Power grid intelligent self-healing and risk prevention and control method and system for cross-site power supply

Abstract

The present application relates to the technical field of power grid intelligent self-healing and risk prevention and control, and is a power grid intelligent self-healing and risk prevention and control method and system for cross-station power supply, which constructs a multi-source data fusion system, establishes a data quality evaluation model after data preprocessing; builds an intelligent auxiliary decision-making framework, establishes a multi-dimensional fault risk real-time quantitative evaluation and early warning model; introduces a transfer reinforcement learning algorithm to construct a multi-objective optimization scheduling model, generates a cross-station differentiated self-healing strategy; constructs a collaborative closed-loop control framework of main station centralized decision-making and sub-station distributed execution, realizes safe remote operation of cross-substation switches, configures a safety locking mechanism; establishes a decision effect evaluation system, optimizes the generalization ability of the algorithm model through the incremental learning mechanism, and customizes differentiated self-healing logic for complex scenarios. The present application realizes cross-station fault rapid self-healing, improves power supply reliability and scheduling intelligent level, and adapts to complex power grid scenarios.

Description

Technical Field

[0001] This invention relates to the field of smart grid self-healing and risk prevention technology, and is a method and system for smart grid self-healing and risk prevention for cross-site power supply. Background Technology

[0002] With rapid socio-economic development, electricity demand continues to grow, the power grid is expanding, and its structure is becoming increasingly complex. Cross-regional power supply has gradually become an important way to ensure power supply coverage, especially in vast, geographically dispersed areas where chain-type power supply and power supply structures connected to renewable energy plants are more common. However, current cross-regional power grids face many prominent problems during operation, severely restricting the improvement of power supply reliability and intelligence.

[0003] In terms of fault self-healing, traditional automatic transfer switches are mostly limited to the operation within a single station, lacking a global grasp of the entire network topology information, making it difficult to achieve cross-station collaborative self-healing. When an N-1 fault occurs, it can easily lead to an expansion of the fault's impact range, triggering multi-station coordinated power outages, causing large-scale power outages, and seriously affecting users' normal power consumption and socio-economic stability. While some domestic literature proposes novel, universal remote automatic transfer switch (ATS) schemes that utilize fiber optic channels for inter-station information exchange and coordinated actions, these schemes still require improvement in terms of switch closing sequence selection and interlocking logic design. Furthermore, most domestic research solutions are limited to specific power supply modes, lacking versatility and global coordination capabilities at the master station level, making it difficult to address the risk of multi-station coordinated power outages in chain-like power grids. Internationally, ATS functions are often integrated into substation integrated automation systems using "virtual logic programming" technology. While this approach offers flexible logic programming and high system integration, it doesn't adequately consider the chain-like structure characteristics of weak power grids in remote areas of my country. It exhibits poor adaptability in scenarios with dispersed loads and scarce power sources, and faces patent barriers, making direct import costly and localization difficult. Therefore, these technical solutions are ill-suited to my country's dual challenges of a "weak grid structure + high renewable energy access" scenario.

[0004] In terms of data processing and scenario adaptation, existing power grid systems aggregate a wide variety of data from multiple sources, lacking effective fusion and preprocessing mechanisms. This makes it difficult to guarantee data quality, resulting in insufficient accuracy in fault diagnosis and risk assessment based on this data, which in turn affects the scientific nature of subsequent dispatch decisions. Meanwhile, the large-scale integration of renewable energy sources makes power grid operation more complex. The volatility and randomness of renewable energy output pose serious challenges to traditional dispatch models and self-healing strategies. Existing research both domestically and internationally reveals that domestic solutions do not adequately consider the adaptability after renewable energy integration, making operational logic susceptible to interference and frequently leading to misjudgments and difficulties in closing circuits. While foreign power grids have high renewable energy penetration rates, their grid structures are relatively robust, differing significantly from the power grid scenarios in some regions of my country. Therefore, their technical solutions are difficult to directly reuse, and neither approach effectively addresses the system stability issues caused by renewable energy integration.

[0005] In terms of decision-making and control systems, the existing intelligent auxiliary decision-making architecture for power grids is not perfect, the functional connections between various links are not smooth, and there is a lack of efficient algorithmic support. Traditional optimization algorithms suffer from slow optimization speed and weak generalization ability when dealing with complex power grid optimization problems, and cannot quickly generate optimal self-healing strategies. The design of safety interlocking mechanisms is not comprehensive enough, and remote operation across substations poses safety risks. Although domestic independent standby automatic transfer devices have improved security through network communication, their hardware and software design and practical application still need optimization. The interlocking logic of foreign products is also difficult to adapt to the complex operating scenarios of my country's power grid. The decision-making effect evaluation system is not sound enough, making it difficult to comprehensively and accurately evaluate the implementation effect of self-healing strategies and failing to provide a strong basis for model optimization. These problems collectively restrict the safe and stable operation of cross-substation power grids and the improvement of their intelligence level. Summary of the Invention

[0006] This invention provides a method and system for intelligent self-healing and risk prevention of power grids for cross-site power supply, which overcomes the shortcomings of the prior art and can effectively solve the problems of insufficient self-healing capability, inaccurate scheduling decisions, and weak adaptability to complex scenarios in cross-site power supply power grids.

[0007] One of the technical solutions of this invention is achieved through the following measures: a smart self-healing and risk prevention method for power grids supplying power across different substations, comprising the following steps: Step S1: Construct a multi-source data fusion system, which aggregates real-time data from the dispatch automation master station, power grid topology data, new energy data, user load data, and equipment parameter data, and establishes a data quality assessment model after data preprocessing; Step S2: Build an intelligent auxiliary decision-making architecture, construct a power grid topology model based on graph theory, use a topology search algorithm to identify power supply structure and path, and establish a multi-dimensional real-time quantitative assessment and early warning model for fault risk. Step S3: Introduce a transfer reinforcement learning algorithm to construct a multi-objective optimization scheduling model, and combine a self-healing-oriented action strategy with an adaptive optimization algorithm adapted to the power grid scheduling scenario to generate a cross-site differentiated self-healing strategy; Step S4: Construct a collaborative closed-loop control architecture with centralized decision-making at the master station and distributed execution at the substations to achieve safe remote operation of switches across substations and configure a safety interlocking mechanism; Step S5: Establish a decision-making effectiveness evaluation system, optimize the generalization ability of the algorithm model through incremental learning mechanism, and customize differentiated self-healing logic for complex scenarios.

[0008] The following are further optimizations and / or improvements to one of the above-mentioned technical solutions: In step S1 above, multi-source data can adopt an edge computing and main station storage mode. The edge side performs preprocessing operations such as filtering, deduplication, and noise reduction on the multi-source data, while the main station side establishes a data quality assessment model to verify the accuracy and integrity of the preprocessed data.

