A photovoltaic power station communication device interference monitoring method

Abstract

This application discloses a method for monitoring interference in communication equipment of a photovoltaic power station. The method includes: collecting synchronous mixed interference signals from multiple acquisition nodes within the photovoltaic power station; performing time-frequency separation on the synchronous mixed interference signals to decompose them into several candidate interference source components; extracting the fingerprint feature vector of each candidate interference source component and matching it with a pre-built interference source fingerprint feature database to determine the physical device identifier corresponding to each candidate interference source; using a multipath cross-correlation analysis mechanism to obtain the propagation path parameters from each identified interference source to each acquisition node, the propagation path parameters including the attenuation coefficient and time delay of at least one propagation mode; inputting the prior operating state parameters and propagation path parameters of each identified interference source into a pre-trained graph neural network model, outputting the contribution of each interference source to each acquisition node, and generating an interference source tracing list. This achieves accurate tracing of interference sources and dynamic tracking of propagation paths.

Description

Technical Field

[0001] This invention relates to the field of communication equipment interference technology, and in particular to a method for monitoring interference in communication equipment of photovoltaic power plants. Background Technology

[0002] The communication system of a photovoltaic power station undertakes key functions such as equipment monitoring, data acquisition, and operation scheduling, and its reliability directly affects the safe and stable operation of the power station. However, there are many sources of electromagnetic interference in the photovoltaic power station environment, including equipment such as inverters, photovoltaic strings, combiner boxes, and DC converters. These devices generate radiated and conducted emissions during operation, which affect the communication equipment through various means such as spatial radiation, cable coupling, and grounding loops, leading to problems such as communication errors, packet loss, and interruptions.

[0003] Currently, assessment methods for interference with communication equipment in photovoltaic power plants mainly fall into the following categories: One category is assessment methods based on on-site measurements. These methods involve placing spectrum analyzers or electromagnetic interference receivers at the installation location of the communication equipment to directly measure parameters such as the amplitude and frequency band of the interference signal. While these methods can obtain accurate interference intensity data, the measurement process is significantly affected by external conditions such as weather, time, and equipment operating status, and it is difficult to achieve long-term continuous monitoring, thus failing to reflect the dynamic changes in interference. Another category is assessment methods based on simulation modeling. These methods predict the interference level at the location of the communication equipment by establishing an electromagnetic field distribution model or circuit conduction model of the photovoltaic power plant. However, these methods rely on precise model parameters and boundary conditions. In actual power plants, equipment layouts are complex, grounding networks are varied, and cable routing is irregular, leading to significant deviations between simulation results and measured values. Furthermore, the computational load increases exponentially with the scale of the power plant, limiting its engineering applicability.

[0004] However, the assessment results in the existing technology can only give the intensity level or trend of the interference to the communication equipment, and cannot provide the most favorable interference monitoring guidance for operation and maintenance personnel.

[0005] Therefore, how to accurately identify interference source devices and quantitatively analyze their propagation characteristics in the complex electromagnetic environment of photovoltaic power plants is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] This application provides a method for monitoring interference in communication equipment in photovoltaic power plants. It can automatically identify the specific interference source equipment and its identity in photovoltaic power plants, and quantitatively analyze the path parameters of interference signals reaching the communication equipment through different propagation modes. This enables accurate source tracing and dynamic tracking of propagation paths, providing a basis for source suppression and path blocking, thereby significantly improving the pertinence and effectiveness of interference control.

[0007] This application provides a method for monitoring interference in communication equipment of a photovoltaic power station, including: S101, collect synchronous mixed interference signals from multiple acquisition nodes in the photovoltaic power station, perform time-frequency separation on the synchronous mixed interference signals, and combine the prior operating state parameters of each potential interference source as guidance to decompose several candidate interference source components. S102, extract the fingerprint feature vector of each candidate interference source component, match and identify it with the pre-built interference source fingerprint feature database, and determine the physical device identifier corresponding to each candidate interference source. S103, using a preset multipath cross-correlation analysis mechanism, obtain the propagation path parameters from each identified interference source to each acquisition node. The propagation path parameters include the attenuation coefficient and time delay of at least one propagation mode. S104: Input the prior operating state parameters and propagation path parameters of each identified interference source into the pre-trained graph neural network model, output the contribution of each interference source to each acquisition node, and generate a list of interference source tracing.

[0008] Preferably, the synchronous mixed interference signal is set as a multi-channel mixed signal matrix within the monitoring window, where each channel corresponds to the received time-domain signal of one acquisition node, and the multi-channel mixed signal matrix is ​​represented as follows: Where M is the total number of data collection nodes. Let be the time-domain signal received by the j-th node.

[0009] Preferably, the prior operating state parameters include at least the switching frequency of each inverter, and at least one auxiliary parameter among the output frequency, operating temperature, and operating time.

[0010] Preferably, the step of performing time-frequency separation on the synchronous mixed interference signal and combining it with the prior operating state parameters of each potential interference source as guidance to decompose it into several candidate interference source components specifically includes: Using the switching frequency and its harmonics to form a set of guiding frequencies, extract time-frequency points with significant energy in the time-frequency domain, calculate the spatial direction vector of each time-frequency point, perform density clustering based on the direction vector, reconstruct the time-domain waveform of the cluster to obtain candidate interference source components, and calculate the correlation degree between each candidate component and the set of guiding frequencies, and sort them according to the correlation degree.

[0011] Preferably, the step of using the switching frequency and its harmonics to construct a set of guiding frequencies, and extracting time-frequency points with significant energy in the time-frequency domain, specifically includes: For multi-channel mixed signal matrix Perform a short-time Fourier transform to obtain the time-frequency representation. Where f is the frequency variable, For time frame indexing; Obtain the switching frequency of each inverter And its integer multiples of harmonic frequencies, forming a set of known switching frequencies and their harmonics. ; For each boot frequency In time-frequency representation Extract the set of time-frequency points with significant energy. , The average energy across all time-frequency points. ,in For frequency points, For time frames, This is the preset energy threshold coefficient.

[0012] Preferably, calculating the correlation between each candidate component and the set of guiding frequencies specifically includes: For each reconstructed candidate interference source component Calculate its spectral peak frequency set Combine it with the set of guiding frequencies Matching to obtain the correlation The degree of correlation is defined as: Wherein, the numerator represents the number of elements in the intersection of the peak frequency and pilot frequency set of the source component, and the denominator represents the total number of peak frequencies of the source component.

[0013] Preferably, the preset multipath cross-correlation analysis mechanism specifically includes: S301, the time-domain waveform separated from the identified interference source is used as the transmitted signal estimate, the received component related to the interference source is extracted from the original received signal of each acquisition node, the cross-correlation function of the two is calculated and multiple effective peaks in the cross-correlation function are detected; S302, each peak corresponds to a propagation path, and its time delay and normalized attenuation coefficient are recorded; S303 classifies the path by mode based on the geographical information and propagation characteristics of the photovoltaic power station.

