Adaptive DC partition protection method based on reconfigurable topology perception fault tree and improved LSTM

Abstract

The invention discloses an adaptive direct-current partition protection method based on a reconfigurable topology perception fault tree and an improved LSTM (Long Short Term Memory), which is suitable for fault diagnosis and partition protection of a direct-current system of a transformer substation. According to the algorithm, improved dynamic fault tree analysis and a long short-term memory (LSTM) network model are combined, and rapid positioning and selective isolation protection of direct-current system faults are achieved. The method comprises the following steps: firstly, constructing an updatable dynamic fault tree model based on fault historical data, and correcting a fault event weight in real time by adopting a Bayesian method to realize dynamic reconstruction of a fault propagation path; secondly, a multi-target particle swarm optimization algorithm is introduced, optimization is carried out between the minimum protection response time and the minimum isolation range, and an optimal partition scheme is generated; multi-dimensional time sequence characteristics such as fault current and bus voltage are extracted through a multi-channel LSTM model, and the fault development trend is predicted; and finally, dynamically adjusting a trigger threshold and a delay parameter of a protection action in combination with an online reinforcement learning strategy network of an Actor-Critic architecture. The method has the advantages of quick fault response, flexible partition, adaptive protection strategy and the like, and the safety and the stability of the direct current system can be remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of power system protection and intelligent fault identification technology, and proposes an adaptive DC zonal protection method based on reconfigurable topology-aware fault tree and improved long short-time memory network (LSTM). It is applicable to zonal protection and intelligent fault judgment in various DC systems such as flexible DC transmission, high-speed DC distribution network, and DC microgrid. Background Technology

[0002] With the large-scale integration of new energy sources and the rapid development of DC technology, DC transmission and distribution systems are playing an increasingly important role in the modern energy internet. Compared to traditional AC systems, DC systems have significant advantages in energy conversion efficiency, power flow control flexibility, and equipment compactness. In particular, the widespread application of flexible DC transmission technology has made DC systems a key component of smart grids. However, the protection of DC systems has also become more challenging. Because DC current has no zero-crossing point and lacks a natural interruption mechanism, fault currents are continuous and difficult to cut off, requiring protection systems to have rapid response and accurate fault location capabilities; otherwise, equipment damage and system instability can easily occur.

[0003] Currently, DC protection schemes mainly include setting-value protection based on current surges, logic protection based on communication, and intelligent protection methods based on graph models. Traditional fault tree models rely on fixed topology structures for logical reasoning. While they offer good interpretability, they struggle to adapt to grid topology adjustments caused by dynamic changes in circuit breaker or switch states, limiting their flexibility in practical applications. On the other hand, deep learning-based time-series models such as LSTM show potential in identifying complex fault signals, capturing their temporal characteristics. However, their memory gate structures are relatively rigid, lacking dynamic adjustment capabilities and failing to accurately grasp the fault initiation and evolution process. Furthermore, existing protection strategies are mostly based on static thresholds or empirical parameter configurations, lacking dynamic responses to fault risk and classification confidence, which can easily lead to malfunctions or failures to operate, affecting protection performance.

[0004] To address the aforementioned shortcomings, there is an urgent need for an intelligent protection method that combines the advantages of topology dynamic self-sensing and temporal deep learning. This method should possess both the transparency and reliability of logical reasoning, and the ability to dynamically adapt to changes in the power grid structure, thereby improving the accuracy and response speed of fault identification. To this end, this invention proposes an adaptive DC zone protection method based on a reconfigurable topology-aware fault tree and an improved LSTM. This method dynamically reconstructs the fault tree topology by sensing the on / off status of circuit breakers and switches in real time, enabling adaptive adjustment of the logical reasoning model. Simultaneously, the LSTM structure is improved to enhance its dynamic memory and temporal feature capture capabilities, improving the accurate identification of fault initiation and evolution. Furthermore, a dynamic risk assessment mechanism is introduced to intelligently adjust protection action decisions based on real-time fault risk and confidence levels, effectively reducing the probability of false trips and failures to trip. Summary of the Invention

[0005] To overcome the problems of poor topology adaptability, insufficient fault identification accuracy, and slow response of protection strategies in existing DC system protection methods, this invention proposes a DC partition protection method based on reconfigurable topology-aware fault tree and improved LSTM.

