Photovoltaic system island detection method and system based on santlet transform and ridglet probabilistic neural network

Abstract

This invention discloses a photovoltaic system islanding detection method and system based on Santlet transform and Ridglet probabilistic neural network. The method includes: collecting historical voltage data of the photovoltaic inverter, preprocessing it, and then sequentially acquiring the data to be analyzed using a sliding window; performing time-frequency analysis using Santlet continuous wavelet transform to obtain the corresponding time-frequency spectrum; extracting multi-dimensional time-frequency feature vectors and inputting them into a trained Ridglet probabilistic neural network classification model to obtain the corresponding probability values; if the probability value is greater than a probability threshold, it is determined that the photovoltaic system has experienced islanding, and a trip signal is generated and sent to the grid-connected circuit breaker to disconnect it. This invention significantly improves the accuracy, speed, and reliability of islanding detection, effectively reduces the detection blind zone, and has good adaptability to complex power grid environments.

Description

Technical Field

[0001] This invention relates to photovoltaic power generation grid connection technology in power systems, specifically to a photovoltaic system islanding detection method and system based on Santlet transform and Ridglet probabilistic neural network. Background Technology

[0002] As the global energy structure transitions towards cleaner and lower-carbon energy, solar photovoltaic (PV) power generation, as a core component of distributed energy systems, is experiencing a massive surge in grid connection with rapidly increasing penetration rates. However, the inherent method of connecting PV systems to the grid via power electronic inverters introduces the serious safety challenge of the "islanding effect." The islanding effect refers to the situation where, when the main grid experiences a power outage due to a fault or planned power failure, the distributed PV system fails to detect this in time and continues to independently supply power to local loads, forming a self-consistent "power island" uncontrolled by grid dispatch. This phenomenon not only poses a fatal risk of electric shock to on-site maintenance personnel but can also lead to equipment damage and relay protection malfunctions due to unplanned islanding operation and reconnection impacts, seriously threatening the safe and stable operation of the regional distribution network. Therefore, developing an islanding detection method that is highly reliable, adaptable, and does not affect power quality has become a key technological prerequisite for promoting the high-quality development of the PV industry and ensuring the security of smart grids.

[0003] Existing islanding detection methods are mainly divided into active and passive approaches, but both have inherent drawbacks. Passive methods identify islanding by monitoring parameters such as voltage and frequency at the point of common coupling (PCC) to see if they exceed limits. While simple in structure and not affecting power quality, they suffer from a fatal "non-detection zone (NDZ)" problem: when the photovoltaic output power is close to the local load power, the electrical parameters change only slightly after a grid power outage, failing to trigger the threshold alarm and leading to detection failure. Furthermore, this method is susceptible to grid noise and nonlinear load interference, resulting in a high false alarm rate. Active methods, on the other hand, actively detect islanding by injecting small disturbances (such as frequency shifts or power disturbances) into the grid and observing their feedback, effectively reducing NDZ. However, such active disturbances degrade output power quality, causing current harmonic distortion; in multi-inverter parallel operation, disturbance signals may interfere with each other and fail, and there is a risk of system oscillation under weak grid conditions. Although communication-based methods are highly accurate, their high cost and reliance on external links make them difficult to popularize in distributed scenarios.

