determined to be 0.91 per unit.

[0116] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0117] (1) Perform graph segmentation processing on the dynamic distribution network topology structure to obtain segmentation graph data;

[0118] (2) The boundary conditions of the segmentation graph data are set by the Lagrangian relaxation algorithm to obtain the distributed optimization constraints;

[0119] (3) Decompose the coordinated scheduling strategy into time series to obtain multi-time scale scheduling objectives;

[0120] (4) The objective function is constructed and processed for the multi-time scale scheduling objectives and distributed optimization constraints through the alternating direction multiplier method to obtain the distributed optimization problem model;

[0121] (5) Perform safety constraint conversion on the dynamic protection setting value to obtain the safety boundary of power flow calculation;

[0122] (6) The distributed optimization problem model is iteratively solved by a distributed primal-dual algorithm to obtain an initial power allocation solution;

[0123] (7) Perform node voltage sensitivity analysis on the initial power allocation scheme to obtain the voltage control strategy;

[0124] (8) Locally optimize the voltage control strategy through the distributed Newton method to obtain the optimized voltage distribution;

[0125] (9) Coordinate and match the optimized voltage distribution and the initial power allocation scheme to obtain a balanced power allocation scheme;

[0126] (10) The balanced power allocation scheme is globally converged through a distributed consistency algorithm to obtain the power allocation scheme corresponding to the target substation.

[0127] Specifically, the dynamic distribution network topology is firstly processed by graph segmentation. Graph segmentation algorithms, such as spectral clustering or minimum cut method, divide the large-scale network into several sub-networks to obtain segmentation graph data. This segmentation helps to reduce the complexity of the problem and facilitates the subsequent distributed optimization. Subsequently, the boundary conditions of the segmentation graph data are set by the Lagrangian relaxation algorithm. Lagrangian relaxation transforms the coupling constraints between sub-networks into penalty terms in the objective function, thereby obtaining distributed optimization constraints. The coordinated scheduling strategy is decomposed into time series. Using methods such as wavelet decomposition or empirical mode decomposition, the scheduling strategy is decomposed into components of different time scales to obtain multi-time scale scheduling objectives. This decomposition helps to capture the dynamic characteristics of the system at different time scales. Then, the alternating direction multiplier method is used to construct the objective function of the multi-time scale scheduling objectives and distributed optimization constraints. The alternating direction multiplier method combines the advantages of dual decomposition and augmented Lagrangian method, can effectively handle optimization problems with complex constraints, and finally obtain a distributed optimization problem model.

[0128] The dynamic protection setting value is transformed into a safety constraint, and the protection setting is converted into a constraint condition in the power flow calculation to obtain the power flow calculation safety boundary. This step ensures that the optimization result meets the system safety operation requirements. The distributed optimization problem model is iteratively solved by the distributed primal-dual algorithm. The algorithm alternately updates the primal variables and dual variables in each subnetwork, achieves global optimization through information exchange, and obtains the initial power allocation plan. The node voltage sensitivity analysis is performed on the initial power allocation plan. The voltage control strategy is obtained by calculating the impact of power changes on node voltage. This step helps to identify the key nodes and effective measures for voltage control. The voltage control strategy is locally optimized by the distributed Newton method. The distributed Newton method performs Newton iterations in each subnetwork and achieves global coordination through boundary information exchange to obtain the optimized voltage distribution.

[0129] The optimized voltage distribution and the initial power allocation scheme are coordinated and matched. This step adjusts the power allocation to meet the optimized voltage requirements and obtains a balanced power allocation scheme. Finally, the distributed consistency algorithm is used to globally converge the balanced power allocation scheme. The distributed consistency algorithm uses iterative information exchange to gradually converge the solutions of each subnetwork and finally obtain the power allocation scheme corresponding to the target substation.

[0130] For example, a substation manages a distribution network with 100 nodes. First, the network is divided into 5 subnetworks by spectral clustering algorithm, each of which contains about 20 nodes. The Lagrangian relaxation algorithm transforms the power exchange constraints between subnetworks into penalty terms in the objective function, and the penalty coefficient is set to 1000. The collaborative scheduling strategy obtains scheduling targets at three time scales of 15 minutes, 1 hour and 4 hours through wavelet decomposition. The alternating direction multiplier method constructs an optimization model with three objectives: minimizing operating costs, maximizing renewable energy utilization and meeting load demand.