[0006] To address the aforementioned technical problems, this invention provides a method for adaptive DC zoning protection based on reconfigurable topology-aware fault tree and improved LSTM, comprising:

[0007] S1. Real-time acquisition of voltage signals, current signals, and circuit breaker switch status of each zone of the DC system;

[0008] S2. Automatically load or partially reconstruct the fault tree template based on real-time operating condition indicators, and establish logical gate relationships on the two-level topology nodes of the partition layer and the whole station layer for basic events E. i Apply Bayesian online updates;

[0009] S3. Perform feature preprocessing on the acquired signal to extract the half-cycle current RMS, multi-scale spectral energy sequence {x}, etc. t The input is a bidirectional LSTM with embedded temporal attention, and time-varying gating units are used to process the instantaneous features. Adaptive adjustment of the forget gate and input gate;

[0010] S4. Calculate the failure risk probability of each partition and output it as a vector p = [p1, ..., p2]. N ];

[0011] S5. Define the fault type and confidence level, and output the classification results. and its confidence level c;

[0012] S6. According to p, And c, and considering the current operating conditions and zone risk levels, an adaptive protection strategy is adopted, utilizing the particle swarm optimization algorithm to optimize the overcurrent / grounding threshold {θ} online. i} and action delay {Δt i To minimize the protection cost function, and based on the optimization results, switch the corresponding strategy template, such as the action delay template and the backup channel template, and perform partitioned differentiated configuration;

[0013] S7. Output the updated action threshold set {θ} i} and the corresponding action delay {Δt} i};

[0014] S8. Protection device execution: Send each protection parameter and action command in {θ} and {Δt} to the high-speed low-voltage DC circuit breaker to implement current limiting or current interruption;

[0015] S9. Monitoring Log and Model Feedback: Records protection execution results and newly acquired electrical characteristic data, and uses a weighted sliding window mechanism to jointly correct the parameters of the reconfigurable fault tree model and the improved LSTM model, so as to achieve collaborative model iteration and accuracy improvement.

[0016] The method as described in claim 1, characterized in that the fault tree model in S2 is a reconfigurable hierarchical model with topology adaptation capability, supporting dynamic logic gate expansion and condition-driven structure reconfiguration, and its basic event node E i The risk probability update follows the following Bayesian online formula:

[0017]

[0018] Where D represents the current observed data feature, P(E) i Let P(D|E) be the prior probability. i ) is the likelihood function, and its update period is synchronized with topological changes.

[0019] The fault tree employs a dynamic "AND / OR / voting" hybrid logic gate structure at the partition level, and nests "sequence gate + negation gate + gate control template" at the global level, supporting dynamic switching of the following structures:

[0020]

[0021] Where p j W represents the risk probability of child nodes. ij ∈[0,1] represents the state-aware weight, which is dynamically generated based on the real-time activation / deactivation status and the electrical switch topology.

[0022] The improved LSTM network in S3 includes a bidirectional recurrent unit, a temporal attention module, and a time-varying gating function module; its gating structure includes a forgetting gate f. tWith input gate i t Designed as:

[0023]

[0024] in It represents a time-varying modulation quantity driven by the rate of change of the input signal and the spectral energy, which adaptively adjusts the memory update rate to adapt to the dynamic response of different types of faults.

[0025] The temporal attention module uses an embedded multi-scale mechanism to construct a temporal attention vector α. t The calculation method is as follows:

[0026]

[0027] e t =v T tanh(W h h t +W x x t )

[0028] Where h t x is the current hidden state. t For the input features, v, W h W x These are trainable parameters. This attention mechanism highlights short-term spike responses and suppresses ineffective periodic perturbations.

[0029] The adaptive protection strategy in S6 adopts the following multi-objective optimization function:

[0030]

[0031] Where, p i For the risk of partition i, c i C represents the confidence level for fault identification. i The cost of tripping in this zone is Δt. i The action delay is λ1 and λ2 are empirical weights, and the protection strategy converges to the minimum cost solution in the particle swarm algorithm iteration.

