Photovoltaic power station intelligent remote operation and maintenance method and system based on relay protection

Abstract

The invention discloses a photovoltaic power station intelligent remote operation and maintenance method and system based on relay protection, and the method comprises the steps: collecting the multi-mode characteristic quantity of a photovoltaic array region relay protection device in real time through a distributed optical fiber sensing network, so as to construct a dynamic protection characteristic matrix; performing feature decoupling on the dynamic protection feature matrix by adopting a space-time convolution adversarial encoder to generate a fault feature vector set; inputting the fault feature vector set into a physically constrained LSTM-GAN hybrid prediction model, and predicting a multi-dimensional failure probability tensor of the key relay protection node in a future preset duration; generating an optimal operation and maintenance strategy decision tree based on the multi-dimensional failure probability tensor; and analyzing the operation and maintenance strategy decision tree through a remote execution gateway driven by the digital twinborn, and synchronously updating the topological connection relationship of the digital twinborn bodies of the power station. According to the embodiment of the invention, the operation and maintenance efficiency and reliability of photovoltaic power station protection can be improved.

Description

Technical Field

[0001] The present invention belongs to the field of operation and maintenance technology, and in particular to an intelligent remote operation and maintenance method and system for a photovoltaic power station based on relay protection. Background Art

[0002] With the continuous expansion of photovoltaic power plants and the increasing complexity of grid connection, traditional relay protection operation and maintenance models face severe challenges. Existing methods mainly rely on regular inspections and threshold alarm mechanisms, which make it difficult to capture multimodal fault characteristics (such as high-frequency arc oscillations and insulation degradation) in the photovoltaic array area in real time. Fault prediction is mostly based on the judgment of a single physical quantity threshold, lacking comprehensive consideration of the grid topology constraints and the coupling relationship between multiple physical fields. In addition, the formulation of operation and maintenance strategies often considers equipment status in isolation, without taking into account multi-dimensional factors such as maintenance costs, power generation losses, and the risk of cascading failures, resulting in delayed responses or excessive maintenance. Although some systems have introduced remote monitoring technology, data collection accuracy is insufficient, feature decoupling is not thorough, and there is a lack of dynamic synchronization between strategy execution and digital twins, which restricts operation and maintenance efficiency and reliability. Summary of the Invention

[0003] The purpose of the present invention is to provide a method and system for intelligent remote operation and maintenance of photovoltaic power plants based on relay protection, so as to solve the deficiencies in the existing technology and improve the operation and maintenance efficiency and reliability of photovoltaic power plant protection.

[0004] An embodiment of the present application provides a method for intelligent remote operation and maintenance of a photovoltaic power station based on relay protection, the method comprising:

[0005] The multimodal characteristics of the photovoltaic array area relay protection device are collected in real time through a distributed optical fiber sensor network to construct a dynamic protection characteristic matrix, wherein the multimodal characteristics include transient impedance spectrum, fault arc high-frequency oscillation waveform and insulation degradation gradient value;

[0006] A spatiotemporal convolutional adversarial encoder is used to perform feature decoupling on the dynamic protection feature matrix, separating three orthogonal feature subsets: inherent equipment aging features, environmental disturbance features, and latent fault features, and generating a set of fault feature vectors.

[0007] The fault feature vector set is input into the physically constrained LSTM-GAN hybrid prediction model, and the multi-dimensional failure probability tensor of the key relay protection nodes within a preset time period in the future is predicted based on the grid topology constraints and the multi-physical field coupling relationship;

[0008] Based on the multi-dimensional failure probability tensor, a three-dimensional optimization space is constructed that includes equipment maintenance costs, power generation loss risks, and failure chain reaction suppression factors. A multi-agent reinforcement learning algorithm is used to conduct a strategic competition game in a virtual power plant simulation environment to generate an optimal operation and maintenance strategy decision tree.

[0009] The operation and maintenance strategy decision tree is parsed through a remote execution gateway driven by the digital twin, automatically reconstructing the relay protection setting group, switching the fault isolation path, activating the virtual inertia support function of the energy storage system, and synchronously updating the topological connection relationship of the power station digital twin.

[0010] Optionally, the multimodal characteristic quantities of the photovoltaic array area relay protection device are collected in real time through the distributed optical fiber sensing network to construct a dynamic protection characteristic matrix, wherein the multimodal characteristic quantities include transient impedance spectrum, fault arc high-frequency oscillation waveform and insulation degradation gradient value, including:

[0011] The wavelet packet energy spectrum is decomposed based on the optical fiber vibration sensor signal to extract the high-frequency oscillation waveform fragments of the fault arc in the frequency range of 10kHz-1MHz to generate the original arc oscillation sequence;

[0012] Based on the original arc oscillation sequence and combined with the temperature field distribution of the photovoltaic panel, the impedance phase correction is performed, the real part gradient change rate of the transient impedance spectrum is calculated through the frequency domain differential operator, and the impedance spectrum gradient vector is output;

[0013] The dielectric response model of the insulation layer is driven by the impedance spectrum gradient vector, and the electric field intensity attenuation inversion algorithm is used to calculate the aging rate of the insulation material per square meter to generate the insulation degradation gradient tensor;

[0014] The arc oscillation original sequence, impedance spectrum gradient vector, and insulation degradation gradient tensor are spliced ​​into a three-dimensional feature cube according to the time-space alignment rule, and normalized through a sliding time window to output a dynamic protection feature matrix.

[0015] Optionally, the spatiotemporal convolution adversarial encoder is used to perform feature decoupling on the dynamic protection feature matrix, separating three orthogonal feature subsets of equipment inherent aging features, environmental disturbance features, and latent fault features, and generating a set of fault feature vectors, including:

[0016] The dynamic protection feature matrix is ​​input into the spatiotemporal convolution encoder, and the spatial dependency across time steps is extracted using the 3D hole convolution kernel to generate a multi-scale spatiotemporal feature map.

[0017] Taking the multi-scale spatiotemporal feature map as input, the adversarial decoder and encoder play a Min-Max game, and use the gradient reversal layer to separate the low-frequency component map of the device's inherent aging characteristics;

[0018] Apply an environmental noise mask to the low-frequency component spectrum, use the spectral clustering algorithm to identify the feature clusters affected by temperature / humidity disturbances, and output the environmental disturbance feature subset;

[0019] The low-frequency component map and the environmental disturbance feature subset are subtracted from the multi-scale spatiotemporal feature map, and the unrepresented abnormal patterns are extracted through the residual orthogonalization layer to generate a set of hidden fault feature vectors.

[0020] Optionally, the fault feature vector set is input into a physically constrained LSTM-GAN hybrid prediction model, and a multi-dimensional failure probability tensor of key relay protection nodes within a preset time period in the future is predicted based on the grid topology constraints and the multi-physical field coupling relationship, including:

[0021] The fault feature vector set is mapped to the nodes of the power grid topology graph, and the status information of adjacent protection devices is aggregated based on the graph convolutional network to generate a topology constraint feature vector.

[0022] The LSTM-GAN generator is initialized with the topological constraint feature vector, and the field strength-current constraint operator is constructed by combining the electromagnetic-thermal multi-physics field coupling equation to output the physically compliant potential state;

[0023] Input the potential state of physical compliance into the Monte Carlo failure simulator, iteratively calculate the event trigger conditions including short-circuit current exceeding the standard and insulation breakdown within a preset time period, and generate the failure event probability distribution;

[0024] According to the probability distribution of failure events, a Markov state transition matrix is ​​constructed along the time dimension. Multi-dimensional risk factors are integrated through tensor contraction operations to output the initial failure probability tensor.

[0025] The Bayesian correction of historical equipment failure data is applied to the initial failure probability tensor, and the KL divergence is used to constrain the distribution shift to generate an optimized multi-dimensional failure probability tensor.

[0026] Optionally, based on the multi-dimensional failure probability tensor, a three-dimensional optimization space including equipment maintenance cost, power generation loss risk, and failure chain reaction suppression factor is constructed, and a strategic competition game is conducted in a virtual power plant simulation environment through a multi-agent reinforcement learning algorithm to generate an optimal operation and maintenance strategy decision tree, including:

[0027] Projecting the multi-dimensional failure probability tensor into a three-dimensional optimization space to generate a risk cost hypersurface, wherein the three-dimensional optimization space includes an X-axis: maintenance cost weight, a Y-axis: power generation loss risk value, and a Z-axis: chain reaction inhibition factor;

[0028] Deploy three types of intelligent agents: a maintenance robot, a dispatch controller, and a fault isolator in a virtual power plant simulation environment. These agents are given reward functions to optimize X / Y / Z-axis targets, initiating multi-agent strategy competition.

[0029] The Nash equilibrium solver is used to coordinate the agent strategy, and a deep Q-network is used to explore the combination of maintenance path switching, fixed value adjustment, and virtual inertia activation, and the candidate strategy evaluation matrix is ​​output.

[0030] Based on the Pareto frontier solution of the candidate strategy evaluation matrix, a decision branch rule for cost-benefit trade-off is constructed to generate an operation and maintenance strategy decision tree with confidence score.

[0031] Optionally, the remote execution gateway driven by the digital twin parses the operation and maintenance strategy decision tree, automatically reconstructs the relay protection setting group, switches the fault isolation path, activates the virtual inertia support function of the energy storage system, and synchronously updates the topological connection relationship of the power station digital twin, including:

[0032] According to the leaf node instructions of the operation and maintenance strategy decision tree, a dynamic fixed value parameter package containing overcurrent threshold and action delay is injected into the relay protection device through the safety tunnel, and the reconstruction status code fed back by the device is collected in real time;

[0033] The fault path switching engine is activated based on the reconstructed status code, and the optimal disconnector combination is calculated based on the real-time topology of the power grid. A topology switching instruction sequence is sent to the circuit breaker, and the switch position feedback matrix is ​​obtained synchronously.

[0034] The energy storage virtual inertia controller is triggered according to the switch position feedback matrix, the droop coefficient and virtual impedance parameters are dynamically adjusted, the inertia response eigenvector is output to the energy storage converter PCS, and an inertia activation verification waveform is generated;

[0035] The reconstruction status code, switch position feedback matrix, and inertia activation verification waveform are integrated and reconstructed, and the engine is driven by topology differences to update the power plant digital twin, outputting a synchronously verified twin topology connection diagram that includes relay protection settings, fault isolation paths, and inertia status.