[0014] Preferably, the propagation path parameters are further defined as: for each interference source-acquisition node pair All detected path parameters are organized into a vector set to form the propagation path parameters: ,in For path modal labels, For the number of paths, The normalized attenuation coefficient, Time delay.

[0015] Preferably, step S104 specifically includes: inputting the prior operating state parameters corresponding to the physical device identifiers of each interference source identified in step S102, and the propagation path parameters from each interference source to each acquisition node estimated in step S103, into a pre-trained graph neural network model. The graph neural network model is established based on the interference propagation topology graph constructed from the propagation path parameters, and outputs the contribution of each interference source to each acquisition node, thereby generating an interference source tracing list.

[0016] Preferably, S104 specifically includes: For the current monitoring window, execute steps S101 to S103 to obtain the identified interference sources within the current window, along with their prior operating state parameters and propagation path parameters. These parameters are then input into the pre-trained graph neural network model to obtain the predicted received power at each acquisition node j. And the contribution power of each interference source n to the acquisition node j ; Define the contribution of interference source n to acquisition node j: divide the contribution power of each interference source to the node by the sum of the contribution power of all interference sources to obtain the contribution of each interference source.

[0017] One or more technical solutions provided in this application have at least the following technical effects or advantages: By combining time-frequency sparse blind separation with fingerprint feature matching, it is possible to separate each candidate interference source component from a multi-source mixed signal and uniquely determine its corresponding physical device (such as a specific inverter or combiner box), thus achieving interference localization. Utilizing a multipath cross-correlation analysis mechanism, the propagation path from each interference source to each acquisition node is decomposed into multiple modes to obtain the attenuation coefficients and time delays of different paths, such as direct radiation, reflected radiation, conducted common mode, conducted differential mode, and near-field coupling, thereby revealing the spatial propagation characteristics of the interference signal. Based on a pre-trained graph neural network model, the prior operating state parameters and propagation path parameters of the interference source are input, and the contribution weight of each interference source to each communication device is output, generating a quantifiable source tracing list to provide a priority basis for interference mitigation.

[0018] Maintenance personnel can directly locate specific interference source devices and dominant propagation modes based on the source tracing list, avoiding blindly replacing communication equipment or installing filters throughout the site, thus reducing unnecessary maintenance costs. Through real-time contribution monitoring and multi-level early warning (level 1 warning suggests source filtering, level 2 warning suggests path shielding, and level 3 warning prompts investigation of unknown sources), precise measures can be taken before interference worsens, ensuring the continuous and reliable operation of the photovoltaic power station's communication system. Time-frequency sparse blind separation uses prior switching frequencies as guidance, solving the underdetermined mixing problem. Graph neural networks are trained offline and inferred online, avoiding the large computational load of real-time simulation. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the interference monitoring method for photovoltaic power plant communication equipment according to an embodiment of the present invention. Detailed Implementation

[0020] To facilitate understanding of the present invention, a more complete description of this application will be given below with reference to the accompanying drawings, which illustrate preferred embodiments of the invention. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to enable a more thorough and complete understanding of the disclosure of the present invention.

[0021] It should be noted that the terms "vertical," "horizontal," "up," "down," "left," "right," and similar expressions used in this article are for illustrative purposes only and do not represent the only possible implementation.

[0022] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention; the term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0023] Current interference assessment technologies for photovoltaic power plant communication equipment primarily focus on the quantification and classification of interference intensity, enabling the determination of the magnitude and trend of interference affecting communication equipment. However, in practical engineering, maintenance personnel not only need to know the magnitude of the interference but also its source and the path it takes to propagate to the communication equipment for reference. Because photovoltaic power plants contain numerous potential interference sources (such as inverters, photovoltaic strings, and combiner boxes), and interference signals can propagate through multiple paths including spatial radiation, conductive coupling, and near-field induction, the signals received by communication equipment are a nonlinear mixture of multiple sources and paths. Existing technologies cannot perform source separation and path inversion of the mixed signals, leaving interference mitigation at a passive anti-interference level, lacking precise basis for source suppression and path blocking.

[0024] Therefore, this application proposes a method for monitoring interference in communication equipment of photovoltaic power plants. This method acquires mixed interference signals through multi-node synchronous acquisition, and uses prior operating state parameters (especially switching frequencies) of potential interference sources to guide time-frequency sparse blind separation, decomposing candidate interference source components. Then, it identifies the physical identity of the interference sources through fingerprint feature matching. A pre-set multipath cross-correlation analysis mechanism is used to estimate the propagation path parameters (including attenuation coefficients and time delays of at least one propagation mode) from each identified interference source to each acquisition node. Finally, the prior operating state parameters and propagation path parameters of each identified interference source are input into a pre-trained graph neural network model, outputting the contribution of each interference source to each acquisition node and generating an interference source tracing list. This method achieves automatic identification of interference sources and dynamic tracking of propagation paths, thus providing a technical path for interference source mitigation.

[0025] Example 1: Figure 1 This is a flowchart illustrating the interference monitoring method for photovoltaic power station communication equipment according to an embodiment of the present invention.

[0026] like Figure 1 As shown, a method for monitoring interference in communication equipment of a photovoltaic power station includes the following steps: S101, based on a pre-deployed multi-node acquisition network, collects synchronous mixed interference signals from multiple acquisition nodes within the photovoltaic power station, performs time-frequency separation on the synchronous mixed interference signals, and combines the prior operating state parameters of each potential interference source as guidance to decompose several candidate interference source components.

[0027] The pre-deployed multi-node acquisition network is defined as follows: interference signal acquisition terminals are pre-deployed at multiple key locations within the photovoltaic power station. For example, key locations include the installation point of each communication device, the outlet of each main DC combiner line, and the AC output side of each inverter. All acquisition terminals achieve microsecond-level time synchronization through the GPS second pulse or the IEEE 1588 precise time protocol, with a synchronization error of less than one-tenth of the sampling period.

[0028] Specifically, there are M data collection nodes, numbered as follows: In the same continuous monitoring window Inside (e.g., monitoring window) (Set to 60s, but can be adjusted adaptively based on actual monitoring needs). All nodes synchronously acquire interference signals within the communication frequency band, generating synchronous mixed interference signals from multiple acquisition nodes, resulting in a multi-channel mixed signal matrix. Each channel corresponds to the received time-domain signal of one acquisition node. The multi-channel mixed signal matrix is ​​represented as follows: Where M is the total number of data collection nodes. Let be the time-domain signal received by the j-th node. It should be noted that since the signals emitted by each interference source arrive at each node via different paths, and there is multi-source overlap, It can be represented as: Where N is the total number of potential interference sources. For the transmitted signal of the nth interference source, Let n be the path decay coefficient from source n to node j. To delay the transmission time, For the background noise at node j, only the background noise at node j is obtained in this step. And N, , , All of these are unknown quantities to be evaluated.