[0009] In step S2 above, the intelligent assisted decision-making architecture may include a five-layer architecture: data layer, perception layer, decision layer, control layer, and feedback layer. Among them, the data layer is used for multi-source data aggregation and storage, the perception layer is used for power grid operation status perception and risk warning, the decision layer is used for optimal self-healing strategy generation, the control layer is used for decision command execution, and the feedback layer is used for quantitative evaluation of decision effect and iterative optimization of algorithm model.

[0010] In step S3 above, the transfer reinforcement learning algorithm can reuse historical optimization information through a three-level mechanism of knowledge extraction, knowledge transfer, and knowledge adaptation. The knowledge extraction stage extracts the optimal policy model parameters from the solved source task. The knowledge transfer stage uses the source task parameters as the initial parameters of the new task. The knowledge adaptation stage adjusts the parameters through a small number of new task samples to accurately adapt to the scheduling requirements of the new scenario.

[0011] In step S3 above, the self-healing-oriented action strategy can be a pre-simulation-greedy action strategy. First, invalid and duplicate actions are filtered out by set comparison. Then, the remaining actions are simulated to eliminate actions that exceed the limit. Finally, the optimal action is selected according to a three-level ranking of load recovery amount first, line loss value second, and strategy evaluation value supplement.

[0012] In step S4 above, the safety interlocking mechanism may include three layers of protection: hardware interlocking, software verification, and manual intervention. Hardware interlocking restricts operation based on the physical state of the equipment, software verification verifies the feasibility of operation through logical algorithms, and manual intervention reserves a confirmation step for emergency operations.

[0013] In step S5 above, the complex scenario may include at least one of the following: serial power supply structure scenario, new energy access scenario, and extreme low temperature scenario; the evaluation indicators of the decision-making effect evaluation system include at least one of the following: fault recovery time, load recovery rate, and fault risk pass rate.

[0014] In step S2 above, multi-source data can adopt an edge computing and main station storage mode. The edge side performs preprocessing operations such as filtering, deduplication, and noise reduction on the multi-source data, while the main station side establishes a data quality assessment model to verify the accuracy and integrity of the preprocessed data.

[0015] The second technical solution of the present invention is achieved through the following measures: a smart self-healing and risk prevention system for power grids supplying power across substation areas, comprising: The data acquisition and processing module is used to aggregate data from multiple sources, perform data cleaning, verification and storage, and establish a data quality assessment model after preprocessing. The situational awareness and risk warning module is used to identify power grid structure and conduct real-time quantitative assessment of fault risks based on graph theory topology models and topology search algorithms, and to generate risk warning information. The intelligent decision-making module is used to deploy transfer reinforcement learning algorithms and multi-objective optimization scheduling models, integrate self-healing-oriented action strategies and adaptive optimization algorithms adapted to power grid scheduling scenarios, and generate cross-site differentiated self-healing strategies. The collaborative control module is used to build a collaborative control architecture that enables centralized decision-making at the main station and distributed execution at the substations, realizes the issuance, execution and feedback of control commands, and configures a security interlocking mechanism. The effect evaluation and model optimization module is used to establish an evaluation index system, quantify the evaluation of decision-making effects, optimize algorithm models through feedback data, and adapt to complex power grid scenarios.

[0016] The following are further optimizations and / or improvements to the second technical solution of the above invention: The multi-source data collected by the aforementioned data acquisition and processing module may include at least one of the following: real-time remote signaling data from the dispatch automation master station, telemetry data, power grid GIS topology data, power output prediction data from new energy power plants, user load characteristic data, and equipment rated parameter data.

[0017] This invention, by constructing a multi-source data fusion system, ensures the reliability and accuracy of data, providing solid data support for subsequent risk assessment and decision generation. It establishes a comprehensive intelligent auxiliary decision-making architecture, enabling full perception of the power grid's operational status and accurate risk warnings, allowing dispatchers to promptly grasp the power grid's operational situation. By introducing advanced technologies such as transfer reinforcement learning algorithms, it constructs a multi-objective optimized scheduling model, generating cross-site domain differentiated self-healing strategies, effectively improving the scientific rigor and relevance of decision-making, and enabling rapid adaptation to different complex operating scenarios. It constructs a master-substation collaborative closed-loop control architecture, configuring multiple safety interlocking mechanisms to ensure the safety and timeliness of remote operation of cross-substation switches, achieving rapid fault self-healing. Finally, it establishes a decision-making effect evaluation system, continuously optimizing the algorithm model through an incremental learning mechanism to improve the system's generalization ability and operational performance. This invention overcomes the limitations of traditional power grid self-healing and risk prevention technologies, realizing intelligent and efficient operation of cross-site domain power grids, significantly improving power supply reliability and stability, and reducing losses caused by faults. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the intelligent self-healing and risk prevention method for power grids with cross-site power supply according to an embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram of the structure of a smart self-healing and risk prevention system for power grids oriented towards cross-site power supply, according to an embodiment of the present invention.

[0020] Figure 3 This is a schematic diagram of the series power supply wiring method in Embodiment 2 of the present invention.

[0021] Figure 4 This is a flowchart illustrating the pre-simulation-greedy action strategy of Embodiment 2 of the present invention.

[0022] Figure 5 This is a flowchart illustrating the improved algorithm based on the pre-simulation-greedy action strategy and Adadelta optimization in Embodiment 2 of the present invention. Detailed Implementation

[0023] The present invention is not limited to the following embodiments, and specific implementation methods can be determined according to the technical solutions and actual conditions of the present invention.

[0024] The present invention will be further described below with reference to embodiments: Example 1: As Figure 1 As shown, this embodiment provides a smart self-healing and risk prevention method for power grids with cross-site power supply, including the following steps: Step S1: Construct a multi-source data fusion system, aggregating real-time data from the dispatch automation master station, power grid topology data, new energy data, user load data, and equipment parameter data. After data preprocessing, establish a data quality assessment model. By comprehensively aggregating various types of data related to power grid operation and performing preprocessing and quality assessment, ensure that the data can truly reflect the power grid status and provide a reliable foundation for subsequent analysis and decision-making. This can guarantee the effectiveness of data support and avoid decision-making biases caused by data problems.

[0025] Step S2: Build an intelligent auxiliary decision-making architecture, construct a power grid topology model based on graph theory, use a topology search algorithm to identify power supply structure and path, establish a multi-dimensional real-time quantitative assessment and early warning model for fault risk, clearly present the power grid topology relationship with the help of graph theory tools, accurately identify power supply structure and path through the topology search algorithm, and the multi-dimensional risk assessment model can comprehensively consider various influencing factors to achieve real-time monitoring and early warning of risks. In this way, dispatchers can keep abreast of the power grid risk status and make preparations in advance.