[0036] Another embodiment of the present application provides a photovoltaic power station intelligent remote operation and maintenance system based on relay protection, the system comprising:

[0037] An acquisition module is used to collect multimodal characteristic quantities of the relay protection device in the photovoltaic array area in real time through a distributed optical fiber sensing network to construct a dynamic protection characteristic matrix, wherein the multimodal characteristic quantities include transient impedance spectrum, fault arc high-frequency oscillation waveform and insulation degradation gradient value;

[0038] a separation module, configured to perform feature decoupling on the dynamic protection feature matrix using a spatiotemporal convolution adversarial encoder, separate three orthogonal feature subsets, namely, inherent equipment aging features, environmental disturbance features, and latent fault features, and generate a set of fault feature vectors;

[0039] A prediction module is used to input the set of fault feature vectors into a physically constrained LSTM-GAN hybrid prediction model, and predict the multi-dimensional failure probability tensor of key relay protection nodes within a preset time period in the future based on the grid topology constraints and the coupling relationship between multiple physical fields;

[0040] A generation module is used to construct a three-dimensional optimization space including equipment maintenance cost, power generation loss risk, and failure chain reaction suppression factor based on the multi-dimensional failure probability tensor, and to generate an optimal operation and maintenance strategy decision tree by conducting a strategic competition game in a virtual power plant simulation environment through a multi-agent reinforcement learning algorithm;

[0041] The reconstruction module is used to parse the operation and maintenance strategy decision tree through the remote execution gateway driven by the digital twin, automatically reconstruct the relay protection setting group, switch the fault isolation path, activate the virtual inertia support function of the energy storage system, and synchronously update the topological connection relationship of the power station digital twin.

[0042] Yet another embodiment of the present application provides a storage medium, wherein the storage medium stores a computer program, wherein the computer program is configured to execute any of the above methods when run.

[0043] Yet another embodiment of the present application provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute any of the above methods.

[0044] Compared with the existing technology, the present invention provides an intelligent remote operation and maintenance method for photovoltaic power stations based on relay protection. The method collects the multimodal feature quantities of the relay protection devices in the photovoltaic array area in real time through a distributed optical fiber sensor network to construct a dynamic protection feature matrix; a spatiotemporal convolution adversarial encoder is used to decouple the dynamic protection feature matrix to generate a set of fault feature vectors; the set of fault feature vectors is input into a physically constrained LSTM-GAN hybrid prediction model to predict the multidimensional failure probability tensor of key relay protection nodes within a preset time in the future; based on the multidimensional failure probability tensor, an optimal operation and maintenance strategy decision tree is generated; the operation and maintenance strategy decision tree is parsed by a remote execution gateway driven by a digital twin, and the topological connection relationship of the power station digital twin is synchronously updated, thereby improving the operation and maintenance efficiency and reliability of photovoltaic power station protection. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A hardware structure block diagram of a computer terminal for an intelligent remote operation and maintenance method for a photovoltaic power station based on relay protection provided by an embodiment of the present invention;

[0046] Figure 2 A schematic diagram of a flow chart of a method for intelligent remote operation and maintenance of a photovoltaic power station based on relay protection provided by an embodiment of the present invention;

[0047] Figure 3 A schematic structural diagram of an intelligent remote operation and maintenance system for a photovoltaic power station based on relay protection provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0048] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limiting the present invention.

[0049] The embodiment of the present invention first provides a method for intelligent remote operation and maintenance of a photovoltaic power station based on relay protection. The method can be applied to electronic devices such as computer terminals, specifically ordinary computers.

[0050] The following describes it in detail by taking running on a computer terminal as an example. Figure 1 The hardware structure block diagram of a computer terminal for a photovoltaic power station intelligent remote operation and maintenance method based on relay protection provided by an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, a memory, and a network interface connected via a system bus, wherein the memory may include a non-volatile storage medium and an internal memory.

[0051] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions that, when executed, enable a processor to execute any one of the intelligent remote operation and maintenance methods for photovoltaic power plants based on relay protection.

[0052] The processor is used to provide computing and control capabilities and support the operation of the entire computer equipment.

[0053] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor, the processor can execute any intelligent remote operation and maintenance method of the photovoltaic power station based on relay protection.

[0054] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art will understand that Figure 1 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0055] It should be understood that the processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0056] See also Figure 2 The embodiment of the present invention provides a method for intelligent remote operation and maintenance of a photovoltaic power station based on relay protection, which may include the following steps:

[0057] S201, collecting multimodal feature quantities of relay protection devices in a photovoltaic array area in real time through a distributed optical fiber sensor network to construct a dynamic protection feature matrix, wherein the multimodal feature quantities include a transient impedance spectrum, a fault arc high-frequency oscillation waveform, and an insulation degradation gradient value;

[0058] Specifically, the wavelet packet energy spectrum can be decomposed according to the optical fiber vibration sensing signal to extract the high-frequency oscillation waveform fragments of the fault arc in the frequency range of 10kHz-1MHz to generate the original arc oscillation sequence;

[0059] ‌Fiber optic vibration signal acquisition and preprocessing‌

[0060] Distributed fiber optic vibration sensors (such as those based on Φ-OTDR phase-sensitive optical time-domain reflectometry) are installed along the cable trenches in the photovoltaic array area, capturing abnormal cable vibration signals at a sampling rate of 2 million times per second (the sampling rate is the number of data points collected per second). When a ground fault occurs on the photovoltaic DC side (such as a short circuit caused by water ingress to the module junction box), mechanical vibration waves caused by arc discharge are transmitted through the cable armor to the optical fiber, generating micro-vibration signals with frequencies ranging from 10 kHz to 1 megahertz (MHz). For example, a vibration pulse with a frequency of 85 kHz and a duration of 3 milliseconds (ms) is detected in a certain cable section at timestamp T1, triggering a fault warning.

[0061] Application of wavelet packet energy spectrum decomposition technology:

[0062] The Db10 wavelet basis function (a discrete wavelet with 10th-order vanishing moment) is used to perform an 8-layer decomposition of the original vibration signal, dividing the full frequency band (0-1 MHz) into 256 sub-bands of equal width.

[0063] Calculate the energy proportion of each sub-band: When the energy of the 120th sub-band (corresponding to 85-90kHz) suddenly increases to 15% of the full-band energy (normal background noise is only 0.3%), it is determined that a fault arc feature exists.

[0064] Waveform segment extraction and sequence generation:

[0065] The vibration waveform of the ±5 millisecond time window (total length 10 milliseconds) is intercepted with the fault point as the center, and the standardized waveform segment is output after filtering with the Hanning window (a smooth window function).

[0066] Continuously record all fault fragments within 24 hours (e.g., 38 in total) and sort them by timestamp to generate the original arc oscillation sequence (data structure: timestamp-frequency-amplitude).

[0067] Based on the original arc oscillation sequence and combined with the temperature field distribution of the photovoltaic panel, the impedance phase correction is performed, the real part gradient change rate of the transient impedance spectrum is calculated through the frequency domain differential operator, and the impedance spectrum gradient vector is output;

[0068] Correction of the Effect of Temperature Field Distribution on Impedance Measurement

[0069] The temperature field of the photovoltaic panel is scanned in real time by an infrared thermal imager (spatial resolution 2 cm) to establish a temperature-impedance compensation model:

[0070] When the component temperature rises from 25°C to 65°C, the real part of the DC impedance (resistance component) decreases by about 12% (temperature coefficient -0.4% / °C).

[0071] Example: Arc oscillation is detected in a string at time T1, and its surface temperature gradient is 52°C (center) - 45°C (edge). Impedance compensation needs to be performed in different zones.

[0072] Impedance Phase Correction Logic: The impedance phase angle (unit: radian) in the original arc oscillation sequence is deducted from the temperature drift. For example, if the measured phase angle θ = 0.75 radians at a given moment, corresponding to a zone temperature of 48°C (reference temperature: 25°C), the corrected phase angle θ' = θ - (48 - 25) × 0.0023 (temperature drift coefficient) = 0.70 radians.

[0073] Transient Impedance Spectrum Gradient Calculation

[0074] Frequency domain differential operator (a method for calculating the rate of change in the frequency domain) execution process:

[0075] The corrected impedance spectrum is subjected to fast Fourier transform (FFT, a method for converting time domain signals to frequency domain) to obtain the complex impedance spectrum Z(f) = R(f) + jX(f), where the real part R(f) represents the resistance component.

[0076] In the 10kHz-1MHz frequency band, the real part difference of adjacent frequency points is calculated with a step size of 1kHz: ΔR = R( ) - R( ).

[0077] Example: At 85kHz, R(85kHz)=0.18Ω, R(84kHz)=0.22Ω → ΔR = -0.04Ω / kHz.

[0078] Gradient vector generation: Arrange the ΔR values ​​of 981 frequency points (from 10 kHz to 1 MHz in 1 kHz increments) in ascending frequency order to construct an impedance spectrum gradient vector of dimension 981 (e.g., [-0.02, -0.03, ..., +0.01] Ω / kHz).

[0079] Key diagnostic indicator: If the gradient value of 5 consecutive frequency points is less than -0.05Ω / kHz (a sharp drop), it indicates dendritic discharge in the insulation layer.

[0080] The dielectric response model of the insulation layer is driven by the impedance spectrum gradient vector, and the electric field intensity attenuation inversion algorithm is used to calculate the aging rate of the insulation material per square meter to generate the insulation degradation gradient tensor;

[0081] ‌Dielectric response model construction and parameter mapping‌

[0082] The dielectric properties of the photovoltaic cable insulation layer (XLPE cross-linked polyethylene) are described by frequency domain dielectric spectroscopy (FDS):

[0083] A database of complex dielectric constant ε*(f) =ε'(f) - jε''(f) is established, where the real part ε' represents the energy storage capacity and the imaginary part ε'' represents the loss.

[0084] The ΔR value in the impedance gradient vector is reversely mapped to the dielectric loss tangent value tanδ: when ΔR <-0.03Ω / kHz, the increase in tanδ is ≥0.015 (for example, from 0.02 to 0.035).

[0085] Electric field intensity attenuation inversion algorithm (a method for estimating material aging by electric field changes):

[0086] Based on Maxwell's electromagnetic field theory, the electric field strength E and the dielectric constant in the insulating layer satisfy ▽·(ε▽φ) = 0 (φ is the electric potential).

[0087] The local electric field attenuation rate is inferred based on the change in impedance gradient: if ΔR = -0.05Ω / kHz in a certain section, the corresponding annual attenuation rate of the electric field strength is 8.5% (normal aging rate <2%).

[0088] ‌Aging rate quantization and tensor generation‌

[0089] Spatial discretization: Divide the PV array area into 1m x 1m grid cells (e.g., a 100kW array is divided into 200 cells).