[0029] In some embodiments, the prior operating state parameters include at least the switching frequency of each inverter, and at least one auxiliary parameter selected from output frequency, operating temperature, and operating time. Specifically, the prior operating state parameters are obtained by acquiring the prior operating state parameters of each potential interference source (including each inverter, photovoltaic string, and combiner box) in real time from the photovoltaic power plant monitoring system. The prior operating state parameters include at least the switching frequency of each inverter. (n is the index of the interference source) and its harmonic order, and may also include auxiliary parameters such as output power, operating temperature, and operating time, which are determined according to the actual operating scenario. Align the prior parameters with the acquired mixed interference signal in time.

[0030] The mixed interference signals of photovoltaic power plants have the following special characteristics that distinguish them from general blind source separation problems: First, the number of potential interference sources is much greater than the number of acquisition nodes, which is an underdetermined mixed problem; Second, the switching frequency of each inverter-type interference source can be obtained in real time in the power plant monitoring system, which constitutes some usable prior information; Third, the number of interference sources that are actually active at any given time is much smaller than the total number of potential sources, that is, the source signal has time-domain sparsity.

[0031] To address the aforementioned unique characteristics, this embodiment employs a blind separation method that integrates sparse constraints and prior frequency guidance. A short-time Fourier transform is performed on the multi-channel mixed signal matrix to obtain its time-frequency representation. Since the interference source signals in a photovoltaic power station exhibit sparsity in the time-frequency domain, the number of dominant sources at each time-frequency point is far less than the total number of sources. Simultaneously, a set of guiding frequencies is constructed using the known inverter switching frequency and its harmonics. In some embodiments, the synchronous mixed interference signal undergoes time-frequency separation, and the prior operating state parameters of each potential interference source are used as guidance to decompose several candidate interference source components, including: Using the switching frequency and its harmonics to form a set of guiding frequencies, extract time-frequency points with significant energy in the time-frequency domain, calculate the spatial direction vector of each time-frequency point, perform density clustering based on the direction vector, reconstruct the time-domain waveform of the cluster to obtain candidate interference source components, and calculate the correlation degree between each candidate component and the set of guiding frequencies, and sort them according to the correlation degree.

[0032] Specifically, time-frequency separation includes: S201, for multi-channel mixed signal matrix Perform a short-time Fourier transform to obtain the time-frequency representation. Where f is the frequency variable, For time frame index.

[0033] It should be noted that the window function uses a Hanning window, with the window length set according to the frequency resolution requirements (typically 1024 sampling points), and an overlap rate of 50%. In the time-frequency domain, due to the sparsity of interference source signals in photovoltaic power plants, each time-frequency point... The number of source signals that play a dominant role is much smaller than the total number of sources. This property is called time-frequency sparsity, which is the key to solving the underdetermined blind separation problem.

[0034] S202, obtain the switching frequency of each inverter in real time from the power plant monitoring system. and its integer multiples of harmonic frequencies (Where L is determined based on the energy attenuation characteristics of the inverter's switching frequency harmonics: calculate the percentage of each harmonic amplitude relative to the fundamental frequency, and take the highest harmonic order whose cumulative energy share reaches more than 95%. In typical scenarios, the value of L ranges from 5 to 10.) All known switching frequencies and their harmonics constitute the pilot frequency set. .

[0035] The guiding frequency set is used to constrain the frequency components of the source components during the separation process; that is, each separated source component should contain at least some frequency components from the guiding frequency set. This constraint transforms the originally unsupervised blind separation problem into a semi-supervised problem, significantly improving the stability and physical interpretability of the separation.

[0036] S203, for each boot frequency In time-frequency representation Extract the set of time-frequency points with significant energy and calculate the energy at each time-frequency point. The spatial direction vectors are used to group time-frequency points with similar direction vectors into the same cluster using the density-based spatial clustering algorithm (DBSCAN), and the time-domain waveform of each cluster is reconstructed to obtain candidate interference source components.

[0037] Specifically: Extract the set of time-frequency points with significant energy. , The average energy across all time-frequency points. ,in For frequency points, For time frames, The preset energy threshold coefficient (which can be determined by collecting the background noise time-frequency energy distribution during periods of no interference at the power station (such as the inverter standby state at night), taking 3 times the average noise energy as the threshold, i.e., λ=3, which can be adjusted in the range of 2 to 4 according to the signal-to-noise ratio in practical applications). Since time-frequency points corresponding to the same pilot frequency often originate from the same interference source, a density-based spatial clustering algorithm is used to divide the time-frequency points into several clusters, with each cluster corresponding to the time-frequency support domain of a candidate interference source component. During clustering, the spatial direction vector of the time-frequency points is used as a feature, and the spatial direction vector is calculated as follows: This vector is an M-dimensional complex vector, reflecting the amplitude ratio and phase difference of the same signal received by different acquisition nodes, and is closely related to the spatial location of the source. For example, the density clustering in the time-frequency separation uses the angle between direction vectors as a distance metric, where two direction vectors... The included angle is defined as: Clustering distance threshold The method for determining the threshold can be as follows: Calibrate the inverter at a known location, calculate the angle between its direction vector and the theoretical direction vector, and take the maximum value of the angles among all calibration samples as the threshold, typically 15°, with a minimum cluster size of 10 points. It should be noted that when the average angle between the direction vectors of two candidate clusters is less than this clustering distance threshold... The time-frequency support domains of the interference source components are merged into the same source, and each cluster corresponds to a candidate interference source component.

[0038] Specifically, reconstructing the time-domain waveform of each cluster yields candidate interference source components, including: For each cluster Construct the corresponding time-frequency mask In the time-frequency domain, retain the time-frequency points within the cluster and set other time-frequency points to zero; The time-domain waveform of the source component is reconstructed using the inverse short-time Fourier transform. The final Q source components These are the separated candidate interference source signals, each component of which corresponds to a time-domain waveform and an estimated direction vector.

[0039] Specifically, the determination of the number of sources Q (the number of clusters retained for clustering) no longer relies on the general information criterion, but instead utilizes the physical constraints of the photovoltaic power plant, and is constrained by the upper limit of the total number of inverters in the photovoltaic power plant (i.e., , This refers to the number of inverters. The number of other potential sources, including combiner boxes, DC-DC converters, etc., can be obtained from the power plant design drawings. A distance threshold is set during clustering. When the average angle between the direction vectors of two candidate clusters is less than a threshold, they are merged into the same source. After clustering, Q is taken as the number of clusters. Clusters with too low energy are discarded. For example, the criterion for judging too low energy can be: the power of the reconstructed signal is lower than the background noise power by 3dB. The background noise power can be obtained by selecting the signal collected during a period of no interference at the power station (such as the inverter standby state at night) and calculating its time-domain mean square power as the background noise power. .

[0040] S204, for each reconstructed candidate interference source component Calculate its spectral peak frequency set (Take the frequencies corresponding to the top 10 largest amplitude values) and combine them with the set of pilot frequencies. The correlation is obtained by matching, and then sorted in descending order according to the correlation between the peak values ​​of each component spectrum and the set of guiding frequencies.

