A distributed photovoltaic diagnosis method and system based on multi-modal time series data

Abstract

The present application relates to the technical field of photovoltaic device diagnosis, and particularly relates to a distributed photovoltaic diagnosis method and system based on multi-modal time series data. The distributed photovoltaic diagnosis system based on multi-modal time series data comprises a multi-modal time series data acquisition module, an electrical characteristic vector extraction module, a single fault source judgment module and a fault source combination judgment module. The present application matches the electrical characteristic vector with the electrical characteristic fingerprint library, calculates the similarity and sets a threshold value, can directly judge the fault source in the single fault source case, and in the multi-fault source case, carries out linear splitting through a group optimization algorithm, combines the contribution component and the appearance frequency of the fault source, calculates the appearance probability of the fault source and analyzes the set fault source combination probability threshold value, ensures that only when the contribution of multiple fault sources reaches a certain threshold value, the fault source combination information is output, and the diagnosis precision of the fault source is further optimized.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic equipment diagnostic technology, and specifically to a distributed photovoltaic diagnostic method and system based on multimodal time series data. Background Technology

[0002] Traditional photovoltaic (PV) system fault diagnosis often relies on single electrical parameters (such as current, voltage, and temperature) to determine the cause of a fault. However, these traditional diagnostic methods have significant limitations. In complex operating environments, multiple fault sources often occur simultaneously, and their effects on the electrical characteristics of the PV system are often overlapping. For example, poor contact can cause a drop in current, as can damage to the solar panel, and insufficient sunlight due to panel shading can also affect the current. This makes it impossible to accurately distinguish which fault source is causing the problem using a single electrical parameter. Furthermore, when multiple fault sources occur simultaneously, their combined impact on the system's electrical parameters becomes even more complex. Traditional diagnostic methods often fail to effectively identify and pinpoint the specific contribution of each fault source, making accurate fault location and clear determination of the fault cause extremely difficult. This inaccurate fault diagnosis not only affects the timely detection of PV system faults but also severely restricts subsequent maintenance and repair work. Summary of the Invention

[0003] This invention matches electrical feature vectors with an electrical feature fingerprint database, calculates similarity, and sets a threshold. In the case of a single fault source, it can directly identify the fault source. In the case of multiple fault sources, it uses a swarm optimization algorithm to perform linear decomposition, combines the contribution components and occurrence frequency of the fault sources, calculates the occurrence probability of the fault sources, and analyzes it with a set fault source combination probability threshold. This ensures that fault source combination information is only output when the contributions of multiple fault sources reach a certain threshold, thereby optimizing the diagnostic accuracy of fault sources, reducing false alarms and false negatives, and improving the fault location and maintenance efficiency of distributed photovoltaic systems.

[0004] This invention provides a distributed photovoltaic diagnostic method based on multimodal time series data, comprising:

[0005] Regularly acquire multimodal time-series data from distributed photovoltaic systems;

[0006] The acquired multimodal time-series data is fed into the distributed photovoltaic diagnostic model for processing, and the corresponding electrical feature vector is output.

[0007] The output electrical feature vector is sequentially matched with the electrical feature fingerprints in the electrical feature fingerprint database. Specifically, the matching operation involves calculating the similarity between the electrical feature vector and the electrical feature fingerprint, and determining whether the similarity between the electrical feature vector and the electrical feature fingerprint is higher than the similarity threshold. If the similarity between the electrical feature vector and the electrical feature fingerprint is higher than the similarity threshold, the fault source information corresponding to the electrical feature fingerprint is directly output. If the similarity between the electrical feature vector and the electrical feature fingerprint is not higher than the similarity threshold, a fault source combination judgment operation is performed, and the corresponding fault source combination information is output. The electrical feature fingerprint database contains a one-to-one correspondence between fault source information and electrical feature fingerprints.

[0008] The fault source combination judgment operation is as follows: using the electrical feature vector as the target vector and the electrical feature fingerprint as the basis vector, a linear split operation is performed through a population optimization algorithm to obtain the contribution component of the fault source information corresponding to each electrical feature fingerprint. Then, the occurrence frequency of each fault source is obtained. Based on the occurrence probability of all fault source information and the set fault source combination probability threshold, the corresponding fault source combination information is analyzed and output.

[0009] As a preferred aspect, the electrical fingerprint database is constructed in the following manner:

[0010] For each fault source information, several training samples labeled by the fault source information are obtained. The training samples include multimodal time series data. The training samples are sent to the distributed photovoltaic diagnostic model for processing, and the electrical feature vectors corresponding to the training samples are output. Then, the electrical feature vectors corresponding to all training samples are clustered and the cluster center is used as the electrical feature fingerprint corresponding to the fault source information.