[0090] Each unit is associated with a dedicated impedance spectrum gradient vector (such as the gradient vector of unit G07 ).

[0091] Aging rate calculation: Input the unit gradient vector to the dielectric response model and output the dielectric loss increment Δtanδ (e.g. +0.02). Calibrated by accelerated aging testing: every 0.01 increase in Δtanδ is equivalent to 120 days of aging.

[0092] Example: Δtanδ of element G07 = 0.025 → aging rate = 120×(0.025 / 0.01) = 300 days / year (normal value 365 days / year).

[0093] ‌Degenerate gradient tensor (3D data structure) construction‌:

[0094] Dimension 1: grid cell coordinates (such as X / Y axis index); Dimension 2: aging rate value (unit: equivalent aging days / year); Dimension 3: confidence level (calculated based on the signal-to-noise ratio, >80% is highly reliable).

[0095] Output example: The tensor element T(7,12) = [rate=300, confidence=92%] indicates that the annual aging rate of the cell in row 7 and column 12 is equivalent to 300 days.

[0096] The arc oscillation original sequence, impedance spectrum gradient vector, and insulation degradation gradient tensor are spliced ​​into a three-dimensional feature cube according to the time-space alignment rule, and normalized through a sliding time window to output a dynamic protection feature matrix.

[0097] Multi-source data spatiotemporal alignment mechanism

[0098] Time alignment: The time when the fault event is triggered is used as the reference point (such as the starting point of arc oscillation), align the three types of data to ±50 ms time window.

[0099] Example: The timestamp of an arc oscillation segment is -2ms to +8ms, impedance spectrum calculation timestamp T0+1ms, degradation tensor update timestamp +3ms → Reinterpolate for the origin.

[0100] Spatial Mapping: Arc oscillation sequences are associated with specific string numbers (e.g., String-05), impedance spectrum gradient vectors are associated with corresponding combiner boxes (e.g., Combiner-3), and degradation tensors are associated with grid cells (e.g., G12). → Create a spatial correlation matrix using the plant topology.

[0101] ‌3D Feature Cube Construction‌

[0102] ‌Data structure design‌:

[0103] Dimension 1 (time axis): 1001 time points at 0.1 ms intervals (covering ±50 ms);

[0104] Dimension 2 (characteristic axis): contains three layers: Layer 1: arc oscillation amplitude (unit: microstrain με, reflecting vibration intensity); Layer 2: impedance real part gradient (unit: ohm / kilohertz, Ω / kHz); Layer 3: grid unit aging rate (unit: equivalent aging days / year);

[0105] Dimension 3 (spatial axis): Encoded according to the string-combiner box-grid unit mapping relationship (e.g., String05-Comb3-G12).

[0106] ‌Data filling rules‌:

[0107] Time point t= At -2ms: Layer 1 is filled with an arc amplitude value of 1.8με; Layer 2 is filled with an impedance gradient of -0.04Ω / kHz; Layer 3 is filled with a unit G12 aging rate of 300 days / year.

[0108] Sliding Window Normalization

[0109] Dynamic Normalization Window: Normalizes data within a window of ±10 milliseconds (201 sampling points in total), centered at the current time point.

[0110] Calculate the mean μ and standard deviation σ of the data in the window and perform normalization: (x -μ) / σ.

[0111] Example: Arc amplitude window μ = 0.5 με, σ = 0.2 → the original value 1.8 με is normalized to (1.8-0.5) / 0.2=6.5.

[0112] ‌Feature Matrix Generation‌:

[0113] The normalized three-dimensional cube is expanded along the spatial axis, and each spatial unit is converted into a 1001×3 matrix (1001 time points × 3 features).

[0114] Output dynamic protection feature matrix: total dimension = number of spatial units × 1001 × 3 (e.g. 200 units → 200 × 1001 × 3 matrix).

[0115] Matrix annotation space-time label: fault number F202405001 occurred at location String05, time 2024-05-01 14:23:05.876.

[0116] This method utilizes a distributed fiber-optic sensor network to comprehensively monitor the relay protection devices in photovoltaic power plants. By collecting multiple characteristic parameters reflecting the operating status of the equipment, including transient impedance spectra reflecting electrical characteristics, high-frequency arc waveforms characterizing faults, and insulation degradation gradients reflecting the degree of equipment aging, a dynamically updated protection characteristic matrix is ​​constructed. This multi-dimensional data acquisition method overcomes the limitations of traditional single-electrochemical quantity monitoring, achieving comprehensive awareness of the operating status of photovoltaic power plant relay protection devices and providing a rich data foundation for subsequent fault diagnosis and prediction. Through the collaborative analysis of multimodal characteristic quantities, early signs of potential equipment failures can be more accurately captured, significantly improving the sensitivity and reliability of fault detection.

[0117] S202, using a spatiotemporal convolutional adversarial encoder to perform feature decoupling on the dynamic protection feature matrix, separating three orthogonal feature subsets: inherent equipment aging features, environmental disturbance features, and latent fault features, and generating a set of fault feature vectors;

[0118] Specifically, the dynamic protection feature matrix can be input into the spatiotemporal convolution encoder, and the spatial dependency across time steps can be extracted using the 3D hole convolution kernel to generate a multi-scale spatiotemporal feature map.

[0119] ‌Hardware Acceleration Architecture for Spatiotemporal Convolutional Encoders‌

[0120] The dynamic protection feature matrix (example dimensions: 200 spatial units × 1001 time points × 3 features) is processed in parallel using an FPGA (field programmable gate array, a hardware acceleration chip). The encoder uses a three-level convolutional structure:

[0121] The first level uses a 3×3×3 3D convolution kernel (3 time steps × 3 spatial neighborhoods × 3 feature channels) and slides with a step size of 1. For example, at time point T0, the convolution kernel covers -1ms to +1ms time window, and simultaneously associate three adjacent spatial units (such as the fault point unit G12 and its left and right units G11 and G13).

[0122] The second level enables Dilated Convolution and sets the dilation rate (i.e., the interval between convolution kernel elements) to 2. This means that the convolution kernel operates once every 2 sample points on the time axis, extending the coverage of a single convolution to ±5 milliseconds (for example, from -5ms to +5ms), significantly enhancing the ability to capture long-term dependencies.

[0123] Multi-scale feature fusion mechanism: The first level outputs high-resolution detail features (e.g., capturing arc oscillation mutations within 0.1 milliseconds), while the second level outputs macro-trend features (e.g., impedance gradients within 5 milliseconds). Skip connections are used to concatenate these two levels of features in the channel dimension to generate a fused spatiotemporal feature map (example dimensions: 200 × 100 × 1 × 64, where 64 represents the number of feature channels).

[0124] ‌Extraction logic for cross-space dependencies‌

[0125] Spatial dependency extraction targets the electrical topology of a photovoltaic array. For example, a combiner box (Combiner-03) manages 20 strings (String01-20). The system assigns each string an independent spatial unit code. When the convolution kernel spans adjacent units in the spatial dimension, it automatically associates the strings with the closest electrical distance:

[0126] If an abnormality occurs in unit G12 (corresponding to String12), the convolution kernel simultaneously collects data from its upstream unit G11 (String11) and downstream unit G13 (String13) to identify whether there are signs of fault propagation.

[0127] Typical case analysis:

[0128] When it is detected that the arc amplitude of unit G12 suddenly increases to 6.5 (normalized value) at time T0, and the adjacent unit G11 Similar fluctuations (amplitude 5.8) appear at +0.3ms, and spatiotemporal convolution will generate a strong spatial correlation feature vector (e.g., spatial dependence coefficient 0.92), indicating the risk of fault propagation along the cable.

[0129] Taking the multi-scale spatiotemporal feature map as input, the adversarial decoder and encoder play a Min-Max game, and use the gradient reversal layer to separate the low-frequency component map of the device's inherent aging characteristics;

[0130] ‌Game-theoretic architecture design for adversarial decoders‌

[0131] ‌Min-Max Game Mechanism‌:

[0132] Encoder goal: Generate mixed features whose sources cannot be distinguished by the decoder (confusing aging and environmental features).

[0133] Decoder goal: accurately separate and reconstruct the aging feature components in the original input matrix.

[0134] The two are alternately optimized through adversarial training to form a game balance.

[0135] The core function of the Gradient Reversal Layer is:

[0136] This layer multiplies the encoder gradient by a negative coefficient (e.g., -0.5) during backpropagation, forcing the encoder to learn feature representations that are opposite to the decoder's objectives. For example, when the decoder attempts to enhance aging features, the reversed gradient forces the encoder to weaken the representation of such features.

[0137] ‌Separation process of low-frequency component spectrum‌

[0138] Intrinsic equipment aging (such as cable insulation degradation) manifests as a low-frequency signal (typically <1kHz). The adversarial decoder extracts the low-frequency component through a three-step operation:

[0139] Frequency domain filtering: Perform a fast Fourier transform (FFT) on the spatiotemporal feature map to retain the 0-1kHz frequency band components (for example, filter out arc high-frequency noise > 1kHz).

[0140] ‌Inverse transform reconstruction‌: Convert the filtered frequency domain signal back to the time domain to generate a preliminary low-frequency spectrum (dimension 200×1001×16).

[0141] Feature decorrelation: By using an orthogonal constraint loss function, the correlation between the low-frequency spectrum and the environmental disturbance features is ensured to be less than 0.1 (close to complete independence).

[0142] An example of an industrial field case study: After three years of operation, a photovoltaic power plant experienced aging of its cable insulation, causing the real part of its impedance to continuously decrease at a frequency of 0.5 kHz. The adversarial decoder isolated the monotonically decreasing low-frequency curve (with an average annual decrease rate of 8.5%) from the mixed features, while the high-frequency arc oscillations were relegated to other feature subsets.

[0143] Apply an environmental noise mask to the low-frequency component spectrum, use the spectral clustering algorithm to identify the feature clusters affected by temperature / humidity disturbances, and output the environmental disturbance feature subset;

[0144] ‌Generation and Application of Ambient Noise Mask‌

[0145] Noise Source Modeling:

[0146] Temperature disturbance: The daily fluctuation range of the photovoltaic panel surface temperature is -10°C to +65°C. The temperature distribution matrix (with the same dimension as the spatial unit) is constructed using the measured data of the infrared thermal imager.

[0147] Humidity disturbance: After a rainstorm, relative humidity (RH) surged from 30% to 85%, and a network of humidity sensors recorded the spatiotemporal variations.

[0148] ‌Mask construction logic‌:

[0149] Calculate the correlation coefficient between temperature / humidity and low-frequency characteristics: if a 1°C temperature increase in a certain area causes an impedance change of 0.05Ω, it is marked as a temperature-sensitive area.

