Power supply network cooperative override trip prevention decision-making system

Abstract

The invention discloses a cooperative override trip prevention decision-making system for a power supply network, and relates to the technical field of power system protection and control, and the system comprises a data collection module which is connected with a multi-level protection device, a monitoring terminal and an environment sensor in the power supply network, and is used for obtaining electrical quantity data, non-electrical quantity data and environment parameters in real time. According to the cooperative override trip prevention decision-making system for the power supply network, cross-protocol analysis and space-time fusion of multi-source heterogeneous data are realized through a multi-dimensional space-time alignment engine, a dynamic semantic gateway and a wavelet packet decomposition technology are combined, a power frequency quantity, a transient quantity and an environmental parameter are uniformly mapped to a fault characteristic spectrum, and the fault characteristic spectrum is analyzed and analyzed. The problem of incomplete fault feature extraction caused by data islands in the prior art is solved, key fault features can be accurately screened, the Pareto optimal action sequence can be generated, the misjudgment rate of override trip is reduced, meanwhile, the unnecessary power failure duration is shortened, and the fault isolation accuracy and the power supply reliability are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power system protection and control, and in particular to a power supply network collaborative anti-over-tripping decision-making system. Background Art

[0002] Traditional power grid over-tripping prevention technologies typically achieve fault isolation based on the independent action logic of local protection devices or coordinated mechanisms with limited communication. These methods rely on single electrical quantity characteristics to determine the fault location, using preset current time-step settings or regional logic criteria. These methods are primarily designed for traditional radial power grids with relatively simple structures and stable load distributions. However, with the significant increase in the penetration of distributed energy resources (DGEs) in new power systems, the widespread integration of flexible loads, and the growing demand for dynamic network topology reconfiguration, power grid operation has become highly complex and uncertain. Factors such as the random output characteristics of DGEs, the frequent switching between isolated and grid-connected microgrid states, and the bidirectional flow of power in ring networks result in multi-timescale fluctuations in fault current amplitude and direction. Traditional protection strategies that rely on local characteristics or fixed rules struggle to accurately capture the dynamic fault characteristics of the entire network. Furthermore, due to differences in communication protocols, mismatched timescales, and inconsistent semantic descriptions, the electrical and non-electrical quantity data collected by protection devices, monitoring terminals, and environmental sensing devices at different levels cannot be integrated and correlated within a unified framework in real time. This results in a lack of global perspective and dynamic adaptability in fault diagnosis. Existing collaborative decision-making algorithms are often based on static topology assumptions or simplified models, and are unable to effectively balance the conflicts between fault removal speed, selectivity, and power supply reliability. This is especially true in scenarios where network topology changes in real time or disturbances are superimposed. Information silos or delayed decisions can easily lead to over-tripping or unnecessary power outages. The essence of this problem lies in the difficulty of traditional technology systems to meet the challenges of real-time integration of multi-source heterogeneous data and the dynamic collaborative decision-making requirements of new power systems, resulting in a significant decrease in the reliability and accuracy of over-trip prevention. Summary of the Invention

[0003] (1) Technical problems solved

[0004] In response to the shortcomings of the existing technology, the present invention provides a power supply network collaborative anti-overtripping decision-making system, which solves the problem that traditional technical systems are difficult to cope with the real-time integration challenges of multi-source heterogeneous data in new power systems and the dynamic collaborative decision-making needs, resulting in a significant decrease in the reliability and accuracy of anti-overtripping.

[0005] (2) Technical solution

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a power supply network collaborative anti-overtripping decision system, including a data acquisition module, a multi-dimensional spatiotemporal alignment engine, a multimodal game decision module, a dual-channel verification module and a collaborative execution module;

[0007] The data acquisition module is connected to the multi-level protection devices, monitoring terminals, and environmental sensors in the power supply network through a physical interface to obtain real-time electrical quantity data, non-electrical quantity data, and environmental parameters. The protection devices include FTUs and DTUs. The electrical quantity data includes the power frequency and transient high-frequency components of voltage and current. The non-electrical quantity data includes equipment temperature and mechanical status. The environmental parameters include temperature, humidity, and wind speed. The collected data is transmitted to the multi-dimensional spatiotemporal alignment engine through a communication interface.