[0004] In summary, existing technologies have not fundamentally solved the core challenges of passive intelligent detection methods in terms of feature engineering robustness, model generalization ability, computational real-time performance, and adaptive decision thresholds, resulting in a dilemma of "insufficient reliability of passive methods and poor power quality of active methods." The core problem lies in the fact that the threshold judgment mechanism relied upon by passive methods struggles to capture weak fault characteristics under power balance and lacks intelligent discrimination capabilities against interference signals; while active methods sacrifice power quality for improved reliability. Therefore, there is an urgent need in this field to develop a completely new technical approach that must, within a passive framework completely free of grid disturbances, overcome the limitations of traditional threshold criteria and fundamentally solve the NDZ and disturbance immunity problems through more refined feature extraction and a more intelligent decision-making mechanism. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a photovoltaic system islanding detection method and system based on Santlet transform and Ridglet probabilistic neural network, which significantly improves the accuracy, speed and reliability of islanding detection, effectively reduces the detection blind zone and has good adaptability to complex power grid environments.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for detecting islands in photovoltaic systems based on Santlet transform and Ridglet probabilistic neural network includes the following steps: S101) Collect historical voltage data of the photovoltaic inverter; S102) The voltage history data is preprocessed. A sliding window is used to sequentially obtain data of a specified time window from the preprocessed voltage history data as the data to be analyzed. The data to be analyzed in the current time window is subjected to time-frequency analysis using Santlet continuous wavelet transform to obtain the corresponding time-frequency spectrum. Multi-dimensional time-frequency feature vectors are extracted from the time-frequency spectrum. S103) The extracted multi-dimensional time-frequency feature vector is input into the trained Ridglet probabilistic neural network classification model to obtain the probability value corresponding to the current time window; (S104) If the probability value is greater than the preset probability threshold, it is determined that the photovoltaic system has an islanding effect, and a trip signal is generated and sent to the grid-connected circuit breaker to disconnect it; if the probability value is less than the preset probability threshold, it is determined that the system is in normal grid-connected operation, the data to be analyzed in the next time window is obtained, and time-frequency analysis is performed using Santlet continuous wavelet transform to obtain the corresponding time-frequency spectrum. After extracting the multi-dimensional time-frequency feature vector from the time-frequency spectrum, the process jumps to step S103 to obtain the probability value of the next time window.

[0007] Furthermore, when using a sliding window to sequentially obtain data for a specified time window from the preprocessed historical voltage data as the data to be analyzed, specifically, the preprocessed historical voltage data is stored in a first-in-first-out (FIFO) buffer queue of a specified size in chronological order until the buffer queue is full, thus obtaining the data to be analyzed for the current time window. The probability value of the data to be analyzed for the current time window is then waited for and obtained. If the probability value is less than a preset probability threshold, the earliest data stored in the buffer queue is removed according to a preset ratio, and the preprocessed historical voltage data is stored in the buffer queue in chronological order again until the buffer queue is full, thus obtaining the data to be analyzed for the next time window.

[0008] Furthermore, when performing time-frequency analysis using the Santlet continuous wavelet transform to obtain the corresponding time-frequency spectrum, a scale sequence consisting of multiple scale values ​​is established based on the frequency range of the data to be analyzed and the center frequency of the Santlet mother wavelet. For each scale in the scale sequence, the Santlet transform coefficients of the data to be analyzed are calculated on the entire time axis, forming a two-dimensional complex coefficient matrix as the time-frequency spectrum. The mathematical expression is as follows:

[0009] Where CWT(s,τ) is the time-frequency coefficient matrix of the data to be analyzed in the current time window, s is the scaling factor, and τ is the time shift. This represents the data to be analyzed within the current time window. It is the complex conjugate of the Santlet wavelet, which is the mother wavelet. It is a time variable; the time-domain expression of the Santlet wavelet is:

[0010] in, Here, 'a' is the Sinc function, and 'a' is a frequency adjustment parameter used to control the main lobe width of the Sinc function. For bandwidth adjustment parameters, This represents the convolution operation.

[0011] Furthermore, when extracting multi-dimensional time-frequency feature vectors from the time-spectrum graph, the process includes: Calculate the coefficient energy in the fundamental frequency band and the main odd harmonic frequency band of the time-frequency coefficient matrix CWT(s,τ) to obtain the frequency band energy characteristics; Calculate the Shannon entropy of the time-frequency coefficient matrix CWT(s,τ) to obtain the time-frequency entropy characteristics; The instantaneous frequency is estimated by extracting the modulus maxima at each time point in the scale domain of the time-frequency coefficient matrix CWT(s,τ), and the variance or standard deviation of the instantaneous frequency is calculated to obtain the instantaneous frequency stability characteristics. Calculate the energy ratio of odd harmonics to even harmonics in the time-frequency coefficient matrix CWT(s,τ), or calculate the rate of change of odd and even harmonics relative to the fundamental wave energy to obtain the characteristics of odd and even harmonic variation.

[0012] Furthermore, after extracting the multi-dimensional time-frequency feature vector from the time-frequency spectrum, the method further includes: using principal component analysis or linear discriminant analysis to reduce the dimensionality of the extracted multi-dimensional time-frequency feature vector, and selecting the feature combination most sensitive to the island state as the final multi-dimensional time-frequency feature vector.