[0026] Step S3: Introduce a transfer reinforcement learning algorithm to construct a multi-objective optimization scheduling model. Combine self-healing-oriented action strategies with adaptive optimization algorithms adapted to power grid scheduling scenarios to generate cross-site differentiated self-healing strategies. The transfer reinforcement learning algorithm can reuse historical optimization knowledge to improve decision-making efficiency in new scenarios. The self-healing-oriented action strategies and adaptive optimization algorithms ensure the optimality and adaptability of the strategies. In this way, effective self-healing strategies for different fault scenarios can be quickly generated to improve fault handling efficiency.

[0027] Step S4: Construct a collaborative closed-loop control architecture with centralized decision-making at the master station and distributed execution at the substations to achieve safe remote operation of switches across substations. Configure a safety interlocking mechanism, with the master station responsible for global decision-making and the substations executing specific operations. The collaborative closed-loop control ensures smooth instruction transmission and execution, and multiple safety interlocking mechanisms guarantee the safety and reliability of the operation process. This enables efficient collaborative operation across substation domains and avoids operational risks.

[0028] Step S5: Establish a decision-making effectiveness evaluation system, optimize the generalization ability of the algorithm model through an incremental learning mechanism, customize differentiated self-healing logic for complex scenarios, and the decision-making effectiveness evaluation system can comprehensively evaluate the implementation effect of the strategy. The incremental learning mechanism enables the model to be continuously optimized and upgraded, and the differentiated self-healing logic adapts to different complex scenarios. In this way, the system's operating performance can be continuously improved, ensuring effective prevention and control in various scenarios.

[0029] In step S1 of this embodiment, multi-source data can be processed using an edge computing and main station storage model. The edge side performs preprocessing operations such as filtering, deduplication, and noise reduction on the multi-source data. The main station side establishes a data quality assessment model to verify the accuracy and completeness of the preprocessed data. The edge side processes multi-source data locally, quickly filtering out irrelevant data, removing duplicate and redundant data, reducing noise interference, and decreasing data transmission pressure. The main station side uses the quality assessment model to verify the accuracy (whether it matches the actual power grid state) and completeness (whether it covers the key information required for decision-making) of the preprocessed data against data standards. This improves data processing efficiency, reduces interference from invalid data in subsequent stages, ensures the reliability of input data, and lays a precise data foundation for core stages such as risk assessment and strategy generation.

[0030] In step S2 of this embodiment, the intelligent auxiliary decision-making architecture may include a five-layer architecture: a data layer, a perception layer, a decision layer, a control layer, and a feedback layer. The data layer is used for multi-source data aggregation and storage; the perception layer is used for power grid operation status perception and risk warning; the decision layer is used for optimal self-healing strategy generation; the control layer is used for decision command execution; and the feedback layer is used for quantitative evaluation of decision effects and iterative optimization of algorithm models. The data layer acts as a data hub, aggregating and storing various processed data. Based on the data provided by the data layer, the perception layer achieves real-time power grid status perception and risk identification through topology analysis and risk models. The decision layer calls the algorithm model to generate self-healing strategies; the control layer executes decision commands; and the feedback layer transmits the execution effects back to each layer, driving model optimization. This achieves closed-loop collaboration between data flow, status perception, decision generation, command execution, and model optimization, improving the consistency and efficiency of the architecture's operation and providing systematic support for intelligent decision-making.

[0031] In step S3 of this embodiment, the transfer reinforcement learning algorithm can reuse historical optimization information through a three-level mechanism of knowledge extraction, knowledge transfer, and knowledge adaptation. The knowledge extraction stage extracts the optimal policy model parameters from the solved source task; the knowledge transfer stage uses the source task parameters as the initial parameters for the new task; and the knowledge adaptation stage adjusts the parameters using a small number of new task samples to accurately adapt to the scheduling requirements of the new scenario. The knowledge extraction stage mines the core information of model parameters from historically effective decisions; the knowledge transfer stage directly reuses the core parameters, avoiding training from scratch for new tasks; and the knowledge adaptation stage fine-tunes the parameters using a small number of new samples to adapt to the differences in the new scenario. This significantly shortens the algorithm training time in new scenarios, improves decision-making efficiency, enhances the algorithm's adaptability to different power grid operation scenarios, and avoids resource waste caused by repeated learning.

[0032] In step S3 of this embodiment, the self-healing-oriented action strategy can be a pre-simulation-greedy action strategy. First, invalid and repetitive actions are filtered out through set comparison. Then, simulation is performed on the remaining actions to eliminate those that exceed limits. Finally, the optimal action is selected based on a three-level ranking: load restoration amount first, line loss value second, and strategy evaluation value supplemented. First, set comparison quickly eliminates meaningless and repetitive action options, narrowing the decision-making scope. Then, simulation verifies whether the remaining actions will cause grid parameters to exceed limits, ensuring action safety. Finally, actions with high load restoration, low line loss, and excellent strategy scores are prioritized. This reduces invalid action attempts, lowers decision complexity, avoids secondary faults caused by exceeding limits, and ensures that the generated self-healing actions balance effectiveness, economy, and safety.

[0033] In step S4 of this embodiment, the safety interlocking mechanism may include three layers of protection: hardware interlocking, software verification, and manual intervention. Hardware interlocking restricts operation based on the physical state of the equipment; software verification verifies the feasibility of operation through logical algorithms; and manual intervention reserves an emergency operation confirmation step. Hardware interlocking restricts operations that do not meet safety requirements at the physical level of the equipment (such as locking the operating mechanism during equipment maintenance); software verification verifies whether the operation complies with power grid operation rules and safety constraints through logical algorithms; and manual intervention is for extreme cases, requiring manual confirmation before execution. This constructs multiple safety defenses, preventing the risk of misoperation from physical, algorithmic, and manual levels, ensuring the safety and reliability of remote operations across substations, and preventing the expansion of power grid faults caused by operational errors.

[0034] In step S5 of this embodiment, the complex scenario may include at least one of the following: a series power supply structure scenario, a new energy access scenario, and an extreme low temperature scenario; the evaluation indicators of the decision-making effect evaluation system include at least one of the following: fault recovery time, load recovery rate, and fault risk pass rate. Differentiated self-healing logic is customized for the core characteristics of different complex scenarios (multi-station linkage in a series power supply structure, output fluctuation of new energy access, and equipment parameter changes in extreme low temperatures); the fault recovery time measures the handling speed, the load recovery rate measures the recovery effect, and the fault risk pass rate measures the grid safety level after self-healing. This improves the adaptability of the self-healing strategy to complex scenarios, ensuring effective fault handling under different operating conditions; and the multi-dimensional indicators comprehensively evaluate the decision-making effect, providing accurate basis for algorithm model optimization and continuously improving the grid's self-healing and risk prevention capabilities.