[0008] The multi-dimensional spatiotemporal alignment engine, whose input end is in communication with the data acquisition module and whose output end is in communication with the multimodal game decision module, is used to perform cross-protocol parsing, time base synchronization and spatial topology mapping on the electrical quantity data, non-electrical quantity data and environmental parameters to generate a dynamic fault feature map. After receiving the data, the multi-dimensional spatiotemporal alignment engine first parses the device data from different communication protocols through a dynamic semantic gateway and maps it to a unified semantic space. In view of the time scale difference between the power frequency quantity and the transient high-frequency component data, the wavelet packet decomposition technology is used to reconstruct the data stream in the frequency domain and extract the fault feature vector of the unified time base. At the same time, according to the physical location of the equipment, the electrical connection relationship and the output characteristics of the distributed power supply, a dynamic weight matrix is ​​constructed, and the spatiotemporal correlation map reflecting the fault propagation path is generated in combination with the digital twin model.

[0009] The multimodal game decision module receives the dynamic fault feature map at its input end and is connected to the dual-channel verification module at its output end, and is used to screen key fault features based on a causal reasoning network and generate collaborative action instructions through a multi-objective game model;

[0010] The dual-channel verification module includes a physical layer verification unit and a logical layer verification unit. The input end receives the collaborative action instruction, and the output end is connected to the collaborative execution module, which is used to perform physical simulation verification and formal logic verification on the instruction; the physical layer verification unit injects the action instruction into the digital twin system for simulation execution to verify whether the voltage and current changes of the key nodes meet the preset threshold range; the logical layer verification unit uses formal verification tools to detect whether the action sequence violates the topology connectivity protection rules and the minimum power supply duration rules for key loads; only when both the physical layer and the logical layer verification pass, the instruction is marked as valid;

[0011] The collaborative execution module includes an edge agent and a cloud hub. The edge agent is deployed at the edge node of the power supply network. The cloud hub communicates with the edge agent and the protection device through the blockchain network, and is used to control the action timing of the protection device according to the verified instructions.

[0012] Preferably, the multidimensional spatiotemporal alignment engine includes a dynamic semantic gateway and a spatiotemporal correlation analysis unit;

[0013] The dynamic semantic gateway has a built-in protocol converter and semantic tag library, which is used to map device data with different communication protocols into a unified semantic space;

[0014] The spatiotemporal correlation analysis unit integrates a wavelet packet decomposition algorithm and a digital twin model to synchronize the time base of the power frequency quantity and transient high-frequency component data streams, and generates a spatiotemporal correlation map of the fault propagation path based on the dynamic coding rules of the power grid topology.

[0015] Preferably, the spatiotemporal correlation analysis unit performs the following operations:

[0016] A dynamic weight matrix is ​​constructed based on the physical location and electrical connection relationship of the equipment. The weight values ​​of the dynamic weight matrix are adjusted in real time according to the output characteristics of the distributed power supply. The physical location of the equipment is the GPS coordinate, and the electrical connection relationship is the impedance parameter.

[0017] Based on the timestamp synchronization mechanism triggered by fault events, wavelet packet decomposition technology is used to reconstruct the multi-time scale data stream in the frequency domain to generate the fault feature vector with a unified time base;

[0018] The dynamic weight matrix and the fault feature vector are input into the digital twin model to generate a spatiotemporal correlation map including the fault location, propagation direction and energy distribution.

[0019] Preferably, the multimodal game decision module includes a causal reasoning network and a multi-objective game decision tree;

[0020] The causal reasoning network uses a Bayesian causal graph model to calculate causal weights for voltage phase difference, transient energy distribution, and equipment health index, eliminating noise features that are irrelevant to the fault area.

[0021] The multi-objective game decision tree integrates the Monte Carlo tree search algorithm and the deep Q network, quantifies speed as an action response time benefit function, quantifies selectivity as a fault area isolation accuracy function, and quantifies power supply reliability as a load power outage loss function, and generates a Pareto optimal action sequence through dynamic game.

[0022] Preferably, the multi-objective game decision tree performs the following steps:

[0023] Initialize the payoff function weights of the game participants based on the key fault features output by the causal reasoning network;

[0024] Through Monte Carlo tree search, candidate action branches are traversed to calculate the comprehensive score of each branch's speed, selectivity and power supply reliability;

[0025] A deep Q-network is used to optimize the scoring strategy and generate action sequences and priority rankings that meet multi-objective balance.

[0026] Preferably, in the dual-channel verification module:

[0027] The physical layer verification unit is connected to the digital twin model and is used to inject collaborative action instructions into the digital twin system, namely the PSCAD / EMTDP simulation platform. After the simulation is executed, it verifies whether the voltage change of key nodes is within ±10% of the rated value and whether the current is within the circuit breaker's breaking capacity. Key nodes include busbars and distributed power access points.