[0013] Furthermore, the Ridglet probabilistic neural network classification model includes an input layer, a pattern layer, a summation layer, and an output decision layer. The input layer receives feature data through neurons with the same dimension as the multi-dimensional time-frequency feature vector. The pattern layer uses neurons that correspond one-to-one with different categories, and uses the radial basis function of the Ridglet function as the activation function to capture the feature direction from the feature vector passed from the input layer. The summation layer sums the outputs of all pattern layer neurons belonging to the same category. The output decision layer calculates the posterior probability of the input vector belonging to each category based on the output of the summation layer, and takes the category corresponding to the maximum posterior probability as the output. The category includes an isolated category or a normal category.

[0014] Furthermore, the Ridglet probabilistic neural network classification model employs supervised learning during training, using the cluster center of each class's time-frequency feature vector as the weight vector of the neurons in the pattern layer for that class, and optimizing the smoothing parameter in the activation function through cross-validation to maximize the network's classification accuracy on the validation set.

[0015] The present invention also proposes a photovoltaic system islanding detection system based on Santlet transform and Ridglet probabilistic neural network, including a processor and a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, which is executed by the processor to implement the steps of the photovoltaic system islanding detection method based on Santlet transform and Ridglet probabilistic neural network.

[0016] The present invention also proposes a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the photovoltaic system island detection method based on Santlet transform and Ridglet probabilistic neural network.

[0017] The present invention also proposes a computer program product, including a computer program that, when executed by a processor, implements the steps of the photovoltaic system island detection method based on Santlet transform and Ridglet probabilistic neural network.

[0018] Compared with the prior art, the advantages of the present invention are as follows: This invention constructs a fully passive, high-precision detection system through the deep synergy of Santlet transform and Ridglet probabilistic neural network, fundamentally solving the NDZ and anti-interference problems of traditional methods. Specifically, by performing Santlet transform on the voltage signal through a sliding window to achieve continuous time-frequency analysis, it can accurately extract the weak harmonic distortion and instantaneous frequency jitter characteristics of the voltage under power balance conditions, significantly eliminating detection blind spots and improving accuracy. Combined with the sliding window overlapping buffer mechanism, it achieves millisecond-level continuous monitoring and rapid response, meeting the speed requirements. Furthermore, the Ridglet probabilistic neural network utilizes its sensitivity to high-dimensional feature directions to effectively distinguish between islanded states and grid disturbances such as load switching. It quantifies uncertainty through probability values ​​and combines this with adjustable thresholds to achieve reliable decision-making, significantly reducing false alarm rates and enhancing adaptability to complex operating conditions. The entire process requires no interference to the grid, ensuring zero power quality degradation and achieving a balance between accuracy, speed, and reliability. Attached Figure Description

[0019] Figure 1 This is a flowchart of a method according to an embodiment of the present invention. Detailed Implementation

[0020] The present invention will be further described below with reference to the accompanying drawings and specific preferred embodiments, but this does not limit the scope of protection of the present invention.

[0021] This embodiment proposes a photovoltaic system islanding detection method based on Santlet transform and Ridglet probabilistic neural network, aiming to solve the problems of detection blind zone and low detection accuracy under nonlinear load interference in existing islanding detection methods.

[0022] This method utilizes the Santle transform, which has excellent time-frequency localization characteristics, to perform time-frequency analysis on voltage signals and extract multi-dimensional time-frequency feature vectors that can sensitively reflect islanding states. Then, the feature vectors are input into a pre-trained Ridglet probabilistic neural network for pattern recognition. Finally, the network outputs a probability value belonging to an "islanding state." By comparing whether this probability value exceeds a set threshold, it is determined whether islanding has occurred, thereby improving the accuracy, speed, and reliability of islanding detection, effectively reducing the detection blind zone, and demonstrating good adaptability to complex power grid environments.