[0011] As a preferred approach, using the electrical feature vector as the target vector and the electrical feature fingerprint as the basis vector, a linear decomposition operation is performed through a swarm optimization algorithm to obtain the contribution component of the fault source information corresponding to each electrical feature fingerprint, specifically including the following:

[0012] Set up simulated weight component individuals, each of which includes the simulated weight component corresponding to each electrical feature fingerprint. Then, combine all simulated weight component individuals into a population set and set the maximum number of iterations.

[0013] The fitness of each simulated weight component individual is calculated by performing a weighted summation of all electrical feature fingerprints according to the simulated weight components in the simulated weight component individual to obtain a simulated vector. The similarity between the simulated vector and the target vector is used as the fitness of the simulated weight component individual.

[0014] Based on the fitness of the simulated weight component individuals, the population set is iteratively updated using a population optimization algorithm until the maximum number of iterations is reached. After normalizing the simulated weight component individuals with the highest fitness, the contribution component of the fault source information corresponding to each electrical feature fingerprint is obtained.

[0015] As a preferred aspect, the distributed photovoltaic diagnostic model consists of an LSTM layer, a fully connected layer, and an identification layer. The LSTM layer is used to receive multimodal time-series data. In the fully connected layer, the outputs of all LSTM layers undergo fully connected processing, which unifies and normalizes the feature dimensions. The identification layer processes the output of the fully connected layer and outputs a probability vector. The probability vector includes the probability corresponding to each fault source, and the output of the fully connected layer is used as the corresponding electrical feature vector.

[0016] As a preferred aspect, the analysis is based on the occurrence probability of all fault source information and the set fault source combination probability threshold. Specifically, the occurrence probabilities of all fault source information are added to the judgment set in descending order. The judgment set is initially empty. Whenever a fault source occurrence probability is added to the judgment set, the sum of the occurrence probabilities of all fault sources in the judgment set is calculated. This continues until the sum of the occurrence probabilities of all fault sources in the judgment set is higher than the fault source combination probability threshold. Then, the fault source information corresponding to the occurrence probabilities of all fault sources in the judgment set is combined into the corresponding fault source combination information.

[0017] As a preferred aspect, training the distributed photovoltaic diagnostic model specifically includes the following steps: forming a training set by assembling all training samples labeled with fault source information, training the distributed photovoltaic diagnostic model using the training set, using the labeled fault source information as the target output during training, calculating the loss value based on the target output and the predicted output of the distributed photovoltaic diagnostic model, and then updating the parameters of the distributed photovoltaic diagnostic model using the gradient descent method based on the direction of minimizing the loss value.

[0018] The present invention also provides a distributed photovoltaic diagnostic system based on multimodal time series data, comprising:

[0019] The multimodal time-series data acquisition module is used to periodically acquire multimodal time-series data from distributed photovoltaic power generation.

[0020] The electrical feature vector extraction module is used to input the currently acquired multimodal time series data into the distributed photovoltaic diagnostic model for processing and output the corresponding electrical feature vector.

[0021] The single fault source determination module is used to sequentially match the output electrical feature vector with the electrical feature fingerprints in the electrical feature fingerprint database. The matching operation specifically involves calculating the similarity between the electrical feature vector and the electrical feature fingerprint, and determining whether the similarity between the electrical feature vector and the electrical feature fingerprint is higher than the similarity threshold. If the similarity between the electrical feature vector and the electrical feature fingerprint is higher than the similarity threshold, the fault source information corresponding to the electrical feature fingerprint is directly output. If the similarity between the electrical feature vector and the electrical feature fingerprint is not higher than the similarity threshold, the fault source combination determination operation is performed, and the corresponding fault source combination information is output. The electrical feature fingerprint database includes a one-to-one correspondence between fault source information and electrical feature fingerprints.

[0022] The fault source combination judgment module is used to judge the fault source combination operation. Specifically, it uses the electrical feature vector as the target vector and the electrical feature fingerprint as the basis vector. It performs a linear split operation through a group optimization algorithm to obtain the contribution component of the fault source information corresponding to each electrical feature fingerprint. Then, it obtains the occurrence frequency of each fault source. Based on the occurrence probability of all fault source information and the set fault source combination probability threshold, it analyzes and outputs the corresponding fault source combination information.