[0032] Step S9 employs a weighted sliding window mechanism to jointly correct the parameters of the LSTM and fault tree model, with the following update rules:

[0033]

[0034] L = L CE +γ·||θ-θ ref || 2

[0035] Where θ is the model parameter, L CEFor cross-entropy loss, θ ref The sliding reference parameter is γ, which is the weight of the regularization term and controls the stability of the model. The online update frequency is dynamically synchronized with the protection execution cycle.

[0036] The adaptive DC zonal protection method based on reconfigurable topology-aware fault tree and improved long short-time memory network provided in this application has the following beneficial effects:

[0037] This application dynamically reconstructs the fault tree topology by sensing the status of circuit breakers and switches in real time, enabling the protection logic to automatically adjust with changes in system topology, thereby eliminating misjudgments and protection blind spots caused by topology changes. It improves the internal memory gate structure of the LSTM network, enhancing sensitivity to the initiation and evolution of fault signals, achieving accurate capture and rapid response to complex timing characteristics, and significantly reducing fault detection latency. It introduces an online fault risk assessment and confidence calculation mechanism, dynamically adjusting the protection action threshold based on real-time risk indicators, effectively reducing the probability of false action and failure to act, and significantly improving system reliability. Finally, it integrates topological logic reasoning with deep timing learning. The dual discrimination mechanism can quickly and accurately locate fault zones, meeting the refined protection needs of multi-zone scenarios such as flexible DC transmission, high-speed DC distribution networks, and DC microgrids. At the same time, the logical reasoning of the fault tree ensures the traceability and transparency of the decision, while the deep learning model plays a role in improving the accuracy of the decision, balancing security and intelligence. In addition, this method does not rely on large-scale communication networks or complex field calibration, and has good applicability to DC systems with various topologies, which helps to reduce deployment costs and simplify operation and maintenance, ultimately providing a high-efficiency, intelligent, and interpretable zone protection solution for DC transmission and distribution systems. Attached Figure Description

[0038] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein:

[0039] Figure 1 This application provides an overall architecture diagram of a DC partition protection method based on reconfigurable topology-aware fault tree and improved LSTM.

[0040] Figure 2 This is a schematic diagram of a reconfigurable topology-aware fault tree provided in this application;

[0041] Figure 3 A schematic diagram of the improved LSTM provided in this application;

[0042] Figure 4 A schematic diagram of adaptive DC zone protection provided in this application;

[0043] Figure 5A comparison table of protective action responses under the improved method provided in this application. Detailed Implementation

[0044] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0045] Reference Figure 1 As shown, the overall architecture diagram of the adaptive DC zoning protection method based on reconfigurable topology-aware fault tree and improved LSTM provided in this application includes:

[0046] S1. Data Acquisition Module: First, it collects electrical quantities of each protection zone, such as DC current, DC voltage, switch and circuit breaker status, and environmental parameters, such as temperature and humidity, by using current and voltage sensors and status monitoring devices deployed at each DC bus and feeder. The collected data is then uploaded to the control center in real time via fieldbus or Ethernet.

[0047] S2. Reconfigurable Topology-Aware Fault Tree Model: The system logically divides each substation into zones using a reconfigurable topology-aware fault tree model. Based on the real-time on / off states of circuit breakers and switches, the system dynamically reconstructs the power grid topology and automatically generates a fault tree model. Unlike traditional fixed-topology fault trees, this model has adaptive capabilities in node merging, branching, and logic gate configuration. It can respond promptly to zone reorganization caused by changes in circuit breaker states, ensuring the accuracy and completeness of the fault tree reasoning logic. The model adopts a multi-level tree structure, pre-modeling possible short-circuit, open-circuit, grounding, and component failure scenarios within each zone, and using Boolean operations to quickly calculate the fault risk probability of each zone.

[0048] S3. Improved LSTM Fault Identification: The improved LSTM fault identification module models the timing current and voltage waveforms acquired in step S1. Compared to traditional LSTM, this invention introduces a variable gate control mechanism and a timing attention mechanism internally.

