Photovoltaic arcing detection method, product, medium and system based on artificial intelligence

Abstract

The invention discloses a photovoltaic arcing detection method based on artificial intelligence, a product, a medium and a system. The method comprises the following steps: S1, acquiring photovoltaic data of a photovoltaic system; s2, performing feature extraction on the photovoltaic data by adopting a full-connection neural network to obtain feature data; wherein a neural network parameter optimization method based on a fractal extension optimization algorithm is adopted to optimize parameters of the full-connection neural network; s3, performing dimension reduction on the feature data to obtain feature data after dimension reduction; and S4, inputting the feature data after dimension reduction into a preset photovoltaic arc discharge detection and identification model for photovoltaic arc discharge detection and identification. The problem of gradient disappearance or explosion can be effectively avoided, and the stability and generalization ability of model training are enhanced.

Description

Technical Field

[0001] The present invention mainly relates to the field of photovoltaic technology, and specifically to an artificial intelligence-based photovoltaic arc detection method, product, medium and system. Background Art

[0002] During the operation of photovoltaic systems, arc faults may occur due to factors such as circuit aging, environmental impacts or component defects. Such faults will lead to reduced energy efficiency and even cause safety accidents such as fires in severe cases. The occurrence of arc faults is not easy to be detected promptly and accurately by traditional monitoring methods, because such faults often manifest as tiny current and voltage anomalies in the early stages, and these signals are easily masked by normal fluctuations in the system. Therefore, it is particularly important to develop a technology that can effectively identify and predict arc faults in photovoltaic systems. Existing technologies mainly rely on hardware equipment for fault monitoring, which is not only costly, but also often not sensitive enough to faults or has a high false alarm rate. In addition, these methods need to be improved in the intelligence of data processing and fault diagnosis, especially when facing the complex and changing operating environment of photovoltaic systems, their accuracy and stability are often challenged. For example, the current existing technologies mainly have the following deficiencies:

[0003] 1. Traditional neural network feature extraction methods rely on gradient descent and may encounter gradient vanishing or exploding problems, which limits the stability of model training.

[0004] 2. The feature dimensionality reduction method in the existing technology may fail to effectively retain the key features of the data, reducing the accuracy of arc detection and the adaptability of the model. Summary of the Invention

[0005] In response to the technical problems existing in the prior art, the present invention provides an artificial intelligence-based photovoltaic arc detection method, product, medium and system that effectively avoid the problem of gradient disappearance or explosion.

[0006] In order to solve the above technical problems, the technical solution proposed by the present invention is:

[0007] A photovoltaic arc detection method based on artificial intelligence comprises the following steps:

[0008] S1. Obtaining photovoltaic data of the photovoltaic system;

[0009] S2. Using a fully connected neural network to extract features from the photovoltaic data to obtain feature data; wherein a neural network parameter optimization method based on a fractal expansion optimization algorithm is used to optimize the parameters of the fully connected neural network;

[0010] S3, reducing the dimension of the feature data to obtain the reduced-dimensional feature data;

[0011] S4. Inputting the feature data after dimensionality reduction into a preset photovoltaic arc detection and recognition model to perform photovoltaic arc detection and recognition.

[0012] Preferably, in step S2, the specific process of using the neural network parameter optimization method based on the fractal expansion optimization algorithm to optimize the parameters of the fully connected neural network is:

[0013] S201, setting initial parameters of the neural network, including weights and biases;

[0014] S202, performing an expansion operation on the parameters in each iteration cycle;

[0015] S203, performing a local search in the expanded parameter space;

[0016] S204: Evaluate the expanded parameter set, select the best performing parameter combination, and then perform a contraction operation to focus the search space on the most promising area.

[0017] S205, refining the search in the contracted parameter space;

[0018] S206. After each optimization step, evaluate the network to monitor the training progress and effect;

[0019] S207. Dynamically adjust the fractal expansion and contraction strategy based on the performance evaluation results;

[0020] S208: Repeat iterative steps S201-S207 until a preset stopping condition is met.

[0021] Preferably, in step S201, a standard Gaussian distribution is used for parameter initialization, wherein the weight w and bias b are initialized as follows:

[0022]

[0023] Where w (0) and b (0) are the initial weights and biases of the neural network, represents a normal distribution, and σ is the standard deviation.