[0041] The correlation index is defined as follows: In the formula, the numerator represents the number of intersection elements between the peak frequency set of the source component and the pilot frequency set, the denominator represents the total number of peak frequencies of the source component (not exceeding 10), and the correlation degree. The higher the value, the more likely the source component corresponds to a real physical interference source. This can be understood as: if... If a source component contains the fundamental or harmonic frequency of a certain inverter's switching frequency, then the correlation between that source component and the inverter increases.

[0042] S204 may further include: sorting candidate interference source components from high to low correlation, temporarily marking components with correlation below the lower correlation threshold as suspected noise or unknown sources, and entering them into a buffer for subsequent iterative optimization. For example, the method for determining the lower correlation threshold may be: calculating the correlation for known interference sources and pure noise samples respectively, and taking the threshold that can distinguish between the two (i.e., the highest classification accuracy). For example, through ROC curve analysis, the typical value is 0.2.

[0043] In summary, for each pilot frequency, time-frequency points with energy significantly exceeding the average background noise are extracted from the time-frequency representation, and the spatial direction vector (i.e., the normalized vector composed of the amplitude ratios of each channel) of each time-frequency point is calculated. A density-based clustering algorithm is used to group time-frequency points with similar direction vectors into the same cluster, with each cluster corresponding to the time-frequency support domain of a candidate interference source component. The time-domain waveform of each cluster is reconstructed using short-time Fourier transform, yielding several candidate interference source components. During clustering, the total number of inverters in the photovoltaic power station is used as an upper limit constraint on the number of sources; when the average angle between the direction vectors of two clusters is less than a preset threshold, they are merged into the same source. For each reconstructed candidate interference source component, the matching degree between its spectral peak frequency set and the pilot frequency set is calculated. The candidate interference source components are sorted from high to low matching degree, and components with too low a matching degree are temporarily marked as suspected noise or unknown sources. Thus, the time-frequency sparse blind separation utilizes prior switching frequencies as guidance, transforming the originally unsupervised separation problem into a semi-supervised problem, significantly improving the stability and physical interpretability of the separation.

[0044] S102, extract the fingerprint feature vector of each candidate interference source component, match and identify it with the pre-built interference source fingerprint feature database, and determine the physical device identifier corresponding to each candidate interference source.

[0045] In some embodiments, the pre-built interference source fingerprint feature database is obtained in the following way: A1. In advance, during the construction or regular inspection phase of the photovoltaic power station, conduct independent excitation tests on each potential interference source device (including inverters, photovoltaic strings, combiner boxes, and DC converters) under different operating conditions and collect interference signals.

[0046] For example, the device under test is electrically isolated from other parts of the power station and operated under three typical operating conditions: no-load, rated load, and half-load. Its radiated emission signal is acquired using a broadband near-field probe and a spectrum analyzer, while its conducted emission signal is acquired simultaneously using a current probe and a voltage probe. The acquisition duration T is set to 10 seconds, and the sampling rate and online monitoring data are known. The acquired raw signal (time-domain signal) is denoted as... , where the subscript i represents the index of the i-th interference source device, and the subscript k represents the index of the k-th working state.

[0047] A2. For the acquired time-domain signal Feature extraction is performed, including the root mean square value in the time domain, harmonic amplitude sequence, sideband width, and modulation depth, to form a fingerprint feature vector. : Calculate the root mean square value of the signal amplitude: Where T is the sampling duration; Harmonic amplitude sequence: The frequency domain representation is obtained by performing a fast Fourier transform. Extract the amplitude sequence of its main harmonic orders. Where L is the highest harmonic order; Sideband width: Calculates the width of the signal's sideband distribution near the switching frequency. , can be defined as the frequency range corresponding to a 3dB decrease in the spectral amplitude near the switching frequency; Modulation depth: ,in and These represent the peak and valley values ​​of the sideband near the switching frequency, respectively.

[0048] A3. Match all fingerprint feature vectors with the corresponding device identifiers. The fingerprint feature database is stored together with the working status tag k. .

[0049] In some embodiments, the physical device identifier corresponding to each candidate interference source is determined by matching and identifying the source with a pre-built interference source fingerprint feature database, specifically including: For each candidate interference source component Its fingerprint feature vector ,Will fingerprint feature database Each fingerprint feature vector Similarity calculation is performed using weighted cosine similarity: Where D is the total dimension of the feature vectors. The weights for the d-th dimension feature (determined empirically or by the analytic hierarchy process, with higher weights assigned to harmonic amplitudes; the weight coefficients can be optimized in reverse using historical matching data after initial setting). and These are the d-th eigenvalues, respectively. Specifically, the weights can be determined using the analytic hierarchy process: construct a pairwise comparison matrix of the importance of each eigenvalue dimension, calculate the eigenvector corresponding to the largest eigenvalue and normalize it to obtain the weights, and the weight coefficients can be optimized in reverse using historical matching data after initial setting.

[0050] Preset matching threshold (The determination method can be as follows: collect test samples from known interference sources, calculate their similarity distribution with their own fingerprints and with heterogeneous fingerprints, and take the threshold that maximizes the difference between the correct matching rate and the incorrect matching rate, typically 0.85), calculate all ( Maximum similarity under combination ,like Then determine the component of the candidate interference source. Corresponding to the physical interference source that maximizes similarity and working status Record the confidence level of the match. If all similarities are below the matching threshold, the component is marked as an unknown interference source and its fingerprint feature vector is temporarily stored in the location source buffer. The feature vectors in the unknown source buffer can be manually added to the fingerprint feature library as new devices after accumulation (if the same feature appears more than 3 times) and confirmation by manual inspection.

[0051] S103, using a preset multipath cross-correlation analysis mechanism, obtain the propagation path parameters from each identified interference source to each acquisition node. The propagation path parameters include the attenuation coefficient and time delay of at least one propagation mode.

[0052] In practical photovoltaic power plants, signals from the same interference source can reach the same node simultaneously through multiple modes, such as direct sunlight, ground reflection, and cable conduction, resulting in multipath superposition. Single-path estimation leads to severe distortion of attenuation coefficients and time delays, which in turn misleads the training and source tracing results of subsequent graph neural networks. To address this deeper issue, this embodiment introduces propagation path mode decomposition, which decomposes the portion of the received signal from the same interference source into at least one path component, estimating the parameters of each mode separately, thereby accurately characterizing the multipath propagation characteristics.

[0053] In some embodiments, the preset multipath cross-correlation analysis mechanism specifically includes: S301 uses the time-domain waveform separated from the identified interference source as the transmitted signal estimate, extracts the received component related to the interference source from the original received signal of each acquisition node, calculates the cross-correlation function between the two and detects multiple effective peaks in the cross-correlation function.