[0028] The logic layer verification unit has a built-in formal verification tool and rule library, which is used to detect whether the action sequence violates the topology connectivity protection rules and the minimum power supply duration rules for key loads.

[0029] Preferably, in the collaborative execution module: the edge agent deploys a lightweight fault feature extraction model after knowledge distillation compression, with model parameter volume <1MB and inference delay <10ms, and performs preliminary fault screening locally, i.e. overcurrent detection, and triggers emergency action pre-instructions, including circuit breaker pre-tripping signals; the cloud hub runs a global game decision engine, and sends timing synchronization signals to the edge agent and protection device through the blockchain network to ensure strict timing consistency of multi-terminal actions.

[0030] Preferably, the lightweight model of the edge agent is generated in the following manner: a fault feature extraction model based on a convolutional neural network is trained in the cloud center; the convolutional neural network model is compressed into a low-latency inference model using knowledge distillation technology and deployed to the edge node; and the task allocation ratio of edge and cloud computing resources is adjusted in real time according to changes in the power grid topology through a dynamic weight allocation algorithm.

[0031] Preferably, the multimodal game decision module further includes a dynamic weight allocation unit for adjusting the weights of the profit functions of speed, selectivity and power supply reliability according to the real-time status of the power grid topology. The weight adjustment formula is: W i =α·T response +β·S isolation +γ·R reliability ; Among them, α, β, and γ are dynamic adjustment coefficients, which are related to the grid operation mode and fault severity.

[0032] Preferably, the verification results of the physical layer verification unit and the logical layer verification unit are recorded in an operation log through the blockchain network. The log contains a fault feature map, an action sequence and a verification timestamp, and supports tamper-proof traceability. The fault feature map is a hash value of a spatiotemporal correlation map, the action sequence includes a circuit breaker number and a tripping timestamp, and the verification timestamp includes the completion time of the physical check and the logical check. The log data is stored in a Merkle tree structure, and supports tamper-proof traceability through a timestamp and a device ID.

[0033] (3) Beneficial effects

[0034] The present invention provides a power supply network collaborative anti-overtripping decision-making system, which has the following beneficial effects:

[0035] This power supply network collaborative anti-overtripping decision-making system uses a multi-dimensional spatiotemporal alignment engine to achieve cross-protocol analysis and spatiotemporal fusion of multi-source heterogeneous data. Combined with dynamic semantic gateways and wavelet packet decomposition technology, it uniformly maps power frequency quantities, transient quantities and environmental parameters to fault feature maps, solving the problem of incomplete fault feature extraction caused by data silos in traditional technologies. Based on the collaborative mechanism of causal reasoning networks and multi-objective game decision trees, the system can accurately screen key fault features and generate Pareto optimal action sequences, reducing the misjudgment rate of overtripping, while reducing the duration of unnecessary power outages, and improving the accuracy of fault isolation and power supply reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a schematic diagram of the overall framework of the present invention;

[0037] Figure 2 This is a control logic timing diagram of the present invention. DETAILED DESCRIPTION

[0038] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0039] See also Figure 1 and Figure 2 The present invention provides a technical solution: a power supply network collaborative anti-overtripping decision-making system, comprising:

[0040] The data acquisition module is connected to the multi-level protection devices, monitoring terminals, and environmental sensors in the power supply network. It is used to obtain real-time electrical quantity data, non-electrical quantity data, and environmental parameters through the RS-485 interface, optical fiber interface, and LoRa wireless protocol. The electrical quantity data includes power frequency quantity and transient high-frequency component. The power frequency quantity is 50Hz and sampled at 100 points per second; the transient high-frequency component is 1kHz and sampled at 1000 points per second.

[0041] The input end of the multi-dimensional spatiotemporal alignment engine is connected to the data acquisition module via a data bus, and the output end is connected to the multimodal game decision module. It is used to perform cross-protocol analysis, time base synchronization, and spatial topology mapping of electrical quantity data, non-electrical quantity data, and environmental parameters to generate dynamic fault feature maps. Among them, cross-protocol analysis includes IEC61850 and Modbus protocol conversion, and the time base synchronization error is ±1ms. After receiving the data, the multi-dimensional spatiotemporal alignment engine first parses the device data from different communication protocols through a dynamic semantic gateway and maps it to a unified semantic space. In view of the time scale difference between the power frequency quantity and the transient high-frequency component data, the wavelet packet decomposition technology is used to reconstruct the data stream in the frequency domain and extract the fault feature vector with a unified time base. At the same time, based on the physical location of the equipment, the electrical connection relationship, and the output characteristics of the distributed power supply, a dynamic weight matrix is ​​constructed, and combined with the digital twin model to generate a spatiotemporal correlation map reflecting the fault propagation path.