[0024] Preferably, in step S202, the parameter expansion operation is represented as:

[0025] Δw t =α sa ·(w t-1 ⊙cos(β sa ·t+φ))

[0026] Δb t =α sa ·(b t-1⊙sin(β sa ·t+φ))

[0027] In the formula, Δw t and Δb t is the change in the weight and bias of the neural network at the tth iteration, α sa is the step size of the expansion, β sa controls the frequency of the periodic change, φ is the phase offset, and ⊙ represents the element-wise product.

[0028] Preferably, in step S203, a non-gradient search strategy is used to adjust the parameters, which is expressed as:

[0029]

[0030] Where w t+1 and b t+1 are the weight and bias of the t+1th iteration, L() is the loss function, D is the dataset, is the gradient of the loss function, γ cf is the learning rate.

[0031] Preferably, in step S205, an optimization term based on the Hessian matrix is used to increase the accuracy of the search, which is expressed as:

[0032]

[0033] Where w t+1 ' and b t+1 ' are the weight and bias of the t+1th iteration after the contraction operation; w t+1 ” and b t+1 ” are the weight and bias of the t+1th iteration after the refinement search operation, is the Hessian matrix of the loss function, η gh is the step size for second-order optimization.

[0034] Preferably, in step S206, the performance evaluation is completed by calculating the loss function, which is expressed as:

[0035]

[0036] Where l() is the loss function of a single sample, f is the neural network function, and x i ,y i are data samples and their labels, w and b are the weights and biases of the neural network respectively; N is the total number of samples.

[0037] Preferably, in step S207, the extended step size α is dynamically adjusted. sa , control the frequency of periodic changes β saand phase shift φ to optimize the expansion strategy for future cycles, expressed as:

[0038]

[0039] φ new =φ old +ζ sh ·δL

[0040] Where, and φ new are the adjusted extended step size parameter, the frequency parameter for controlling periodic changes, and the phase offset parameter, and φ old are the extended step size parameter before adjustment, the frequency parameter and phase offset parameter for controlling periodic changes, κ ed ,ξ vh ,ζ sh is the adjustment coefficient, and δL is the loss change.

[0041] Preferably, in step S3, an autoencoder algorithm based on inverse reconstruction is used to perform feature dimensionality reduction.

[0042] Preferably, the specific process of using the autoencoder algorithm based on inverse reconstruction to perform feature dimensionality reduction is:

[0043] S301, initializing the parameters of the autoencoder algorithm, including the weight and bias parameters of the autoencoder;

[0044] S302: The input feature-extracted data is converted into a low-dimensional representation through an encoder, and the low-dimensional representation is reconstructed into an output with the same dimension as the original data through a decoder;

[0045] S303, calculating the inverse reconstruction loss function of the autoencoder;

[0046] S304, using the gradient descent algorithm to update the weights and biases of the encoder and decoder according to the gradient calculated by the loss function;

[0047] S305 , repeating steps S301 - S305 until a preset stopping condition is met.

[0048] Preferably, in step S301, the initialization method is expressed as:

[0049]

[0050] Where W r,init is the initial weight matrix of the autoencoder; b r,init is the initial bias of the autoencoder; n r,in and n r,outRepresent the number of neurons in the input and output layers respectively; U(-1,1) represents a random number uniformly distributed in the interval [-1,1].

[0051] Preferably, in step S302, during the forward propagation process, the encoder output calculation method is expressed as:

[0052] z r =Sig(W r,enc ·x r +b r,enc )

[0053] And, the reconstructed output of the decoder is:

[0054]

[0055] Where z r is the low-dimensional feature representation of the encoding; Sig() is the Sigmoid activation function; W r,enc and W r,dec are the weights of the encoder and decoder respectively; b r,enc and b r,dec are the biases of the encoder and decoder respectively; x r is the input data; is the output of the reconstruction.

[0056] Preferably, in step S303, the calculation method of the inverse reconstruction loss function is expressed as:

[0057]

[0058] Where, L r is the inverse reconstruction loss function; Ntr is the number of samples input in the current batch; ∥∥ represents the L2 norm; λ r is the regularization parameter; D KL (P r ∥Q r ) is the KL divergence.