[0054] Specifically, for each identified physical source of interference (denoted as n), the time-domain waveform separated in step S101 is... As an estimate of the signal transmitted by the interference source (the waveform has been normalized to retain relative amplitude information); for each acquisition node j ( ), from the original received signal of this node Extract the part related to source n. The extraction method can be time-frequency mask projection: construct a mask using the time-frequency support domain corresponding to the source component in step S101. ,right Time-frequency representation Perform mask filtering: Then, the time-domain signal is obtained through short-time inverse Fourier transform. ; Calculate the cross-correlation function between the transmitted signal estimate and the corresponding part of the received signal: This function describes the similarity between the estimated transmitted signal and the corresponding part of the received signal under different time delays. Discrete cross-correlation is used in the actual calculation, and the time delay search range is... ,in Maximum possible propagation delay of the power station (e.g., the time it takes for an electromagnetic wave to travel the longest distance across the power station, typically 10 microseconds).

[0055] In actual photovoltaic power plants, because signals may reach the same acquisition node through multiple paths (such as direct sunlight, ground reflection, and cable conduction), the cross-correlation function typically exhibits multiple significant peaks; therefore, the cross-correlation function... Perform peak detection: Find the global maximum value The peak detection threshold is set to 0.3 times the global maximum value of the cross-correlation function, i.e., the peak detection threshold. A preset peak detection algorithm is used to extract all local maxima (points with zero first derivative and negative second derivative), which are then sorted by amplitude from largest to smallest. Peaks with amplitudes greater than the peak detection threshold are retained. One effective peak value.

[0056] S302, each peak corresponds to a propagation path, and its time delay and normalized attenuation coefficient are recorded.

[0057] Specifically, for the p-th effective peak (p= Record its latency: And the normalized attenuation coefficient: ,in The cross-correlation value at zero delay (i.e. (Related to its own zero latency), representing the reference energy without decay.

[0058] S303, the path is classified into modes based on the geographical information and propagation characteristics of the photovoltaic power station. The mode classification includes at least one of direct radiation, reflected radiation, conducted common mode, conducted differential mode, and near-field coupling.

[0059] Specifically, the theoretical delay of direct light is calculated, and each path is classified into modes based on the difference between the measured delay and the theoretical delay, as well as the attenuation variation with distance.

[0060] For example, modality classification is performed on each detected path, specifically including: Using the geographic information system (equipment coordinates, cable wiring diagram) of the photovoltaic power station, the spatial coordinates of the interference source n and the acquisition node j are obtained, and the straight-line distance is calculated. The theoretical time delay of direct radiation is obtained as follows: c is the known speed of light, which is matched with the actual detected peak value for classification; Set delay tolerance error ,in Given the sampling frequency (i.e., half a sampling period), for each detected valid peak value p: If the difference between the detection delay and the theoretical direct-shot delay is less than the preset tolerance error, then... and attenuation coefficient If the distance is approximately inversely proportional to the square of the distance, it is determined to be a direct radiation mode; If the delay is significantly greater than (Right now And attenuation coefficient If it is approximately inversely proportional to the fourth power of the distance, it is determined to be a reflected radiation mode; If the delay is extremely small ( If the frequency characteristics are similar to the cable length divided by the speed of electrical signal propagation in the cable, then it can be classified as conducted common mode or conducted differential mode (if the frequency characteristics show common mode interference (significant low-frequency components), it is determined to be conducted common mode; if the frequency characteristics show differential mode (significant high-frequency components), it is determined to be conducted differential mode). If the delay is negligible If it exists between adjacent devices (for example, the distance between devices is less than 1 meter), it is determined to be a near-field coupling mode; Finally, for each interference source-collection node pair All detected path parameters are organized into a vector set to form propagation path parameters (including multipath information): ,in The path mode label is encoded as an integer: 1-direct, 2-reflection, 3-common mode, 4-differential mode, 5-near field. For the number of paths, if =0 (no significant path), then An empty set indicates that the interference source has no significant impact on the acquisition node; if only a single path exists ( If the propagation path parameter is 1, then the propagation path parameter contains only the attenuation coefficient and time delay of one propagation mode; if multipath exists, then it contains the parameters of multiple propagation modes.

[0061] Therefore, this multipath cross-correlation analysis mechanism can accurately characterize the propagation characteristics of interference signals, providing accurate topological edge information for subsequent graph neural networks.

[0062] S104: Input the prior operating state parameters and propagation path parameters of each identified interference source into the pre-trained graph neural network model, output the contribution of each interference source to each acquisition node, and generate a list of interference source tracing.

[0063] In some embodiments, step S104 specifically includes: inputting the prior operating state parameters corresponding to the physical device identifiers of each interference source identified in step S102, and the propagation path parameters from each interference source to each acquisition node estimated in step S103, into a pre-trained graph neural network model. The graph neural network model is established based on the interference propagation topology graph constructed by the propagation path parameters, and outputs the contribution of each interference source to each acquisition node, thereby generating an interference source tracing list.

[0064] Specifically, the structure of the pre-trained graph neural network model is constructed based on the following method: Using the identified interference sources as source nodes, the collection nodes as the collection node set, and the propagation path parameters as directed edge features, a directed bipartite graph is constructed. The prior operating state parameters of each interference source are used as source node features, and the location coordinates and device type codes of the acquisition nodes are used as acquisition node features. A message-passing architecture is adopted. The source node concatenates the features and edge features to generate a message and passes it to the acquisition node. The acquisition node aggregates the message and updates the node state through the attention mechanism, and outputs the predicted received power of each acquisition node and the contribution power of each interference source. The model is trained using the actual acquired received power as the supervision signal.

[0065] For structural design, as a specific example: All physical interference sources identified in step S102 are used as the source node set. Using all data collection nodes as the data collection node set The propagation path obtained in step S103 is used as the directed edge set. There exists a directed edge from source node n to collection node j if and only if (i.e., at least one propagation path exists); the initial features of the edges include the attenuation coefficient vector, time delay vector, and modality label vector of all paths between the pair; construct a directed bipartite graph. ; Define the source node feature vector This includes prior operating status parameters (including switching frequency, output power, and temperature) obtained from the monitoring system, as well as the operating status labels matched in step S102. All features need to be normalized. Define the feature vector of the acquisition node : Contains the location coordinates of the data acquisition node And device type encoding (e.g., switch is encoded as 1, RTU is encoded as 2, etc.); Define edge characteristics: For each edge Its initial features are the multipath parameters obtained in step S103, organized as a fixed-length vector. Since the number of paths may differ between different pairs, the dimension is fixed in the following way: the path with the largest attenuation coefficient among all paths is selected. Path ( The maximum number of paths to retain between a single pair of nodes is preset, set based on the number of significant multipaths observed in field tests. In this embodiment, it is set to 3. If the actual number of paths is insufficient... (If insufficient, fill with zeros). The characteristics of each path are: After splicing, we get 3D edge feature vector; Design a graph neural network model using a message-passing architecture, comprising the following layers: Input layer: receives source node features, collected node features, and edge features; Message passing layer: For each edge , source node features The message vector is generated by concatenating the edge features with the data and passing it through a two-layer fully connected network (with ReLU activation function). The network weights are initialized uniformly using Xavier.

