Method and system for guaranteeing full-period operation and maintenance of direct current system

Abstract

The invention discloses a full-period operation and maintenance guarantee method and system for a direct-current system, and relates to the technical field of operation and maintenance of power systems. The method comprises the following steps: S1, acquiring a multi-dimensional operation state data stream through multiple types of sensors; s2, generating a standardized time sequence data set through synchronous alignment and normalization processing; s3, calling a pre-trained fault prediction neural network model to evaluate a future fault risk; s4, when the fault probability exceeds a threshold value, a fault isolation decision engine is triggered; s5, the intelligent execution unit executes millisecond-level isolation operation; s6, the closed-loop feedback evaluation module calculates parameter deviation; s7, on-line fine adjustment of model and algorithm parameters is carried out; and S8, periodically generating a preventive maintenance work order. The system correspondingly comprises a multi-source state sensing layer, a data preprocessing layer and other functional layers. According to the method, full-period intelligent closed-loop operation and maintenance of'prediction-isolation-feedback-optimization 'are realized, the traditional manual dependence bottleneck is broken through, the reliability and safety of a direct current system are remarkably improved, and the method is suitable for the fields of electric power, data centers and the like.

Description

Technical Field

[0001] This invention relates to the field of power system operation and maintenance technology, specifically to a method and system for full-cycle operation and maintenance support of DC systems; applicable to DC systems in fields such as power and data centers. Background Technology

[0002] DC systems serve as crucial supports in key areas such as power systems, rail transportation, data centers, new energy power generation, and industrial automation. Their operational stability and fault response capabilities have become core elements in ensuring infrastructure safety. As the core carrier of energy transmission and distribution, the architecture of a DC system typically encompasses power modules, distribution units, load equipment, and protection devices, relying on precise voltage control and current isolation mechanisms to maintain system stability. However, current mainstream operation and maintenance systems still rely primarily on manual intervention as the core response mode. Even if the system possesses basic monitoring capabilities, its function is limited to triggering fault alarms, lacking proactive prediction and closed-loop handling capabilities for abnormal states. This leads to delayed operation and maintenance responses, increased risk of fault propagation, and severely restricts system availability and safety redundancy.

[0003] The core requirements for full-cycle operation and maintenance (O&M) support of DC systems focus on automatic fault isolation and proactive prediction of operational anomalies. An ideal O&M system should be able to identify potential risks at the fault's nascent stage and autonomously execute isolation strategies the instant a fault occurs, blocking the fault propagation path. Simultaneously, it should coordinate with upstream control units to adjust operating parameters, achieving system-level self-healing. However, existing technical solutions generally place the focus of O&M on post-event response and manual inspection, failing to construct an intelligent closed loop covering the entire chain of "prediction-diagnosis-isolation-recovery." This results in the system still heavily relying on the experience and manual operation of O&M personnel when facing complex conditions such as sudden short circuits, insulation degradation, or load changes, leading to both response delays and the risk of operational errors.

[0004] While some existing technologies attempt to enhance equipment status awareness by introducing intelligent inspection terminals or robots—for example, by verifying inspection progress through image comparison or optimizing path planning based on twin space—their fundamental technology remains at the level of data acquisition and visualization. They lack both fault evolution models to support anomaly prediction and the ability to issue control commands and execute switching actions for automatic isolation. When faced with the rapid fault propagation characteristics and millisecond-level protection requirements unique to DC systems, these solutions fall short in terms of response granularity and decision-making closed-loop capabilities, failing to meet the maintenance upgrade demands of modern high-reliability DC systems for "zero human intervention and full-cycle self-sufficiency." Therefore, there is an urgent need for a DC system full-cycle maintenance and support method and system that deeply integrates status prediction, fault diagnosis, and automatic control capabilities to overcome the structural bottlenecks of existing technologies in terms of proactivity, closed-loop operation, and intelligence.

[0005] In the prior art, Chinese Patent Publication No. CN107069948A discloses a remote detection and monitoring system and method for a DC power supply system. The DC power supply system mainly includes a charger, a battery, and an insulation monitoring device. The remote monitoring system comprises: a charger monitoring module, an insulation detection module, a battery inspection module, a battery storage room environmental detection unit, a system server, and a switch serial communication unit. The method involves using a remote PC via the server to detect the battery, charger, insulation monitoring system, battery room temperature and humidity, and the working environment of the DC power supply system. This enables fault diagnosis and the generation of fault handling strategies, and also includes a battery bank leakage alarm function to prevent DC system collapse when mains power is lost. This invention allows maintenance personnel to monitor the status of DC system equipment through this remote fault diagnosis method, improving DC system fault handling capabilities and work efficiency, and ensuring safe equipment operation. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to address the shortcomings of the prior art by disclosing a method and system for full-cycle operation and maintenance of DC systems, which solves the problems of DC systems relying on manual labor and having a delayed response.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method and system for full-cycle operation and maintenance support of a DC system, comprising:

[0008] S1 Multi-dimensional Real-time Operation Status Data Acquisition: Through voltage sensors, current sensors, temperature sensors, insulation monitoring devices, and communication status monitoring modules deployed at key nodes of the DC system, real-time data acquisition of DC bus voltage, branch current, equipment temperature rise, insulation resistance to ground, and communication link status is generated, forming a multi-dimensional real-time operation status data stream covering the entire system topology.

[0009] S2 Standardized Time Series Dataset Generation: The multi-dimensional real-time running status data stream is synchronized and aligned according to the time series, and data normalization processing is performed according to the preset sampling period to generate a standardized time series dataset containing core features such as voltage fluctuation feature sequence and current harmonic distortion sequence.

[0010] S3 Fault Risk Probabilistic Assessment: Based on the standardized time-series dataset, a pre-trained fault prediction neural network model is invoked to probabilistically assess the potential fault risks of the DC system within a preset time window in the future, and output multi-time-granularity fault occurrence probability values.

[0011] S4 Fault Isolation Decision Trigger: When the probability of failure of any critical equipment exceeds a preset threshold, the fault isolation decision engine is triggered.

[0012] S5 Optimal Isolation Path Execution: The fault isolation decision engine generates an optimal fault isolation path instruction that includes the circuit breaker operation sequence, bypass switch control logic, and load transfer priority. This instruction is then sent to an intelligent execution unit consisting of a programmable logic controller, a solid-state relay array, and a mechanical disconnect switch drive mechanism to perform millisecond-level electrical disconnection and physical isolation operations.

[0013] S6 Closed-Loop Feedback Deviation Calculation: After fault isolation is completed, the closed-loop feedback evaluation module is started to collect the steady-state operating parameters of the system and calculate the deviation from the expected performance index using the weighted Euclidean distance formula;

[0014] Online fine-tuning of S7 model and algorithm parameters: Based on the deviation calculation results, the weight parameters of the fault prediction neural network model and the path optimization algorithm parameters of the fault isolation decision engine are fine-tuned online;

[0015] S8 Preventive Maintenance Work Order Generation: During normal system operation, the preventive maintenance planning module is periodically activated to generate an equipment-level maintenance priority ranking table based on equipment operation data and push encrypted digital work orders.