[0059] Preferably, in step S304, the updating method is expressed as:

[0060]

[0061] Where η r is the learning rate; is the updated encoder weight; is the updated encoder bias; is the updated decoder weight; is the updated decoder bias.

[0062] Preferably, in step S4, the photovoltaic arc detection and recognition model is a preset trained support vector machine classifier, or a preset trained random forest classifier.

[0063] Preferably, in step S1, the photovoltaic data includes one or more of current data, voltage data, frequency data, waveform peak value, temperature data, ambient humidity, light intensity, system load, time and equipment operation status code.

[0064] The present invention also discloses a computer program product, comprising a computer program, which executes the steps of the above method when executed by a processor.

[0065] The present invention further discloses a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the method described above are executed.

[0066] The present invention also discloses an artificial intelligence-based photovoltaic arc detection system, comprising a memory and a processor connected to each other, wherein a computer program is stored on the memory, and when the computer program is run by the processor, the steps of the method described above are executed.

[0067] Compared with the prior art, the advantages of the present invention are:

[0068] The photovoltaic arc detection method of the present invention adopts a 6-layer fully connected neural network and optimizes network parameters using a fractal expansion optimization algorithm. It does not rely on the traditional gradient descent method, but instead adopts an iterative expansion and contraction mode to optimize parameters, effectively avoiding the problem of gradient disappearance or explosion. The feature extraction model training strategy based on the fractal expansion optimization algorithm effectively solves the gradient problem that may occur in the traditional neural network training process and enhances the stability of model training.

[0069] This paper uses an autoencoder based on inverse reconstruction to perform feature dimensionality reduction. The encoder compresses high-dimensional data into a low-dimensional representation, and the decoder reconstructs the data, maximizing the preservation of key features. The inverse reconstruction autoencoder algorithm optimizes the feature dimensionality reduction process, enabling the model to better adapt to data characteristics under various operating conditions and improving the model's generalization capabilities.

[0070] By integrating multi-dimensional data acquisition and intelligent data processing and analysis technologies, the present invention can accurately identify and classify arc faults in complex photovoltaic system operating environments, effectively improve the safety and stability of photovoltaic systems, and achieve sensitive detection of early signs of arc faults. It also optimizes the identification process through intelligent algorithms, reduces the false alarm rate, and thus ensures the efficient and safe operation of photovoltaic systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 The flowchart of the photovoltaic arc detection method in an embodiment of the present invention.

[0072] Figure 2 This is a training flowchart of the autoencoder algorithm based on inverse reconstruction in the present invention. DETAILED DESCRIPTION

[0073] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0074] like Figure 1 As shown, the photovoltaic arc detection method based on artificial intelligence provided by the embodiment of the present invention mainly comprises the following steps:

[0075] S1. Data collection and annotation

[0076] The data for the training model of the present invention is collected from the photovoltaic system. In one embodiment, the data collected from the photovoltaic system mainly includes:

[0077] Current data (A): Current readings under different working conditions, especially the current values when an arc fault occurs and during normal operation;

[0078] Voltage data (V): Voltage readings collected simultaneously, focusing on voltage changes before and after the arc occurs;

[0079] Frequency data (Hz): Perform spectrum analysis on current and voltage signals and record the high-frequency components at the time of fault;

[0080] Waveform Peak (V): Captures and records the peak value of voltage and current waveforms, especially the highest point when high-frequency oscillations or pulses are detected;

[0081] Temperature data (°C): Temperature data of circuits and components to assist in identifying arc faults caused by overheating;

[0082] Ambient humidity (%): Ambient humidity information, as humidity may affect the formation and continuity of the arc;

[0083] Light intensity (Lux): records the light intensity received by the photovoltaic panel. Light changes may affect the photovoltaic output;

[0084] System load (kW): The load of the PV system, as load changes may cause fluctuations in voltage and current;

[0085] Time (Unix timestamp): ensures data synchronization and can be associated with the time information of a specific event;

[0086] Equipment operating status code (integer): The equipment status of the PV system, including the code of any alarm or system status indication.