[0066] Aggregation layer: Node j uses an attention mechanism to aggregate all incoming edge messages. ,in The query vector is learned and uniformly initialized using Xavier; Let j be the set of all source nodes pointing to node j; Update layer: This layer updates the aggregated messages. Characteristics of the data acquisition node itself The nodes are spliced ​​together, and their states are updated via gated loop units (GRUs). Output layer: A linear layer maps the state of the acquisition node to a scalar, and outputs the predicted received power at that node. Meanwhile, the contribution power of each source node to the node is obtained through attention weight decomposition. .

[0067] It should be noted that in the message passing layer of the graph neural network, for each edge from the source node n to the acquisition node j, the model calculates an attention weight. This attention weight reflects the relative contribution of source n to the total interference of node j under the current input (between 0 and 1, and for a fixed j, the sum of the attention weights of all n is 1); then, the model predicts the total power. This attention weight is assigned to each source. The contribution power is not directly trained by the supervision signal, but is obtained by multiplying the total power by the attention weights. The attention weights are part of the model's learnable parameters. During training, in order to minimize the prediction error of the total power, the model will automatically adjust the attention weights so that the sum of the contributions of each source after allocation equals the total prediction value, which conforms to the physical constraint of total energy conservation.

[0068] In some embodiments, the pre-trained graph neural network model is trained as follows: Collect historical operational data, including source node characteristics, acquisition node characteristics, edge characteristics, and actual acquired received power within multiple historical monitoring windows. (Depend on The mean square power is calculated, and the actual collected interference power (received power) is used as the supervision signal. The loss function is the mean square error between the predicted value and the true value plus a regularization term for the path attenuation coefficient. After training, a pre-trained graph neural network model is obtained. For example, the training set size is... For each sample, the loss function is defined as: The first term is the mean squared error, and the second term is the L1 regularization term. The regularization coefficient (preferably 0.001) is used to encourage sparse propagation paths; the Adam optimizer is used, with a learning rate of 0.001, a batch size of 32, and 200 training epochs. After training, a pre-trained graph neural network model is obtained.

[0069] In some embodiments, S104 specifically includes: For the current monitoring window, steps S101 to S103 are executed to obtain the identified interference sources within the current window, along with their prior operating state parameters and propagation path parameters. These parameters are organized into a graph structure and input into a pre-trained graph neural network model. The model then performs forward computation to obtain the predicted received power at each acquisition node j. And the contribution power of each source node n to the acquisition node j ; Define the contribution of interference source n to acquisition node j: ,satisfy The contribution rate represents the proportion of the total interference power contributed by interference source n to the acquisition node j. It is obtained by dividing the contribution power of each interference source to the node by the sum of the contribution power of all interference sources. For each data acquisition node (usually located around the communication equipment), an interference source tracing list is generated, sorted by contribution from largest to smallest. List items include: interference source device identifier. Contribution percentage Propagation path parameters (attenuation coefficient of the dominant mode, i.e., the path with the largest contribution) and latency and modal tags ), matching confidence If a source of interference has multiple propagation paths, the contribution percentage of each path will be listed separately. It should be noted that this list is dynamically updated every monitoring window, forming a continuous time-series source tracing record.

[0070] In some embodiments, the method further includes: Based on contribution and propagation path parameters, set up multi-level early warning systems: When the contribution of a certain interference source to critical communication equipment exceeds the first threshold, a level one warning is triggered, and it is recommended to conduct electromagnetic compatibility retesting or install a filter on the interference source. When the attenuation coefficient of a certain propagation path continues to decrease (for multiple consecutive monitoring windows) and the rate of change exceeds the second threshold, a level two warning is triggered, indicating that there may be damage to the shielding layer or an increase in grounding impedance. It is recommended to physically shield the path or reroute it. When the cumulative power ratio of the unknown source components in step S101 (i.e., the ratio of the sum of the power of all unknown sources to the total interference power) exceeds the third threshold, a level three warning is triggered, indicating that there may be a new interference source or equipment abnormality, and manual investigation is recommended.

[0071] For example, the first threshold is determined according to the power plant safety level, with a typical value set to 30%, the second threshold is a decrease of 5% per day (meaning that the decrease rate exceeds 5% / day for multiple consecutive monitoring windows), the third threshold is set to 15%, and the multiple consecutive monitoring windows are set to three consecutive monitoring windows. Each monitoring window is set according to the actual monitoring scenario requirements, for example, each monitoring window is set to 60 seconds.

[0072] It should be noted that all source tracing results, early warning information, and heat maps of transmission routes are included. Figure 1 And push it to the power plant operation and maintenance center visualization interface. The method for drawing the propagation path heatmap is as follows: mark the locations of all interference sources and communication devices on the power plant geographic information system base map, and for each pair (n,j), if If the value is greater than 0.05, then a directed line is drawn between the two points, and the line thickness is equal to the value of the line drawn between the two points. Proportional, with color indicating the dominant mode (red - radiation, blue - conduction, green - near field).

[0073] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: In the field of communication interference in photovoltaic power plants, a complete technical closed loop has been achieved, from interference source separation, identification, multipath propagation parameter estimation to contribution tracing. First, by incorporating prior switching frequencies as guiding information into time-frequency sparse blind separation, the problem of interference source separation under underdetermined mixed conditions is solved, avoiding the failure of general blind separation algorithms in low signal-to-noise ratio environments. Second, by using fingerprint feature databases and weighted cosine similarity matching, physically meaningless separation components are mapped to specific actual devices such as inverters and string devices, achieving reliable identification of interference sources. Third, by utilizing a multipath cross-correlation analysis mechanism, through cross-correlation multi-peak detection and modal classification, the propagation path of communication interference in photovoltaic power plants is decomposed into multiple physical modes for the first time, solving the problem of severe distortion of the single-path assumption in complex electromagnetic environments. It not only estimates a single propagation path but also decomposes the attenuation coefficients and time delays of multiple modes such as direct, reflected, and conducted propagation, and performs modal classification based on geographical information, accurately characterizing the multipath propagation characteristics in complex electromagnetic environments. Fourth, by combining graph neural networks with propagation path parameters and prior operating state parameters, a nonlinear mapping from source state to interference contribution of communication devices is achieved through a pre-trained model, enabling rapid output of source tracing results without online training.

[0074] In summary, time-frequency separation provides clean source components for identity recognition, identity recognition identifies the target source for multipath analysis, multipath analysis provides accurate topological edge information for graph neural networks, and graph neural networks ultimately output contribution and a source list.

[0075] Example 2: The activity of interference sources in photovoltaic power plants often has regularity. The output power of inverters varies with the daily light intensity, there are coordinated fluctuations between multiple inverters, and some interferences are aggravated under specific weather conditions. If these spatiotemporal activity patterns can be explored, it is possible to predict interference before it occurs and take preventive measures in advance.

[0076] In some embodiments, the method further includes: S401, for each identified interference source, obtain its activity sequence within a preset time period during a continuous monitoring window, where the activity is the maximum contribution of the interference source to all acquisition nodes.