[0016] Furthermore,

[0017] In the S3 fault risk probabilistic assessment, the standardized time series dataset includes voltage fluctuation characteristic sequence, current harmonic distortion sequence, temperature rise gradient sequence, insulation degradation trend sequence, and communication packet loss rate sequence.

[0018] Furthermore,

[0019] In the online fine-tuning of the S7 model and algorithm parameters, the fault prediction neural network model adopts a multi-layer long short-term memory network structure. Its input layer receives the standardized time series dataset, and the hidden layer extracts voltage sag correlation features, current mutation coupling features, temperature rise cumulative effect features, and insulation gradual failure features through a gating mechanism. The output layer outputs the failure probability values ​​of each key device at three time granularities in the future.

[0020] Furthermore,

[0021] In the execution of the S5 optimal isolation path, the optimal fault isolation path instruction includes the sequence of circuit breaker numbers to be disconnected, the sequence of bypass switch numbers to be closed, and the load transfer priority list.

[0022] Furthermore,

[0023] In the execution of the S5 optimal isolation path, the intelligent execution unit includes a programmable logic controller, a solid-state relay array, and a mechanical isolation switch drive mechanism;

[0024] The solid-state relay array is constructed using silicon carbide power devices, with a turn-off time of <20μs.

[0025] The mechanical disconnect switch drive mechanism uses a permanent magnet synchronous motor in conjunction with a worm gear reducer, achieving a position feedback accuracy of ±0.5 degrees.

[0026] Furthermore,

[0027] In the S6 closed-loop feedback deviation calculation, the closed-loop feedback evaluation module uses the weighted Euclidean distance formula to calculate the deviation between the steady-state operating parameters and the expected performance indicators;

[0028] Among them, the weighting coefficient for bus voltage recovery accuracy is 0.4, the weighting coefficient for load power supply continuity index is 0.3, the weighting coefficient for standby path temperature rise rate is 0.2, and the weighting coefficient for total system loss increment is 0.1.

[0029] Furthermore,

[0030] In the generation of the S8 preventive maintenance work order, the preventive maintenance planning module generates an equipment-level maintenance priority ranking table based on the equipment's cumulative running time, historical fault records, current health status score, and manufacturer-recommended maintenance cycle, and pushes the digital work order to the operation and maintenance management platform.

[0031] The digital work orders are transmitted via an encrypted communication protocol, using the national cryptographic algorithm SM4 for encryption and SM2 for signature.

[0032] A full-cycle operation and maintenance support system for a DC system includes: a multi-source state perception layer, a data preprocessing layer, a fault prediction model layer, an isolation decision execution layer, a closed-loop feedback optimization layer, and a preventive maintenance planning layer;

[0033] The multi-source state perception layer is deployed at key nodes of the DC system and integrates voltage sensors, current sensors, temperature sensors, insulation monitoring devices, and communication state monitoring modules.

[0034] The multi-source status sensing layer is used to collect DC bus voltage, branch current, equipment temperature rise, insulation resistance to ground value and communication link status data in real time, forming a multi-dimensional real-time operating status data stream covering the entire topology of the system;

[0035] The data preprocessing layer interacts with the multi-source state perception layer.

[0036] The data preprocessing layer is used to perform time series synchronization and alignment operations on multi-dimensional real-time running status data streams, and to complete data normalization processing according to a preset sampling period to generate a standardized time series dataset containing core features such as voltage fluctuation feature sequences and current harmonic distortion sequences.

[0037] The fault prediction model layer calls a pre-trained fault prediction neural network model, and based on the standardized time series dataset, performs a probabilistic assessment of the potential fault risks of the DC system within a future preset time window, and outputs the probability values ​​of key equipment faults at multiple time granularities.

[0038] The isolation decision execution layer includes a fault isolation decision engine and an intelligent execution unit;

[0039] The isolation decision execution layer is used to trigger the fault isolation decision engine to generate the optimal fault isolation path instruction when the failure probability of any critical equipment exceeds a preset threshold, and send it to the intelligent execution unit to perform millisecond-level electrical disconnection and physical isolation operations.

[0040] After the closed-loop feedback optimization layer fault isolation operation is completed, the steady-state operating parameters of the system are collected, the deviation from the expected performance index is calculated using the weighted Euclidean distance formula, and the weight parameters of the fault prediction neural network model and the path optimization algorithm parameters of the fault isolation decision engine are fine-tuned online based on the deviation results.

[0041] During normal system operation, the preventive maintenance planning layer periodically activates the equipment status assessment logic. Based on the equipment's cumulative runtime, historical fault records, current health status score, and manufacturer-recommended maintenance cycle, it generates an equipment-level maintenance priority ranking table and pushes a digital work order encrypted with national cryptographic algorithms to the operation and maintenance management platform.

[0042] Furthermore,

[0043] The system also includes a human-machine interface for graphically displaying the real-time topology of the DC system, a heatmap of equipment health status scores, a fault risk prediction curve, the execution status of isolation operations, and historical maintenance work order records.

[0044] Furthermore,

[0045] The insulation monitoring device adopts the unbalanced bridge method principle, with a sampling resistor value of 100kΩ, an accuracy class of 0.1, and a voltage measurement resolution of 1mV.

[0046] The optimal fault isolation path instructions include: circuit breaker operation sequence, bypass switch control logic, and load transfer priority.

[0047] Compared with the prior art, the present invention has the following advantages:

[0048] This invention transforms the traditional operation and maintenance model, which relies on manual inspection and passive response, into a fully intelligent closed-loop operation and maintenance system that automatically completes the entire lifecycle of "prediction-diagnosis-isolation-recovery-optimization." This significantly reduces reliance on manpower and the risk of human error. By employing a multi-layer long short-term memory network model to perform in-depth analysis of multi-dimensional time-series data, it can accurately identify potential risks and quantify their probabilities before a fault occurs, achieving a leap from "post-fault maintenance" to "pre-fault early warning" and providing a valuable time window for proactive intervention. Through a fault isolation decision engine, it dynamically generates the optimal isolation path and, with the help of an intelligent execution unit that coordinates solid-state relays based on silicon carbide devices with high-precision mechanical mechanisms, it can complete the electrical and physical isolation of faulty branches within hundreds of milliseconds. Isolation effectively curbs the spread of faults and ensures power continuity in non-faulty areas. Through a closed-loop feedback evaluation module, the system can fine-tune the prediction model and decision algorithm parameters online based on the deviation between the actual operating effect and expected indicators after isolation operations. This enables the system to adaptively evolve, and the operation and maintenance strategy continuously optimizes over time, becoming increasingly "intelligent." The preventative maintenance planning module integrates multi-dimensional information such as equipment runtime, health status, and environmental factors to dynamically generate maintenance priorities and digital work orders, making maintenance work more scientific and targeted, effectively reducing unplanned downtime and extending equipment lifespan. The entire process employs national cryptographic algorithms for data encryption and instruction signing, ensuring the security of data transmission and execution. Simultaneously, comprehensive visualization through a human-machine interface and complete audit logs empower maintenance personnel with global monitoring capabilities and post-event traceability, achieving a perfect combination of intelligent operation and maintenance and manual supervision. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the overall technical architecture of a DC system full-cycle operation and maintenance guarantee method and system according to the present invention;

[0050] Figure 2 This is a schematic diagram of the core principle framework of the fault prediction neural network model of the present invention;

[0051] Figure 3 This is a flowchart illustrating the logical flow of the multi-source state perception and data preprocessing layer of this invention.