[0087] All the data collected above are stored in a structured JSON format. Specifically, in this embodiment, one of the data is represented as:

[0088] {

[0089] "time":1652946060,

[0090] "Current Reading": 15.2,

[0091] "voltage reading": 224,

[0092] "Frequency Component": 52,

[0093] "Waveform Peak": 230,

[0094] "Temperature":77,

[0095] "Humidity": 40,

[0096] "Light Intensity": 800,

[0097] "System Load": 95,

[0098] "Device operation status code": 1

[0099] }

[0100] It should be noted that this embodiment is only intended to illustrate one data format and type of the present invention. In actual applications, the attributes of data are usually more than 10 attributes, and the number of attributes of data may reach dozens or even hundreds.

[0101] Furthermore, the collected data is labeled. The labeling method of the present invention is manual labeling. In one embodiment, the labeled categories include: normal operation (labeled 0), series arc fault (labeled 1), parallel arc fault (labeled 2), and suspected fault (labeled 3), a total of 4 categories.

[0102] S2. Feature extraction model training

[0103] The collected data is input into the feature extraction model to train the feature extraction model. The present invention uses a 6-layer fully connected neural network for feature extraction. In the prior art, some solutions use neural networks for feature extraction. In some neural network structures, they may encounter problems such as gradient vanishing, gradient explosion, or falling into local optimal solutions, which affects the stability of training and the performance of the model. Therefore, the present invention adopts a neural network parameter optimization method based on a fractal expansion optimization algorithm, which is inspired by fractal structures in nature, such as snowflakes, tree branches, and river branches. Fractal structures generally show self-similarity in mathematics and nature. The iterative method in fractal theory is used to simulate the multi-scale exploration of parameter space. It does not rely on the traditional gradient descent method, but simulates the formation process of fractals in nature through repeated expansion and contraction patterns to optimize the weights and biases of the network.

[0104] Specifically, the training process of the neural network based on the fractal expansion optimization algorithm is as follows:

[0105] S201. Set the initial weights and biases of the neural network. In one embodiment, a standard Gaussian distribution is used for parameter initialization, where the weights w and biases b are initialized as follows:

[0106]

[0107] Where w (0) and b (0) are the initial weights and biases of the neural network, represents a normal distribution, and σ is the standard deviation; preferably, σ is set to 0.1.

[0108] S202: In each iteration cycle, the parameters are first expanded, that is, the variation range of the parameters is increased according to the preset fractal rules to simulate the fractal expansion mode, which is expressed as:

[0109] Δw t =α sa ·(w t-1 ⊙cos(β sa ·t+φ))

[0110] Δb t =α sa ·(b t-1 ⊙sin(β sa ·t+φ))

[0111] In the formula, Δw t and Δb t is the change in the weight and bias of the neural network at the tth iteration, α sa is the step size of the expansion, β saControls the frequency of periodic changes, φ is the phase offset, and ⊙ represents the element-wise product. Preferably, the parameter α sa The initial value of β is set to 0.05, sa The initial value of is π / 4, and the initial value of φ is π / 2.

[0112] S203. Perform a local search in the expanded parameter space. In one embodiment, a non-gradient search strategy is used to adjust the parameters, which is expressed as:

[0113]

[0114] Where w t+1 and b t+1 are the weight and bias of the t+1th iteration, L() is the loss function, D is the dataset, is the gradient of the loss function, γ cf is the learning rate.

[0115] Among them, the learning rate γ cf Set the dynamic adaptive adjustment method. The adjustment method is expressed as:

[0116]

[0117] Where D f is the fractal dimension, is the learning rate of the tth iteration, ΔL t is the change in the loss function at the tth iteration; k sa is a hyperparameter that controls the sensitivity of learning rate changes. Preferably, k sa Set to 0.1.

[0118] Among them, the fractal dimension D f It is an indicator to measure the complexity of loss change, and the calculation method is expressed as:

[0119]

[0120] Furthermore, the gradient Calculated by the back propagation algorithm, in one embodiment, the loss function L() is the mean square error, and the gradient is calculated as follows:

[0121]

[0122] Where, f(x i ; w) is the output of the network model, y i is the true label, is the gradient of the model output with respect to the weights, calculated using the chain rule.