[0150] Generate a binary mask matrix: assign 1 to temperature-sensitive cells and 0 to the rest (for example, 87 temperature-sensitive cells are marked out of 200 cells).

[0151] The mask matrix is ​​multiplied element-by-element with the low-frequency component map to shield non-environmental disturbance signals (such as slow changes caused by aging).

[0152] ‌Environmental Feature Separation by Spectral Clustering‌

[0153] Similarity Matrix Construction: Based on the masked low-frequency spectrum, the similarity of the environmental responses between spatial units is calculated. For example, if units G12 and G15 both show an impedance increase of 0.2Ω when subjected to a temperature change, their similarity is set to 0.95 (maximum 1.0).

[0154] Clustering Process: Construct a similarity matrix S (200×200 dimensions), where the matrix element S(i,j) represents the correlation between the environmental responses of cells i and j. Calculate the Laplacian matrix and perform eigendecomposition. Perform K-means clustering on the first three eigenvectors.

[0155] Output cluster labels: such as "high temperature sensitive cluster" (including 32 units such as G07 and G12) and "high humidity sensitive cluster" (including 41 units such as G18 and G22).

[0156] Typical perturbation pattern extraction:

[0157] Characteristics of high-temperature sensitive clusters: For every 10°C increase in temperature, the real part of the impedance decreases by 0.15Ω (linear negative correlation).

[0158] High humidity sensitive cluster characteristics: When the humidity is greater than 70%, the impedance phase angle shift increases by 0.1 radians (non-linear jump).

[0159] These patterns are encoded as a subset of environmental perturbation features (data structure: cluster label-perturbation type-quantization coefficient).

[0160] The low-frequency component map and the environmental disturbance feature subset are subtracted from the multi-scale spatiotemporal feature map, and the unrepresented abnormal patterns are extracted through the residual orthogonalization layer to generate a set of hidden fault feature vectors.

[0161] ‌Residual signal calculation and enhancement‌

[0162] Feature culling operation: Subtract the low-frequency component map (200×1001×16) and the environmental disturbance feature subset (200×1001×16) from the multi-scale spatiotemporal feature map (200×1001×64) to obtain a preliminary residual feature map (dimension 200×1001×32).

[0163] Noise Suppression Technology:

[0164] Perform wavelet threshold denoising on the residual signal:

[0165] Performs a 5-layer wavelet decomposition using the Sym8 wavelet basis functions.

[0166] Soft thresholding is applied to the high-frequency coefficients (detail components), and the threshold is set to twice the noise standard deviation σ (e.g. σ = 0.08, threshold = 0.16).

[0167] After reconstruction, the denoised residual image is output (the signal-to-noise ratio is improved by more than 12dB).

[0168] Anomaly detection mechanism of residual orthogonalization layer

[0169] Orthogonalization constraint implementation: Through the Gram-Schmidt orthogonalization process, the residual features are forced to be linearly independent of the extracted aging and environmental features (inner product < For example, the cosine similarity between a residual feature vector and the temperature perturbation feature is originally 0.25, but it drops to 0.002 after orthogonalization.

[0170] Hidden fault mode identification:

[0171] Transient event detection: Identify microsecond pulses lasting 0.1-2 milliseconds (such as precursors to partial discharge in insulation), with an amplitude threshold set at 6 times the background noise (e.g., amplitude > 0.48 με).

[0172] Frequency Domain Anomaly Location: Scans for energy surge points in the 100kHz-500kHz frequency range. If the energy at a frequency point exceeds 5 standard deviations from the mean (for example, if the energy at 85kHz surges to 300 times the normal value), it is determined to be a hidden arc fault.

[0173] Case description: A 185kHz oscillation pulse with a duration of 0.8 milliseconds and an amplitude of 0.52με was detected in the residual signal of a power station. After orthogonalization, a hidden fault vector was generated:

[0174] {

[0175] Fault type: "Partial discharge",

[0176] Location: Unit G09,

[0177] Frequency characteristics: 185±5kHz,

[0178] Confidence level: 94%

[0179] }.

[0180] Feature vector set output: Summarizes the abnormal patterns of all spatial units and generates a set of fault feature vectors (data structure: N×5 matrix, N is the number of faults, and columns contain timestamp, spatial coordinates, fault type, feature parameters, and confidence).

[0181] Using deep learning technology, a spatiotemporal convolutional adversarial encoder intelligently decouples collected multidimensional features. This model effectively distinguishes between three distinct types of features: natural equipment aging, environmental interference, and true faults. This avoids the mutual interference between features found in traditional methods, extracting more representative fault feature vectors. This solves the challenge of fault feature extraction in the complex operating environments of photovoltaic power plants and significantly improves the accuracy of fault diagnosis. Feature decoupling enables differentiated processing strategies for different types of features, providing a more accurate basis for subsequent prediction and maintenance decisions.

[0182] S203, inputting the set of fault feature vectors into a physically constrained LSTM-GAN hybrid prediction model, and predicting the multi-dimensional failure probability tensor of key relay protection nodes within a preset time period in the future based on the grid topology constraints and the multi-physical field coupling relationship;

[0183] Specifically, the set of fault feature vectors can be mapped to the nodes of the power grid topology graph, and the status information of adjacent protection devices can be aggregated based on the graph convolutional network to generate a topology constraint feature vector;

[0184] Grid topology node mapping mechanism:

[0185] The electrical connections of a PV power plant are abstracted as a weighted directed graph. Each relay protection device (such as a combiner box protection unit and inverter output relay) is represented as a graph node, while cables and electrical connections are represented as edges. Each feature vector in the fault feature vector set (such as an insulation partial discharge vector) is associated with a specific device node. For example, combiner box protection unit No. 12 corresponds to node N12, and its downstream string protection nodes S12-01 to S12-20 are connected. The system uses a topology table (such as the adjacency matrix Adj) to accurately map the implicit fault vector (location label "Combiner12-String15") to node S12-15.

[0186] ‌State aggregation of Graph Convolutional Network (GCN):

[0187] A three-layer graph convolutional layer aggregates neighborhood states. The first layer aggregates first-order neighbors (directly connected nodes). For example, node S12-15 aggregates the states of its upstream combiner node N12 and its adjacent string nodes S12-14 and S12-16. The second layer extends this to second-order neighbors (e.g., connecting to other string nodes through N12). The convolution kernel weights are dynamically adjusted based on electrical distance: nodes within a physical distance of 10 meters receive a weight of 0.6, while those greater than 30 meters receive a weight reduction of 0.1. The final output topology constraint feature vector contains three pieces of information: the target node's own fault characteristics, the average risk value of its first-order neighbors, and the maximum risk value of its second-order neighbors. In a typical case, when insulation aging characteristics (92% confidence level) were detected on a string node, the risk value of its upstream combiner node increased by 37%, triggering an update to the topology constraint vector.

[0188] The LSTM-GAN generator is initialized with the topological constraint feature vector, and the field strength-current constraint operator is constructed by combining the electromagnetic-thermal multi-physics field coupling equation to output the physically compliant potential state;

[0189] ‌LSTM-GAN (Long Short-Term Memory Generative Adversarial Network) Synergistic Architecture‌:

[0190] The generator uses a stacked LSTM structure (three hidden layers, 128 neurons per layer). After inputting a topologically constrained feature vector, it predicts a sequence of future states at a time step. The discriminator, a convolutional neural network (CNN), distinguishes between real historical data and generated data. A key innovation lies in the embedded "Field-Current Constraint Operator": this operator solves the coupled electromagnetic and thermal field equations in real time to ensure that the predicted values ​​conform to physical laws. For example, when predicting short-circuit current, the conductor temperature rise ΔT (in degrees Celsius) is simultaneously calculated. The logical consistency between the current I and the temperature rise ΔT is verified using the Joule heating formula Q = I²Rt. If the predicted current is 10 kiloamperes (kA) but the temperature rise is only 5°C (actually >80°C), the prediction is considered a violation and the prediction is re-generated.

[0191] ‌Process for generating potential status of physical compliance‌:

[0192] The LSTM generator outputs an initial prediction sequence (e.g., the probability trend of a relay failure in the next 30 minutes).

[0193] The constraint operator uses electromagnetic equations to calculate the electric field strength E (unit: kilovolts per meter, kV / m) and current density J (unit: amperes per square millimeter, A / mm²), and combines them with the heat conduction equation to verify the rationality of the temperature distribution.

[0194] When the predicted value violates physical laws (e.g., a sudden increase in the probability of insulation breakdown but the electric field strength does not exceed the specified value), the generator receives a gradient penalty signal and re-outputs a corrected sequence. In one power plant case, the initial prediction for relay K23 failure at T+25 minutes was 85%. However, the constraint operator detected that the current density of the line in question was only 1.2 A / mm² (the safety threshold is 3.0 A / mm²), forcing the probability to be corrected to 42%.

[0195] Input the potential state of physical compliance into the Monte Carlo failure simulator, iteratively calculate the event trigger conditions including short-circuit current exceeding the standard and insulation breakdown within a preset time period, and generate the failure event probability distribution;

[0196] Monte Carlo Failure Simulator random sampling mechanism:

[0197] Using the physical compliance status as input, the simulator performs 5,000 random iterations over a preset duration (default 72 hours). Each iteration randomly generates four types of disturbance parameters: ambient temperature fluctuation (±15°C), humidity jump (30%→85%), grid voltage surge (+10%), and equipment aging acceleration factor (0.8-1.2x). The simulator includes twelve built-in failure event models, including:

[0198] Short-circuit current exceeding limit model: When the randomly sampled current value exceeds the relay setting (such as the overcurrent threshold of 2.1kA), the probability of triggering action.

[0199] Insulation breakdown model: Based on the inversion results of electric field strength attenuation, if the local E>35kV / mm (the breakdown threshold of XLPE material), the breakdown risk is accumulated.

[0200] Probability distribution generation and key indicator quantification:

[0201] After the iterations are complete, the frequency of each failure event is counted. For example, in 5,000 simulations, relay K09 failed due to insulation breakdown 1,200 times, resulting in a 24% failure probability; combiner box CB05 failed due to excessive short-circuit current 850 times, resulting in a 17% probability. The output probability distribution matrix contains three dimensions: device node number, failure event type, and probability value (in percentage). Data from a coastal power station shows that on days with high humidity (RH > 80%), the probability of inverter node insulation failure is 2.3 times the normal value.

[0202] According to the probability distribution of failure events, a Markov state transition matrix is ​​constructed along the time dimension. Multi-dimensional risk factors are integrated through tensor contraction operations to output the initial failure probability tensor.