[0042] The multimodal game decision module receives a dynamic fault feature map at its input and is connected to the dual-channel verification module at its output. It is used to screen key fault features based on a causal reasoning network and generate coordinated action instructions through a multi-objective game model. The coordinated action instructions include a circuit breaker tripping priority sequence. After receiving the spatiotemporal correlation map, the multimodal game decision module uses a Bayesian causal graph model to perform a causal correlation analysis on the voltage phase difference, transient energy distribution, and equipment health index, eliminating noise features unrelated to the fault and generating a set of candidate fault areas. Subsequently, the candidate action branches are traversed through a Monte Carlo tree search, and a deep Q network is combined to dynamically assign weights to the speed, selectivity, and power supply reliability targets to generate a Pareto optimal action sequence, which is then sorted by priority and output to the dual-channel verification module.

[0043] In the multi-objective game decision-making process, the weights of the benefit functions of speed, selectivity and power supply reliability are dynamically adjusted according to the real-time changes in the power grid topology; all fault feature maps, action sequences and verification results are recorded in operation logs through the blockchain network, supporting tamper-proof traceability and ensuring the auditability of the decision-making process.

[0044] The dual-channel verification module includes a physical layer verification unit and a logical layer verification unit. Its input receives collaborative action instructions, and its output is connected to the collaborative execution module. It is used to perform physical simulation verification of the instructions based on the PSCAD simulation platform with a step size of 10μs and formal logic verification using the UPPAAL tool.

[0045] The collaborative execution module includes an edge agent and a cloud hub. The edge agent is deployed at the edge node of the power grid, and the cloud hub communicates with the edge agent and protection devices via the blockchain network. It controls the timing of protection device actions based on verified instructions, with a timestamp accuracy of ±1ms. In the collaborative execution module, the edge agent deploys a lightweight fault feature extraction model compressed by knowledge distillation, with a parameter size of <1MB and an inference latency of <10ms, enabling local fault screening and emergency action pre-triggering. The cloud hub dynamically adjusts strategy weights based on a real-time updated grid image and sends timing synchronization signals to edge nodes and protection devices via the blockchain network to ensure strict timing consistency of multi-terminal actions. Finally, verified instructions are converted into control signals for the protection devices, completing fault isolation.

[0046] The multi-dimensional spatiotemporal alignment engine includes a dynamic semantic gateway and a spatiotemporal correlation analysis unit;

[0047] The dynamic semantic gateway, with a built-in protocol converter and semantic tag library, is used to convert raw data collected by different devices into standardized data frames in a unified semantic space. The protocol converter supports IEC61850, DNP3, and MQTT protocol conversion, and the semantic tag library contains standardized rules for defining device types, data categories, and units.

[0048] During the specific implementation process, the dynamic semantic gateway, with a built-in protocol converter and semantic tag library, is used to parse the MMS message of the IEC61850 protocol into the effective current value MMXU.A.phsA.cVal.mag.f and the fault flag XCBR.Pos.stVal, and convert the Modbus register data into floating-point values, while defining the device type tag and data unit.

[0049] In the semantic tag library definition rule example, the device type "photovoltaic inverter" corresponds to the label DG_PV, and the data category "output" corresponds to the unit kW.

[0050] The spatiotemporal correlation analysis unit integrates a wavelet packet decomposition algorithm and a digital twin model. The wavelet packet decomposition algorithm uses the Daubechies wavelet basis for a five-layer decomposition. The digital twin model is a virtual mirror based on real-time topology. It is used to reconstruct the power frequency data and transient high-frequency component data streams in the time and frequency domain, generate time-aligned fault feature vectors, and combine the dynamic encoding rules of the power grid topology, namely the node-branch weight matrix, to generate a spatiotemporal correlation map of the fault propagation path. Among them, the power frequency data is a 50Hz fundamental wave, and the transient high-frequency component data stream is greater than 1kHz.