[0123] S204. Evaluate the expanded parameter set, select the best performing parameter combination, and then perform a contraction operation to focus the search space on the most promising area to improve the accuracy of the search and ensure that the parameters do not deviate too much from the optimal solution, which is expressed as:

[0124] w t+1 '=λ vg w t+1 +(1-λ vg )·w *

[0125] b t+1 '=λ vg b t+1 +(1-λ vg )·b *

[0126] Where w t+1 ' and b t+1 'are the weights and biases of the t+1th iteration after the contraction operation, w * and b * is the best performing parameter in the current evaluation cycle, λ vg is the shrinkage factor, usually set to 0.9.

[0127] S205, refine the search in the contracted parameter space to further optimize network performance. Specifically, the present invention uses an optimization term based on the Hessian matrix to increase the accuracy of the search, which is expressed as:

[0128]

[0129] Where w t+1 ” and b t+1 ” are the weight and bias of the t+1th iteration after the refinement search operation, is the Hessian matrix of the loss function, η gh is the step size of the second-order optimization. Preferably, η gh Typically set to 0.005.

[0130] Among them, the Hessian matrix Represents the second-order partial derivative matrix of the loss function. Taking the weight at the tth iteration as an example, the calculation method is expressed as:

[0131]

[0132] Where, is the second-order derivative of the model output with respect to the two weights.

[0133] S206. After each optimization step, evaluate the network to monitor the training progress and effect. Specifically, performance evaluation is completed by calculating the loss function, which is expressed as:

[0134]

[0135] Where, e() is the loss function of a single sample, f is the neural network function, and x i ,y i are data samples and their labels, and N is the total number of samples.

[0136] S207, dynamically adjust the fractal expansion and contraction strategy according to the performance evaluation results, and dynamically adjust the parameter α sa , β sa and φ, to optimize the expansion strategy for future cycles, expressed as:

[0137]

[0138] φ new =φ old +ζ sh ·δL

[0139] Where, and φ new are the adjusted extended step size parameter, the frequency parameter for controlling periodic changes, and the phase offset parameter, and φ old are the extended step size parameter before adjustment, the frequency parameter and phase offset parameter for controlling periodic changes, κ ed ,ξ vh ,ζ sh is the adjustment coefficient, δL is the loss change. Preferably, κ ed ,ξ vh ,ζ sh Set to 0.5, 0.3, and 0.2 respectively.

[0140] S208. Repeat the above steps until a preset stop iteration condition is met, indicating that the model training is complete. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0141] S3. Feature dimensionality reduction model training

[0142] The feature-extracted data is input into a feature dimensionality reduction model for training. The present invention employs an autoencoder algorithm based on inverse reconstruction as the feature dimensionality reduction model. The autoencoder comprises two parts: an encoder and a decoder. The encoder compresses the input data into a low-dimensional representation, while the decoder reconstructs the data from this low-dimensional representation. To enhance the model's adaptability to feature dimensionality reduction, the present invention employs an inverse reconstruction strategy to process the input high-dimensional data after feature extraction, minimizing the dependencies between different features and thereby improving encoding quality and decoding accuracy.

[0143] Specifically, such as Figure 2 As shown in Figure 2, the training process of the autoencoder algorithm based on inverse reconstruction is as follows:

[0144] S301. Initialize the parameters of the autoencoder algorithm, including the weight and bias parameters of the autoencoder. In one embodiment, the initialization method is expressed as follows:

[0145]

[0146] b r,init =0

[0147] Where W r,init is the initial weight matrix of the autoencoder; b r,init is the initial bias of the autoencoder; n r,in and n r,out Represent the number of neurons in the input and output layers respectively; U(-1,1) represents a random number uniformly distributed in the interval [-1,1].