[0203] Markov State Transition Matrix construction logic:

[0204] The preset duration is discretized into time slices (15 minutes each), and the device state set is defined as: normal (State_N), warning (State_W), and failure (State_F). The state transition rules are statistically analyzed based on historical data:

[0205] If the current time is State_N, the probability P_W of changing to State_W in the next time slice (e.g., the average P_W of the relay is 0.3%);

[0206] The probability P_F of transitioning from State_W to State_F (e.g., P_F = 1.2% / 15min in a humid environment);

[0207] The matrix dimensions are 3 × 3 × T (T is the number of time slices). For example, in the transfer matrix of relay K12 for the next 72 hours (288 time slices), P_W→F=0.8% when T=120 time slices.

[0208] Tensor Contraction integrates multiple risk sources:

[0209] The initial failure probability tensor contains five dimensions: device node × failure type × time slice × spatial location × environmental level. Redundant dimensions are compressed by the contraction operation:

[0210] Merge similar failure types (e.g., merge "insulation breakdown" and "partial discharge" into the "insulation failure" dimension);

[0211] The spatial location dimension is aggregated into regional levels (e.g., string area, confluence area, inverter area);

[0212] The environmental levels are simplified into three categories: dry (RH < 50%), humid (50-80%), and high humidity (> 80%).

[0213] An example of a three-dimensional tensor output after contraction: number of relay nodes × number of time slices × combined failure probability. Contraction results for a mountain power station show that the combined failure probability of equipment in the confluence area in a high-humidity environment after T+48 hours is 7.8%, significantly higher than the 1.2% in a dry environment.

[0214] The Bayesian correction of historical equipment failure data is applied to the initial failure probability tensor, and the KL divergence is used to constrain the distribution shift to generate an optimized multi-dimensional failure probability tensor.

[0215] ‌Prior Data Fusion for Bayesian Correction‌:

[0216] Collect three years of historical power plant failure records (such as relay operation logs and insulation test reports) to construct a specific prior probability of equipment failure. The correction formula is essentially a weighted average: posterior probability = (initial probability × new prior probability × historical data weight) / normalization factor.

[0217] Weighting principle: New evidence (real-time monitoring data) is weighted 0.7, and historical data is weighted 0.3. For example, if the initial predicted failure probability of a relay is 18%, but its historical average annual failure rate is only 2%, the probability drops to 6.3% after correction.

[0218] ‌KL Divergence (Kullback-Leibler Divergence) Distribution Shift Control‌:

[0219] KL divergence quantifies the difference in probability distribution before and after correction. Set the threshold D_KL_max=0.05 (i.e. the distribution difference does not exceed 5%):

[0220] Calculate the KL divergence between the modified distribution Q and the initial distribution P: D_KL =ΣQ(i) log(Q(i) / P(i)).

[0221] If D_KL>0.05 (e.g., severe light fluctuations leading to inaccurate predictions), adaptive adjustment is initiated: the weight of historical data is increased to 0.5, and the data is recalibrated until D_KL≤0.05.

[0222] Consider a power plant scenario where a thunderstorm causes the initial predicted insulation failure probability to soar to 35%. However, historical data for the same operating conditions shows a maximum of only 12%, with a KL divergence of 0.11. After two iterations of correction, the final output probability is 22% (D_KL = 0.04).

[0223] This method combines the advantages of long-short-term memory networks (LSTMs) and generative adversarial networks (GANs), and incorporates physical constraints of the power grid to construct a hybrid prediction model. This model not only considers the temporal correlation of historical data but also fully accounts for the grid topology and multi-physics coupling effects. It can predict the probability distribution of relay protection devices under different future time periods and failure modes, achieving multi-dimensional and accurate prediction of relay protection device failure risks, providing a scientific basis for preventive maintenance. By considering physical constraints, the prediction results are ensured to conform to the actual operating laws of the power system, greatly improving the credibility of the predictions.

[0224] S204: Based on the multi-dimensional failure probability tensor, a three-dimensional optimization space is constructed that includes equipment maintenance costs, power generation loss risks, and failure chain reaction suppression factors. A strategic competition game is conducted in a virtual power plant simulation environment using a multi-agent reinforcement learning algorithm to generate an optimal operation and maintenance strategy decision tree.

[0225] Specifically, the multi-dimensional failure probability tensor can be projected into a three-dimensional optimization space to generate a risk cost hypersurface, wherein the three-dimensional optimization space includes an X-axis: maintenance cost weight, a Y-axis: power generation loss risk value, and a Z-axis: chain reaction inhibition factor;

[0226] ‌Construction logic of 3D optimization space‌

[0227] The multi-dimensional failure probability tensor (including three-dimensional data of device nodes, failure types, and time slices) is compressed into three-dimensional space through nonlinear mapping. The mapping rules are as follows:

[0228] X-axis (Repair Cost Weight): Quantifies costs based on device type (e.g., relay, inverter) and repair complexity (e.g., replacement time, spare part price). For example, if the relay repair cost weight is 1.0 (base unit), the cost of replacing the entire combiner box would be weighted 8.5 (equivalent to the cost of 8.5 relays).

[0229] Y-axis (Generation Loss Risk): Calculated by combining the probability of equipment failure and the associated generation capacity. For example, if the failure probability of an inverter is 15% and its rated capacity is 250 kilowatts (kW), the loss risk value = 15% × 250 = 37.5 (unit: kW risk).

[0230] The Z-axis (chain reaction suppression factor) measures the extent of fault propagation. If a relay failure could cause three adjacent devices to trip, the suppression factor would initially be 3.0. This factor could be reduced to 0.5 (indicating an 80% chance of chain reactions being blocked) through topological isolation.

[0231] ‌Generation Process of Risk-Cost Hypersurface‌

[0232] The three-dimensional coordinates (X, Y, Z) of each device node are input into a Gaussian process regression model (a nonlinear fitting algorithm based on probability) to generate a continuous and smooth hypersurface. The height of the surface represents the overall risk value, and the peak area corresponds to the high-risk node. For example, the hypersurface of a power station shows:

[0233] The coordinates (8.2, 42.6, 2.8) correspond to combiner box CB09 (high maintenance cost, high loss risk, strong chain reaction), with a comprehensive risk value of 92 points (out of 100);

[0234] The coordinates (1.3, 8.7, 0.4) correspond to the string protection node S05-12, and the comprehensive risk value is only 18 points.

[0235] Hypersurface color mapping rules: red (risk > 70 points), yellow (30-70 points), green (<30 points), operation and maintenance personnel can intuitively identify high-risk areas.

[0236] Deploy three types of intelligent agents: a maintenance robot, a dispatch controller, and a fault isolator in a virtual power plant simulation environment. These agents are given reward functions to optimize X / Y / Z-axis targets, initiating multi-agent strategy competition.

[0237] ‌Virtual simulation environment and agent initialization‌

[0238] Build a power plant digital twin based on the Unity3D engine, importing actual topology and real-time data. Three types of intelligent agent deployment rules:

[0239] Maintenance Robot: This robot is a mobile maintenance vehicle responsible for inspecting and repairing physical equipment. Initially located in the warehouse, it operates throughout the entire site. Optimization goal: Minimize X-axis maintenance costs (e.g., selecting low-cost spare parts and minimizing routing).

[0240] Dispatch Controller: This central control console manages power generation dispatch. Optimization objective: Minimize Y-axis power generation losses (e.g., prioritizing the removal of inefficient strings when isolating a fault).

[0241] Fault Isolator: This entity is a circuit breaker control system that disconnects the fault path. Optimization goal: Maximize the Z-axis chain reaction suppression factor (e.g., rapid isolation to prevent inverter impact).

[0242] ‌Reward function design and strategy competition mechanism‌

[0243] Each type of agent has its own reward function, and its strategy is trained through reinforcement learning:

[0244] The reward for repairing a robot = 100 base points - actual repair cost (X). For example, choosing helicopter transport (cost weight + 15) will result in a 15 point deduction, while choosing land transport (cost weight + 5) will result in a 95 point reward.

[0245] The dispatch controller's reward = 100 base points - power generation loss (Y value / 10). If the loss is 250kW (Y=25), the reward is 75 points.

[0246] The reward for a fault isolator is equal to the chain suppression factor (Z-score) × 20. When the Z-score decreases from 3.0 to 0.5, the reward decreases from 60 points to 10 points (to encourage efficient isolation).

[0247] After the competition is initiated, the agents execute actions in parallel in the simulation. A typical example: A maintenance robot chooses to repair relay K07 (cost weight 1.5), but the fault isolator preemptively cuts off the power to K07 (Z value drops from 2.1 to 0.3), rendering the robot's action invalid. The system records this conflict and generates a competition log.

[0248] The Nash equilibrium solver is used to coordinate the agent strategy, and a deep Q-network is used to explore the combination of maintenance path switching, fixed value adjustment, and virtual inertia activation, and the candidate strategy evaluation matrix is ​​output.

[0249] Conflict Resolution with the Nash Equilibrium Solver

[0250] Coordination of agent strategy conflicts is a two-step process:

[0251] Policy Preference Quantification: Calculates each agent's satisfaction with the current state. For example, a maintenance robot might prefer to prioritize low-cost nodes (80% satisfaction), while a fault isolator might prefer to immediately disconnect high-risk nodes (95% satisfaction).

[0252] Nash equilibrium solution: Find a strategy combination that no agent is willing to change unilaterally. For example, in a sandstorm, the three-party compromise solution is:

[0253] Maintenance robot: repairs risky nodes (cost weight 4.2); scheduling controller: removes faulty strings (loss value 18.3); fault isolator: isolates only local areas (suppression factor 1.2).

[0254] The total reward value of this solution (80+81.7+24=185.7) is higher than any single-party optimal strategy (for example, the total reward for the robot's mandatory maintenance of high-risk nodes is only 142).

[0255] ‌Combinatorial Action Exploration with Deep Q-Network (DQN)‌

[0256] The DQN algorithm (a reinforcement learning model combined with deep learning) explores three types of action combinations:

[0257] Maintenance Path Switching: Optimal path planning for a robot from the warehouse to its destination. If node A (risk value 65) is 50 meters away from node B (risk value 70), the DQN estimates that the path "A→B" (total time 25 minutes) is better than "B→A" (total time 38 minutes).

[0258] Setting Adjustment: The dispatch controller modifies relay protection thresholds. For example, the overcurrent setting can be reduced from 2.1 kiloamperes (kA) to 1.8 kA, reducing fault duration (loss reduction by 12%).