[0051] The spatiotemporal correlation analysis unit performs the following operations:

[0052] According to the GPS coordinates of the physical location of the equipment and the impedance parameters in the electrical connection relationship, a dynamic weight matrix is ​​generated. The weight values ​​of the matrix elements are updated in real time as the power output fluctuates. Among them, the output characteristics of the distributed power source are the photovoltaic / wind power fluctuation curve. When the dynamic weight matrix is ​​constructed, the matrix dimension is N×N, where N is the number of grid nodes and the weight value W is ij Calculation formula: Among them, Z ij P is the impedance from node i to j, in Ω, which represents the physical characteristics of the electrical connection; DG (t) is the proportion of distributed power output to total load at time t, unit: %, reflecting the impact of power output fluctuation on the weight; W ij : The dynamic weight value from node i to node j, reflecting the correlation strength of fault propagation between the two nodes;

[0053] Based on a timestamp synchronization mechanism triggered by fault events, wavelet packet decomposition technology is used to perform frequency band segmentation and reconstruction on multi-time-scale data streams, extracting fault feature vectors with a unified time base. The multi-time-scale data streams contain 1kHz sampled transient data and 50Hz power frequency data.

[0054] The dynamic weight matrix and fault feature vector are input into the digital twin model. Through topological relationship calculation and energy propagation simulation, a spatiotemporal correlation map is generated, which includes the fault location (coordinate grid), the propagation direction (tidal flow vector), and the energy distribution (transient energy heat map). The fault location is the coordinate grid, the propagation direction is the tidal flow vector, and the energy distribution is the transient energy heat map.

[0055] The multimodal game decision module includes a causal reasoning network and a multi-objective game decision tree;

[0056] The causal inference network uses a Bayesian causal graph model, with nodes representing voltage phase difference, transient energy distribution, and equipment health index, and edges representing conditional probability relationships. The causal weights between each feature and the fault area are calculated, and noise features with causal weights below the threshold are eliminated. The calculation of the causal weights between each feature and the fault area includes P(fault|Δφ), and noise features with causal weights below the threshold include ambient temperature and humidity fluctuations. In the specific implementation process, the causal inference network uses a Bayesian causal graph model. Based on 1,000 sets of data from a historical fault case library, the causal probabilities of voltage phase difference, transient energy distribution, and equipment health index are calculated, and features with causal probabilities P(Fault}|·)>0.7 are screened for entry into the decision tree. The causal weight calculation formula is:

[0057] ;

[0058] P(Fault|·): Given the voltage phase difference Δφ and transient energy E transient The posterior probability of a fault under the condition; Δφ: voltage phase difference (unit: °), one of the fault characteristics; E transient : Transient energy distribution, unit: J, reflects the energy intensity of the fault transient process; P(Fault): Prior probability of fault, based on historical statistical data.

[0059] Multi-objective game decision tree, integrating Monte Carlo tree search algorithm and deep Q network, quantifies the speed as the action response time benefit function T response ≤100%, selectivity quantified as the fault area isolation accuracy function S isolation ≥95\%, power supply reliability is quantified as load outage loss function R reliability ≤5\% Pareto optimal action sequences, i.e., non-dominated solution sets, are generated through dynamic game. The Monte Carlo tree search algorithm randomly generates 200 sets of candidate action branches, and the deep Q network is a 3-layer fully connected structure with an input dimension of 128.

[0060] The multi-objective game decision tree in the dual-channel verification module performs the following steps:

[0061] Based on the key fault characteristics output by the causal reasoning network, the payoff function weights of the game participants are initialized; the key fault characteristics include voltage phase difference mutation and abnormal transient energy distribution. Voltage phase difference mutation: voltage phase difference > 30°, abnormal transient energy distribution: transient energy > 1.5 times the baseline value; game participants include speed, selectivity, and reliability.

[0062] By traversing candidate action branches through Monte Carlo tree search, 200 sets of candidate action branches can be randomly generated, such as "tripping circuit breaker A with a delay of 50ms, and tripping circuit breaker B immediately". For each set of actions, 100 virtual deductions are performed and a comprehensive score is calculated, that is, the average speed is calculated, which includes response time ≤ 100ms, selectivity, that is, isolation accuracy ≥ 95%, and reliability, that is, power outage loss ≤ 5% load. The comprehensive score calculation formula is: Score = αT response +βS isolation +γR reliability ,Score: Comprehensive score of candidate action branches, used for Pareto optimal action screening,T response : Response time benefit function, required to be ≤100ms; S isolation : Isolation accuracy function, required to be ≥95%; R reliability : Power outage loss function, requiring ≤5% load; among them, the candidate action branches include priority tripping of the circuit breaker in area A and delayed tripping of the circuit breaker in area B.