[0131] The dynamic protection setting value is converted into the constraint condition in the power flow calculation, such as converting the overcurrent protection setting value of a key line of 480A into a line power limit of 120MW. After 100 iterations, the distributed primal-dual algorithm obtains the initial power allocation scheme, in which the total output of renewable energy is 50MW. The node voltage sensitivity analysis found that node 25 is most sensitive to the system voltage with a sensitivity coefficient of 0.05pu / MW. The distributed Newton method is executed in parallel in 5 subnetworks, and each subnetwork performs 10 Newton iterations to obtain the optimized voltage distribution, which increases the voltage of node 25 from 0.95pu to 0.98pu. The coordinated matching process adjusts the output of renewable energy to 48MW to meet the optimized voltage requirements. After 50 iterations, the distributed consensus algorithm makes the solutions of the 5 subnetworks converge to the global optimal solution. In the final power allocation scheme, the proportion of renewable energy reaches 40%, the system loss is reduced by 5%, and the voltage of all nodes is kept in the range of 0.95pu-1.05pu. This example shows how to decompose a complex distribution network optimization problem into sub-problems that can be solved in parallel through a series of distributed algorithms, so as to efficiently obtain the global optimal solution that satisfies multiple constraints.

[0132] The above describes the substation data collaborative identification method in the embodiment of the present application. The following describes the substation data collaborative identification device in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of the substation data collaborative identification device includes:

[0133] The processing module 201 is used to perform topology identification processing on the real-time voltage and current data of the target substation to obtain a dynamic distribution network topology structure;

[0134] The regulating module 202 is used to dynamically regulate the output characteristics of the distributed energy units in the dynamic distribution network topology structure through a virtual inertia control algorithm to obtain a power regulation signal simulating the characteristics of a synchronous generator;

[0135] The coordination module 203 is used to coordinately optimize the power regulation signal and the pre-acquired power supply and demand data through a multi-agent reinforcement learning algorithm to obtain a coordinated scheduling strategy;

[0136] An evaluation module 204 is used to perform health status evaluation processing on the pre-collected multi-source heterogeneous data of the target substation and the collaborative scheduling strategy through a deep transfer learning algorithm to obtain a device health status evaluation result;

[0137] The optimization module 205 is used to dynamically optimize the dynamic distribution network topology and the equipment health status evaluation results through an adaptive fuzzy logic algorithm to obtain a dynamic protection setting value;

[0138] The allocation module 206 is used to optimize and allocate the dynamic distribution network topology, the coordinated dispatching strategy and the dynamic protection setting value through a distributed optimal power flow algorithm to obtain a power allocation solution corresponding to the target substation.

[0139] Through the collaboration of the above components, the topology recognition and processing of real-time voltage and current data by adaptive graph structure analysis algorithm can quickly and accurately capture the dynamic changes of distribution network topology, greatly improving the system's response speed and adaptability to changes in network structure. Secondly, the introduction of virtual inertia control algorithm solves the problem of insufficient system inertia caused by distributed energy access. By dynamically adjusting the output characteristics of distributed energy units, the effective simulation of synchronous generator characteristics is achieved, which significantly enhances the stability and reliability of the system. Furthermore, the application of multi-agent reinforcement learning algorithm enables the system to autonomously learn the optimal collaborative scheduling strategy in a complex and changing environment, greatly improving the flexibility and efficiency of energy scheduling, effectively balancing the supply and demand relationship, and reducing operating costs. In addition, the introduction of deep transfer learning algorithm solves the adaptability problem of traditional health status assessment methods in the face of data scarcity or new equipment. Through transfer learning technology, the system can use limited data to quickly build an accurate health status assessment model, improving the efficiency of equipment management and the accuracy of predictive maintenance. At the same time, the application of adaptive fuzzy logic algorithm enables the system to adjust the protection setting value in real time according to the dynamically changing network topology and equipment health status, greatly improving the safety and reliability of the system. Finally, the introduction of the distributed optimal power flow algorithm solves the complex optimization problems in large-scale distribution networks. Through distributed computing, the solution efficiency is significantly improved. At the same time, network constraints, equipment limitations and safety boundaries are taken into account to achieve more accurate and reliable power distribution. This comprehensive collaborative identification method not only improves the operating efficiency and reliability of the substation, but also enhances the intelligence level and adaptive ability of the system. By integrating a variety of advanced algorithms, this method can better cope with the complexity and dynamics of the power grid. The substation data collaborative identification method proposed in the present invention has achieved remarkable results in improving system performance, reducing operating costs, and enhancing reliability and safety.

[0140] The present application also provides a substation data collaborative identification device, which includes a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the substation data collaborative identification method in the above-mentioned embodiments.