[0035] In step S2 of this embodiment, multi-source data can be processed using an edge computing and main station storage model. The edge side performs preprocessing operations such as filtering, deduplication, and noise reduction on the multi-source data. The main station establishes a data quality assessment model to verify the accuracy and completeness of the preprocessed data. The edge side completes data preprocessing locally, quickly filtering invalid and redundant data and reducing the amount of data transmitted to the main station. The main station focuses on data quality verification, using standardized rules to verify the accuracy and completeness of the data, ensuring that the data meets the usage requirements of the intelligent auxiliary decision-making architecture. This reduces data transmission latency and bandwidth consumption, improves data supply efficiency, ensures the reliability of data input to the intelligent auxiliary decision-making architecture, provides accurate data support for power grid operation status perception and risk early warning, and improves the overall operational efficiency of the architecture.

[0036] The transfer reinforcement learning model in this invention is a combination of transfer reinforcement learning (TRL) and deep neural network architecture. Its internal layers include an input layer (a 24-dimensional power grid operating state vector), two hidden layers (containing 64 and 32 neurons respectively), and an output layer (an 8-dimensional action space). Signal transmission between layers is achieved through the ReLU activation function (hidden layers) and the Softmax function (output layer). The model input data comes from multiple sources, including remote signaling, telemetry, and power grid topology data from the D5000 dispatch platform, and is preprocessed into standardized vectors. The output is the power grid self-healing action strategy (such as switch opening and closing commands, load transfer path selection, etc.). The training steps follow the process of "initialization-iterative training-parameter update-model convergence." The training samples cover 12 N-1 fault data from a power grid in Xinjiang from 2022 to 2024, as well as various complex scenario simulation data. The sample labels correspond to the optimal actions and effect evaluation indicators for fault handling. An experience replay pool mechanism breaks data correlation and improves training stability. Meanwhile, by reusing historical fault handling experience through the "knowledge extraction-transfer-adaptation" mechanism, the model parameter weights are adjusted for scenarios such as new energy fluctuations and extreme low temperatures to ensure the effectiveness of the model in specific fields.

[0037] The pre-simulation-greedy action strategy adopted in this invention is an action selection strategy designed to meet the real-time requirements of power grid fault self-healing. This strategy is executed in three core steps: set comparison and filtering, simulation exclusion of exceeding limits, and three-level sorting selection. A greedy action retention mechanism is also incorporated to ensure algorithm convergence. The first step, set comparison and filtering, pre-constructs a set of invalid actions under power grid fault conditions (such as closing actions when the fault point is not isolated, and automatic transfer actions without power supply support) and a set of repetitive actions (such as continuous opening and closing commands of the same switch). The initial action set generated by the algorithm is then compared with these two sets using a difference operation. Invalid and repetitive actions can be directly excluded without simulation, significantly reducing the solution space. In this invention, invalid actions include actions prohibited by power grid operation, and the threshold for repetitive actions is three consecutive identical switching operations. The second step, simulation exclusion of exceeding limits, uses a power grid power flow calculation simulation model to process the remaining actions after the difference operation. The first step involves a short-term simulation, with a simulation duration of 5 seconds. During the simulation, core indicators such as line load rate, bus voltage, and equipment rated parameters are verified in real time. Actions that would cause these indicators to exceed safety limits are eliminated. The safety limit for line load rate is 80%, and the allowable fluctuation range for bus voltage is ±5% of the rated voltage. The third step involves a three-level ranking selection. Valid actions that have passed the simulation verification are ranked according to priority: load restoration amount first, line loss value second, and strategy evaluation value supplementation. The weights for load restoration amount are 70%, line loss value 20%, and strategy evaluation value 10%. The action ranked first is selected as the candidate optimal action. Simultaneously, to avoid algorithm convergence problems caused by fluctuations in grid operation data, this invention sets the greedy action retention ratio at 15%. That is, out of every 100 action selections, 15 non-optimal but safety-constrained greedy actions are randomly retained and included in the action experience base, ensuring the diversity of model training and convergence stability.

[0038] The Adadelta optimization algorithm used in this invention is an adaptive learning rate optimization algorithm for deep neural network training, which can be directly applied to the neural network training process of the power grid self-healing model. This algorithm overcomes the drawback of traditional optimization algorithms requiring manual setting of the initial learning rate. By constructing historical gradient sum-of-squares cumulative variables and historical parameter update sum-of-squares cumulative variables, it achieves dynamic adaptive adjustment of the learning rate, eliminating the need for repeated manual trial and error. This perfectly solves the defects of traditional optimization algorithms, such as the learning rate easily approaching 0, slow convergence, and susceptibility to local optima. The core calculation parameters of the algorithm include the decay factor ρ and the numerical stability factor ε. Based on the training requirements of the power grid intelligent self-healing model, this invention, through multiple experimental verifications, has determined the optimal parameter settings as ρ=0.95 and ε=10⁻. 6The attenuation factor ρ controls the proportion of historical gradient information retained, while the numerical stability factor ε avoids computational anomalies where the denominator is zero. In its adaptation to the power grid self-healing model, this algorithm, considering the real-time characteristics of power grid fault data, binds the parameter update frequency to the model learning frequency. It sets a parameter update to be completed every P actions (P=5 in this invention), avoiding repeated parameter oscillations caused by high-frequency updates while ensuring the model's rapid learning of power grid fault characteristics. This improves the neural network training convergence speed by over 40%, and after training, the model's accuracy in power grid fault self-healing decision-making remains stable at over 98%.

[0039] This invention first aggregates various multi-source data through a data acquisition phase, preprocesses and assesses its quality, and then stores it for later use. Subsequently, leveraging the perception layer in the intelligent auxiliary decision-making architecture, it identifies the power supply structure and path based on topology models and search algorithms. A multi-dimensional risk assessment model monitors the power grid's operational status in real time and issues timely risk warnings. When a power grid fault occurs, the intelligent decision-making module utilizes transfer reinforcement learning algorithms and related optimization algorithms to quickly generate differentiated self-healing strategies across substation domains. The collaborative control module, following the master station's decision instructions and protected by a safety interlocking mechanism, remotely operates switches across substations to complete fault self-healing. Finally, the self-healing effect is evaluated through a decision-making effectiveness evaluation system, and the evaluation results are fed back to the model optimization module. An incremental learning mechanism continuously optimizes the algorithm model, improving the system's generalization ability and adaptability, ensuring the safe and stable operation of the power grid in various complex scenarios. Overall, this invention achieves intelligent self-healing and risk prevention for cross-substation power grids, significantly improving power supply reliability and intelligence.

[0040] Example 2: This example provides a specific application of a smart self-healing and risk prevention method for power grids with cross-site power supply, which can be applied to typical chain-structured power grids. The specific steps are as follows: Step A1: Construct a multi-source data fusion system to provide basic data support for the smart self-healing and risk prevention and control of the power grid.

[0041] Step A1.1: Clarify the categories and sources of basic data to ensure the comprehensiveness and reliability of the data.

[0042] Step A1.1.1: Collect real-time operational data, including substation remote signaling (switch status, protection action signals, etc.) and telemetry (voltage, current, power, frequency, etc.) data, sourced from the D5000 dispatch automation master station platform, with a sampling frequency of 25 frames / second, covering all 220 kV, 110 kV, and 35 kV substations in a power grid in Xinjiang, as well as 63 peak winter power supply lines, 37 key substations, and major new energy power plants.