[0063] A deep Q network is used to optimize the scoring results. The input layer inputs the fault feature vector dimension of 128, the hidden layer is a fully connected network with 256-128-64 nodes, and the output layer outputs the Q value of the action sequence. The top three action sequences with the highest Q value are selected to enter the Pareto front and are prioritized in ascending order of response time.

[0064] The physical layer verification unit in the dual-channel verification module is connected to the digital twin model to inject collaborative action instructions into the PSCAD simulation platform and verify whether the voltage changes at key nodes are within ±10% of the rated value and whether the current is within 1.2 times the circuit breaker's breaking capacity.

[0065] The logic layer verification unit in the dual-channel verification module uses the formal verification tool UPPAAL to detect whether the action sequence violates the preset rules, including the topology connectivity protection rule and the minimum power supply time rule for critical loads. The topology connectivity protection rule is to retain at least one power supply path, and the minimum power supply time rule for critical loads is greater than or equal to 15 minutes. During the specific implementation process, the logic layer verification unit has a built-in formal verification tool to detect whether the action sequence violates the topology connectivity rule CTL formula AG (EF (LoadConnected)), that is, there is always a path to connect the load; whether it violates the critical load power supply time rule TCTL formula A [≥15min], that is, the power supply must be maintained for at least 15 minutes.

[0066] In the collaborative execution module: the edge agent integrates a lightweight model after knowledge distillation compression, with a parameter size of 0.8MB and a delay of 8ms, which is used to locally perform overcurrent detection and trigger the circuit breaker pre-tripping signal; the cloud center runs a global game decision engine, which dynamically adjusts the strategy weight based on the real-time updated power grid image, that is, the digital twin data, and sends a timing synchronization signal to the edge agent and protection device through the blockchain network Hyperledger Fabric framework, with a timestamp accuracy of ±1ms, to ensure strict timing consistency of multi-terminal actions. In the specific implementation process, the cloud center broadcasts the timing synchronization signal through the Orderer node of Hyperledger Fabric, including the action instruction hash value, the global timestamp ±1ms and the edge node ID list to ensure the timing consistency of multi-terminal actions.

[0067] The lightweight model of the edge agent is generated in the following way:

[0068] A ResNet-50-based fault feature extraction model is trained in the cloud hub, with the input being the current waveform and the output being the fault type probability.

[0069] Using knowledge distillation technology, that is, the loss function L KD , compress the convolutional neural network model into a low-latency inference model and deploy it to the edge node; in the specific implementation process, a teacher-student network architecture is adopted to transfer the knowledge of the teacher network, namely the feature attention distribution, to the student network, where the teacher network parameter volume is 100MB and the student network parameter volume is 1MB; during the teacher network training, the input current waveform segment is 200ms in length and the sampling rate is 1kHz, and the output fault type probability includes overcurrent, short circuit, and grounding; the network structure is ResNet-50, and during the student network compression, the attention transfer loss function is adopted: L KD =α·L CE (y, yT) + β·L MSE (A S , A T ); where A S 、A T is the attention map of the student and the teacher, L KD : total loss value of knowledge distillation, used for compression model training, α, β: weight coefficients, controlling the balance between cross entropy loss and attention transfer loss, L CE : Cross entropy loss function, which measures the difference between the student network output y and the teacher network output yT, L MSE : Mean square error loss function, measuring the student network attention map A S With the teacher network attention map A T ; α=0.7, β=0.3; the compressed model parameter size is 0.8MB, and the inference delay is 8ms.

[0070] Through the dynamic weight allocation algorithm, the task allocation ratio of edge and cloud computing resources can be adjusted in real time according to the changes in the power grid topology. In the specific implementation process, the task ratio of edge and cloud computing resources can be adjusted through the dynamic weight allocation algorithm according to the switching of microgrid islands, that is, the edge processes 70% of local data and the cloud processes 30% of global data.