[0052] Figure 4 This is a schematic diagram of the collaborative control framework between the fault isolation decision engine and the intelligent execution unit of the present invention;

[0053] Figure 5 This is a framework diagram of the adaptive parameter adjustment mechanism of the closed-loop feedback optimization layer of the present invention;

[0054] Figure 6 This is a logical framework diagram of the multi-dimensional scoring and work order generation of the preventive maintenance planning module of the present invention;

[0055] Figure 7 This is a schematic diagram illustrating the interaction between the DC system topology and the dynamic analysis of the isolation boundary in this invention.

[0056] Figure 8 This is a schematic diagram illustrating the data linkage relationship between the human-computer interaction interface and various functional layers of the system in this invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0058] Example 1

[0059] This invention provides a method and system for full-cycle operation and maintenance of DC systems. Its core lies in constructing a closed-loop intelligent operation and maintenance system covering six major links: perception, prediction, decision-making, execution, feedback, and maintenance, which completely breaks through the limitations of traditional DC systems that rely on manual response and passive fault handling.

[0060] S1 Multi-dimensional Real-time Operation Status Data Acquisition: Through voltage sensors, current sensors, temperature sensors, insulation monitoring devices, and communication status monitoring modules deployed at key nodes of the DC system, real-time data acquisition of DC bus voltage, branch current, equipment temperature rise, insulation resistance to ground, and communication link status is generated, forming a multi-dimensional real-time operation status data stream covering the entire system topology.

[0061] Specifically, a PT1000-110kV high-precision voltage sensor is deployed between the positive and negative terminals of the DC bus, with a sampling frequency of 1000 times / s, a measurement range covering 80%-120% of the rated voltage (88-132kV), and a resolution of 0.05V; an HEC-3000A Hall effect current sensor is installed at the outlet of each branch, with a synchronous sampling frequency of 1000 times / second, a measurement range of 0-3000A, and a resolution of 0.01A; DS18B20 digital temperature sensors are attached to heat-sensitive parts such as power device heat sinks and capacitor casings, with a sampling frequency of 50 times / second, a temperature measurement range of -40℃ to 150℃, and an accuracy of ±0.5℃.

[0062] The insulation monitoring device employs the unbalanced bridge method. Sampling resistors with a resistance of 100kΩ and an accuracy class of 0.1 are connected between the positive and negative terminals of the busbar and ground. When the busbar voltage is 110kV, the voltage difference between the midpoint of the two resistors and ground is measured to be 2.3mV. This difference is then calculated using the formula: ;

[0063] The calculated insulation resistance to ground is 239 MΩ. The communication status monitoring module is built into the IEC61850 standard smart terminal with a sampling period of 1 second. Within a 10-second window, 3000 data packets are sent and 2997 are received, with a calculated packet loss rate of 0.3%. All sensors and monitoring modules are connected to the S7-1500 central data acquisition unit via Profinet industrial Ethernet. The unit has a built-in IEEE1588PTP time synchronization module to ensure that the timestamp error of each channel is ≤8μs, forming a multi-dimensional real-time operating status data stream covering 20 branches and 15 core devices of the system.

[0064] S2 Standardized Time Series Dataset Generation: The multi-dimensional real-time running status data stream is synchronized and aligned according to the time series, and data normalization processing is performed according to the preset sampling period to generate a standardized time series dataset containing core features such as voltage fluctuation feature sequence and current harmonic distortion sequence.

[0065] Specifically, after receiving the raw data, the central data acquisition unit uses cubic spline interpolation to complete the missing sampling points and resamples at a 10ms sampling period; it then performs normalization processing using the maximum-minimum scaling method (voltage normalization range 99-121kV, current 0-2250A, temperature -10℃-100℃), with the normalization formula as follows: ,

[0066] The generated standardized time-series dataset contains 5 core sequences. Example data includes: voltage fluctuation feature sequence [0.82, 0.85, 0.91, 0.89, 0.87] (corresponding to actual voltages of 99.2kV, 99.5kV, 100.1kV, 99.9kV, and 99.7kV); insulation degradation trend sequence with a fitting slope of -0.002 over 10 sampling periods, corresponding to a decrease in insulation resistance of 1MΩ per 100ms; and abnormal data exceeding the threshold (such as a current of 3500A at a certain moment) are marked and stored in the MySQL abnormal event database and do not participate in subsequent model inference.

[0067] S3 Fault Risk Probabilistic Assessment: Based on the standardized time-series dataset, a pre-trained fault prediction neural network model is invoked to probabilistically assess the potential fault risks of the DC system within a preset time window in the future, and output multi-time-granularity fault occurrence probability values.

[0068] Specifically, a pre-trained fault prediction neural network model deployed on a cluster of 4 NVIDIA A100 GPUs is invoked. The model is a three-layer LSTM structure with 128 memory units in each layer. The input layer receives 10 seconds of historical data (1000 sampling points × 5 types of features).

[0069] Model training process: 5000 historical samples (including 800 fault samples, covering four types of faults: voltage collapse, current overload, insulation breakdown, and communication interruption) were used. The prediction bias was measured using the cross-entropy loss function. A stochastic gradient descent algorithm with a driving term (initial learning rate 0.001, learning rate × 0.9 when accuracy improvement is <0.5% after 3 consecutive rounds of verification) was used for 1000 iterations to converge. The test set accuracy was 97.7% and the false alarm rate was 3.2%. The inference result for a certain charger: the probability of a fault in the next 10 seconds is 0.62, 0.71 in 30 seconds, and 0.78 in 60 seconds. Because the probability in 30 seconds exceeds the preset threshold of 0.7, the S4 fault isolation decision is triggered.

[0070] S4 Fault Isolation Decision Trigger: When the probability of failure of any critical equipment exceeds a preset threshold, the fault isolation decision engine is triggered.