[0049] (1) The gate control unit is extended to an adaptively adjustable double-gate structure to enhance the sensitivity to fault initiation abrupt changes;

[0050] (2) Attention weights based on temporal correlation are added during the hidden state update process to highlight fault characteristics at key time points. Simultaneously, to suppress noise interference, a wavelet packet decomposition preprocessing submodule is configured at the input end to perform multi-scale decomposition on the original waveform, extracting high-frequency fault pulses and low-frequency trend components. The improved LSTM, after training, can output fault types, such as short circuits and grounding faults, along with their classification confidence scores.

[0051] After generating the fault risk probability, fault type, and confidence level for each partition in S4 and S5, the adaptive protection strategy module in S6 performs a fusion decision on the two types of outputs. This module, based on Bayesian inference and online risk assessment algorithms, weights and fuses the fault risk probability and classification confidence level to dynamically calculate the protection action priority and action threshold for each partition. The adaptive protection strategy features the following innovations: automatically adjusting overcurrent, instantaneous trip, and other action thresholds according to risk changes, and supporting reserved time and step-by-step action logic to balance response speed and selectivity.

[0052] S7 Output 3: Protection Parameters: Zone protection parameters calculated according to the adaptive strategy, including operating current threshold, operating time setpoint and backup threshold, etc., are sent to the protection devices of each zone through the protection controller.

[0053] S8. Protection Device Execution: Upon receiving the parameters, each protection device executes the corresponding zone tripping or alarm function. For high-risk zones, instantaneous tripping or forced tripping can be triggered; for medium- and low-risk zones, a time-delay protection strategy is implemented to prevent malfunction.

[0054] S9. Monitoring Logs and Model Feedback: The feedback module summarizes the protection execution results and subsequent power grid operation data, monitors the accuracy and response time of protection actions in real time, and feeds back the actual fault handling results to the topology-aware fault tree and improved LSTM model for online model updates and parameter optimization, forming a closed-loop self-learning mechanism to further improve the intelligence level and robustness of the protection system.

[0055] Reference Figure 2 As shown, the schematic diagram of the reconfigurable topology-aware fault tree provided in this application includes:

[0056] The reconfigurable topology-aware fault tree module consists of four parts: a bottom event acquisition unit, a dynamic topology reconfiguration unit, a fault tree logic reasoning unit, and a risk probability output unit. First, the bottom event acquisition unit periodically scans the set of switch / circuit breaker states S of the primary equipment. (t) ={s i}, where s i ∈{0,1} represents the opening / closing of the i-th circuit breaker, and transmits topology-related information such as secondary circuit signals to the reconfiguration unit; the dynamic topology reconfiguration unit, based on the historical partition coordination matrix T (t-1) and current state S (t) by

[0057] T (t) =R(T) (t-1) ,S (t) )

[0058] Complete the new generation matrix T (t) ∈{0,1} n×nThe calculation, where R(·) can be implemented in O(nlogn) time using breadth-first search or disjoint-set data structure algorithms. This formula embodies the reconfigurable topology awareness of the fault tree structure of this invention, ensuring that each protection zone can be automatically reassembled after a circuit breaker or switch is put into operation or removed. Following this, the fault tree logic reasoning unit will reconstruct the T... ( t) is mapped to a tree-like node structure, representing the underlying events ε={E} such as secondary circuit stability maintenance faults, transformer malfunctions, and device malfunctions. k Calculate the failure probability using a combination of AND and OR gates:

[0059]

[0060] The risk P of the top-level node in each partition is fed into the adaptive protection strategy through the risk probability output unit. Based on this, the invention provides dedicated modeling for "topology sensing module failure" and "algorithm recognition module (improved LSTM) failure," further improving the accuracy of inference and dynamic adaptability. Topology sensing failure is defined as the "AND gate" aggregation of three types of bottom events: dynamic reconstruction failure E... a Data synchronization delay E b Model online update failed. c Its failure probability is

[0061] P 拓扑失败 =P a P b P c

[0062] An adaptive weighting factor α = 1 + β(1 - R) is introduced. sync ), where R sync β is an adjustment coefficient for data synchronization reliability. Risk amplification is applied.