[0141] The present application also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the substation data collaborative identification method.

[0142] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0143] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for collaborative identification of substation data, characterized in that: The substation data collaborative identification method comprises: Perform topology recognition processing on the real-time voltage and current data of the target substation to obtain the dynamic distribution network topology structure; Dynamically adjusting the output characteristics of the distributed energy units in the dynamic distribution network topology structure by using a virtual inertia control algorithm to obtain a power adjustment signal simulating the characteristics of a synchronous generator; The method dynamically adjusts the output characteristics of the distributed energy units in the dynamic distribution network topology structure by using a virtual inertia control algorithm to obtain a power adjustment signal simulating the characteristics of a synchronous generator, including: performing node classification processing on the dynamic distribution network topology structure to obtain a distributed energy unit set; performing output characteristic parameter extraction processing on each unit in the distributed energy unit set to obtain an initial characteristic parameter matrix; performing parameter optimization processing on the initial characteristic parameter matrix by using an adaptive particle swarm optimization algorithm to obtain an optimized characteristic parameter set; performing virtual rotational inertia calculation processing on the optimized characteristic parameter set to obtain a virtual inertia distribution diagram; and performing a wavelet transform algorithm on the power grid frequency. The rate signal is decomposed to obtain a multi-scale frequency feature; the multi-scale frequency feature and the virtual inertia distribution diagram are matched with frequency response characteristics to obtain a virtual inertia control gain coefficient; the virtual inertia control gain coefficient is dynamically adjusted by a fuzzy adaptive PID algorithm to obtain an adaptive control parameter; the adaptive control parameter is mapped to a virtual synchronous machine model to obtain a virtual synchronous machine state variable; the virtual synchronous machine state variable is feedforward compensated by a predictive control algorithm to obtain a compensation control amount; the compensation control amount and the virtual synchronous machine state variable are subjected to power synthesis processing to obtain a power regulation signal simulating the characteristics of a synchronous generator; The power regulation signal and the pre-acquired power supply and demand data are collaboratively optimized through a multi-agent reinforcement learning algorithm to obtain a collaborative scheduling strategy; Performing health status assessment processing on the pre-collected multi-source heterogeneous data of the target substation and the collaborative dispatching strategy through a deep transfer learning algorithm to obtain a health status assessment result of the equipment; The health status assessment process is performed on the multi-source heterogeneous data of the target substation collected in advance and the collaborative scheduling strategy through the deep transfer learning algorithm to obtain the equipment health status assessment result, including: performing outlier detection and interpolation processing on the multi-source heterogeneous data of the target substation collected in advance to obtain a corrected data set; performing high-order feature extraction processing on the corrected data set through a tensor decomposition algorithm to obtain a multi-dimensional feature tensor; performing spatiotemporal correlation analysis processing on the multi-dimensional feature tensor and the collaborative scheduling strategy to obtain a spatiotemporal feature matrix; performing probability distribution estimation processing on the spatiotemporal feature matrix through a variational inference algorithm , obtain the latent variable distribution; perform migration mapping on the latent variable distribution to obtain a cross-domain feature representation; perform topological structure learning on the cross-domain feature representation through a graph attention network algorithm to obtain a device association graph; perform spectral clustering on the device association graph to obtain a device state group; perform domain invariant feature extraction on the device state group through an adversarial training algorithm to obtain a robust feature set; perform multi-task learning on the robust feature set to obtain a multidimensional health indicator; perform uncertainty quantification on the multidimensional health indicator through a Bayesian inference algorithm to obtain the device health status assessment result; Dynamically optimizing the dynamic distribution network topology and the equipment health status assessment results by using an adaptive fuzzy logic algorithm to obtain a dynamic protection setting value; The dynamic distribution network topology, the coordinated dispatching strategy and the dynamic protection setting value are optimized and allocated through a distributed optimal power flow algorithm to obtain a power allocation scheme corresponding to the target substation.