[0259] ‌Virtual Inertia Activation‌: The fault isolator triggers the energy storage system to simulate generator inertia. The droop coefficient is increased from 3% to 5%, enhancing grid stability (the damping factor is increased by 0.4).

[0260] Output candidate strategy evaluation matrix (dimension example: 50 rows × 5 columns), each row records the three-dimensional coordinates (X, Y, Z), total reward value, and execution time of a combined action. For example, the optimal strategy:

[0261] Action: Path S03 → K09, constant value 1.9kA, inertia droop 4%] → (X=5.1, Y=15.2, Z=0.8), reward 192 points, time taken 42 minutes.

[0262] Based on the Pareto frontier solution of the candidate strategy evaluation matrix, a decision branch rule for cost-benefit trade-off is constructed to generate an operation and maintenance strategy decision tree with confidence score.

[0263] ‌Pareto Frontier Solution Screening Rules‌

[0264] Filter non-dominated solutions from the evaluation matrix (i.e., no other strategy is better in all three dimensions X / Y / Z). For example, three candidate strategies:

[0265] Strategy A: (X=3.0, Y=10.5, Z=1.2) rewards 185 points; Strategy B: (X=5.8, Y=8.3, Z=0.9) rewards 188 points; Strategy C: (X=4.1, Y=12.7, Z=1.5) rewards 182 points.

[0266] Strategy B is better than A and C on the Y-axis (generation losses), but worse than A on the X-axis (cost) and worse than C on the Z-axis (chain inhibition) - all three are included in the Pareto frontier.

[0267] Decision branching rules and confidence scoring mechanism

[0268] Decision tree construction logic:

[0269] First-level branches are based on environmental conditions: Branch 1: Normal weather → Choose the solution with the highest overall reward (such as Strategy B); Branch 2: Extreme weather (such as heavy rain) → Prioritize suppressing chain reactions (Choose Strategy C, as it has the highest Z value of 1.5).

[0270] Secondary branching based on risk level: If the target node risk is greater than 60, enable quick response action (fixed value adjustment + virtual inertia activation); if the risk is less than 30, only arrange the maintenance path.

[0271] Confidence Score Calculation: The base score is 80, which increases or decreases based on the strategy's historical success rate in similar scenarios. Example: A strategy succeeded 8 out of 10 thunderstorm simulations → Confidence = 80 + (8 / 10) × 20 = 96.

[0272] Decision tree output example (sandstorm scenario for a 100MW power plant)

[0273] Root node: Environmental status = sandstorm (wind speed > 15m / s)

[0274] │

[0275] ├─ Branch 1: If the highest risk node value is ≥ 75 → Execute Strategy C (adjust the fixed value to 1.8kA + isolate the high-risk area);

[0276] │ Confidence level 92% Estimated loss value Y=14.2;

[0277] │

[0278] └─ Branch 2: If the highest risk node value is less than 75 → Execute Strategy B (Maintenance Robot Path Optimization);

[0279] Confidence level 88% Repair cost X=5.3.

[0280] The decision tree is stored in JSON format, including branch conditions, action instructions, expected three-dimensional coordinates and confidence scores for subsequent remote execution gateway calls.

[0281] This method constructs a three-dimensional optimization space that comprehensively considers economic efficiency, safety, and reliability. It then uses multi-agent reinforcement learning to simulate the effects of various O&M strategies in a virtual environment. Different agents represent different optimization objectives, and through competitive game play, they find the optimal solution that balances the interests of all parties. Ultimately, a structured O&M strategy decision tree is formed, enabling intelligent O&M strategy formulation that considers both economic efficiency and safe and stable system operation. Through strategy game play in a virtual simulation environment, the effectiveness of strategies can be evaluated across a variety of possible scenarios, significantly improving the scientific nature and robustness of decision-making.

[0282] S205, parsing the operation and maintenance strategy decision tree through the remote execution gateway driven by the digital twin, automatically reconstructing the relay protection setting group, switching the fault isolation path, and activating the virtual inertia support function of the energy storage system, and synchronously updating the topological connection relationship of the power station digital twin.

[0283] Specifically, according to the leaf node instructions of the operation and maintenance strategy decision tree, a dynamic fixed value parameter package containing overcurrent threshold and action delay can be injected into the relay protection device through the safety tunnel, and the reconstructed status code fed back by the device can be collected in real time;

[0284] Decision instruction analysis and secure tunnel establishment

[0285] After leaf node instructions in the O&M strategy decision tree (e.g., "Lower the overcurrent threshold of relay K07") are parsed by the digital twin system, a secure tunnel is established using industrial-grade encryption protocols (e.g., IEC 62351). This tunnel utilizes a dual-channel redundant design: the primary channel is a dedicated fiber-optic network (with a transmission latency of less than 5 milliseconds), and the backup channel is a 4G / 5G wireless network (encrypted with AES-256). Bidirectional authentication is performed during tunnel establishment: the O&M platform sends a digital certificate containing a timestamp and device fingerprint. The relay protection device (e.g., Siemens 7SJ82) verifies the certificate's legitimacy and returns a dynamic key (valid for 60 seconds).

[0286] ‌Generation and injection of dynamic valued parameter packs‌

[0287] The Dynamic Setting Packet contains three key parameters:

[0288] Overcurrent Threshold: Adjusted according to the decision tree instructions, for example, from the default 2.1 kiloamperes (kA) to 1.8 kiloamperes (kA), with an accuracy of 0.01 kA.

[0289] Operation Delay: The operation time modified according to the threshold, for example, shortened from 0.5 seconds to 0.3 seconds (in 0.02 second increments).

[0290] ‌Setting Group ID‌: This identifies the currently active setting scheme (e.g., “Rainstorm Mode - Group 3”).

[0291] The parameter packet is transmitted through the tunnel in a fixed frame format (frame header 0xAA55 + data area + CRC32 checksum). After injection, the relay protection device immediately performs parameter programming (taking 200 milliseconds) and returns a reconfiguration status code: 0x00 indicates success, 0x01 indicates parameter out of range, and 0x02 indicates hardware rejection of the write.

[0292] ‌Fault-tolerance handling of feedback mechanisms‌

[0293] If the received status code is an error type (such as 0x01), the system automatically initiates the three-level response:

[0294] Level 1 response: resend the parameter packet (up to 3 times, with an interval of 500 milliseconds);

[0295] Secondary response: switch to redundant control of adjacent equipment (if relay K07 fails, K08 will be enabled to take over);

[0296] Level 3 response: Trigger the local backup setting group (such as switching to factory preset group 1).

[0297] Typical case: A power station experienced status code 0x02 due to electromagnetic interference. The system completed the transfer of control from K07 to K08 within 1.2 seconds, ensuring continuous strategy execution.

[0298] The fault path switching engine is activated based on the reconstructed status code, and the optimal disconnector combination is calculated based on the real-time topology of the power grid. A topology switching instruction sequence is sent to the circuit breaker, and the switch position feedback matrix is ​​obtained synchronously.

[0299] Fault Path Switching Engine triggering logic

[0300] When the reconstruction status code is 0x00 (success), the engine starts path calculation based on the real-time topology of the power grid. The topology data contains three types of dynamic elements:

[0301] Electrical connection relationship: For example, the hierarchical path of string S05-12 → combiner box CB05 → inverter INV03;

[0302] Equipment status mark: circuit breaker open / closed status (e.g. QF09 is "closed"), disconnector position (e.g. QS12 is "open");

[0303] Fault impact domain: High-risk areas marked by the failure probability tensor (such as the eight strings associated with CB07).

[0304] The engine generates candidate solutions with the goal of minimizing the outage scope, combining electrical distance (wiring length) and equipment reliability (historical failure rate).

[0305] ‌Decision-making process for optimal disconnector combination‌

[0306] Use Dijkstra's algorithm (a dynamic programming algorithm for solving the shortest path) to search for the optimal isolation path:

[0307] ‌Construct a switch operation cost model‌: the cost weight of operating a circuit breaker is 3.0 (due to the long operation time); the cost weight of operating an isolating switch is 1.5; the additional cost of cross-region switching is +2.0.

[0308] Minimum Cut Set Calculation: For example, when isolating the faulty combiner box CB07, Option 1: Disconnect upstream QF07 and downstream QS07-09 (total cost 3.0 + 1.5 × 3 = 7.5); Option 2: Disconnect QF03 and QS07-12 (total cost 3.0 + 1.5 × 6 = 12.0) → Option 1 is preferred.

[0309] The final output is a topology switching command sequence (Topology Switching Command Sequence), such as: [QF07: open, QS07: open, QS08: open, QS09: open].

[0310] ‌Synchronous Verification of Switch Position Feedback Matrix‌

[0311] After the command is issued, the switch position is collected through a dual channel of hard contact signal (Dry Contact Signal) and GOOSE message (a high-speed industrial communication protocol). The rows of the feedback matrix (Switch Position Matrix) correspond to the device number, and the columns contain three elements:

[0312] Instruction target position (such as "open"); actual detection position (such as "closed"); time mark (action delay time, required to be ≤120 milliseconds).

[0313] When it is detected that the actual position of QS08 does not match the instruction (for example, the instruction is "open" but the actual position is still "closed"), the engine will reissue the single-step operation instruction within 300 milliseconds until the matrix is ​​all green (100% compliance rate).

[0314] The energy storage virtual inertia controller is triggered according to the switch position feedback matrix, the droop coefficient and virtual impedance parameters are dynamically adjusted, the inertia response eigenvector is output to the energy storage converter PCS, and an inertia activation verification waveform is generated;

[0315] Starting conditions for the energy storage virtual inertia controller

[0316] When the switch position feedback matrix shows at least one circuit breaker opening action (such as QF07 status changes to "open"), the controller is automatically activated. Its core function is to simulate the synchronous generator inertia response (SynchronousGenerator Inertia Response) by adjusting two types of parameters:

[0317] Droop Coefficient: This controls the power compensation strength during frequency drops. The default value is 3% (i.e., for every 1 Hz drop in frequency, the power increases by 33%). This value is dynamically adjusted based on the severity of the grid disturbance: for small disturbances (<0.2Hz), it is reduced to 1.5%, and for large disturbances (>0.5Hz), it is increased to 5%.

[0318] Virtual Impedance: Simulates the internal resistance of a generator, with an initial value of 0.02 ohms. When multiple inverters are connected in parallel, the value is automatically increased to 0.05 ohms to suppress circulating current.