[0148] S302: The input feature-extracted data is converted into a low-dimensional representation through the encoder, and the low-dimensional representation is reconstructed into an output with the same dimension as the original data through the decoder. In the forward propagation process, the encoder output calculation method is expressed as:

[0149] z r =Sig(W r,enc ·x r +b r,enc )

[0150] And, the reconstructed output of the decoder is:

[0151]

[0152] Where z r is the low-dimensional feature representation of the encoding; Sig() is the Sigmoid activation function; W r,enc and W r,dec are the weights of the encoder and decoder respectively; b r,enc and b r,decare the biases of the encoder and decoder respectively; x r is the input data; is the output of the reconstruction.

[0153] S303. Calculate the inverse reconstruction loss of the autoencoder. The calculation method of the inverse reconstruction loss function is defined as:

[0154]

[0155] Where, L r is the inverse reconstruction loss function; Ntr is the number of samples input in the current batch; ∥∥ represents the L2 norm; λ r is the regularization parameter; D KL (P r ∥Q r ) is the KL divergence.

[0156] In one embodiment, KL divergence, which measures the original data distribution P r and reconstruct the data distribution Q r The difference between them is used to ensure that the key features of the data are not lost during the dimensionality reduction process. The calculation method is expressed as:

[0157]

[0158] Where p r,i and q r,i are the probability distributions of the original data and the reconstructed data in the i-th dimension, respectively, and Mtr is the feature dimension before feature dimensionality reduction.

[0159] S304. Based on the gradient calculated by the loss function, the weights and biases of the encoder and decoder are updated using the gradient descent algorithm. The update method is expressed as:

[0160]

[0161] Where η r is the learning rate; is the updated encoder weight; is the updated encoder bias; is the updated decoder weight; is the updated decoder bias.

[0162] The gradient For example, the calculation method is expressed as:

[0163]

[0164] Where, represents the partial derivative of the loss function with respect to the reconstructed output, is the partial derivative during decoding, is the partial derivative during encoding.

[0165] S305: Repeat the above steps until a preset stop iteration condition is met, indicating that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1200 times.

[0166] S4. Photovoltaic arc detection and recognition model training

[0167] The reduced-dimensional data is input into a preset photovoltaic arc detection and recognition model for photovoltaic arc detection and recognition. The photovoltaic arc detection and recognition model can be a preset trained support vector machine classifier or a preset trained random forest classifier. In other words, a conventional classifier can be used for classification processing to achieve photovoltaic arc detection and recognition.

[0168] During the photovoltaic arc detection and identification phase, in one embodiment, the collected raw data is input into a trained feature extraction and feature dimensionality reduction model for feature processing. Furthermore, the processed features are input into a classifier model for classifier training to obtain classification results. In this embodiment, the classification categories include normal operation (labeled 0), series arc fault (labeled 1), parallel arc fault (labeled 2), and suspected fault (labeled 3), for a total of four categories.

[0169] The photovoltaic arc detection method of the present invention adopts a 6-layer fully connected neural network and optimizes network parameters using a fractal expansion optimization algorithm. It does not rely on the traditional gradient descent method, but instead adopts an iterative expansion and contraction mode to optimize parameters, effectively avoiding the problem of gradient disappearance or explosion. The feature extraction model training strategy based on the fractal expansion optimization algorithm effectively solves the gradient problem that may occur in the traditional neural network training process and enhances the stability of model training.

[0170] This paper uses an autoencoder based on inverse reconstruction to perform feature dimensionality reduction. The encoder compresses high-dimensional data into a low-dimensional representation, and the decoder reconstructs the data, maximizing the preservation of key features. The inverse reconstruction autoencoder algorithm optimizes the feature dimensionality reduction process, enabling the model to better adapt to data characteristics under various operating conditions and improving the model's generalization capabilities.

[0171] By integrating multi-dimensional data acquisition and intelligent data processing and analysis technologies, the present invention can accurately identify and classify arc faults in complex photovoltaic system operating environments, effectively improve the safety and stability of photovoltaic systems, and achieve sensitive detection of early signs of arc faults. It also optimizes the identification process through intelligent algorithms, reduces the false alarm rate, and thus ensures the efficient and safe operation of photovoltaic systems.

[0172] The present invention also discloses a computer program product, including a computer program that, when executed by a processor, performs the steps of the above-described method. The present invention further discloses a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, performs the steps of the above-described method. The present invention also discloses an artificial intelligence-based photovoltaic arc detection system, including an interconnected memory and a processor, the memory storing a computer program that, when executed by the processor, performs the steps of the above-described method. The products, media, and systems of the present invention correspond to the above-described method and also possess the advantages described by the above-described method.