[0071] Specifically, the fault isolation decision engine (integrated into S7-1500 PLC) identifies the faulty electrical island (including charger and branch circuit breakers Q1-Q3) through the adjacency matrix traversal method. After evaluating three sets of candidate isolation boundaries, it generates the optimal isolation path instruction (corresponding to claim 4): the circuit breaker number sequence to be disconnected [Q1,Q3,Q5], the bypass switch number sequence to be closed [B2,B4], and the load transfer priority list [Level 1 load: main control system power supply branch > Level 2 load: auxiliary equipment power supply branch > Level 3 load: backup lighting branch]. After being digitally signed by SM2, the instruction is sent to the intelligent execution unit. The SSR-200A / 1200V silicon carbide solid-state relay array in the unit completes the electrical disconnection within 15μs (conduction resistance 4.8mΩ, withstands 10 times the rated current surge). Subsequently, the permanent magnet synchronous motor drives the isolating switch (rated speed 3000rpm, transmission ratio 1:100) to perform physical isolation within 160ms. The built-in Hall position sensor provides position accuracy of ±0.3°. The total isolation time is 175ms, ensuring that the fault does not spread to the non-faulty area.

[0072] S5 Optimal Isolation Path Execution: The fault isolation decision engine generates an optimal fault isolation path instruction that includes the circuit breaker operation sequence, bypass switch control logic, and load transfer priority. This instruction is then sent to an intelligent execution unit consisting of a programmable logic controller, a solid-state relay array, and a mechanical disconnect switch drive mechanism to perform millisecond-level electrical disconnection and physical isolation operations.

[0073] S6 Closed-Loop Feedback Deviation Calculation: After fault isolation is completed, the closed-loop feedback evaluation module is started to collect the steady-state operating parameters of the system and calculate the deviation from the expected performance index using the weighted Euclidean distance formula;

[0074] Specifically, after isolation is completed, the closed-loop feedback evaluation module waits for the system to enter a steady state. The judgment criteria are: bus voltage fluctuation ≤0.5% for 5 seconds, and four key parameters are collected: bus voltage recovery accuracy 98.8%, load power supply continuity 97.2%, standby path temperature rise rate 1.2℃ / 10s, and total system loss increment 6%.

[0075] The deviation is calculated using the weighted Euclidean distance formula:

[0076] ,

[0077] The parameter update was triggered because the deviation was greater than the tolerance threshold of 0.15.

[0078] Online fine-tuning of S7 model and algorithm parameters: Based on the deviation calculation results, the weight parameters of the fault prediction neural network model and the path optimization algorithm parameters of the fault isolation decision engine are fine-tuned online;

[0079] The deviation is backpropagated as an error signal to the LSTM model, and the weight gradient of each layer is calculated using the chain rule with a step size of 0.0001. The weight of the voltage sag feature extraction layer is fine-tuned by 0.45%. At the same time, the load influence weight coefficient of the fault isolation decision engine is updated from 0.3 to 0.32, and the path selection logic is optimized.

[0080] S8 Preventive Maintenance Work Order Generation: During normal system operation, the preventive maintenance planning module is periodically activated to generate an equipment-level maintenance priority ranking table based on equipment operation data and push encrypted digital work orders.

[0081] Specifically, during the period of lowest system load each day, the preventive maintenance planning module is activated, and data of a certain circuit breaker is read: cumulative operation of 17520h, operation stability score: 100-(17520 / 100000)×20=96.496 points, 1 failure in the past year (deduction of 5 points), ambient temperature and humidity of 25℃ / 45% (environmental adaptability score of 80 points), the last maintenance was delayed by 1 day (maintenance response score of 99 points), and the inventory of key spare parts is sufficient (spare parts availability score of 100 points).

[0082] The weighted summation (0.3 / 0.25 / 0.2 / 0.15 / 0.1) yields a score of 90.298, generating a planned maintenance work order. This work order is then encrypted using SM4 (encrypted data: E8A3F29D7C1B4E96...) and pushed to the operations and maintenance management platform. The work order includes the maintenance items (cleaning contacts, inspecting mechanical characteristics), spare parts model (VS1-12 / 630), and suggested execution time.

[0083] Example 2

[0084] In an embodiment of the present invention, voltage sensors, current sensors, temperature sensors, insulation monitoring devices, and communication status monitoring modules deployed at key nodes of the DC system are used to collect DC bus voltage, branch current, equipment temperature rise, insulation resistance to ground value, and communication link status data in real time, forming a multi-dimensional real-time operating status data stream covering the entire topology of the system.

[0085] Specifically, this includes: installing high-precision voltage sensors between the positive and negative terminals of the DC bus, with a sampling frequency set to 1000 times / second, a measurement range covering 80% to 120% of the rated voltage, and a resolution of not less than 0.1V; installing Hall effect current sensors at the outlet of each branch, with a sampling frequency synchronously set to 1000 times / second, a measurement range configured according to 50% to 200% of the branch's rated current, and a resolution of not less than 0.01A; and attaching digital temperature sensors to heat-sensitive parts such as power device heat sinks, capacitor casings, and busbar connection points, with a sampling frequency of 50 times / second and a temperature measurement range of -40℃ to 15℃. The insulation monitoring device operates at 0℃ with an accuracy of ±0.5℃. It employs the unbalanced bridge method, connecting high-precision sampling resistors (100kΩ, 0.1 accuracy class) between the positive and negative terminals of the DC bus and ground. The insulation resistance to ground is calculated in real-time by measuring the voltage difference between the midpoints of the two resistors and the bus voltage value. The voltage measurement resolution reaches 1mΩ, and the insulation resistance calculation range is 0.1MΩ-1000MΩ. The communication status monitoring module is built into each intelligent terminal device, continuously recording the timestamps of data packet transmission and reception, counting the number of lost packets per unit time, and calculating the packet loss rate. The sampling period is 1 second. All sensors and monitoring modules are connected to the central data acquisition unit via industrial Ethernet or fiber optic ring network. The data acquisition unit has a built-in hardware time synchronization module, ensuring that the timestamp error of all channel data is less than 10µs, thus forming a strictly synchronized multi-dimensional real-time operating status data stream.

[0086] Specifically, the model structure parameters are as follows: the input layer has a dimension of 1000×5 (1000 sampling points × 5 types of features), the first LSTM (voltage sag correlation feature extraction), the second LSTM (current mutation coupling feature extraction), and the third LSTM (multi-source heterogeneous feature fusion) each contain 128 memory units, the forget gate, input gate, and output gate use the sigmoid function for activation, and the cell state update uses the tanh function; the output layer has 3 neurons (corresponding to the failure probability in the next 10s, 30s, and 60s), and the activation function is sigmoid.

[0087] Feature extraction example:

[0088] The first layer of LSTM filters invalid historical data with a forget gate weight of 0.6 and filters effective voltage features with an input gate weight of 0.3. It identifies the transient event of "110kV bus voltage transiently dropping from 110kV to 99kV (amplitude 10%) for 0.5s" and calculates the correlation between this event and capacitor aging fault as 0.82.