[0043] Step A1.1.2: Collect power grid topology data, including substation wiring methods, line parameters, power source distribution, load node information, etc., from the power grid GIS system and PMS system, and verify and update it in combination with field measurement data. Establish a dynamic topology update mechanism for power grid structure changes during peak winter demand.

[0044] Step A1.1.3: Obtain new energy data, including output data, forecast data, and operating status information of wind power and photovoltaic power stations. These data are sourced from the new energy monitoring system. The forecast data is generated using a model that integrates meteorological factors, achieving a short-term forecast accuracy of over 92%.

[0045] Step A1.1.4: Aggregate load data, including load characteristics of various types of users, historical load curves, classification of important load levels, etc., which are sourced from the electricity information collection system and load management system, with a focus on collecting winter heating load data.

[0046] Step A1.1.5: Organize equipment parameter data, including rated parameters, operating limits, maintenance records, etc. of equipment such as circuit breakers, main transformers, and lines, from the PMS system and equipment ledger, and supplement the equipment's low-temperature operating parameters.

[0047] Step A1.2: Use an edge computing and main station storage model to process multi-source data. The edge side performs preprocessing operations such as filtering, deduplication, and noise reduction on the multi-source data. The main station side establishes a data quality assessment model to verify the accuracy and integrity of the preprocessed data to ensure that the data accuracy rate reaches more than 99.9%.

[0048] Step A1.3: Build a network backup automatic transfer model. A certain operating mode in a region is represented by a backup automatic transfer. Multiple backup automatic transfers can be established in a region. Whether the backup automatic transfer is charged is determined by the set conditions. Each backup automatic transfer can contain multiple control strategies. In actual operation, the control strategy that first meets the action conditions is executed first.

[0049] Step A1.4: Clarify the model criteria. The power loss criterion is used to determine whether the expected power loss equipment has actually lost power after a fault. Powered equipment refers to equipment that remains on standby and powered after a fault. The triggering condition is based on the protection signal and switch position to determine whether the fault point is within the expected range. After a fault, the bus voltage loss must simultaneously meet the topology criterion (the bus is not connected to any power source) and the telemetry criterion (the bus voltage is less than the set value).

[0050] Step A1.5: Conduct basic data analysis, selecting the 35kV Yushihalasu substation and the Jieleagashi substation as pilot substations, such as... Figure 3As shown, under normal operating conditions, the grid disconnection point is located on the Yu side of the 35kV Yujie line. Both substations are configured for local automatic transfer switching and do not have remote automatic transfer switching capabilities. Under the current operating conditions, if the 35kV Huijie line fails, it will cause a voltage loss at the 35kV Jieleagashi substation and the 169th Regiment substation, constituting a level 7 grid event. If the 35kV I busbar at the 110kV Huimin substation fails, it will cause a voltage loss at the entire 35kV Jieleagashi substation, the Shanghu substation, and the 169th Regiment substation, constituting a level 6 grid event. This configuration allows for the comprehensive collection of grid operation-related data and ensures data quality, providing a solid and reliable data foundation for subsequent risk assessments and decision-making processes, and avoiding decision-making biases caused by missing or inaccurate data.

[0051] Step A2: Build an intelligent auxiliary decision-making architecture to achieve comprehensive perception and efficient decision-making of the power grid's operating status.

[0052] Step A2.1: Construct a five-layer architecture consisting of "data layer - perception layer - decision layer - control layer - feedback layer", with each layer having a clear function and working in synergy.

[0053] Step A2.1.1: The data layer serves as the foundation of the architecture, enabling the aggregation, cleaning, and storage of multi-source data, and integrating real-time data from the D5000 platform, GIS topology data, new energy data, load data, etc.

[0054] Step A2.1.2: The perception layer constructs a power grid topology model based on graph theory, uses a breadth-first search algorithm to identify chain structures and series power supply paths, establishes a multi-dimensional risk assessment model, and integrates equipment status, power grid load level, and external environmental factors (such as low temperature and icing) to achieve real-time quantitative assessment and early warning of N-1 fault risk.

[0055] Step A2.1.3: The decision-making layer, as the core decision-making unit, introduces the transfer reinforcement learning algorithm to construct a multi-objective optimization scheduling model. The optimization objectives are to minimize the loss of pressure and load, the lowest accident level, the fewest number of operations, and the optimal load rate. Combined with the winter load characteristics and the fluctuation of new energy sources, a differentiated self-healing strategy is generated.

[0056] Step A2.1.4: The control layer constructs a "master station-substation" collaborative control mode. The master station issues control commands, and the substation deploys intelligent execution terminals to realize remote operation of the switch. Multiple interlocking mechanisms are set to ensure operational safety.

[0057] Step A2.1.5: The feedback layer quantitatively evaluates the effectiveness of fault handling and feeds the evaluation results back to the decision layer as samples for continuous training and optimization of the algorithm model, thereby improving the model's generalization ability. This architectural design enables closed-loop collaboration between data flow, state awareness, decision generation, instruction execution, and model optimization, breaking down the problems of poor coordination between different stages in traditional architectures and providing systematic support for intelligent decision-making.

[0058] Step A3: Optimize the algorithm model to improve the speed, accuracy and adaptability of decision-making.

[0059] Step A3.1: Optimize the deep neural network architecture. The input layer uses a power grid operation state vector containing more than 20 parameters such as voltage, current, and power. The hidden layer uses the ReLU activation function, and the output layer uses the Softmax function to output the action probability distribution. The optimal architecture is determined to be "Input layer (24-dimensional) - Hidden layer 1 (64 neurons) - Hidden layer 2 (32 neurons) - Output layer (8-dimensional action space)".

[0060] Step A3.2: Using a self-game reinforcement learning method, a collaborative model of "master agent-sub-agent" is constructed. The master agent is responsible for global decision-making, and the sub-agent is responsible for local optimization within the region. An "experience replay pool" mechanism is introduced to store the action experience of the agents. Data correlation is broken through random sampling, and the training weight of low-temperature fault samples is increased for extreme winter scenarios.

[0061] Step A3.3: Design a transfer reinforcement learning algorithm that reuses historical optimization information through a three-level mechanism of knowledge extraction, knowledge transfer, and knowledge adaptation. In the knowledge extraction stage, the optimal policy model parameters are extracted from the solved source task. In the knowledge transfer stage, the source task parameters are used as the initial parameters for the new task. In the knowledge adaptation stage, the parameters are adjusted through a small number of new task samples to accurately adapt to the scheduling requirements of the new scenario, thereby shortening the training time of the new task by more than 60%.