[0071] The multimodal game decision module also includes a dynamic weight allocation unit, which is used to adjust the weights of the profit functions of speed, selectivity and power supply reliability according to the real-time status of the power grid topology. The weight adjustment formula is: W i =α·T response +β·S isolation +γ·R reliability ; Among them, α, β, γ are dynamic adjustment coefficients with a value range of 0-1, which are related to the grid operation mode and fault severity. i : The comprehensive weight value of the i-th goal, including speed, selectivity, and reliability, T response : Speed ​​index, quantified as action response time, unit: ms; S isolation : Selectivity index, quantified as the fault area isolation accuracy, unit: %, R reliability : Power supply reliability index, quantified as load power outage loss, unit: %;

[0072] In the specific implementation process, the real-time state of the power grid topology includes ring network / radial operation mode, wherein the value rules of the dynamic coefficients α, β, and γ are as follows:

[0073] In ring network mode, when the ring network is detected to be closed, that is, the bus voltage phase difference is less than 5°, set α=0.4, β=0.4, and γ=0.2. The fault propagation path is calculated using the Dijkstra algorithm, and the branch with the largest impedance is isolated first.

[0074] In radial mode, when a single power supply is detected, that is, the power flow direction is unique, set α=0.6, β=0.3, and γ=0.1, and the action sequence will prioritize triggering the protection device close to the fault point.

[0075] The verification results of both the physical and logical layer verification units are recorded in operation logs via the blockchain network. These logs contain fault signature maps, action sequences, and verification timestamps, and support tamper-proof traceability. In specific implementations, log data is stored using a Merkle tree, where leaf nodes are the SHA256 hash values ​​of individual log entries for tamper-proof storage. This allows for querying tamper-proof traceability paths by device ID and time range. Non-leaf nodes are cascaded hashes of child node hash values, forming a tree-like verification structure.

[0076] It should be further explained that this system uses a multi-dimensional spatiotemporal alignment engine to achieve cross-protocol parsing and spatiotemporal fusion of multi-source heterogeneous data. Combined with dynamic semantic gateways and wavelet packet decomposition technology, it uniformly maps power frequency quantities, transient quantities, and environmental parameters to fault feature maps, resolving the incomplete fault feature extraction problem caused by data silos in traditional technologies. Based on the collaborative mechanism of causal reasoning networks and multi-objective game decision trees, the system can accurately screen key fault features and generate Pareto optimal action sequences, reducing the misjudgment rate of over-tripping while shortening the duration of unnecessary power outages, thereby improving the accuracy of fault isolation and power supply reliability.

[0077] Through an edge-cloud collaborative architecture and dual-channel verification mechanism, the system ensures real-time decision-making while leveraging blockchain technology to achieve trusted synchronization and tamper-proof traceability of action sequences, ensuring that multi-terminal protection devices strictly follow the optimal strategy. Lightweight model deployment and dynamic weight allocation algorithms further optimize resource utilization efficiency and adapt to dynamic changes in grid topology. This solution not only resolves the technical bottlenecks of multi-objective decision-making conflicts and dynamic response lags in new power systems, but also provides scalable technical support for the safe and stable operation of power grids with a high proportion of renewable energy access.

[0078] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a..." does not exclude the presence of additional identical elements in the process, method, article, or device comprising the element.

[0079] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A power supply network collaborative anti-overtripping decision-making system, characterized in that: include: The data acquisition module is connected to the multi-level protection devices, monitoring terminals and environmental sensors in the power supply network to obtain electrical quantity data, non-electrical quantity data and environmental parameters in real time; A multi-dimensional spatiotemporal alignment engine, whose input end is in communication with the data acquisition module and whose output end is connected to the multimodal game decision module, is used to perform cross-protocol parsing, time base synchronization, and spatial topology mapping on the electrical quantity data, non-electrical quantity data, and environmental parameters to generate a dynamic fault feature map; A multimodal game decision module, whose input end receives the dynamic fault feature map and whose output end is connected to the dual-channel verification module, is used to screen key fault features based on a causal reasoning network and generate coordinated action instructions through a multi-objective game model; A dual-channel verification module, comprising a physical layer verification unit and a logical layer verification unit, with an input end receiving the collaborative action instruction and an output end connected to the collaborative execution module, for performing physical simulation verification and formal logic verification on the instruction; The collaborative execution module includes an edge agent and a cloud hub. The edge agent is deployed at the edge node of the power supply network. The cloud hub communicates with the edge agent and the protection device through the blockchain network, and is used to control the action timing of the protection device according to the verified instructions.