[0063] P′ 拓扑失败 =αP 拓扑失败

[0064] The improved LSTM module's "algorithm-based failure identification" also uses the same "AND gate" format, multiplying the failure probabilities of the three sub-functions—variable gate adjustment, wavelet packet preprocessing, and timing attention—to calculate the results. Finally, in the fault tree, the aforementioned topology failures, algorithm-based failure identification, and traditional hardware (DI / AI / CPU / DO) and software failure events are coupled at a higher level using OR / AND gates as needed, outputting accurate partitioned fault risk probabilities. This, in turn, links with the adaptive protection strategy, achieving the third innovation—dynamic risk assessment and threshold adaptation. The organic integration of these three innovations (reconfigurable topology sensing, improved timing identification, and adaptive risk assessment) comprehensively enhances the real-time performance, reliability, and interpretability of DC partitioned protection.

[0065] Reference Figure 3As shown, the improved LSTM schematic diagram provided in this application includes:

[0066] Figure 3 The improved LSTM cell structure proposed in this invention is shown. The structure mainly includes three stages: feature preprocessing stage, improved LSTM time series modeling stage, and attention-enhanced output stage. It effectively optimizes the time-frequency complexity and time-series dependency characteristics of power system fault signals.

[0067] In the feature preprocessing stage, the input raw fault timing signal x is... t The signal is input to the wavelet packet decomposition module for multi-scale time-frequency analysis. Wavelet packet decomposition divides the signal into multiple frequency bands, extracting high-frequency components (containing information about sudden changes) and low-frequency components (containing information about trend changes and background), enabling finer-grained feature acquisition. Subsequently, the high-frequency and low-frequency components are concatenated and fused to form a fused feature vector. The fused feature vector is used as input to the improved LSTM. This step fully preserves the local abrupt changes and global trend information of the fault signal, improving the accuracy of subsequent time series modeling.

[0068] After entering the improved LSTM temporal modeling stage, feature vectors are fused. Compared to the hidden state h of the previous time step t-1 The inputs are fed into a dynamic gating mechanism module. Unlike traditional LSTM structures, this invention introduces a dynamic gating mechanism, dynamically adjusting the weight distribution of the three gates—the input gate, forget gate, and output gate—based on the current input features and historical states. The calculation method is as follows:

[0069] Input Gate:

[0070]

[0071] Forgotten Gate:

[0072]

[0073] Output gate:

[0074]

[0075] Where σ(·) represents the Sigmoid activation function, W t W f W o Let b represent the trainable parameter matrix. i ,b f ,b o For the bias term, [] indicates vector concatenation.

[0076] Then, based on the above gating signal and candidate memory state

[0077]

[0078] Update the current memory cell state C t :

[0079]

[0080] And further obtain the current hidden state output:

[0081]

[0082] A temporal attention mechanism is introduced to enhance the model's ability to focus on information from key time slices. The hidden state h output at the current time step is then... t The input is fed into the attention module, along with all historical states {h1,h2,…,h...} t} Perform similarity calculations to generate attention weights:

[0083]

[0084] This leads to the context representation vector:

[0085]

[0086] Attention context vector c t The ability to weightedly fuse effective information from different time slices helps the model focus on key feature mutation regions, thereby improving its ability to distinguish fault types. Finally, the attention-enhanced representation c... t The input is fed into the output layer, where fault type classification and confidence output are achieved through fully connected layers and a softmax function. This improved LSTM network structure effectively enhances the sensitivity and robustness of temporal modeling through wavelet packet decomposition feature fusion and dynamic gating + attention mechanisms, demonstrating good adaptability and generalization ability in complex power system fault diagnosis tasks.