[0319] ‌Generation and Transmission of Inertia Response Eigenvector‌

[0320] The controller calculates the feature vector (Inertia Response Feature Vector) once per second, which contains four-dimensional parameters:

[0321] Active power regulation slope (unit: megawatts per hertz, MW / Hz); reactive power voltage coefficient (unit: megavars per kilovolt, MVar / kV); virtual inertia time constant (unit: seconds, s);

[0322] Damping coefficient (dimensionless).

[0323] This vector is sent via the CAN bus (an industrial fieldbus protocol) to the energy storage converter's power conversion system (PCS), the AC / DC converter within the energy storage system. Upon receiving it, the PCS immediately switches control modes (for example, from constant power mode to virtual synchronous mode).

[0324] Closed-loop test of inertia activation verification waveform

[0325] After the PCS executes parameter injection, it actively injects a set of test disturbances: simulating a 0.3Hz frequency step (lasting 200 milliseconds); collecting the energy storage system output power change curve (verification waveform); and calculating key indicators: response delay (required to be less than 80 milliseconds) and power regulation accuracy (error less than 5%).

[0326] A qualified waveform must meet two characteristics:

[0327] Frequency drop phase: The power rises to 90% of the target value within 100 milliseconds (e.g., from 0 to 900 kilowatts (kW)); Frequency recovery phase: The power decays according to an exponential curve (time constant 1.2±0.2 seconds).

[0328] In a test at a power station, when the droop coefficient was 4%, the output power reached 93% of the theoretical value within 85 milliseconds, which was verified to be successful.

[0329] The reconstruction status code, switch position feedback matrix, and inertia activation verification waveform are integrated and reconstructed, and the engine is driven by topology differences to update the power plant digital twin, outputting a synchronously verified twin topology connection diagram that includes relay protection settings, fault isolation paths, and inertia status.

[0330] Multi-source data fusion and consistency verification

[0331] The Topology Difference-Driven Engine receives three types of input:

[0332] Reconstruction status code of the relay protection device (such as K07 returns 0x00); switch position feedback matrix (such as QF07: open, QS07-09: open); inertia verification waveform (such as response delay of 82 milliseconds, qualified).

[0333] The engine performs cross-validation: if the setting value is successfully reconstructed but the associated circuit breaker does not operate (for example, QF07 is still "closed"), a logic conflict is marked; if the inertia response is out of tolerance (for example, the delay is greater than 100 milliseconds) but the setting value does not take effect, the communication link failure is traced back.

[0334] Digital Twin Update Mechanism

[0335] The update process is divided into three steps:

[0336] ‌Topology connection relationship reconstruction‌:

[0337] Delete the fault path: for example, disconnect the virtual connection line between CB07 and INV03; add a backup path: for example, switch string S07-09 to the backup inverter INV09; update electrical parameters: set the isolation zone impedance to infinity (10,000 ohms).

[0338] ‌Device Status Sync‌:

[0339] Relay protection setting: Displays K07 overcurrent threshold = 1.8 kiloamperes (kA) (changed value highlighted in red); Energy storage inertia status: Marked droop coefficient = 4%, virtual impedance = 0.03 ohms; Circuit breaker position: QF07 icon turns gray (open state).

[0340] Timing alignment: All changes are bound to precise time stamps (accuracy ±1 millisecond), supporting historical backtracking.

[0341] Output and verification of twin topology connection diagram

[0342] The final output Twin Topology Connection Diagram consists of three layers: a physical layer showing device location, cable routing, and real-time temperature distribution; an electrical layer annotating node voltages (e.g., INV09 output voltage 405 volts (V)) and branch currents (e.g., string S07-06 current 8.2 amperes (A)); and a functional layer highlighting protection setting groups, fault isolation zones (flashing red), and virtual inertia scopes (overlaid in blue).

[0343] Synchronization verification compares the twin's predicted values ​​with actual SCADA (Supervisory Control and Data Acquisition) data. If the current prediction error is less than 2% and the voltage error is less than 1%, synchronization is considered valid. In one case, the twin predicted that the INV09 load factor would rise to 82% after isolation, while the actual value was 79.3%, with an error of 3.3%, triggering automatic calibration (correcting model parameters).

[0344] This approach uses digital twin technology to automate the execution of O&M strategies. The remote gateway accurately interprets decision instructions, coordinates and controls multiple processes, including relay protection setting adjustment, fault isolation path switching, and activation of energy storage system support functions. It also provides real-time feedback on execution results and updates the digital twin model to maintain consistency between the virtual and the real. This achieves closed-loop management from decision-making to execution, significantly improving O&M efficiency and responsiveness. The application of digital twin technology ensures visualization and traceability of the execution process, accumulating valuable experience for subsequent O&M optimization.

[0345] It can be seen that the multimodal feature quantities of the relay protection devices in the photovoltaic array area are collected in real time through the distributed fiber optic sensing network to construct a dynamic protection feature matrix; the spatiotemporal convolution adversarial encoder is used to decouple the dynamic protection feature matrix to generate a set of fault feature vectors; the set of fault feature vectors is input into the physically constrained LSTM-GAN hybrid prediction model to predict the multi-dimensional failure probability tensor of the key relay protection nodes within a preset time in the future; based on the multi-dimensional failure probability tensor, the optimal operation and maintenance strategy decision tree is generated; the operation and maintenance strategy decision tree is parsed through the digital twin-driven remote execution gateway, and the topological connection relationship of the power station digital twin is synchronously updated, thereby improving the operation and maintenance efficiency and reliability of photovoltaic power station protection.

[0346] Another embodiment of the present invention provides a photovoltaic power station intelligent remote operation and maintenance system based on relay protection, see Figure 3 , the system may include:

[0347] The acquisition module 301 is used to collect multimodal characteristics of the relay protection device in the photovoltaic array area in real time through a distributed optical fiber sensor network to construct a dynamic protection characteristic matrix, wherein the multimodal characteristics include transient impedance spectrum, fault arc high-frequency oscillation waveform, and insulation degradation gradient value;

[0348] A separation module 302 is configured to perform feature decoupling on the dynamic protection feature matrix using a spatiotemporal convolution adversarial encoder, thereby separating three orthogonal feature subsets: inherent device aging features, environmental disturbance features, and latent fault features, and generating a set of fault feature vectors.

[0349] Prediction module 303, configured to input the set of fault feature vectors into a physically constrained LSTM-GAN hybrid prediction model, and predict the multi-dimensional failure probability tensor of key relay protection nodes within a preset time period in the future based on the grid topology constraints and the multi-physical field coupling relationship;

[0350] A generation module 304 is configured to construct a three-dimensional optimization space including equipment maintenance costs, power generation loss risks, and failure chain reaction suppression factors based on the multi-dimensional failure probability tensor, and generate an optimal operation and maintenance strategy decision tree by conducting a strategic competition game in a virtual power plant simulation environment using a multi-agent reinforcement learning algorithm;

[0351] The reconstruction module 305 is used to parse the operation and maintenance strategy decision tree through the remote execution gateway driven by the digital twin, automatically reconstruct the relay protection setting group, switch the fault isolation path and activate the virtual inertia support function of the energy storage system, and synchronously update the topological connection relationship of the power station digital twin.

[0352] An embodiment of the present invention further provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps of any one of the above method embodiments when running.

[0353] Specifically, in this embodiment, the above-mentioned storage medium may be configured to store a computer program for performing the following steps:

[0354] S201, collecting multimodal feature quantities of relay protection devices in a photovoltaic array area in real time through a distributed optical fiber sensor network to construct a dynamic protection feature matrix, wherein the multimodal feature quantities include a transient impedance spectrum, a fault arc high-frequency oscillation waveform, and an insulation degradation gradient value;

[0355] S202, using a spatiotemporal convolutional adversarial encoder to perform feature decoupling on the dynamic protection feature matrix, separating three orthogonal feature subsets: inherent equipment aging features, environmental disturbance features, and latent fault features, and generating a set of fault feature vectors;

[0356] S203, inputting the set of fault feature vectors into a physically constrained LSTM-GAN hybrid prediction model, and predicting the multi-dimensional failure probability tensor of key relay protection nodes within a preset time period in the future based on the grid topology constraints and the multi-physical field coupling relationship;

[0357] S204: Based on the multi-dimensional failure probability tensor, a three-dimensional optimization space is constructed that includes equipment maintenance costs, power generation loss risks, and failure chain reaction suppression factors. A strategic competition game is conducted in a virtual power plant simulation environment using a multi-agent reinforcement learning algorithm to generate an optimal operation and maintenance strategy decision tree.

[0358] S205, parsing the operation and maintenance strategy decision tree through the remote execution gateway driven by the digital twin, automatically reconstructing the relay protection setting group, switching the fault isolation path, and activating the virtual inertia support function of the energy storage system, and synchronously updating the topological connection relationship of the power station digital twin.

[0359] An embodiment of the present invention further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any one of the above method embodiments.

[0360] Specifically, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0361] Specifically, in this embodiment, the processor may be configured to execute the following steps through a computer program:

[0362] S201, collecting multimodal feature quantities of relay protection devices in a photovoltaic array area in real time through a distributed optical fiber sensor network to construct a dynamic protection feature matrix, wherein the multimodal feature quantities include a transient impedance spectrum, a fault arc high-frequency oscillation waveform, and an insulation degradation gradient value;

[0363] S202, using a spatiotemporal convolutional adversarial encoder to perform feature decoupling on the dynamic protection feature matrix, separating three orthogonal feature subsets: inherent equipment aging features, environmental disturbance features, and latent fault features, and generating a set of fault feature vectors;

[0364] S203, inputting the set of fault feature vectors into a physically constrained LSTM-GAN hybrid prediction model, and predicting the multi-dimensional failure probability tensor of key relay protection nodes within a preset time period in the future based on the grid topology constraints and the multi-physical field coupling relationship;

[0365] S204: Based on the multi-dimensional failure probability tensor, a three-dimensional optimization space is constructed that includes equipment maintenance costs, power generation loss risks, and failure chain reaction suppression factors. A strategic competition game is conducted in a virtual power plant simulation environment using a multi-agent reinforcement learning algorithm to generate an optimal operation and maintenance strategy decision tree.

[0366] S205, parsing the operation and maintenance strategy decision tree through the remote execution gateway driven by the digital twin, automatically reconstructing the relay protection setting group, switching the fault isolation path, and activating the virtual inertia support function of the energy storage system, and synchronously updating the topological connection relationship of the power station digital twin.

[0367] The above describes in detail the structure, features and effects of the present invention based on the embodiments shown in the drawings. The above is only a preferred embodiment of the present invention, but the scope of implementation of the present invention is not limited to what is shown in the drawings. Any changes made in accordance with the concept of the present invention, or modifications to equivalent embodiments with equivalent changes, which do not exceed the spirit covered by the description and drawings, should be within the scope of protection of the present invention.