[0023] like Figure 1 As shown, the method includes the following steps: S101) Collect historical voltage data of the photovoltaic inverter; S102) The voltage history data is preprocessed. A sliding window is used to sequentially obtain data of a specified time window from the preprocessed voltage history data as the data to be analyzed. The data to be analyzed in the current time window is subjected to time-frequency analysis using Santlet continuous wavelet transform to obtain the corresponding time-frequency spectrum. Multi-dimensional time-frequency feature vectors are extracted from the time-frequency spectrum. S103) The extracted multi-dimensional time-frequency feature vector is input into the trained Ridglet probabilistic neural network classification model to obtain the probability value corresponding to the current time window; (S104) If the probability value is greater than the preset probability threshold, it is determined that the photovoltaic system has an islanding effect, and a trip signal is generated and sent to the grid-connected circuit breaker to disconnect it; if the probability value is less than the preset probability threshold, it is determined that the system is in normal grid-connected operation, the data to be analyzed in the next time window is obtained, and time-frequency analysis is performed using Santlet continuous wavelet transform to obtain the corresponding time-frequency spectrum. After extracting the multi-dimensional time-frequency feature vector from the time-frequency spectrum, the process jumps to step S103 to obtain the probability value of the next time window.

[0024] The following provides a detailed explanation of each step.

[0025] In step S101 of this embodiment, the three-phase voltage signal or single-phase voltage signal u(t) at the common connection point (PCC) of the photovoltaic inverter grid connection point is collected in real time through a voltage sensing device.

[0026] In step S102 of this embodiment, the acquired voltage signal u(t) is first preprocessed, including signal filtering to eliminate high-frequency noise and signal normalization to eliminate the influence of amplitude fluctuations. Then, the preprocessed voltage signal is subjected to time-frequency analysis using Santlet continuous wavelet transform to obtain the time-frequency spectrum of the signal. From the time-frequency spectrum, a set of multi-dimensional time-frequency feature vectors that can sensitively characterize changes in system state are extracted. ,in n As a feature dimension, the specific process includes: S1021) Signal preprocessing: In this embodiment, when preprocessing the historical voltage data, the signal conditioning circuit and digital filtering algorithm are used to preprocess u(t), mainly including: Anti-aliasing filtering: A low-pass filter is used to filter out high-frequency noise above the Nyquist frequency.

[0027] Power frequency bandpass filtering: retains a specific frequency band signal centered on the power frequency (50Hz) to highlight core features and suppress irrelevant interference.

[0028] Signal normalization: Normalize the voltage amplitude to a uniform range to eliminate the impact of amplitude fluctuations on subsequent feature extraction.

[0029] S1022) Sliding window caching mechanism: In this embodiment, when using a sliding window to sequentially obtain data for a specified time window from preprocessed historical voltage data as the data to be analyzed, specifically, the preprocessed historical voltage data is stored sequentially into a first-in-first-out (FIFO) buffer queue of a specified size until the buffer queue is full, thus obtaining the data to be analyzed for the current time window. The probability value of the data to be analyzed for the current time window is then waited for and obtained. If the probability value is less than a preset probability threshold, the earliest stored data in the buffer queue is removed according to a preset ratio. The preprocessed historical voltage data is then stored sequentially into the buffer queue again until the buffer queue is full, thus obtaining the data to be analyzed for the next time window. The specific process is as follows: Window parameter initialization: Set a fixed-length data cache window as the analysis window, and the time length of the window... Corresponding to N power frequency cycles, where N is an integer between 10 and 50; based on the system standard power frequency and sampling frequency The total number of sampling points L in the window is determined by the following formula: (1) Data caching and window filling: The discrete sequence u(n) of the voltage signal acquired in real time is preprocessed and stored in a first-in-first-out buffer queue of length L in sequence until the queue is filled for the first time, forming the first complete analysis window; Window data triggering processing: Whenever the cache queue is updated by new data, a subsequent time-frequency feature extraction step and the following process are triggered; the complete data sequence within the current window is then processed. As the signal block to be analyzed, it is sent to the subsequent processing module; Window sliding update: After completing data processing for the current window, the earliest M data points stored in the cache queue are removed, and space is made up for the newly collected M data points, thus achieving a sliding update of the window; where M is the step size of each window update, satisfying... With a step size M (usually...) (Achieving an overlap rate of 50%~80%) Slide the window, and each time it reaches a new window position, the voltage signal sequence within the window is... The output serves as the data block for the current analysis. This design ensures continuous and rapid detection, and effectively captures the complete characteristics of transient processes.

[0030] S1023) Santlet transform time-frequency analysis: This embodiment uses a specially designed composite wavelet—the Santlet wavelet—as the mother wavelet. Its time-domain expression is: (2) in, Here, 'a' is the Sinc function, and 'a' is a frequency adjustment parameter used to control the main lobe width of the Sinc function. This is a bandwidth adjustment parameter used to control the decay rate of the Gaussian function. This indicates a convolution operation; the wavelet combines the ideal frequency domain bandpass characteristics of the Sinc function with the excellent time domain attenuation characteristics of the Gaussian function.