[0089] The second layer LSTM: Cell state records current harmonic accumulation effect, extracts the mutation feature of "branch current fundamental amplitude 1500A, third harmonic amplitude 300A (distortion rate 20%)", and couples the IGBT module overheating fault coefficient 0.78;

[0090] The third layer of LSTM integrates the cumulative temperature rise feature (power device temperature rise of 8℃ in 10min, cumulative effect coefficient 0.65), insulation degradation feature (slope -0.003), and communication packet loss rate feature (0.5%), and outputs a comprehensive feature vector [0.72, 0.68, 0.81].

[0091] Example of inference output: Based on the model inference, a battery pack outputs the probability of failure in the next 10 seconds as 0.58, 30 seconds as 0.73, and 60 seconds as 0.81. The probability in the next 30 seconds exceeds the threshold of 0.7, triggering the fault isolation process.

[0092] Example 3

[0093] In an embodiment of the present invention, the multi-dimensional real-time running status data stream is synchronized and aligned according to the time series, and the data is normalized according to the preset sampling period to generate a standardized time series dataset.

[0094] Specifically, this includes: After receiving the raw data, the central data acquisition unit first interpolates and aligns the data from each channel according to a unified time reference. The interpolation method uses a cubic spline function to ensure the continuity and smoothness of the data at missing sampling points. The aligned data is then resampled every 10ms to generate a time series with a fixed frequency. Subsequently, each physical quantity is normalized using a maximum-minimum scaling method. This involves subtracting the minimum operating value specified in the equipment's technical specifications from the value of each sampling point, and then dividing by the difference between the maximum and minimum operating values. For voltage data, the minimum operating value is 90% of the rated voltage, and the maximum operating value is 110% of the rated voltage. For current data, the minimum operating value is 0A, and the maximum operating value is 150% of the branch's rated current. For temperature data, the minimum operating value is -10℃, and the maximum operating value is 100℃. For insulation resistance data, the minimum operating value is 1MΩ. The maximum working value is 500MΩ; for communication packet loss rate data, the minimum working value is 0 and the maximum working value is 1; the normalized values ​​are compressed into the 0-1 range to form a dimensionless standardized time series dataset; if the original value of a sampling point exceeds the above maximum and minimum value range, it is marked as abnormal data and will not participate in the subsequent model inference process, but will be stored separately in the abnormal event database as auxiliary evidence for fault diagnosis; the standardized time series dataset is divided into five independent sequences according to the type of physical quantity: the voltage fluctuation characteristic sequence records the instantaneous deviation amplitude and duration of the bus voltage relative to the rated value; the current harmonic distortion sequence extracts the ratio of the fundamental amplitude to the amplitude of each harmonic through fast Fourier transform; the temperature rise gradient sequence calculates the ratio of the temperature difference between adjacent sampling points to the time interval; the insulation degradation trend sequence performs linear fitting on the insulation resistance values ​​of ten consecutive sampling periods and extracts the slope as the degradation rate; the communication packet loss rate sequence directly records the proportion of packet loss per unit time.

[0095] Specifically, stability score: A fuse has accumulated 26280 hours of operation, according to the formula:

[0096] The score is calculated as "100 - (cumulative time / 100000) × 20", resulting in:

[0097] 100 - (26280 / 100000) × 20 = 94.744 points;

[0098] Historical fault deduction: There were 3 faults in the past year. The deduction is calculated as "deduction = number of faults × 5". The deduction is 3 × 5 = 15 points. The corrected score is 94.744 - 15 = 79.744 points.

[0099] Environmental adaptability score: Ambient temperature 35℃ and humidity 60%, corresponding to the "medium" level, score 60 points;

[0100] Maintenance response score: The last maintenance was delayed by 2 days. The score is calculated as "score = 100 - number of days delayed × 1", so the score is 100 - 2 × 1 = 98 points.

[0101] Spare parts availability score: Critical spare parts need to be ordered and are expected to arrive in 3 days. The score is calculated as "score = 100 - estimated arrival days × 2", so the score is 100 - 3 × 2 = 94 points.

[0102] Weighted summation score: Calculated with weights of 0.3 (operational stability), 0.25 (historical failures), 0.2 (environmental adaptability), 0.15 (maintenance response), and 0.1 (spare parts availability). The total score is 94.744×0.3+79.744×0.25+60×0.2+98×0.15+94×0.1=82.108 points, generating a planned maintenance work order.

[0103] Example of core work order content: Maintenance item (replace fuse assembly, test insulation resistance), spare part model (RT16-100A), suggested execution time (within 3 days 00:00-02:00 during off-peak load period), safety operation procedure number (DL / T408-2010), risk warning (before operation, disconnect the upstream circuit breaker, test for voltage and connect the grounding wire).

[0104] Example 4

[0105] In an embodiment of the present invention, the step of calling a pre-trained fault prediction neural network model based on the standardized time-series dataset to probabilistically assess the potential fault risks of the DC system within a preset future time window specifically includes: the fault prediction neural network model is deployed in a central processing server, and its structure is a three-layer long short-term memory network, each layer containing 128 memory units; the input layer receives a standardized time-series dataset with a length of 1000 sampling points, corresponding to 10 seconds of historical operating data; the first layer of the long short-term memory network extracts voltage sag correlation features, its forget gate controls the influence weight of historical voltage fluctuations on the current state, the input gate filters whether the current voltage deviation constitutes a valid feature, and the output gate determines the transmission strength of the feature to the next stage; the second layer of the long short-term memory network extracts current mutation coupling features, its cell state records the cumulative effect of current harmonic components over time, and the hidden state output reflects the dynamic correlation between current distortion and load changes; the third layer of the long short-term memory network integrates temperature rise gradient, insulation degradation trend, and communication packet loss rate features, and coordinates the time dependency relationship of the three types of heterogeneous data through a gating mechanism; the output layer is a fully connected layer containing three neurons, divided into three layers. The model is designed to detect fault occurrence probabilities at three time granularities: 10s, 30s, and 60s. During training, a historical fault sample database is used for supervised learning. This database contains complete time-series data of voltage collapse, current overload, insulation breakdown, and communication interruption events recorded in similar DC systems over the past five years. Each sample is labeled with a fault type, the time of occurrence, and the extent of impact. The cross-entropy loss function is used to measure the difference between the predicted probability and the true label during training. The optimizer uses a stochastic gradient descent algorithm with a driving term, and the initial learning rate is set to... The learning rate decay coefficient is set to 0.001. After each round of training, the learning rate decay coefficient is dynamically adjusted based on the accuracy of the validation set. If the accuracy improvement is less than 0.5% for three consecutive rounds of validation, the learning rate is multiplied by 0.9. After the model converges, its fault prediction accuracy on the test set reaches 97.7%, and the false alarm rate is controlled within 3.5%. In actual operation, the model receives a new standardized time series dataset every ten seconds and outputs fault probability values ​​at three time granularities. If any probability value exceeds the preset threshold of 0.7, the device is determined to have a high-risk fault hazard, triggering the subsequent isolation decision process.