Claims

1. A method for intelligent remote operation and maintenance of photovoltaic power stations based on relay protection, characterized in that: The method comprises: The multimodal characteristics of the photovoltaic array area relay protection device are collected in real time through a distributed optical fiber sensor network to construct a dynamic protection characteristic matrix, wherein the multimodal characteristics include transient impedance spectrum, fault arc high-frequency oscillation waveform and insulation degradation gradient value; A spatiotemporal convolutional adversarial encoder is used to perform feature decoupling on the dynamic protection feature matrix, separating three orthogonal feature subsets: inherent equipment aging features, environmental disturbance features, and latent fault features, and generating a set of fault feature vectors. The fault feature vector set is input into the physically constrained LSTM-GAN hybrid prediction model, and the multi-dimensional failure probability tensor of the key relay protection nodes within a preset time period in the future is predicted based on the grid topology constraints and the multi-physical field coupling relationship; Based on the multi-dimensional failure probability tensor, a three-dimensional optimization space is constructed that includes equipment maintenance costs, power generation loss risks, and failure chain reaction suppression factors. A multi-agent reinforcement learning algorithm is used to conduct a strategic competition game in a virtual power plant simulation environment to generate an optimal operation and maintenance strategy decision tree. The operation and maintenance strategy decision tree is parsed through a remote execution gateway driven by the digital twin, automatically reconstructing the relay protection setting group, switching the fault isolation path, activating the virtual inertia support function of the energy storage system, and synchronously updating the topological connection relationship of the power station digital twin.

2. The method according to claim 1, characterized in that The multimodal characteristic quantities of the photovoltaic array area relay protection device are collected in real time through the distributed optical fiber sensor network to construct a dynamic protection characteristic matrix, wherein the multimodal characteristic quantities include transient impedance spectrum, fault arc high-frequency oscillation waveform and insulation degradation gradient value, including: The wavelet packet energy spectrum is decomposed based on the optical fiber vibration sensor signal to extract the high-frequency oscillation waveform fragments of the fault arc in the frequency range of 10kHz-1MHz to generate the original arc oscillation sequence; Based on the original arc oscillation sequence and combined with the temperature field distribution of the photovoltaic panel, the impedance phase correction is performed, the real part gradient change rate of the transient impedance spectrum is calculated through the frequency domain differential operator, and the impedance spectrum gradient vector is output; The dielectric response model of the insulation layer is driven by the impedance spectrum gradient vector, and the electric field intensity attenuation inversion algorithm is used to calculate the aging rate of the insulation material per square meter to generate the insulation degradation gradient tensor; The arc oscillation original sequence, impedance spectrum gradient vector, and insulation degradation gradient tensor are spliced ​​into a three-dimensional feature cube according to the time-space alignment rule, and normalized through a sliding time window to output a dynamic protection feature matrix.

3. The method according to claim 2, characterized in that The spatiotemporal convolution adversarial encoder is used to perform feature decoupling on the dynamic protection feature matrix, separating three orthogonal feature subsets: inherent equipment aging features, environmental disturbance features, and latent fault features, and generating a set of fault feature vectors, including: The dynamic protection feature matrix is ​​input into the spatiotemporal convolution encoder, and the spatial dependencies across time steps are extracted using the 3D hole convolution kernel to generate a multi-scale spatiotemporal feature map. Taking the multi-scale spatiotemporal feature map as input, the adversarial decoder and encoder play a Min-Max game, and use the gradient reversal layer to separate the low-frequency component map of the device's inherent aging characteristics; Apply an environmental noise mask to the low-frequency component spectrum, use the spectral clustering algorithm to identify the feature clusters affected by temperature / humidity disturbances, and output the environmental disturbance feature subset; The low-frequency component map and the environmental disturbance feature subset are subtracted from the multi-scale spatiotemporal feature map, and the unrepresented abnormal patterns are extracted through the residual orthogonalization layer to generate a set of hidden fault feature vectors.

4. The method according to claim 3, characterized in that The fault feature vector set is input into the physically constrained LSTM-GAN hybrid prediction model, and the multi-dimensional failure probability tensor of the key relay protection nodes within a preset time period in the future is predicted based on the grid topology constraints and the multi-physical field coupling relationship, including: The fault feature vector set is mapped to the nodes of the power grid topology graph, and the status information of adjacent protection devices is aggregated based on the graph convolutional network to generate a topology constraint feature vector. The LSTM-GAN generator is initialized with the topological constraint feature vector, and the field strength-current constraint operator is constructed by combining the electromagnetic-thermal multi-physics field coupling equation to output the physically compliant potential state; Input the potential state of physical compliance into the Monte Carlo failure simulator, iteratively calculate the event trigger conditions including short-circuit current exceeding the standard and insulation breakdown within a preset time period, and generate the failure event probability distribution; According to the probability distribution of failure events, a Markov state transition matrix is ​​constructed along the time dimension. Multi-dimensional risk factors are integrated through tensor contraction operations to output the initial failure probability tensor. The Bayesian correction of historical equipment failure data is applied to the initial failure probability tensor, and the KL divergence is used to constrain the distribution shift to generate an optimized multi-dimensional failure probability tensor.

5. The method according to claim 4, characterized in that Based on the multi-dimensional failure probability tensor, a three-dimensional optimization space including equipment maintenance cost, power generation loss risk, and failure chain reaction suppression factor is constructed. A multi-agent reinforcement learning algorithm is used to conduct a strategic competition game in a virtual power plant simulation environment to generate an optimal operation and maintenance strategy decision tree, including: Projecting the multi-dimensional failure probability tensor into a three-dimensional optimization space to generate a risk cost hypersurface, wherein the three-dimensional optimization space includes an X-axis: maintenance cost weight, a Y-axis: power generation loss risk value, and a Z-axis: chain reaction inhibition factor; Deploy three types of intelligent agents: a maintenance robot, a dispatch controller, and a fault isolator in a virtual power plant simulation environment. These agents are given reward functions to optimize X / Y / Z-axis targets, initiating multi-agent strategy competition. The Nash equilibrium solver is used to coordinate the agent strategy, and a deep Q-network is used to explore the combination of maintenance path switching, fixed value adjustment, and virtual inertia activation, and the candidate strategy evaluation matrix is ​​output. Based on the Pareto frontier solution of the candidate strategy evaluation matrix, a decision branch rule for cost-benefit trade-off is constructed to generate an operation and maintenance strategy decision tree with confidence score.

6. The method according to claim 5, characterized in that The remote execution gateway driven by the digital twin parses the operation and maintenance strategy decision tree, automatically reconstructs the relay protection setting group, switches the fault isolation path, activates the virtual inertia support function of the energy storage system, and synchronously updates the topological connection relationship of the power station digital twin, including: According to the leaf node instructions of the operation and maintenance strategy decision tree, a dynamic fixed value parameter package containing overcurrent threshold and action delay is injected into the relay protection device through the safety tunnel, and the reconstruction status code fed back by the device is collected in real time; The fault path switching engine is activated based on the reconstructed status code, and the optimal disconnector combination is calculated based on the real-time topology of the power grid. A topology switching instruction sequence is sent to the circuit breaker, and the switch position feedback matrix is ​​obtained synchronously. The energy storage virtual inertia controller is triggered according to the switch position feedback matrix, the droop coefficient and virtual impedance parameters are dynamically adjusted, the inertia response characteristic vector is output to the energy storage converter PCS, and an inertia activation verification waveform is generated; The reconstruction status code, switch position feedback matrix, and inertia activation verification waveform are integrated and reconstructed, and the engine is driven by topology differences to update the power plant digital twin, outputting a synchronously verified twin topology connection diagram that includes relay protection settings, fault isolation paths, and inertia status.

7. An intelligent remote operation and maintenance system for photovoltaic power stations based on relay protection, characterized in that: The system comprises: An acquisition module is used to collect multimodal characteristic quantities of the relay protection device in the photovoltaic array area in real time through a distributed optical fiber sensor network to construct a dynamic protection characteristic matrix, wherein the multimodal characteristic quantities include transient impedance spectrum, fault arc high-frequency oscillation waveform and insulation degradation gradient value; a separation module, configured to perform feature decoupling on the dynamic protection feature matrix using a spatiotemporal convolution adversarial encoder, separate three orthogonal feature subsets, namely, inherent equipment aging features, environmental disturbance features, and latent fault features, and generate a set of fault feature vectors; A prediction module is used to input the set of fault feature vectors into a physically constrained LSTM-GAN hybrid prediction model, and predict the multi-dimensional failure probability tensor of key relay protection nodes within a preset time period in the future based on the grid topology constraints and the coupling relationship between multiple physical fields; A generation module is used to construct a three-dimensional optimization space including equipment maintenance cost, power generation loss risk, and failure chain reaction suppression factor based on the multi-dimensional failure probability tensor, and to generate an optimal operation and maintenance strategy decision tree by conducting a strategic competition game in a virtual power plant simulation environment through a multi-agent reinforcement learning algorithm; The reconstruction module is used to parse the operation and maintenance strategy decision tree through the remote execution gateway driven by the digital twin, automatically reconstruct the relay protection setting group, switch the fault isolation path, activate the virtual inertia support function of the energy storage system, and synchronously update the topological connection relationship of the power station digital twin.

8. The system according to claim 7, characterized in that The acquisition module is specifically used to: The wavelet packet energy spectrum is decomposed based on the optical fiber vibration sensor signal to extract the high-frequency oscillation waveform fragments of the fault arc in the frequency range of 10kHz-1MHz to generate the original arc oscillation sequence; Based on the original arc oscillation sequence and combined with the temperature field distribution of the photovoltaic panel, the impedance phase correction is performed, the real part gradient change rate of the transient impedance spectrum is calculated through the frequency domain differential operator, and the impedance spectrum gradient vector is output; The dielectric response model of the insulation layer is driven by the impedance spectrum gradient vector, and the electric field intensity attenuation inversion algorithm is used to calculate the aging rate of the insulation material per square meter to generate the insulation degradation gradient tensor; The arc oscillation original sequence, impedance spectrum gradient vector, and insulation degradation gradient tensor are spliced ​​into a three-dimensional feature cube according to the time-space alignment rule, and normalized through a sliding time window to output a dynamic protection feature matrix.

9. A storage medium, characterized in that: The storage medium stores a computer program, wherein the computer program is configured to execute the method according to any one of claims 1 to 6 when executed.

10. An electronic device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to perform the method according to any one of claims 1 to 6.

Images