[0062] Step A3.4: Employ a pre-simulation-greedy action strategy. First, filter out invalid and duplicate actions through set comparison. Then, simulate the remaining actions to eliminate those that exceed limits. Finally, select the optimal action based on a three-level ranking: load recovery amount first, line loss value second, and strategy evaluation value as a supplement. To avoid algorithm convergence issues, a certain proportion of greedy actions are retained, specifically as follows: Figure 4 As shown in the figure. This strategy uses a limited number of pre-simulated actions to ensure simulation efficiency within a large action space. The ensemble comparison method can eliminate a large number of invalid actions without simulation, thus narrowing the solution space. At the same time, it retains a certain degree of greedy action mechanism to ensure training effect and convergence under fluctuating data.

[0063] Step A3.5: Adjust the learning frequency of the agent to avoid repeated oscillations of neural network parameters caused by high-frequency learning, determine the most suitable ratio for the algorithm, and improve the algorithm's learning ability.

[0064] Step A3.6: The Adadelta optimization algorithm is adopted to accelerate neural network training. It eliminates the need to set a learning rate and repeatedly try and fail, avoiding the problem of the learning rate approaching 0. It also solves the defects of traditional optimization algorithms, such as slow convergence and easy getting trapped in local optima.

[0065] Step A3.7: Construct a multi-agent collaborative solution mechanism. The upper layer uses game theory and consensus theory to coordinate the behavior between agents and avoid action conflicts. The lower layer uses transfer reinforcement learning or consensus theory to quickly obtain local optimal solutions. For chain-type power grids, a "segmented control" strategy is adopted, dividing the long chain-type power grid into several sub-regions, with one sub-agent configured in each sub-region.

[0066] Step A3.8: Integrate topology optimization algorithms. A hybrid topology optimization algorithm based on the Cplex toolbox and the binary particle swarm optimization (BPSO) algorithm is adopted. First, the BPSO algorithm is used to quickly search for feasible topology structures and generate multiple candidate topology schemes. Then, the Cplex toolbox is used to perform accurate power flow calculations and safety checks on the candidate schemes to select the optimal topology scheme.

[0067] Step A3.9: Define the improved algorithm flow, such as... Figure 5 As shown, the specific steps include: Step 1: Determine the number of iterations N, learning ratio P, target network update frequency M, and initialize neural network parameters, experience replay pool, and total number of actions n=0; Step 2: A fault occurs in the distribution network, initialize the distribution network fault state as the initial state, and the number of actions k=0; Step 3: The environment reads the distribution network state information S and inputs it into the main network to obtain the action evaluation value Q; Step 4: Select the corresponding action according to the "pre-simulation-greedy action strategy"; Step 5: The environment executes the switching action and stores the action process in the experience pool; Step 6: After every P actions, randomly sample from the experience pool, calculate the error, and update the main network; Step 7: After every M actions, update the target network with reference to the main network; Step 8: Determine if it is in the end state. If not, return to Step 3 to continue the action. If yes, the current round of power transfer ends. If the power transfer is successful, output the power transfer control strategy; Step 9: Determine if the total number of actions has reached the set number of iterations. If not, start the next power transfer. If yes, end the training and output the neural network parameters. Through such algorithm optimization, the shortcomings of traditional AI algorithms, such as slow optimization and weak generalization, can be solved, and optimal self-healing strategies adapted to different scenarios can be generated quickly, thereby improving the scientific nature and efficiency of decision-making.

[0068] Step A4: Construct a collaborative closed-loop control architecture with centralized decision-making at the master station and distributed execution at the substations to achieve safe remote operation of switches across substations.

[0069] Step A4.1: The main station deploys the core decision-making system, which is responsible for network-wide risk assessment and strategy generation. The substations deploy intelligent execution terminals to enable rapid command execution and status feedback. Relying on the fiber optic communication network, the command transmission latency is ensured to be less than 100 milliseconds.

[0070] Step A4.2: Configure a safety interlocking mechanism, including hardware interlocking, software verification, and manual intervention for triple protection. Hardware interlocking restricts operation based on the physical state of the equipment, software verification verifies the feasibility of operation through logical algorithms, and manual intervention reserves a confirmation step for emergency operations.

[0071] Step A4.3: Develop specific operating procedures for the pilot area. Through network backup automatic transfer strategy optimization, when a loss of voltage is detected on the 35kV Jieleagashi substation busbar, the program commands the tripping of the Huijie line circuit breaker at the 35kV Jieleagashi substation, followed by the closing of the Jieyu line circuit breaker at the 35kV Yushihalasu substation, thus restoring power supply to the entire 35kV Jieleagashi substation and the 169th Regiment substation. This control architecture and operating procedure enables efficient cross-site collaborative operation, ensuring the safety and timeliness of remote switch operation, rapid fault handling, and preventing the escalation of fault impact.

[0072] Step A5: Establish a decision-making effectiveness evaluation system, optimize the generalization ability of the algorithm model through incremental learning mechanism, and customize differentiated self-healing logic for complex scenarios.

[0073] Step A5.1: Establish a multi-dimensional evaluation index system, including technical indicators (fault recovery time, load recovery rate, N-1 pass rate), economic indicators (power outage losses, operating costs), and safety indicators (action success rate, over-limit occurrence rate). Use the analytic hierarchy process to determine the weight of each index, with technical indicators accounting for 60%, safety indicators accounting for 30%, and economic indicators accounting for 10%.

[0074] Step A5.2: Develop an evaluation software module to achieve automatic quantitative assessment and visualization of decision-making effectiveness. Display the power grid topology and operating status through a GIS map, mark fault areas, undervoltage equipment and recovery status with different colors, display voltage and current changes during fault handling through dynamic curves, and display key indicators using a dashboard.

[0075] Step A5.3: Customize differentiated self-healing logic for complex scenarios, including series power supply structure scenarios, new energy access scenarios, and extreme low temperature scenarios. A total of 12 differentiated self-healing strategies are designed. For example, for new energy fluctuation scenarios, voltage amplitude prediction and power balance adjustment links are added; for single main transformer operation scenarios in winter, a "priority to standby - gradual load restoration" strategy is formulated; for chain-type multi-station voltage loss scenarios, a "segmented isolation - station-by-station recovery" logic is adopted.

[0076] Step A5.4: Introduce "meta-learning" technology and incremental learning mechanism to improve the model's knowledge generalization ability and build a diverse training scenario library covering different power grid structures, fault types, load levels and environmental conditions. When a new scenario appears, there is no need to retrain the entire model; only some parameters of the model need to be updated.

[0077] Step A5.5: Construct a short-term power grid situation prediction model, integrate LSTM (Long Short-Term Memory) network and attention mechanism, predict the power grid load, new energy output and voltage change trends in the next 15 minutes, with a prediction accuracy of over 90%. Based on the prediction results, generate pre-decision schemes for the time period in advance, and focus on predicting the operating status of the period from 18:00 to 22:00 during the winter load peak.