[0173] The present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by hardware related to computer program instructions. The computer program can be stored in a computer-readable storage medium, and when the computer program is executed by the processor, it can implement the steps of the above-mentioned method embodiment. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. Computer-readable storage media include: any entity or device that can carry computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc. The memory is used to store computer programs and / or modules, and the processor implements various functions by running or executing computer programs and / or modules stored in the memory, and calling data stored in the memory. The memory may include high-speed random access memory and non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage device.

[0174] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions based on the principles of the present invention are within the scope of protection of the present invention. It should be noted that for those skilled in the art, various improvements and modifications that do not depart from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A photovoltaic arc detection method based on artificial intelligence, characterized in that: Including steps: S1. Obtaining photovoltaic data of the photovoltaic system; S2. Using a fully connected neural network to extract features from the photovoltaic data to obtain feature data; wherein a neural network parameter optimization method based on a fractal expansion optimization algorithm is used to optimize the parameters of the fully connected neural network; S3, reducing the dimension of the feature data to obtain the reduced-dimensional feature data; S4. Inputting the feature data after dimensionality reduction into a preset photovoltaic arc detection and recognition model to perform photovoltaic arc detection and recognition.

2. The photovoltaic arc detection method based on artificial intelligence according to claim 1, characterized in that: In step S2, the specific process of using the neural network parameter optimization method based on the fractal expansion optimization algorithm to optimize the parameters of the fully connected neural network is as follows: S201, setting initial parameters of the neural network, including weights and biases; S202, performing an expansion operation on the parameters in each iteration cycle; S203, performing a local search in the expanded parameter space; S204: Evaluate the expanded parameter set, select the best performing parameter combination, and then perform a contraction operation to focus the search space on the most promising area. S205, refining the search in the contracted parameter space; S206. After each optimization step, evaluate the network to monitor the training progress and effect; S207. Dynamically adjust the fractal expansion and contraction strategy based on the performance evaluation results; S208: Repeat iterative steps S201-S207 until a preset stopping condition is met.

3. The photovoltaic arc detection method based on artificial intelligence according to claim 2, characterized in that: In step S201, the standard Gaussian distribution is used for parameter initialization, where the weight w and bias b are initialized as follows: Where w (0) and b (0) are the initial weights and biases of the neural network, represents a normal distribution, and σ is the standard deviation.

4. The photovoltaic arc detection method based on artificial intelligence according to claim 2, characterized in that: In step S202, the parameter expansion operation is represented as follows: Δw t =a sa ·(w t-1 ⊙cos(β sa ·t+φ)) Δb t =a sa ·(b t-1 ⊙sin(β sa ·t+φ)) In the formula, Δw t and Δb t is the change in the weight and bias of the neural network at the tth iteration, α sa is the step size of the expansion, β sa controls the frequency of the periodic change, φ is the phase offset, and ⊙ represents the element-wise product.

5. The photovoltaic arc detection method based on artificial intelligence according to claim 2, 3 or 4, characterized in that: In step S203, a non-gradient search strategy is used to adjust the parameters, which is expressed as: Where w t+1 and b t+1 are the weight and bias of the t+1th iteration, L() is the loss function, D is the dataset, is the gradient of the loss function, γ cf is the learning rate.

6. The photovoltaic arc detection method based on artificial intelligence according to claim 2, 3 or 4, characterized in that: In step S205, an optimization term based on the Hessian matrix is used to increase the accuracy of the search, which is expressed as: Where w t+1 ' and b t+1 ' are the weight and bias of the t+1th iteration after the contraction operation; w t+1 ” and b t+1 ” are the weight and bias of the t+1th iteration after the refinement search operation, is the Hessian matrix of the loss function, η gh is the step size for second-order optimization.

7. The photovoltaic arc detection method based on artificial intelligence according to claim 2, 3 or 4, characterized in that: In step S206, the performance evaluation is completed by calculating the loss function, which is expressed as: Where, is the loss function of a single sample, f is the neural network function, x i ,y i are data samples and their labels, w and b are the weights and biases of the neural network respectively; N is the total number of samples.