[0077] Specifically, the activity level of the interference source n within the monitoring window t is defined. The maximum contribution of this source to all communication devices is: For each source, the accumulated activity sequence is obtained. , where T is the total number of monitoring windows within the preset time period.

[0078] S402, perform time series decomposition on the activity sequence to extract the baseline activity component, periodic component and random residual component.

[0079] Specifically, the time series decomposition method is used to break down the activity series into three components: a baseline activity component, a periodic component, and random residuals. in, The baseline activity component, reflecting the inherent interference level of the equipment (determined by equipment type and installation location), is set as the time-domain average of the activity sequence. For example, the calculation formula is: ; For periodic components, the daily or weekly periodic components are extracted using Fourier transform, and expressed as the sum of the cosine and sine functions of each harmonic. For example, the daily periodic component... , The number of monitoring windows per day, where H is the harmonic order, ranging from 3 to 5, and the coefficient... , Obtained through least squares fitting; The random residual, representing random bursts of disturbance, is set as the remainder after subtracting the baseline component and the periodic component from the original activity: .

[0080] S403 calculates the cross-correlation function between the activity levels of different interference sources, explores spatial cooperation patterns between sources, and constructs an inter-source correlation matrix.

[0081] Specifically, the mining of the inter-source spatial cooperation pattern includes: Calculate the cross-correlation function of the activity sequences of each pair of interference sources under different time delays, and take the maximum absolute value of the cross-correlation function as the correlation strength. When the correlation strength exceeds a preset coordination threshold... When a spatial cooperative relationship exists between the two interference sources, it is determined that such a relationship exists; the elements of the inter-source correlation matrix represent the correlation strength between each pair of interference sources. For example, a preset cooperative threshold is used. The method for determining the threshold can be as follows: For pairs of interference sources in the power plant that have no physical association (e.g., are far apart and have no electrical connection), calculate the association strength distribution and take the 95th percentile as the threshold. In this embodiment, the typical value is 0.7. The cross-correlation function is set as follows: like (e.g., 0.7) indicates that both sources are active simultaneously, possibly driven by the same external factor (e.g., light); if The presence of a significant peak at non-zero latency indicates that the activity of one source can lag behind that of another (e.g., upstream inverter interference is transmitted downstream through the grid).

[0082] The cooperative mode is parameterized, and the total number of currently identified physical interference sources is denoted as . Construct the inter-source correlation matrix The element in the nth row and mth column of the matrix is That is, the correlation strength between the interference source pairs (n, m); when the correlation strength When the interference source n and the interference source m are determined to have a spatial cooperative relationship, it is determined that there is a spatial cooperative relationship between them.

[0083] S404 integrates external environmental features and uses a long short-term memory network to construct a prediction model. It uses historical activity sequences and current external environmental features as inputs to predict the activity of interference sources within a future time window.

[0084] The activity of interference sources in photovoltaic power plants is significantly affected by the external environment. An external environmental feature vector e(t) is introduced, which includes solar irradiance I(t) and ambient temperature. ,humidity The data are obtained from the power plant's weather station, including at least one of the following: wind speed W(t).

[0085] The input layer of the Long Short-Term Memory network receives a concatenated vector of the activity sequence of L historical monitoring windows (continuous monitoring windows within a preset time period, representing the number of past monitoring windows used for prediction) and the current external environment features. Where L is the number of historical windows, set according to the sampling period: if the monitoring window is 60 seconds, then L=24 (corresponding to the past 24 minutes) or L=60 (corresponding to the past 1 hour); the number of hidden layer units is set to 64, and time-dependent information is passed through a gating mechanism; the output layer is the predicted activity level of the next monitoring window. The training loss function is the mean square error between the predicted and measured values, and a constraint with the inter-source correlation matrix as the regularization term is added to maintain the inter-source correlation structure in the prediction results.

[0086] As an example, a predictive model is built based on the activity sequence of L consecutive monitoring windows within a preset time period. Using the current external feature vector as input, predict the activity level of the next monitoring window. The algorithm employs a Long Short-Term Memory (LSTM) network structure, whose gating mechanism can capture long-term dependencies. The input layer of the LSTM receives a concatenated vector of external features and historical activity, the hidden layer states convey temporal information, and the output layer predicts activity. Considering the inter-source synergy effect, the correlation matrix R obtained in step S403 is added as a regularization term to the loss function of the prediction model to maintain the inter-source correlation structure in the prediction results. The second term is a regularization term, which encourages strongly correlated interference sources to keep their predicted values ​​similar, and λreg is the regularization coefficient (typically 0.01). It should be noted that this is only an example; for specific details and principles, please refer to relevant existing technologies, which will not be elaborated upon in this invention.

[0087] Specifically, the prediction model is trained as follows: Select continuous data from the power plant's historical operating data. A monitoring window, for each time point t( ), construct input samples The corresponding label is the measured activity level in the next monitoring window. Actual activity level As calculated in real time in step S401, samples from all interference sources n are merged to form a training set; the training set is then used to train the prediction model. During actual online prediction, the model uses the latest measured activity levels of the L windows and the current external environment features to output the predicted activity level for the next window.

[0088] S405: When the predicted activity level exceeds the preset threshold, a prediction warning is triggered and proactive intervention suggestions are generated.

[0089] Specifically, the triggering condition for the prediction warning is that the lower bound of the confidence interval of the predicted activity exceeds a preset activity threshold.

[0090] The confidence interval is obtained by quantifying the prediction uncertainty using the Monte Carlo dropout method. For example, the Monte Carlo dropout method is used to quantify the prediction uncertainty: the number of forward propagations in the Monte Carlo dropout method is denoted as K, and in this embodiment, K=100. During testing, dropout is enabled (dropout rate set to 0.2), and K forward propagations are performed to obtain the predicted activity set. Calculate the mean and standard deviation: Take the lower bound of the confidence interval as When this lower bound exceeds the preset activity threshold... When the activity threshold is reached, a prediction alert is triggered. To determine whether an alert needs to be triggered, the threshold is determined based on the distribution of interference intensity that requires intervention in historical data. For example, the activity values ​​corresponding to interference events that were actually intervened in by maintenance personnel in historical data are statistically analyzed, and the lower quartiles of these values ​​are taken as the threshold. In this embodiment, the typical value is 0.5. This invention will not elaborate on or limit this.

[0091] The recommended proactive interventions include: Dynamically adjust the inverter switching frequency: temporarily shift the inverter switching frequency with high predictive activity. To avoid the resonant points of the operating frequency band of communication equipment; Actively activate backup communication links: During periods when interference is predicted to occur, switch critical control signals to anti-interference links such as fiber optic cables; Scheduled maintenance tasks: High-interference periods are scheduled as low-priority communication task windows to avoid critical operations.