[0078] Step A5.6: Design an objective function to evaluate the power transfer / series power supply scheme, using a scoring system, as shown in Table 1. The main considerations are the load of stations at risk of total power loss due to an N-1 fault under a certain topology, the number of switching operations in the scheme, and the maximum load rate. The total score of the objective function is F = F1 (power loss risk score) + F2 (switching operation score) + F3 (load rate score); where F1 = (1-S1) × 100 (S1 is the ratio of the load of stations at risk of total power loss to the total load of all stations at risk of total power loss), and the winter critical load... The weight of the risk of power loss is 1.5 times that of the normal load; F2 is determined according to the number of switching operations N. When N∈[0,4], F2=0, and when N>4, F2=-(N-4)×1; F3 is determined according to the maximum load rate of the line P_Li / P_Li_max. When ∈[0,80%], F3=0, when ∈(80%,95%], F3=-(P_Li / P_Li_max-80%)×0.4, when ∈(95%,100%], F3=-(P_Li / P_Li_max-80%)×0.2, and when >100%, the scheme is invalid.

[0079] Step A5.7: When evaluating the power transfer / series power supply scheme, a scoring system is adopted, as shown in Table 2. The overall objective function is the sum of the three function values, with a focus on controlling the risk of station-wide power outage in the N-1 case. Switching operation and load rate indicators are only used for comparison under similar power outage risk scores. Through this evaluation system and optimization mechanism, the decision-making effect can be comprehensively assessed, the algorithm model can be continuously optimized, the system's adaptability to complex scenarios can be improved, and effective fault self-healing and risk prevention and control can be achieved under various operating conditions.

[0080] During operation, step A1 first completes the collection, processing, and model building of multi-source data to ensure the reliability of basic data. Then, step A2 activates the intelligent auxiliary decision-making architecture, with the perception layer monitoring the power grid's operating status in real time, quantifying the risk of N-1 faults, and issuing early warnings. When a fault occurs in the power grid, step A3's algorithm model quickly generates cross-site differentiated self-healing strategies. Step A4's collaborative closed-loop control architecture, under the protection of a triple safety interlocking mechanism, executes remote switch operations to complete fault self-healing. Finally, step A5's decision-making effect evaluation system assesses the handling effect and feeds the evaluation results back to the model optimization stage, continuously improving the model's generalization ability through incremental learning. This embodiment, relying on the Tacheng Dispatch Technical Support System, successfully solved the problem of multiple substations losing power due to N-1 faults in the chain structure of a power grid in Xinjiang. It broke through the limitations of traditional automatic transfer switches, reduced fault recovery time from minutes to seconds, increased the N-1 pass rate of the power grid to over 98%, reduced secondary equipment investment by 30%, eliminated the risk of a 220 kV substation operating on a single line in a five-level power grid, and achieved fault handling within seconds under extreme low temperatures, ensuring stable power supply for residential heating and important users. It can be promoted to remote power grid areas such as Xinjiang.

[0081] Example 3: As Figure 2 As shown, this embodiment provides a smart self-healing and risk prevention system for power grids supplying power across different substation areas, including: The data acquisition and processing module is used to aggregate data from multiple sources, perform data cleaning, verification, and storage, and establish a data quality assessment model after preprocessing. This module is specifically responsible for the collection, processing, and management of various types of power grid operation data. Through cleaning, verification, and other operations, it improves data quality and establishes a quality assessment model to ensure data reliability. In this way, it can provide high-quality data support for the operation of the entire system.

[0082] The situational awareness and risk warning module is used to identify power grid structure and conduct real-time quantitative assessment of fault risks based on graph theory topology models and topology search algorithms, and generate risk warning information. This module uses professional algorithms to analyze the power grid structure and operating status, accurately assess fault risks and issue timely warnings. This allows staff to quickly grasp the power grid risk situation and deploy prevention and control measures in advance.

[0083] The intelligent decision-making module is used to deploy transfer reinforcement learning algorithms and multi-objective optimization scheduling models. It integrates self-healing-oriented action strategies and adaptive optimization algorithms adapted to power grid scheduling scenarios to generate cross-site differentiated self-healing strategies. As the core decision-making unit of the system, this module uses advanced algorithms to generate the optimal self-healing strategy, ensuring the strategy's relevance and effectiveness. In this way, it can provide scientific decision support for fault handling.

[0084] The collaborative control module is used to build a collaborative control architecture that enables centralized decision-making at the master station and distributed execution at the substations. It realizes the issuance, execution, and feedback of control commands and configures a security interlocking mechanism. This module is responsible for coordinating the work of the master station and the substations, ensuring the smooth transmission and execution of commands, and preventing operational risks through the security interlocking mechanism. In this way, the safe and efficient operation across the domain can be ensured.

[0085] The effect evaluation and model optimization module is used to establish an evaluation index system, quantify the evaluation of decision-making effects, optimize algorithm models through feedback data, and adapt to complex power grid scenarios. This module comprehensively evaluates the decision-making effects and optimizes the model based on the evaluation results to improve the system's adaptability. In this way, the system performance can be continuously improved to meet the complex operation requirements of the power grid.

[0086] In this embodiment, the multi-source data collected by the data acquisition and processing module includes at least one of the following: real-time remote signaling data from the dispatch automation master station, telemetry data, power grid GIS topology data, power output prediction data from new energy power plants, user load characteristic data, and equipment rated parameter data. By expanding the data acquisition scope, the comprehensiveness of the data is ensured, making subsequent analysis and decision-making more comprehensive and accurate. This further improves the reliability and effectiveness of system operation.

[0087] During operation, the data acquisition and processing module starts first, aggregating various real-time data from the dispatch automation master station, power grid topology data, and new energy-related data from multiple sources. It performs preprocessing operations such as data cleaning and verification, and establishes a data quality assessment model to ensure data quality before storing it. The situational awareness and risk warning module calls upon the data stored in the data acquisition and processing module, and based on graph theory topology models and topology search algorithms, identifies the power grid structure and power supply paths, quantitatively assesses fault risks in real time, generates risk warning information, and feeds it back to relevant personnel and the intelligent decision-making module. When a risk warning or fault signal is received, the intelligent decision-making module activates a transfer reinforcement learning algorithm and a multi-objective optimization scheduling model, combined with a self-healing-oriented action strategy and adaptive... The system optimizes algorithms to generate differentiated self-healing strategies across different substation domains and sends these strategies to the collaborative control module. The collaborative control module then constructs a master-substation collaborative control architecture based on these instructions, issues control commands, and the substations execute relevant operations. Simultaneously, a safety interlocking mechanism ensures operational safety, and the operation results are fed back to the master station in real time. The effect evaluation and model optimization module collects various data during the fault handling process, quantifies the decision-making effect based on an evaluation index system, and uses the evaluation results as feedback data to optimize the algorithm model in the intelligent decision-making module, improving the model's adaptability to complex scenarios. Through the collaborative cooperation of these modules, the entire system achieves intelligent self-healing and risk prevention for cross-substation power grids, providing strong support for the safe and stable operation of the power grid.