[0087] Reference Figure 4 As shown, the adaptive DC zone protection schematic diagram provided in this application includes:

[0088] The system architecture includes a real-time topology sensing module, a reconfigurable fault tree generation and analysis module, an improved LSTM adaptive discrimination module, a partition decision logic module, and a protection execution module, forming a feedback closed-loop intelligent protection system. The system first monitors the DC network operating status through the real-time topology sensing module, including the on / off status of buses, power nodes, energy storage units, and circuit breakers, and updates the system topology in real time. Once a topology change is detected, such as a circuit breaker tripping or load switching, the fault tree module is immediately activated for structural reconstruction. This invention employs a reconfigurable fault tree mechanism, dynamically generating a fault tree structure based on the current topology relationships, forming a deduction diagram of nodes—logic gates—failure paths. The probability of the fault path is quantified through Bayesian inference and denoted as:

[0089]

[0090] Where F represents a system failure event, and M... i Let P(M) represent the i-th path in the minimum cut set. i Let P(F|M) be the probability of its occurrence. i The conditional probability of this path causing system failure is given. This quantitative analysis result, as a priori risk assessment information, will be input into the improved LSTM model along with the real-time measured current / voltage sequence.

[0091] Compared to traditional temporal neural networks, this invention introduces an improved LSTM model with optimized structure for performing fault identification and partitioning tasks. Its main innovations include three aspects: First, fault tree analysis output is introduced as auxiliary features at the LSTM input, i.e., in the input vector X... t Add topology code T and fault path probability vector P(M) to form extended input:

[0092] X t ′=[I t V t [P(M)]

[0093] Among them I t and V t The data consists of current and voltage time-series data, respectively. Secondly, an attention mechanism is introduced to enhance the network's response to local abrupt changes, enabling dynamic focusing on signals at critical moments. Thirdly, a gating structure is used to dynamically weight and adjust the memory cell states based on the protection characteristics of different zones, improving the LSTM's adaptability to different fault types and regional characteristics. The confidence score vector S = [s1, s2, ..., s...] output by the LSTM is improved. k The value ] represents the tripping tendency for each partition, which is then passed to the partition decision logic module. This module combines a dynamic threshold θ. i Perform the following checks on each partition:

[0094]

[0095] Finally, the protection execution module issues a control command to drive the circuit breaker or DC solid-state switch to complete rapid isolation. The execution result is synchronously fed back to the topology sensing module, realizing dynamic closed-loop update of the protection strategy.

[0096] Reference Figure 5 As shown, the comparison table of protection action responses under the improved method provided in this application includes:

[0097] To assess the actual protection performance and response effect of the "Adaptive DC Zone Protection Method," multiple sets of typical fault scenarios were designed for comparative experiments. Specific tests included various types of fault events (such as three-phase short circuit abc and three-phase ground fault abcg), with the fault impedance set to 0.01Ω to simulate severe fault conditions in actual operation. Simultaneously, fault points were set between multiple DC distribution line nodes (such as 1-4DL, 1-5DL, 2-3DL, 5-6DL, etc.) to fully cover electrical paths in different areas and topological locations, obtaining more universal and representative comparative data. In each test, the relay numbers and their operating ranges in response to faults were recorded and compared between existing traditional methods and the improved method proposed in this invention.

[0098] like Figure 5 The comparison table clearly shows that, under the same fault location and fault type,

[0099] Traditional protection methods often have a wide operating range, involving protection relays in multiple non-faulty areas. For example, when an abcg type three-phase ground fault occurs on line 1-4DL (numbered R6), the traditional method would cause four relays (R3, R8, R10, and R15) to operate, resulting in a wide range of tripping behavior. This not only expands the fault isolation range but may also cause maloperation in non-faulty areas, reducing power supply reliability. The improved method proposed in this invention, by introducing a reconfigurable topology-aware fault tree, perceives real-time changes in the system's topology and combines an improved LSTM neural network model that integrates network topology features and fault electrical quantity sequences, enabling more accurate fault location and zone judgment. Combined with the confidence scoring mechanism and adaptive threshold judgment strategy proposed in this method, the system ultimately selects only relay R6, directly associated with the faulty line, to operate, achieving rapid isolation of the smallest protection unit.