2. The method for collaborative identification of substation data according to claim 1, characterized in that: The topology identification process is performed on the real-time voltage and current data of the target substation to obtain a dynamic distribution network topology structure, including: Performing data cleaning processing on the real-time voltage and current data of the target substation to obtain a preprocessed data set; Performing time series segmentation processing on the preprocessed data set to obtain multiple time window data subsets; Performing feature extraction processing on the plurality of time window data subsets by a sliding window algorithm to obtain a time-varying feature vector; Performing principal component analysis on the time-varying feature vector to obtain a reduced-dimensional feature set; Performing similarity calculation on the dimension reduction feature set by using an adaptive kernel function algorithm to obtain an inter-node association matrix; Performing threshold segmentation processing on the inter-node association matrix to obtain an initial topological connection relationship; Performing community detection processing on the initial topological connection relationship by using a graph clustering algorithm to obtain a subnet division result; Performing boundary node identification processing on the subnet division result to obtain a set of key connection points; Performing connection optimization processing on the key connection point set by a recursive minimum spanning tree algorithm to obtain an optimized topological structure; The optimized topology structure is dynamically time-warped to obtain the dynamic distribution network topology structure.

3. The method for collaborative identification of substation data according to claim 1, characterized in that: The collaborative optimization processing of the power regulation signal and the pre-acquired power supply and demand data by a multi-agent reinforcement learning algorithm to obtain a collaborative scheduling strategy includes: Performing time series alignment processing on the power regulation signal and the pre-acquired power supply and demand data to obtain a synchronized data set; Perform multi-scale analysis on the synchronous data set by wavelet decomposition algorithm to obtain a feature subset; perform principal component analysis on the feature subset to obtain a reduced-dimensional feature vector; Performing agent division processing on the dimension-reduced feature vector by a clustering algorithm to obtain a multi-agent system structure; Performing state space modeling on the multi-agent system structure to obtain a state transfer function of each agent; The state transfer function of each agent is processed through action strategy learning through the deep deterministic policy gradient algorithm to obtain the initial strategy set; Performing multi-agent collaboration mechanism design processing on the initial strategy set to obtain an inter-agent communication protocol; Performing strategy optimization processing on the communication protocol between intelligent agents by using a distributed dominant actor-critic algorithm to obtain an optimized strategy set; Performing reward function design processing on the optimization strategy set to obtain a global value evaluation index; The global value assessment index is subjected to a comprehensive strategic processing through a hierarchical reinforcement learning algorithm to obtain the collaborative scheduling strategy.

4. The method for collaborative identification of substation data according to claim 1, characterized in that: The method of dynamically optimizing the dynamic distribution network topology and the equipment health status evaluation result by using an adaptive fuzzy logic algorithm to obtain a dynamic protection setting value includes: Performing sensitivity analysis on the dynamic distribution network topology to obtain a set of key nodes; Performing reliability quantification processing on the equipment health status assessment result to obtain equipment reliability index; Comprehensively evaluating the key node set and the equipment reliability index by using a fuzzy c-means clustering algorithm to obtain a fuzzy rule base; Adaptively adjusting the fuzzy rule base to obtain a dynamic fuzzy reasoning system; Performing membership function optimization processing on the dynamic fuzzy inference system by using a genetic algorithm to obtain an optimized fuzzy set; Performing TS fuzzy model construction processing on the optimized fuzzy set to obtain a nonlinear system model; Performing parameter identification processing on the nonlinear system model by using a particle swarm optimization algorithm to obtain a model parameter set; Performing sensitivity analysis on the model parameter set to obtain a ranking of key parameters; Adaptively adjust the key parameter ranking by using a fuzzy neural network algorithm to obtain a dynamic protection threshold; The dynamic protection threshold is coordinated and optimized to obtain the dynamic protection setting value.

5. The method for collaborative identification of substation data according to claim 1, characterized in that: The method of optimizing and allocating the dynamic distribution network topology, the coordinated dispatching strategy and the dynamic protection setting value by using a distributed optimal power flow algorithm to obtain a power allocation scheme corresponding to the target substation includes: Performing graph segmentation processing on the dynamic distribution network topology structure to obtain segmentation graph data; performing boundary condition setting processing on the segmentation graph data by using a Lagrangian relaxation algorithm to obtain distributed optimization constraint conditions; Performing time series decomposition processing on the collaborative scheduling strategy to obtain multi-time scale scheduling objectives; The objective function is constructed and processed for the multi-time scale scheduling objective and the distributed optimization constraint condition by an alternating direction multiplier method to obtain a distributed optimization problem model; Performing safety constraint conversion processing on the dynamic protection setting value to obtain a power flow calculation safety boundary; The distributed optimization problem model is iteratively solved by a distributed primal-dual algorithm to obtain an initial power allocation solution; Performing node voltage sensitivity analysis on the initial power allocation scheme to obtain a voltage control strategy; The voltage control strategy is locally optimized by a distributed Newton method to obtain an optimized voltage distribution; Coordinating and matching the optimized voltage distribution and the initial power allocation scheme to obtain a balanced power allocation scheme; The balanced power allocation scheme is globally converged by using a distributed consistency algorithm to obtain the power allocation scheme corresponding to the target substation.