8. The photovoltaic arc detection method based on artificial intelligence according to claim 2, 3 or 4, characterized in that: In step S207, the extended step size α is dynamically adjusted. sa , control the frequency of periodic changes β sa and phase shift φ to optimize the expansion strategy for future cycles, expressed as: f new =φ old +g sh ·δL Where, and φ new are the adjusted extended step size parameter, the frequency parameter for controlling periodic changes, and the phase offset parameter, and φ old are the extended step size parameter before adjustment, the frequency parameter and phase offset parameter for controlling periodic changes, κ ed ,ξ vh ,ζ sh is the adjustment coefficient, and δL is the loss change.

9. The photovoltaic arc detection method based on artificial intelligence according to any one of claims 1 to 4, characterized in that: In step S3, an autoencoder algorithm based on inverse reconstruction is used to perform feature dimensionality reduction.

10. The photovoltaic arc detection method based on artificial intelligence according to claim 9, characterized in that: The specific process of using the autoencoder algorithm based on inverse reconstruction to perform feature dimensionality reduction is as follows: S301, initializing the parameters of the autoencoder algorithm, including the weight and bias parameters of the autoencoder; S302: The input feature-extracted data is converted into a low-dimensional representation through an encoder, and the low-dimensional representation is reconstructed into an output with the same dimension as the original data through a decoder; S303, calculating the inverse reconstruction loss function of the autoencoder; S304, using the gradient descent algorithm to update the weights and biases of the encoder and decoder according to the gradient calculated by the loss function; S305 , repeating steps S301 - S305 until a preset stopping condition is met.

11. The photovoltaic arc detection method based on artificial intelligence according to claim 10, characterized in that: In step S301, the initialization method is expressed as: b r,init ＝0 Where W r,init is the initial weight matrix of the autoencoder; b r,init is the initial bias of the autoencoder; n r,in and n r,out Represent the number of neurons in the input and output layers respectively; U(-1,1) represents a random number uniformly distributed in the interval [-1,1].

12. The photovoltaic arc detection method based on artificial intelligence according to claim 11, characterized in that: In step S302, during the forward propagation process, the encoder output is calculated as follows: z r =Sig(W r,enc ·x r +b r,enc ) And, the reconstructed output of the decoder is: Where z r is the low-dimensional feature representation of the encoding; Sig() is the Sigmoid activation function; W r,enc and W r,dec are the weights of the encoder and decoder respectively; b r,enc and b r,dec are the biases of the encoder and decoder respectively; x r is the input data; is the output of the reconstruction.

13. The photovoltaic arc detection method based on artificial intelligence according to claim 12, characterized in that: In step S303, the calculation method of the inverse reconstruction loss function is expressed as: Where, L r is the inverse reconstruction loss function; Ntr is the number of samples input in the current batch; ∥∥ represents the L2 norm; λ r is the regularization parameter; D KL (P r ∥Q r ) is the KL divergence.

14. The photovoltaic arc detection method based on artificial intelligence according to claim 12, characterized in that: In step S304, the update method is expressed as: Where η r is the learning rate; is the updated encoder weight; is the updated encoder bias; is the updated decoder weight; is the updated decoder bias.

15. The photovoltaic arc detection method based on artificial intelligence according to any one of claims 1 to 4, characterized in that: In step S4, the photovoltaic arc detection and recognition model is a preset trained support vector machine classifier, or a preset trained random forest classifier.

16. The photovoltaic arc detection method based on artificial intelligence according to any one of claims 1 to 4, characterized in that: In step S1 , the photovoltaic data includes one or more of current data, voltage data, frequency data, waveform peak value, temperature data, ambient humidity, light intensity, system load, time and equipment operation status code.

17. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 16 are performed.

18. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the computer program performs the steps of the method according to any one of claims 1 to 16.

19. A photovoltaic arc detection system based on artificial intelligence, comprising a memory and a processor connected to each other, wherein a computer program is stored in the memory, characterized in that: When the computer program is executed by a processor, the computer program performs the steps of the method according to any one of claims 1 to 16.

Images