[0106] Specifically, system-level deployment parameters:

[0107] Multi-source status sensing layer: Deployed according to "one set of voltage / current sensors for every two branches and one temperature sensor for each core device", covering 20 branches and 15 devices of the 110kV DC system. All sensors have passed ATEX explosion-proof certification.

[0108] Data preprocessing layer: Deployed on EC-8000 edge computing gateway, equipped with ARM Cortex-A53 quad-core processor, it can process 100,000 raw data per second, with synchronization alignment latency ≤5ms and normalization processing latency ≤3ms;

[0109] Fault prediction model layer: Deployed on a cloud GPU cluster (4 NVIDIA A100s), receiving preprocessed data via RESTful API, with a single inference latency of ≤200ms and supporting 100 concurrent inferences per second;

[0110] Isolation Decision Execution Layer: The fault isolation decision engine is integrated into the S7-1500 PLC and communicates with the intelligent execution unit via the Modbus TCP protocol, with an instruction transmission delay of ≤10ms;

[0111] Human-computer interaction interface: Developed based on the Qt5.15 framework, supporting 1920×1080 resolution display. The main view displays the system's electrical topology, and the device icon color dynamically changes according to the health score (green: >90 points, yellow: 70-90 points, red: <70 points). Clicking on the #1 charger (score 58 points, red icon) will pop up a details window, displaying real-time parameters (voltage 109.8kV, current 1250A, temperature 58℃), historical fault records (2 insulation alarms in the past year), and a fault risk prediction curve for the next 24 hours (updated every 10 minutes). It supports manual triggering of device self-tests, including sensor calibration tests (comparing the error of a standard signal source) and communication link stress tests (sending 1000 test data packets). The self-test results generate a PDF report and are archived.

[0112] Data interaction process standard: multi-source state perception layer → data preprocessing layer (edge ​​gateway) → fault prediction model layer (cloud GPU) → isolated decision execution layer (PLC + execution unit) → closed-loop feedback optimization layer (edge ​​gateway). The entire link adopts SSL / TLS1.3 encrypted transmission to ensure data transmission security.

[0113] Example 5

[0114] In an embodiment of the present invention, when the probability value of failure of any key equipment exceeds a preset threshold, the fault isolation decision engine is triggered.

[0115] Specifically, this includes: a fault isolation decision engine with a built-in DC system topology diagram, which stores the electrical connections between devices in the form of an adjacency matrix. A matrix element value of 1 indicates that the two devices are directly connected, and a value of 0 indicates that there is no direct connection. The engine also maintains a current load distribution status table, recording the percentage of the actual current carried by each branch relative to the rated current. A backup path capacity margin table records the difference between the maximum allowable current of all backup power supply paths and the currently allocated current. A device health status score table is dynamically updated based on the output of the preventive maintenance planning module, with a score range of 0 to 100. When a fault warning signal is received, the engine first traverses the system connection matrix, performing a breadth-first search starting from the faulty device to identify all sets of devices electrically connected to it. This set constitutes the electrical island where the faulty device is located. Then, it calculates the set of all possible isolation boundary nodes within this electrical island. Boundary nodes are defined as those that simultaneously connect the faulty electrical island and the non-faulty electrical island. The system identifies the equipment or switches within the region. For each candidate boundary node, the engine assesses the impact of its disconnection on the continuity of power supply to non-faulty areas. The impact is calculated by dividing the total power of the affected loads by the total power of the system load. Simultaneously, the operational complexity is assessed, defined as the number of switches required to perform the isolation operation. Finally, the combination of boundary nodes with the least impact and lowest operational complexity is selected as the isolation execution point. The isolation path instruction includes a sequence of circuit breaker numbers to be disconnected, a sequence of bypass switch numbers to be closed, and a load transfer priority list. The circuit breaker numbers are determined based on unique identifiers in the equipment ledger database. The bypass switch numbers are determined based on the backup path topology. The load transfer priority list is sorted by load importance level, with first-level loads prioritized, second-level loads next, and third-level loads allowed short-term interruptions. After the instruction is generated, a timestamp and digital signature are added, and it is sent to the corresponding intelligent execution unit via an encrypted communication channel.

[0116] Example 6

[0117] In an embodiment of the present invention, the step of sending the optimal fault isolation path instruction to the corresponding intelligent execution unit specifically includes: the intelligent execution unit is composed of a programmable logic controller, a solid-state relay array, and a mechanical disconnect switch drive mechanism; the programmable logic controller has a built-in instruction parsing module, which, after receiving the encrypted instruction, first verifies the legality of the digital signature, and extracts the instruction content after confirming that it is correct; the parsing module converts the circuit breaker number sequence to be disconnected into the corresponding solid-state relay control signal sequence, and converts the bypass switch number sequence to be closed into the corresponding mechanical drive mechanism control instruction sequence; the control signal sequence is output according to a preset timing sequence, first triggering the solid-state relay array to perform a millisecond-level rapid disconnection action, the solid-state relay is composed of silicon carbide power devices, with an on-resistance of less than 5mΩ, a turn-off time of less than 20us, and can withstand the instantaneous impact of 10 times the rated current; after the solid-state relay completes the electrical disconnection... Within 50ms after the disconnection, the mechanical disconnect switch drive mechanism is activated to perform physical isolation. The drive mechanism uses a permanent magnet synchronous motor with a worm gear reducer, achieving a position feedback accuracy of ±0.5° and a drive torque reserve coefficient of not less than 1.5. The drive mechanism has a built-in Hall position sensor that monitors the position of the moving contact of the disconnect switch in real time. When the moving contact reaches the fully disconnected position, an acknowledgment signal is sent to the programmable logic controller (PLC). After receiving all acknowledgment signals, the PLC generates an isolation operation completion report, which includes the operation start time, the completion time of each switch action, the final isolation status, and abnormal alarm information. The report is uploaded to the central monitoring platform through an independent communication channel for maintenance personnel to view in real time. The total time from the issuance of the command to the completion of physical isolation is controlled within 200ms, ensuring that the faulty branch is completely isolated before triggering a system-wide cascading reaction.