[0031] Therefore, when using the Santlet continuous wavelet transform for time-frequency analysis to obtain the corresponding time-frequency spectrum, a scale sequence consisting of multiple scale values ​​is established based on the frequency range of the data to be analyzed and the center frequency of the Santlet mother wavelet. For each scale in the scale sequence, the Santlet transform coefficients of the data to be analyzed are calculated on the entire time axis, forming a two-dimensional complex coefficient matrix as the time-frequency spectrum. The process is as follows: Scale sequence setting: based on the frequency range to be analyzed and the center frequency of the Santlet mother wavelet Determine a linear or logarithmic scale sequence consisting of m scale values. ; Transformation coefficient calculation: For each scale in the scale sequence Calculate the Santlet transform coefficients of the voltage signal u(t) over the entire time axis, forming a two-dimensional complex coefficient matrix CWT with dimensions m×N, where N is the number of sampling points. The two-dimensional time-frequency coefficient matrix CWT(s,τ) is: (3) Where CWT(s,τ) is the time-frequency coefficient matrix of the data to be analyzed in the current time window, s is the scaling factor, which is inversely proportional to the frequency, and τ is the time shift. This represents the data to be analyzed within the current time window. It is the complex conjugate of the Santlet wavelet, which is the mother wavelet. It is a time variable. This transformation can clearly reveal the joint energy distribution of the signal in the time and frequency domains.

[0032] S1024) Multi-dimensional time-frequency feature extraction: In this embodiment, the extracted multi-dimensional time-frequency feature vector It should include at least the following four types of characteristics: a) Frequency band energy characteristics: Calculate the sum of squares of the coefficient moduli of the time-frequency coefficient matrix CWT(s,τ) within a specific scale band corresponding to the fundamental, third, and fifth harmonics of the power frequency; for a specific scale band... Its corresponding frequency band energy is : (4) The time-frequency feature vector It includes at least two frequency band energy characteristics, namely the fundamental frequency band energy. and one or more harmonic frequency bands of energy .

[0033] b) Time-frequency distribution entropy characteristics: Calculate the Shannon entropy of the entire time spectrum of the time-frequency coefficient matrix CWT(s,τ) or the time spectrum within a specific time window to measure the degree of dispersion of signal energy; First, the square matrix of the modulus of the coefficient matrix. Treat it as a probability distribution graph and normalize it: (5) Then, calculate the Shannon entropy of the normalized matrix: (6) Among them, the Shannon entropy value H is a component of the eigenvector. This entropy value is used to measure the degree of dispersion of signal energy in the time-frequency plane. The entropy value will change significantly when islands occur.

[0034] c) Instantaneous frequency stability characteristics: The instantaneous frequency is estimated by extracting the modulus maxima of τ at each time point in the scale domain of the time-frequency coefficient matrix CWT(s,τ), and the variance or standard deviation of the instantaneous frequency sequence within the detection time window is calculated to capture the frequency jitter characteristics when islands occur. d) Characteristics of odd and even harmonic variations: Calculate the energy ratio of odd harmonics (such as the 3rd and 5th harmonics) to even harmonics (such as the 2nd and 4th harmonics) in the time-frequency coefficient matrix CWT(s,τ), or calculate its rate of change relative to the fundamental energy, as a sensitive indicator reflecting the nonlinear load changes or islanding state of the power grid.

[0035] After calculating the above features, principal component analysis (PCA) or linear discriminant analysis (LDA) can be used to reduce the dimensionality of the initially extracted high-dimensional time-frequency feature set, and select the feature combinations most sensitive to the island state to form the low-dimensional feature vector that is finally input into the Ridglet probabilistic neural network. , which serves as the final multi-dimensional time-frequency feature vector.

[0036] In this embodiment, the time-frequency feature vector is processed through step S103. Or the time-frequency feature vector after dimensionality reduction The input is fed into a pre-trained Ridglet probabilistic neural network (RPNN) classification model, which outputs a probability value representing whether the current system state belongs to an island state. ,in .