2. The power supply network collaborative anti-overtripping decision-making system according to claim 1, characterized in that: The multidimensional spatiotemporal alignment engine includes a dynamic semantic gateway and a spatiotemporal correlation analysis unit; The dynamic semantic gateway has a built-in protocol converter and semantic tag library, which is used to map device data with different communication protocols into a unified semantic space; The spatiotemporal correlation analysis unit integrates a wavelet packet decomposition algorithm and a digital twin model to synchronize the time base of the power frequency quantity and transient high-frequency component data streams, and generates a spatiotemporal correlation map of the fault propagation path based on the dynamic coding rules of the power grid topology.

3. The power supply network collaborative anti-overtripping decision-making system according to claim 2, characterized in that: The spatiotemporal correlation analysis unit performs the following operations: A dynamic weight matrix is ​​constructed based on the physical location and electrical connection relationship of the equipment, and the weight values ​​of the dynamic weight matrix are adjusted in real time according to the output characteristics of the distributed power supply; Based on the timestamp synchronization mechanism triggered by fault events, wavelet packet decomposition technology is used to reconstruct the multi-time scale data stream in the frequency domain to generate the fault feature vector with a unified time base; The dynamic weight matrix and the fault feature vector are input into the digital twin model to generate a spatiotemporal correlation map including the fault location, propagation direction and energy distribution.

4. The power supply network collaborative anti-overtripping decision-making system according to claim 3, characterized in that: The multimodal game decision module includes a causal reasoning network and a multi-objective game decision tree; The causal reasoning network uses a Bayesian causal graph model to calculate causal weights for voltage phase difference, transient energy distribution, and equipment health index, eliminating noise features that are irrelevant to the fault area. The multi-objective game decision tree integrates the Monte Carlo tree search algorithm and the deep Q network, quantifies speed as an action response time benefit function, quantifies selectivity as a fault area isolation accuracy function, and quantifies power supply reliability as a load power outage loss function, and generates a Pareto optimal action sequence through dynamic game.

5. The power supply network collaborative anti-overtripping decision-making system according to claim 4, characterized in that: The multi-objective game decision tree performs the following steps: Initialize the payoff function weights of the game participants based on the key fault features output by the causal reasoning network; Through Monte Carlo tree search, candidate action branches are traversed to calculate the comprehensive score of each branch's speed, selectivity and power supply reliability; A deep Q-network is used to optimize the scoring strategy and generate action sequences and priority rankings that meet multi-objective balance.

6. The power supply network collaborative anti-overtripping decision-making system according to claim 1, characterized in that: In the dual-channel verification module: the physical layer verification unit is connected to the digital twin model, and is used to inject collaborative action instructions into the digital twin system and simulate execution to verify whether the voltage and current changes of key nodes are within the preset threshold range; the logical layer verification unit has built-in formal verification tools and rule bases, and is used to detect whether the action sequence violates the topology connectivity protection rules and the minimum power supply duration rules for critical loads.

7. The power supply network coordinated anti-overtripping decision-making system according to claim 1, characterized in that: In the collaborative execution module: the edge intelligent agent integrates a lightweight fault feature extraction model after knowledge distillation compression, which is used to locally perform fault screening and trigger emergency action pre-instructions; the cloud center runs a global game decision engine and sends timing synchronization signals to the edge intelligent agent and protection device through the blockchain network to ensure strict timing consistency of multi-terminal actions.

8. The power supply network coordinated anti-overtripping decision-making system according to claim 7, characterized in that: The lightweight model of the edge agent is generated in the following way: a fault feature extraction model based on a convolutional neural network is trained in the cloud center; the convolutional neural network model is compressed into a low-latency inference model using knowledge distillation technology and deployed to the edge node; and the task allocation ratio of edge and cloud computing resources is adjusted in real time according to changes in the power grid topology through a dynamic weight allocation algorithm.

9. The power supply network coordinated anti-overtripping decision-making system according to claim 4, characterized in that: The multimodal game decision module also includes a dynamic weight allocation unit for adjusting the weights of the profit functions of speed, selectivity and power supply reliability according to the real-time status of the power grid topology. The weight adjustment formula is: W i =α·T response +β·S isolation +γ·R reliability ; Among them, α, β, and γ are dynamic adjustment coefficients, which are related to the grid operation mode and fault severity.

10. The power supply network coordinated anti-overtripping decision-making system according to claim 6, characterized in that: The verification results of the physical layer verification unit and the logical layer verification unit are recorded in operation logs through the blockchain network. The logs contain fault feature maps, action sequences and verification timestamps, and support tamper-proof traceability.

Images