[0118] Example 7

[0119] In an embodiment of the present invention, the step of activating the closed-loop feedback evaluation module after completing the fault isolation operation specifically includes: the closed-loop feedback evaluation module is activated immediately after the isolation operation completion report is uploaded. First, it waits for the system to enter a new steady-state operation phase. The steady-state determination criterion is that the bus voltage fluctuation amplitude is less than 0.5% of the rated value and the duration exceeds 5 seconds. After the steady state is established, the module collects four key performance parameters: bus voltage recovery accuracy is defined as the absolute difference between the actual bus voltage and the target voltage divided by the target voltage; load power supply continuity index is defined as the percentage of the total load power that remains powered after isolation relative to the original total load power; standby path temperature rise rate is defined as the average rate of temperature rise of the key nodes of the standby path within 10 seconds after isolation; system total loss increment is defined as the increase in the difference between the total input power and the total output power of the system after isolation relative to the difference before isolation; the module calculates the deviation between the above four parameters and the expected performance index output by the fault prediction model before isolation. The deviation calculation uses a weighted Euclidean distance formula, where the weight coefficient of the bus voltage recovery accuracy is... The weighting coefficients are set to 0.4 for the load power supply continuity index, 0.3 for the backup path temperature rise rate, 0.2 for the backup path temperature rise rate, and 0.1 for the system total loss increment. If the calculated weighted Euclidean distance is greater than the preset tolerance threshold of 0.15, the model parameter update process is triggered. The update process first calculates the difference between the actual and expected values ​​of each performance parameter, and uses the difference as an error signal to propagate back to the output layer of the fault prediction neural network model. The gradient of each weight parameter is calculated layer by layer using the chain rule, with a gradient calculation step size of 0.0001. The weight parameters are fine-tuned according to the gradient direction and magnitude, with the fine-tuning amplitude not exceeding 0.5% of the original value. At the same time, the path optimization algorithm parameters in the fault isolation decision engine are updated, and the load weight coefficient and operation complexity penalty coefficient in the impact degree calculation formula are adjusted so that subsequent decisions are more inclined to choose the isolation scheme with less impact on system performance. After the parameter update is completed, an optimization log is generated to record the detailed content of this adjustment, including the parameter values ​​before adjustment, the parameter values ​​after adjustment, the basis for adjustment, and the expected improvement effect.

[0120] Example 8

[0121] In an embodiment of the present invention, the periodic activation of the preventive maintenance planning module during normal system operation specifically includes: the preventive maintenance planning module automatically starts once every 24 hours, with the start time set to 2:00 AM when the system load is lowest; the module first reads the cumulative runtime of each device from the device ledger database, using 1,000 hours as a scoring unit, with a longer runtime resulting in a lower operational stability score, specifically calculated as 100 minus the cumulative runtime divided by 100,000 and then multiplied by 20; and then reads the fault history of each device from the fault history database in the past... For each fault recorded within the year, 5 points are deducted. The deduction for historical fault frequency is equal to the number of faults multiplied by 5. Temperature, humidity, and dust concentration data of the environment in which each piece of equipment is located are read from the environmental monitoring system. An environmental adaptability score is calculated based on the equipment's environmental adaptability rating table, which categorizes environmental parameters into four levels: excellent, good, average, and poor, corresponding to scores of 100, 80, 60, and 40 respectively. The response time of the most recent maintenance for each piece of equipment is read from the maintenance record database. Full marks are awarded if maintenance is completed within 24 hours of a fault occurring; 1 point is deducted for each day of delay. Timely maintenance response is considered a priority. The availability score is calculated as 100 minus the delay days. The inventory status of critical spare parts for each device is retrieved from the spare parts management system. Full marks are awarded if all critical spare parts are in stock; otherwise, points are deducted based on the estimated delivery time. The spare parts availability score is calculated as 100 minus the estimated delivery days multiplied by 2. The scores from the five dimensions are weighted and summed using weights of 0.3, 0.25, 0.2, 0.15, and 0.1 to obtain the final health status score. Devices are sorted from lowest to highest score to generate an equipment-level maintenance priority ranking table. Emergency maintenance work orders are generated for devices with scores below 60. For devices with a score between 60 and 80, a planned maintenance work order is generated; for devices with a score higher than 80, only the status is recorded and no work order is generated; the work order content includes a maintenance item list, the name and quantity of required spare parts, the suggested execution time window, the safety operation procedure number, and risk warning information; the work order is transmitted to the operation and maintenance management platform through an encrypted communication protocol. The encryption protocol uses the national cryptographic algorithm SM4 for data encryption and SM2 for digital signature; after receiving the work order on the mobile terminal, the operation and maintenance personnel must pass fingerprint biometric authentication to view the detailed operation steps, ensuring the security and traceability of the work order execution process.

[0122] Example 9

[0123] In an embodiment of the present invention, the human-computer interaction interface graphically displays the real-time topology of the DC system, a heatmap of the health status scores of each device, a fault risk prediction curve for the next 24 hours, the current isolation operation execution status, and historical maintenance work order execution records. Specifically, the main view of the interface displays the electrical topology of the DC system, with each device presented as an icon. The icon color dynamically changes according to the health status score: green indicates a score above 90, yellow indicates a score between 70 and 90, and red indicates a score below 70. Clicking on any device icon will pop up a detailed information window, displaying the device's real-time operating parameters, historical fault records, maintenance suggestions, and predicted risk curve. The predicted risk curve has the horizontal axis representing future time and the vertical axis representing the probability of fault occurrence. The curve is updated every 10 minutes by a fault prediction neural network model. The isolation operation execution status is displayed in animation. The interface displays isolated branches highlighted in flashing red, isolated branches marked with gray dashed borders, and bypass power supply paths highlighted with bold blue solid lines. Historical maintenance work order execution records are displayed in a list format, with each record including work order number, equipment name, maintenance content, executor, completion time, and acceptance result. The interface allows maintenance personnel to manually trigger equipment self-test processes, including sensor calibration testing, communication link stress testing, switch operation reliability testing, and insulation resistance retesting. Self-test results are automatically generated and archived. The interface also supports temporary adjustments to isolation policy priorities; maintenance personnel can manually specify the priority level of a certain type of load in isolation decisions. This adjustment is only effective for the current decision cycle and automatically reverts to the default settings in the next cycle. All manual operations record the operator's identity, operation time, and operation content, forming a complete audit log.

[0124] Example 10

[0125] This invention achieves a fundamental transformation in the operation and maintenance mode of DC systems through the coordinated operation of the aforementioned modules. The system can accurately predict risks before a fault actually occurs, automatically generate and execute the optimal isolation plan, and effectively curb the spread of faults. Through a closed-loop feedback mechanism, it continuously optimizes the prediction model and decision-making algorithm, enabling the system to have adaptive evolution capabilities. The preventative maintenance module transforms reactive repair into proactive planning, significantly reducing unplanned downtime. The human-machine interface provides comprehensive visual monitoring and flexible manual intervention capabilities, ensuring the system's controllability under extreme conditions. The overall solution covers the entire lifecycle of a DC system, from daily monitoring, risk warning, fault handling to maintenance planning, significantly improving system availability, security, and operation and maintenance efficiency, meeting the stringent requirements of modern industry for highly reliable DC power supply systems.