[0037] In this embodiment, the Ridglet Probabilistic Neural Network (RPNN) classification model is a four-layer feedforward neural network structure, including an input layer, a pattern layer, a summation layer, and an output decision layer. The input layer receives feature data through neurons with the same dimension as the multi-dimensional time-frequency feature vector. The pattern layer uses neurons corresponding one-to-one with different categories, using the radial basis function of the Ridglet function as the activation function to capture the feature direction from the feature vector passed from the input layer. The summation layer sums the outputs of all pattern layer neurons belonging to the same category. The output decision layer calculates the posterior probability of the input vector belonging to each category based on the output of the summation layer, and outputs the category corresponding to the highest posterior probability. The categories include isolated categories or normal categories. The specific hierarchical relationship is as follows: Input layer: composed of feature vectors A group of neurons of the same dimension n is used to receive feature data; Pattern layer: Each class of each training sample corresponds to one neuron. The activation function of this neuron is a radial basis function based on the Ridglet function, and its expression is: (7) in, For the input feature vector, It is the weight vector connecting the input layer and the i-th pattern layer neuron of the j-th class (i.e., the center vector of that class). For smoothing parameters; Summation layer: Each category corresponds to one neuron, and the outputs of all pattern layer neurons for that category are summed and averaged; Output decision layer: Calculates the input vector based on the output of the summation layer. The posterior probability of belonging to each category is calculated, and the category corresponding to the maximum posterior probability is used as the output. In this embodiment, the output is the probability of belonging to the "island" category. .

[0038] The training process of the RPNN model is as follows: Constructing a training sample set: Collect a large amount of historical data covering normal grid connection, various islanding conditions, and common grid disturbances (such as load switching and capacitor switching) to construct a training sample set. This training sample set includes historical voltage signals collected under various conditions (including normal grid connection, islanding, load switching, and motor starting) and has accurately labeled their state categories ("normal" or "islanding"). Feature extraction: For each voltage signal sample in the training sample set, perform the aforementioned steps S101 and S102 to obtain the corresponding time-frequency feature vector; Network parameter determination: Using supervised learning, the mean of the feature vector subset of each category ("normal" or "island") or the centroid obtained through clustering is used as the weight vector of the neurons in the pattern layer for that category. The optimal smoothing parameters were determined through cross-validation. This is to maximize the classification accuracy of the network on the validation set.

[0039] This embodiment implements islanding state decision-making and protection actions through step S104: The probability value... Compared with the preset probability threshold Compare; if If the photovoltaic system experiences an islanding effect, a trip signal is generated and sent to the grid-connected circuit breaker to disconnect it; if If so, the system is determined to be in normal grid-connected operation. The process is as follows: The probability values ​​directly output by the pre-trained RPNN network With a preset decision threshold Compare. If If the condition is not met, it is determined that an islanding state has occurred, and a trip command is immediately issued; otherwise, the system continues to monitor normally.

[0040] Once the decision logic determines that the system is in an islanded state, the microprocessor will immediately generate a trip signal, which will control the grid-connected circuit breaker to quickly disconnect the photovoltaic system from the grid, thereby eliminating the islanding.

[0041] In another specific embodiment, the probability threshold It's not a fixed value, but a configurable parameter based on the system's reliability requirements, ranging from 0.5 to 0.95. A higher threshold increases the reliability of island detection, but slightly increases the risk of missed detections. An adaptive probability threshold can be set based on confidence learning, completely solving the problem of adaptability and determinism in threshold setting, thus achieving high-precision and rapid identification of island states. For example, recording the probability output of the most recent N detection cycles... and the corresponding actual test results feedback; calculate the prediction confidence level. When the average confidence level within the window is lower than the threshold And the false positive rate exceeds At that time, threshold adjustment is triggered: The direction of adjustment is determined by the type of misjudgment (false positive / false negative). This is the preset step size.

[0042] Furthermore, this embodiment also proposes a photovoltaic system islanding detection system based on Santlet transform and Ridglet probabilistic neural network, including a processor and a computer-readable storage medium. The computer-readable storage medium stores a computer program, which is executed by the processor to implement the steps of the photovoltaic system islanding detection method based on Santlet transform and Ridglet probabilistic neural network described in this embodiment.

[0043] Furthermore, this embodiment also proposes a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the photovoltaic system islanding detection method based on Santlet transform and Ridglet probabilistic neural network described in this embodiment.