[0088] It should be noted that, in this invention, cross-site power supply refers to a power supply mode that spans multiple substation areas, involving power transmission and coordinated operation between multiple substations; intelligent self-healing refers to the ability of the power grid to automatically detect and locate faults and quickly take measures to restore power supply when a fault occurs, without or with minimal human intervention; transfer reinforcement learning algorithm is an algorithm that combines transfer learning and reinforcement learning, which can transfer learned knowledge to new tasks, improving learning efficiency and decision-making effectiveness; multi-dimensional fault risk assessment refers to a comprehensive assessment of the potential fault risks of the power grid from multiple different angles and levels, such as equipment, load, and environment; safety interlocking mechanism is a series of safety protection measures set up to prevent misoperation of electrical equipment, ensuring the safety and reliability of equipment operation; topology search algorithm is an algorithm used to traverse and analyze the power grid topology and identify power supply paths and connection relationships.

[0089] The above technical features constitute the embodiments of the present invention, which have strong adaptability and implementation effect. Unnecessary technical features can be added or removed according to actual needs to meet the needs of different situations.

Claims

1. A method for intelligent self-healing and risk prevention of power grids for cross-site power supply, characterized in that, Includes the following steps: Step S1: Construct a multi-source data fusion system, which aggregates real-time data from the dispatch automation master station, power grid topology data, new energy data, user load data, and equipment parameter data, and establishes a data quality assessment model after data preprocessing; Step S2: Build an intelligent auxiliary decision-making architecture, construct a power grid topology model based on graph theory, use a topology search algorithm to identify power supply structure and path, and establish a multi-dimensional real-time quantitative assessment and early warning model for fault risk. Step S3: Introduce a transfer reinforcement learning algorithm to construct a multi-objective optimization scheduling model, and combine a self-healing-oriented action strategy with an adaptive optimization algorithm adapted to the power grid scheduling scenario to generate a cross-site differentiated self-healing strategy; Step S4: Construct a collaborative closed-loop control architecture with centralized decision-making at the master station and distributed execution at the substations to achieve safe remote operation of switches across substations and configure a safety interlocking mechanism; Step S5: Establish a decision-making effectiveness evaluation system, optimize the generalization ability of the algorithm model through incremental learning mechanism, and customize differentiated self-healing logic for complex scenarios.

2. The method for intelligent self-healing and risk prevention of power grids for cross-site power supply according to claim 1, characterized in that, In step S1, the multi-source data adopts an edge computing and main station storage mode. The edge side performs preprocessing operations such as filtering, deduplication, and noise reduction on the multi-source data, while the main station side establishes a data quality assessment model to verify the accuracy and completeness of the preprocessed data.

3. The method for intelligent self-healing and risk prevention of power grids for cross-site power supply according to claim 1, characterized in that, In step S2, the intelligent assisted decision-making architecture includes a five-layer architecture: data layer, perception layer, decision layer, control layer, and feedback layer. Among them, the data layer is used for multi-source data aggregation and storage, the perception layer is used for power grid operation status perception and risk warning, the decision layer is used for optimal self-healing strategy generation, the control layer is used for decision command execution, and the feedback layer is used for quantitative evaluation of decision effects and iterative optimization of algorithm models.

4. The method for intelligent self-healing and risk prevention of power grids for cross-site power supply according to claim 1, 2, or 3, characterized in that, In step S3, the transfer reinforcement learning algorithm reuses historical optimization information through a three-level mechanism of knowledge extraction, knowledge transfer, and knowledge adaptation. The knowledge extraction stage extracts the optimal strategy model parameters from the solved source tasks. The knowledge transfer stage uses the source task parameters as the initial parameters for the new tasks. The knowledge adaptation stage adjusts the parameters using a small number of new task samples to accurately adapt to the scheduling requirements of the new scenarios.

5. The method for intelligent self-healing and risk prevention of power grids for cross-site power supply according to claim 1, 2, or 3, characterized in that, In step S3, the self-healing-oriented action strategy is a pre-simulation-greedy action strategy. First, invalid and duplicate actions are filtered out through set comparison. Then, the remaining actions are simulated to eliminate actions that exceed the limit. Finally, the optimal action is selected according to a three-level ranking system: load recovery amount first, line loss value second, and strategy evaluation value supplemented.

6. The method for intelligent self-healing and risk prevention of power grids for cross-site power supply according to claim 1, 2, or 3, characterized in that, In step S4, the safety interlocking mechanism includes three safeguards: hardware interlocking, software verification, and manual intervention. Hardware interlocking restricts operation based on the physical state of the equipment, software verification verifies the feasibility of operation through logical algorithms, and manual intervention reserves a confirmation step for emergency operations.

7. The method for intelligent self-healing and risk prevention of power grids for cross-site power supply according to claim 1, 2, or 3, characterized in that, In step S5, the complex scenarios include at least one of the following: series power supply structure scenario, new energy access scenario, and extreme low temperature scenario; the evaluation indicators of the decision-making effect evaluation system include at least one of the following: fault recovery time, load recovery rate, and fault risk pass rate.

8. The method for intelligent self-healing and risk prevention of power grids for cross-site power supply according to claim 1, 2, or 3, characterized in that, In step S2, the multi-dimensional fault risk assessment model integrates equipment operating status, power grid load level, and external environmental factors; in step S3, the optimization objectives of the multi-objective optimization scheduling model include at least one of the following: minimum voltage and load loss, lowest accident level, minimum number of operations, and optimal load rate.

9. A smart self-healing and risk prevention system for power grids supporting cross-site power supply, implementing the method of any one of claims 1 to 8, characterized in that, include: The data acquisition and processing module is used to aggregate data from multiple sources, perform data cleaning, verification and storage, and establish a data quality assessment model after preprocessing. The situational awareness and risk warning module is used to identify power grid structure and conduct real-time quantitative assessment of fault risks based on graph theory topology models and topology search algorithms, and to generate risk warning information. The intelligent decision-making module is used to deploy transfer reinforcement learning algorithms and multi-objective optimization scheduling models, integrate self-healing-oriented action strategies and adaptive optimization algorithms adapted to power grid scheduling scenarios, and generate cross-site differentiated self-healing strategies. The collaborative control module is used to build a collaborative control architecture that enables centralized decision-making at the main station and distributed execution at the substations, realizes the issuance, execution and feedback of control commands, and configures a security interlocking mechanism. The effect evaluation and model optimization module is used to establish an evaluation index system, quantify the evaluation of decision-making effects, optimize algorithm models through feedback data, and adapt to complex power grid scenarios.

10. The intelligent self-healing and risk prevention system for power grids oriented towards cross-site power supply according to claim 9, characterized in that, The data acquisition and processing module gathers multi-source data including at least one of the following: real-time remote signaling data from the dispatch automation master station, telemetry data, power grid GIS topology data, power output prediction data from new energy power plants, user load characteristic data, and equipment rated parameter data.

Images