Claims

1. A method for adaptive DC zone protection based on reconfigurable topology-aware fault tree and improved LSTM, characterized in that, include: S1. Real-time acquisition of voltage signals, current signals, and circuit breaker switch status of each zone of the DC system; S2. Automatically load or partially reconstruct the fault tree template based on real-time operating condition indicators, and establish logical gate relationships on the two-level topology nodes of the partition layer and the whole station layer for basic events E. i Apply Bayesian online updates; S3. Perform feature preprocessing on the acquired signal to extract the half-cycle current RMS, multi-scale spectral energy sequence {x}, etc. t The input is a bidirectional LSTM with embedded temporal attention, and time-varying gating units are used to process the instantaneous features. Adaptive adjustment of the forget gate and input gate; S4. Calculate the failure risk probability of each partition and output it as a vector p = [p1, ..., p2]. N ]; S5. Define the fault type and confidence level, and output the classification results. and its confidence level c; S6. According to p, And c, and considering the current operating conditions and zone risk levels, an adaptive protection strategy is adopted, utilizing the particle swarm optimization algorithm to optimize the overcurrent / grounding threshold {θ} online. i } and action delay {Δt i To minimize the protection cost function, and based on the optimization results, switch the corresponding strategy template, such as the action delay template and the backup channel template, and perform partitioned differentiated configuration; S7. Output the updated action threshold set {θ} i } and the corresponding action delay {Δt} i }; S8. Protection device execution: Send each protection parameter and action command in {θ} and {Δt} to the high-speed low-voltage DC circuit breaker to implement current limiting or current interruption; S9. Monitoring Log and Model Feedback: Records protection execution results and newly acquired electrical characteristic data, and uses a weighted sliding window mechanism to jointly correct the parameters of the reconfigurable fault tree model and the improved LSTM model, so as to achieve collaborative model iteration and accuracy improvement.

2. The method as described in claim 1, characterized in that, The fault tree model in S2 is a reconfigurable hierarchical model with topology adaptation capabilities, supporting dynamic logic gate expansion and condition-driven structure reconfiguration. Its basic event node E i The risk probability update follows the following Bayesian online formula: Where D represents the current observed data feature, P(E) i Let P(D|E) be the prior probability. i ) is the likelihood function, and its update period is synchronized with topological changes.

3. The method as described in claim 2, characterized in that, The fault tree employs a dynamic "AND / OR / voting" hybrid logic gate structure at the partition level, and nests "sequence gate + negation gate + gate control template" at the global level, supporting dynamic switching of the following structures: Where p j W represents the risk probability of child nodes. ij ∈[0,1] represents the state-aware weight, which is dynamically generated based on the real-time activation / deactivation status and the electrical switch topology.

4. The method as described in claim 1, characterized in that, The improved LSTM network in S3 includes a bidirectional recurrent unit, a temporal attention module, and a time-varying gating function module; its gating structure includes a forgetting gate f. t With input gate i t Designed as: in It represents a time-varying modulation quantity driven by the rate of change of the input signal and the spectral energy, which adaptively adjusts the memory update rate to adapt to the dynamic response of different types of faults.

5. The method as described in claim 4, characterized in that, The temporal attention module uses an embedded multi-scale mechanism to construct a temporal attention vector α. t The calculation method is as follows: e t =v T fishy(W h h t +W x x t ) Where h t x is the current hidden state. t For the input features, v, W h W x These are trainable parameters. This attention mechanism can highlight short-term spike responses and suppress invalid periodic perturbations.

6. The method as described in claim 1, characterized in that, The adaptive protection strategy in S6 adopts the following multi-objective optimization function: Where, p i For the risk of partition i, c i C represents the confidence level for fault identification. i The cost of tripping in this zone is Δt. i The action delay is λ1 and λ2 are empirical weights, and the protection strategy converges to the minimum cost solution in the particle swarm algorithm iteration.

7. The method as described in claim 1, characterized in that, Step S9 employs a weighted sliding window mechanism to jointly correct the parameters of the LSTM and fault tree model, with the following update rules: L=L CE +γ·||θ-θ ref || 2 Where θ is the model parameter, L CE For cross-entropy loss, θ ref The sliding reference parameter is γ, which is the weight of the regularization term and controls the stability of the model. The online update frequency is dynamically synchronized with the protection execution cycle.

Images