[0044] Furthermore, this embodiment also proposes a computer program product, including a computer program that, when executed by a processor, implements the steps of the photovoltaic system islanding detection method based on Santlet transform and Ridglet probabilistic neural network described in this embodiment.

[0045] In summary, this invention discloses a photovoltaic system islanding detection method and system based on Santlet transform and Ridglet probabilistic neural network. First, the common coupling point voltage signal of the photovoltaic grid-connected point is acquired in real time. Second, the voltage signal is analyzed in time and frequency using the Santlet transform, which has excellent time-frequency localization characteristics, to extract a multi-dimensional time-frequency feature vector that sensitively reflects the islanding state. Then, the feature vector is input into a pre-trained Ridglet probabilistic neural network for pattern recognition. Finally, the network outputs a probability value belonging to the "islanding state." When this probability value exceeds a set threshold, islanding is determined to have occurred, and a grid disconnection command is issued. This invention combines advanced signal processing and artificial intelligence algorithms, significantly improving the accuracy, speed, and reliability of islanding detection, effectively reducing the detection blind zone, and exhibiting good adaptability to complex power grid environments. This invention achieves the following technical effects: Significantly reduced detection blind zone: The excellent frequency selectivity and time-frequency convergence of the Santlet transform can capture the weak characteristic changes of voltage signals under power balance conditions, fundamentally and effectively reducing or even eliminating the non-detection zone (NDZ).

[0046] Enhanced anti-interference capability: By introducing directional sensitivity, the Ridglet probabilistic neural network can accurately distinguish between islanded states and normal power grid disturbances (such as load switching), significantly reducing the false alarm rate and solving the problem that traditional passive methods are susceptible to nonlinear load interference.

[0047] Zero impact on power quality: This method is based entirely on a passive detection framework, which does not require the injection of any disturbance signals into the power grid, thus ensuring that the power quality of the grid-connected power is not degraded.

[0048] Real-time performance and reliability are combined: The sliding window mechanism is combined with the fast feedforward calculation of the probabilistic neural network to meet the standard requirements for detection speed, while providing a quantitative reliability index through the posterior probability output.

[0049] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0050] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for detecting islanding in photovoltaic systems based on Santlet transform and Ridglet probabilistic neural networks, characterized in that, Includes the following steps: S101) Collect historical voltage data of the photovoltaic inverter; S102) The voltage history data is preprocessed. A sliding window is used to sequentially obtain data of a specified time window from the preprocessed voltage history data as the data to be analyzed. The data to be analyzed in the current time window is subjected to time-frequency analysis using Santlet continuous wavelet transform to obtain the corresponding time-frequency spectrum. Multi-dimensional time-frequency feature vectors are extracted from the time-frequency spectrum. S103) The extracted multi-dimensional time-frequency feature vector is input into the trained Ridglet probabilistic neural network classification model to obtain the probability value corresponding to the current time window; (S104) If the probability value is greater than the preset probability threshold, it is determined that the photovoltaic system has an islanding effect, and a trip signal is generated and sent to the grid-connected circuit breaker to disconnect it; if the probability value is less than the preset probability threshold, it is determined that the system is in normal grid-connected operation, the data to be analyzed in the next time window is obtained, and time-frequency analysis is performed using Santlet continuous wavelet transform to obtain the corresponding time-frequency spectrum. After extracting the multi-dimensional time-frequency feature vector from the time-frequency spectrum, the process jumps to step S103 to obtain the probability value of the next time window.

2. The photovoltaic system islanding detection method based on Santlet transform and Ridglet probabilistic neural network according to claim 1, characterized in that, When using a sliding window to sequentially retrieve data from preprocessed historical voltage data for a specified time window as the data to be analyzed, the preprocessed historical voltage data is stored in a first-in-first-out (FIFO) buffer queue of a specified size in chronological order until the buffer queue is full, thus obtaining the data to be analyzed for the current time window. The probability value of the data to be analyzed for the current time window is then waited for and obtained. If the probability value is less than a preset probability threshold, the earliest data stored in the buffer queue is removed according to a preset ratio, and the preprocessed historical voltage data is stored in the buffer queue in chronological order again until the buffer queue is full, thus obtaining the data to be analyzed for the next time window.