6. A substation data collaborative identification device, used to implement the substation data collaborative identification method according to any one of claims 1 to 5, characterized in that: The substation data collaborative identification device comprises: A processing module is used to perform topology identification processing on the real-time voltage and current data of the target substation to obtain a dynamic distribution network topology structure; A regulation module, used for dynamically regulating the output characteristics of the distributed energy units in the dynamic distribution network topology structure through a virtual inertia control algorithm to obtain a power regulation signal simulating the characteristics of a synchronous generator, wherein the dynamic regulation of the output characteristics of the distributed energy units in the dynamic distribution network topology structure through a virtual inertia control algorithm to obtain a power regulation signal simulating the characteristics of a synchronous generator comprises: performing node classification processing on the dynamic distribution network topology structure to obtain a distributed energy unit set; performing output characteristic parameter extraction processing on each unit in the distributed energy unit set to obtain an initial characteristic parameter matrix; performing parameter optimization processing on the initial characteristic parameter matrix through an adaptive particle swarm optimization algorithm to obtain an optimized characteristic parameter set; and performing parameter optimization processing on the optimized characteristic parameter set. The virtual moment of inertia of the set is calculated and processed to obtain a virtual inertia distribution diagram; the power grid frequency signal is decomposed and processed by a wavelet transform algorithm to obtain a multi-scale frequency feature; the multi-scale frequency feature and the virtual inertia distribution diagram are matched with frequency response characteristics to obtain a virtual inertia control gain coefficient; the virtual inertia control gain coefficient is dynamically adjusted by a fuzzy adaptive PID algorithm to obtain an adaptive control parameter; the adaptive control parameter is mapped to a virtual synchronous machine model to obtain a virtual synchronous machine state variable; the virtual synchronous machine state variable is feedforward compensated by a predictive control algorithm to obtain a compensation control amount; the compensation control amount and the virtual synchronous machine state variable are power synthesized to obtain a power regulation signal simulating the characteristics of a synchronous generator; A collaborative module, used to perform collaborative optimization processing on the power regulation signal and the pre-acquired power supply and demand data through a multi-agent reinforcement learning algorithm to obtain a collaborative scheduling strategy; An evaluation module is used to perform health status evaluation processing on the multi-source heterogeneous data of the target substation collected in advance and the collaborative scheduling strategy through a deep transfer learning algorithm to obtain a health status evaluation result of the equipment. The health status evaluation processing on the multi-source heterogeneous data of the target substation collected in advance and the collaborative scheduling strategy through a deep transfer learning algorithm to obtain a health status evaluation result of the equipment includes: performing outlier detection and interpolation processing on the multi-source heterogeneous data of the target substation collected in advance to obtain a corrected data set; performing high-order feature extraction processing on the corrected data set through a tensor decomposition algorithm to obtain a multidimensional feature tensor; performing spatiotemporal correlation analysis processing on the multidimensional feature tensor and the collaborative scheduling strategy to obtain to the spatiotemporal feature matrix; perform probability distribution estimation processing on the spatiotemporal feature matrix through a variational inference algorithm to obtain a latent variable distribution; perform migration mapping processing on the latent variable distribution to obtain a cross-domain feature representation; perform topological structure learning processing on the cross-domain feature representation through a graph attention network algorithm to obtain a device association graph; perform spectral clustering processing on the device association graph to obtain a device state group; perform domain invariant feature extraction processing on the device state group through an adversarial training algorithm to obtain a robust feature set; perform multi-task learning processing on the robust feature set to obtain a multidimensional health indicator; perform uncertainty quantification processing on the multidimensional health indicator through a Bayesian inference algorithm to obtain the device health status assessment result; An optimization module, used for dynamically optimizing the dynamic distribution network topology and the equipment health status evaluation results through an adaptive fuzzy logic algorithm to obtain a dynamic protection setting value; The allocation module is used to optimize and allocate the dynamic distribution network topology, the coordinated scheduling strategy and the dynamic protection setting value through a distributed optimal power flow algorithm to obtain a power allocation plan corresponding to the target substation.

7. A substation data collaborative identification device, characterized in that: The substation data collaborative identification device includes: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the substation data collaborative identification device executes the substation data collaborative identification method as described in any one of claims 1-5.

8. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instruction is executed by the processor, the substation data collaborative identification method as described in any one of claims 1-5 is implemented.

Images