[0092] In some embodiments, the method may further include prediction performance evaluation and online model updates: After each actual monitoring window is completed, the predicted activity level will be... Compared with measured activity In comparison, the prediction error was calculated. An online learning approach was used, and the LSTM model was fine-tuned with data from the most recent period (the last 7 days) in each preset period (24 hours) to adapt it to seasonal changes and equipment aging. The periodic components were also recalculated. Fourier coefficients are used to update the baseline activity. ; In some embodiments, the method may further include long-term trend analysis and equipment health assessment: Extract the baseline activity component of each interference source and its trend over time (calculated by monthly sliding window). When the baseline activity increases by more than 20% of the preset health threshold within a preset time period (three months), it is determined that the electromagnetic compatibility performance of the interference source device has degraded, and preventive maintenance suggestions are generated (such as cleaning heat dissipation, checking filter capacitors, and replacing aging components).

[0093] Therefore, based on multipath decomposition, this paper introduces time series decomposition, spatial cooperative pattern mining, and LSTM prediction models into the field of photovoltaic power plant communication interference monitoring for the first time, achieving a technological leap from passive source tracing to active prediction. It identifies the daily and weekly cycles of interference source activity and extracts periodic components using Fourier decomposition; discovers inter-source cooperative effects through cross-correlation analysis and uses the correlation matrix as a regularization term to constrain the prediction results; integrates external environmental characteristics (sunlight, temperature, humidity) as prediction inputs, enabling the model to adapt to weather changes; and introduces quantification of prediction uncertainty to avoid blind early warnings. Through this embodiment, maintenance personnel can be aware of the risk several hours before interference occurs and take preventative measures such as adjusting switching frequencies and link switching in advance, significantly reducing the probability of communication interruption and improving the continuity and reliability of photovoltaic power plant operation.

[0094] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for monitoring interference in communication equipment of a photovoltaic power station, characterized in that, include: S101, collect synchronous mixed interference signals from multiple acquisition nodes in the photovoltaic power station, perform time-frequency separation on the synchronous mixed interference signals, and combine the prior operating state parameters of each potential interference source as guidance to decompose several candidate interference source components. S102, extract the fingerprint feature vector of each candidate interference source component, match and identify it with the pre-built interference source fingerprint feature database, and determine the physical device identifier corresponding to each candidate interference source. S103, using a preset multipath cross-correlation analysis mechanism, obtain the propagation path parameters from each identified interference source to each acquisition node. The propagation path parameters include the attenuation coefficient and time delay of at least one propagation mode. S104: Input the prior operating state parameters and propagation path parameters of each identified interference source into the pre-trained graph neural network model, output the contribution of each interference source to each acquisition node, and generate a list of interference source tracing.

2. The method for monitoring interference in photovoltaic power station communication equipment according to claim 1, characterized in that, The synchronous mixed interference signal is set as a multi-channel mixed signal matrix within the monitoring window, where each channel corresponds to the received time-domain signal of one acquisition node. The multi-channel mixed signal matrix is ​​represented as follows: Where M is the total number of data collection nodes. Let be the time-domain signal received by the j-th node.

3. The method for monitoring interference in photovoltaic power station communication equipment according to claim 2, characterized in that, The prior operating state parameters include at least the switching frequency of each inverter, and at least one auxiliary parameter among the output frequency, operating temperature, and operating time.

4. The method for monitoring interference in photovoltaic power station communication equipment according to claim 3, characterized in that, The process of performing time-frequency separation on the synchronous mixed interference signal, and using the prior operating state parameters of each potential interference source as guidance, decomposes it into several candidate interference source components, specifically including: Using the switching frequency and its harmonics to form a set of guiding frequencies, extract time-frequency points with significant energy in the time-frequency domain, calculate the spatial direction vector of each time-frequency point, perform density clustering based on the direction vector, reconstruct the time-domain waveform of the cluster to obtain candidate interference source components, and calculate the correlation degree between each candidate component and the set of guiding frequencies, and sort them according to the correlation degree.

5. The method for monitoring interference in photovoltaic power station communication equipment according to claim 4, characterized in that, The step of using the switching frequency and its harmonics to construct a set of guiding frequencies, and extracting time-frequency points with significant energy in the time-frequency domain, specifically includes: For multi-channel mixed signal matrix Perform a short-time Fourier transform to obtain the time-frequency representation. Where f is the frequency variable, For time frame indexing; Obtain the switching frequency of each inverter And its integer multiples of harmonic frequencies, forming a set of known switching frequencies and their harmonics. ; For each boot frequency In time-frequency representation Extract the set of time-frequency points with significant energy. , The average energy across all time-frequency points. ,in For frequency points, For time frames, This is the preset energy threshold coefficient.

6. The method for monitoring interference in photovoltaic power station communication equipment according to claim 5, characterized in that, The calculation of the correlation between each candidate component and the set of guiding frequencies specifically includes: For each reconstructed candidate interference source component Calculate its spectral peak frequency set Combine it with the set of guiding frequencies Matching to obtain the correlation The degree of correlation is defined as: Wherein, the numerator represents the number of elements in the intersection of the peak frequency and pilot frequency set of the source component, and the denominator represents the total number of peak frequencies of the source component.

7. The method for monitoring interference in photovoltaic power station communication equipment according to claim 6, characterized in that, The preset multipath cross-correlation analysis mechanism specifically includes: S301, the time-domain waveform separated from the identified interference source is used as the transmitted signal estimate, the received component related to the interference source is extracted from the original received signal of each acquisition node, the cross-correlation function of the two is calculated and multiple effective peaks in the cross-correlation function are detected; S302, each peak corresponds to a propagation path, and its time delay and normalized attenuation coefficient are recorded; S303 classifies the path by mode based on the geographical information and propagation characteristics of the photovoltaic power station.

8. The method for monitoring interference in photovoltaic power station communication equipment according to claim 7, characterized in that, The propagation path parameters are further defined as follows: for each interference source-acquisition node pair All detected path parameters are organized into a vector set to form the propagation path parameters: ,in For path modal labels, For the number of paths, The normalized attenuation coefficient, For time delay.

9. The method for monitoring interference in photovoltaic power station communication equipment according to claim 8, characterized in that, S104 specifically includes: inputting the prior operating state parameters corresponding to the physical device identifiers of each interference source identified in S102, and the propagation path parameters from each interference source to each acquisition node estimated in S103, into a pre-trained graph neural network model. The graph neural network model is established based on the interference propagation topology graph constructed by the propagation path parameters, and outputs the contribution of each interference source to each acquisition node, thereby generating an interference source tracing list.

10. The method for monitoring interference in photovoltaic power station communication equipment according to claim 9, characterized in that, S104 specifically includes: For the current monitoring window, execute steps S101 to S103 to obtain the identified interference sources within the current window, along with their prior operating state parameters and propagation path parameters. These parameters are then input into the pre-trained graph neural network model to obtain the predicted received power at each acquisition node j. And the contribution power of each interference source n to the acquisition node j ; Define the contribution of interference source n to acquisition node j: divide the contribution power of each interference source to the node by the sum of the contribution power of all interference sources to obtain the contribution of each interference source.

Images