3. The photovoltaic system islanding detection method based on Santlet transform and Ridglet probabilistic neural network according to claim 1, characterized in that, When performing time-frequency analysis using the Santle continuous wavelet transform to obtain the corresponding time-frequency spectrum, a scale sequence consisting of multiple scale values ​​is established based on the frequency range of the data to be analyzed and the center frequency of the Santle mother wavelet. For each scale in the scale sequence, the Santle transform coefficients of the data to be analyzed are calculated along the entire time axis, forming a two-dimensional complex coefficient matrix as the time-frequency spectrum. The mathematical expression is as follows: Where CWT(s,τ) is the time-frequency coefficient matrix of the data to be analyzed in the current time window, s is the scaling factor, and τ is the time shift. This represents the data to be analyzed within the current time window. It is the complex conjugate of the Santlet wavelet, which is the mother wavelet. It is a time variable; the time-domain expression of the Santlet wavelet is: in, Here, 'a' is the Sinc function, and 'a' is a frequency adjustment parameter used to control the main lobe width of the Sinc function. For bandwidth adjustment parameters, This represents the convolution operation.

4. The photovoltaic system islanding detection method based on Santlet transform and Ridglet probabilistic neural network according to claim 3, characterized in that, Extracting multi-dimensional time-frequency feature vectors from the time-spectrum graph includes: Calculate the coefficient energy in the fundamental frequency band and the main odd harmonic frequency band of the time-frequency coefficient matrix CWT(s,τ) to obtain the frequency band energy characteristics; Calculate the Shannon entropy of the time-frequency coefficient matrix CWT(s,τ) to obtain the time-frequency entropy characteristics; The instantaneous frequency is estimated by extracting the modulus maxima at each time point in the scale domain of the time-frequency coefficient matrix CWT(s,τ), and the variance or standard deviation of the instantaneous frequency is calculated to obtain the instantaneous frequency stability characteristics. Calculate the energy ratio of odd harmonics to even harmonics in the time-frequency coefficient matrix CWT(s,τ), or calculate the rate of change of odd and even harmonics relative to the fundamental wave energy to obtain the characteristics of odd and even harmonic variation.

5. The photovoltaic system islanding detection method based on Santlet transform and Ridglet probabilistic neural network according to claim 4, characterized in that, After extracting the multi-dimensional time-frequency feature vector from the time-frequency spectrum, the method further includes: using principal component analysis or linear discriminant analysis to reduce the dimensionality of the extracted multi-dimensional time-frequency feature vector, and selecting the feature combination most sensitive to the island state as the final multi-dimensional time-frequency feature vector.

6. The photovoltaic system islanding detection method based on Santlet transform and Ridglet probabilistic neural network according to claim 1, characterized in that, The Ridglet probabilistic neural network classification model includes an input layer, a pattern layer, a summation layer, and an output decision layer. The input layer receives feature data through neurons with the same dimension as the multi-dimensional time-frequency feature vector. The pattern layer uses neurons that correspond one-to-one with different categories, and uses the radial basis function of the Ridglet function as the activation function to capture the feature direction from the feature vector passed from the input layer. The summation layer sums the outputs of all pattern layer neurons belonging to the same category. The output decision layer calculates the posterior probability of the input vector belonging to each category based on the output of the summation layer, and outputs the category corresponding to the maximum posterior probability. The category includes an isolated category or a normal category.

7. The photovoltaic system islanding detection method based on Santlet transform and Ridglet probabilistic neural network according to claim 6, characterized in that, The Ridglet probabilistic neural network classification model employs supervised learning during training. It uses the cluster center of the time-frequency feature vector of each class as the weight vector of the neurons in the pattern layer of that class, and optimizes and determines the smoothing parameter in the activation function through cross-validation to maximize the classification accuracy of the network on the validation set.

8. A photovoltaic system islanding detection system based on Santlet transform and Ridglet probabilistic neural network, characterized in that, The device includes a processor and a computer-readable storage medium storing a computer program, which is executed by the processor to implement the steps of the photovoltaic system islanding detection method based on Santlet transform and Ridglet probabilistic neural network as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the photovoltaic system islanding detection method based on Santlet transform and Ridglet probabilistic neural network as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, The method includes a computer program that, when executed by a processor, implements the steps of the photovoltaic system island detection method based on Santlet transform and Ridglet probabilistic neural network as described in any one of claims 1 to 7.

Images