[0126] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for full-cycle operation and maintenance support of a DC system. Its features are, include: S1 Multi-dimensional Real-time Operation Status Data Acquisition: Through voltage sensors, current sensors, temperature sensors, insulation monitoring devices, and communication status monitoring modules deployed at key nodes of the DC system, real-time data acquisition of DC bus voltage, branch current, equipment temperature rise, insulation resistance to ground, and communication link status is generated, forming a multi-dimensional real-time operation status data stream covering the entire system topology. S2 Standardized Time Series Dataset Generation: The multi-dimensional real-time running status data stream is synchronized and aligned according to the time series, and data normalization processing is performed according to the preset sampling period to generate a standardized time series dataset containing core features such as voltage fluctuation feature sequence and current harmonic distortion sequence. S3 Fault Risk Probabilistic Assessment: Based on the standardized time-series dataset, a pre-trained fault prediction neural network model is invoked to probabilistically assess the potential fault risks of the DC system within a preset time window in the future, and output multi-time-granularity fault occurrence probability values. S4 Fault Isolation Decision Trigger: When the probability of failure of any critical equipment exceeds a preset threshold, the fault isolation decision engine is triggered. S5 Optimal Isolation Path Execution: The fault isolation decision engine generates an optimal fault isolation path instruction that includes the circuit breaker operation sequence, bypass switch control logic, and load transfer priority. This instruction is then sent to an intelligent execution unit consisting of a programmable logic controller, a solid-state relay array, and a mechanical disconnect switch drive mechanism to perform millisecond-level electrical disconnection and physical isolation operations. S6 Closed-Loop Feedback Deviation Calculation: After fault isolation is completed, the closed-loop feedback evaluation module is started to collect the steady-state operating parameters of the system and calculate the deviation from the expected performance index using the weighted Euclidean distance formula; Online fine-tuning of S7 model and algorithm parameters: Based on the deviation calculation results, the weight parameters of the fault prediction neural network model and the path optimization algorithm parameters of the fault isolation decision engine are fine-tuned online; S8 Preventive Maintenance Work Order Generation: During normal system operation, the preventive maintenance planning module is periodically activated to generate an equipment-level maintenance priority ranking table based on equipment operation data and push encrypted digital work orders.

2. The method according to claim 1, Its features are, In the S3 fault risk probabilistic assessment, the standardized time series dataset includes voltage fluctuation characteristic sequence, current harmonic distortion sequence, temperature rise gradient sequence, insulation degradation trend sequence, and communication packet loss rate sequence.

3. The method according to claim 1, Its features are, In the online fine-tuning of the S7 model and algorithm parameters, the fault prediction neural network model adopts a multi-layer long short-term memory network structure. Its input layer receives the standardized time series dataset, and the hidden layer extracts voltage sag correlation features, current mutation coupling features, temperature rise cumulative effect features, and insulation gradual failure features through a gating mechanism. The output layer outputs the failure probability values ​​of each key device at three time granularities in the future.

4. The method according to claim 1, Its features are, In the execution of the S5 optimal isolation path, the optimal fault isolation path instruction includes the sequence of circuit breaker numbers to be disconnected, the sequence of bypass switch numbers to be closed, and the load transfer priority list.

5. The method according to claim 1, Its features are, In the execution of the S5 optimal isolation path, the intelligent execution unit includes a programmable logic controller, a solid-state relay array, and a mechanical isolation switch drive mechanism; The solid-state relay array is constructed using silicon carbide power devices, with a turn-off time of <20μs. The mechanical disconnect switch drive mechanism uses a permanent magnet synchronous motor in conjunction with a worm gear reducer, achieving a position feedback accuracy of ±0.5 degrees.

6. The method according to claim 1, Its features are, In the S6 closed-loop feedback deviation calculation, the closed-loop feedback evaluation module uses the weighted Euclidean distance formula to calculate the deviation between the steady-state operating parameters and the expected performance indicators; Among them, the weighting coefficient for bus voltage recovery accuracy is 0.4, the weighting coefficient for load power supply continuity index is 0.3, the weighting coefficient for standby path temperature rise rate is 0.2, and the weighting coefficient for total system loss increment is 0.

1.

7. The method according to claim 1, Its features are, In the generation of the S8 preventive maintenance work order, the preventive maintenance planning module generates an equipment-level maintenance priority ranking table based on the equipment's cumulative running time, historical fault records, current health status score, and manufacturer-recommended maintenance cycle, and pushes the digital work order to the operation and maintenance management platform. The digital work orders are transmitted via an encrypted communication protocol, using the national cryptographic algorithm SM4 for encryption and SM2 for signature.

8. A full-cycle operation and maintenance support system for DC systems. The method for full-cycle operation and maintenance of a DC system as described in any one of claims 1 to 7 is adopted. Its features are, It includes: a multi-source state perception layer, a data preprocessing layer, a fault prediction model layer, an isolation decision execution layer, a closed-loop feedback optimization layer, and a preventive maintenance planning layer; The multi-source state perception layer is deployed at key nodes of the DC system and integrates voltage sensors, current sensors, temperature sensors, insulation monitoring devices, and communication state monitoring modules. The multi-source status sensing layer is used to collect DC bus voltage, branch current, equipment temperature rise, insulation resistance to ground value and communication link status data in real time, forming a multi-dimensional real-time operating status data stream covering the entire topology of the system; The data preprocessing layer interacts with the multi-source state perception layer. The data preprocessing layer is used to perform time series synchronization and alignment operations on multi-dimensional real-time running status data streams, and to complete data normalization processing according to a preset sampling period to generate a standardized time series dataset containing core features such as voltage fluctuation feature sequences and current harmonic distortion sequences. The fault prediction model layer calls a pre-trained fault prediction neural network model, and based on the standardized time series dataset, performs a probabilistic assessment of the potential fault risks of the DC system within a future preset time window, and outputs the probability values ​​of key equipment faults at multiple time granularities. The isolation decision execution layer includes a fault isolation decision engine and an intelligent execution unit; The isolation decision execution layer is used to trigger the fault isolation decision engine to generate the optimal fault isolation path instruction when the failure probability of any critical equipment exceeds a preset threshold, and send it to the intelligent execution unit to perform millisecond-level electrical disconnection and physical isolation operations. After the closed-loop feedback optimization layer fault isolation operation is completed, the steady-state operating parameters of the system are collected, the deviation from the expected performance index is calculated using the weighted Euclidean distance formula, and the weight parameters of the fault prediction neural network model and the path optimization algorithm parameters of the fault isolation decision engine are fine-tuned online based on the deviation results. During normal system operation, the preventive maintenance planning layer periodically activates the equipment status assessment logic. Based on the equipment's cumulative runtime, historical fault records, current health status score, and manufacturer-recommended maintenance cycle, it generates an equipment-level maintenance priority ranking table and pushes a digital work order encrypted with national cryptographic algorithms to the operation and maintenance management platform.

9. The system according to claim 8, Its features are, The system also includes a human-machine interface for graphically displaying the real-time topology of the DC system, a heatmap of equipment health status scores, a fault risk prediction curve, the execution status of isolation operations, and historical maintenance work order records.

10. The system according to claim 8, Its features are, The insulation monitoring device adopts the unbalanced bridge method principle, with a sampling resistor value of 100kΩ, an accuracy class of 0.1, and a voltage measurement resolution of 1mV. The optimal fault isolation path instructions include: circuit breaker operation sequence, bypass switch control logic, and load transfer priority.

Images