[0023] The present invention has the following advantages:

[0024] This invention matches electrical feature vectors with an electrical feature fingerprint database, calculates similarity, and sets a threshold. In the case of a single fault source, it can directly identify the fault source. In the case of multiple fault sources, it uses a swarm optimization algorithm to perform linear decomposition, combines the contribution components and occurrence frequency of the fault sources, calculates the occurrence probability of the fault sources, and analyzes it with a set fault source combination probability threshold. This ensures that fault source combination information is only output when the contributions of multiple fault sources reach a certain threshold, thereby optimizing the diagnostic accuracy of fault sources, reducing false alarms and false negatives, and improving the fault location and maintenance efficiency of distributed photovoltaic systems. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the structure of the distributed photovoltaic diagnostic system based on multimodal time-series data used in an embodiment of the present invention. Detailed Implementation

[0026] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of this invention.

[0027] Example 1: A distributed photovoltaic diagnostic method based on multimodal time series data, comprising:

[0028] Regularly acquire multimodal time-series data from distributed photovoltaic (PV) systems. This multimodal time-series data refers to current time-series data, equipment temperature time-series data, and illumination time-series data within a preset time window. The current, equipment temperature, and illumination time-series data are composed of current values, equipment temperature values, and illumination indices sorted by time, respectively. These time-series data are acquired through sensors installed on the distributed PV system. For example, current time-series data is acquired through current sensors, equipment temperature time-series data is acquired through temperature sensors (which reflect the operating temperature of the distributed PV system), and illumination time-series data is acquired through illumination sensors (where illumination refers to the ambient light in the environment where the distributed PV system is located). All these time-series data are marked and aligned using a unified timestamp. Distributed PV refers to photovoltaic power generation devices installed on the user side (such as in homes and businesses). Diagnosis of distributed PV systems is achieved through the analysis of multimodal time-series data.

[0029] The acquired multimodal time-series data is fed into the distributed photovoltaic (PV) diagnostic model for processing, outputting the corresponding electrical feature vectors. It should be noted that the distributed PV diagnostic model consists of an LSTM layer, a fully connected layer, and a recognition layer. The LSTM layer, based on an LSTM model, receives the multimodal time-series data. All current values, equipment temperature values, and illumination values ​​at each timestamp are processed sequentially by the LSTM layer, and the output after each processing is recorded. In the fully connected layer, the outputs of all LSTM layers undergo full connection processing, achieving feature dimension unification and normalization during this process. The recognition layer, implemented using a softmax function, processes the outputs of the fully connected layer, outputting... A probability vector is generated, which includes the probability of each fault source. Fault sources include poor contact of connection terminals, damage to the solar panel, and shading of the solar panel. The output of the fully connected layer is used as the corresponding electrical feature vector. It should be noted that a distributed photovoltaic diagnostic model based on a time series analysis network (LSTM model) is used to diagnose distributed photovoltaic systems and determine whether there are fault sources. It should also be noted that the distributed photovoltaic diagnostic model is trained on training samples labeled with a single fault source. The diagnostic effect is poor for combinations of fault sources. Therefore, feature extraction is performed using the distributed photovoltaic diagnostic model. The resulting electrical feature vector reflects the feature projection of different fault sources in a high-dimensional space.

[0030] The output electrical feature vector is sequentially matched with electrical feature fingerprints in the electrical feature fingerprint database. Specifically, the matching operation involves calculating the similarity between the electrical feature vector and the electrical feature fingerprint, which can be done using a cosine similarity algorithm. The similarity threshold is set by the operator based on experience and is also adjusted in actual operation based on the false alarm rate of a single fault source. A higher false alarm rate for a single fault source necessitates a higher similarity threshold. The false alarm rate for a single fault source refers to the rate obtained from subsequent maintenance feedback on distributed photovoltaic systems, where a combination of fault sources is analyzed as a single fault source. If the similarity between the electrical feature vector and the electrical feature fingerprint exceeds the similarity threshold, the fault source information corresponding to the electrical feature fingerprint is directly output. The system constructs warning messages based on fault source information and sends them to the user side of maintenance personnel, providing information reference for subsequent distributed photovoltaic maintenance operations. If the similarity between the electrical feature vector and the electrical feature fingerprint is not higher than the similarity threshold, a fault source combination judgment operation is performed, and the corresponding fault source combination information is output. The electrical feature fingerprint database includes one-to-one corresponding fault source information and electrical feature fingerprints. It should be noted that since it is impossible to determine whether the currently acquired multimodal time series data is affected by multiple fault sources, the distributed photovoltaic diagnostic model will not be used for diagnostic operations. Instead, the acquired electrical feature vector is matched with the electrical feature fingerprint obtained by statistical analysis. The electrical feature fingerprint can reflect the feature information under the influence of a single fault source. If the corresponding similarity during matching meets the expectations, then the judgment of a single fault source can be directly performed.

[0031] The fault source combination judgment operation specifically involves using the electrical feature vector as the target vector and the electrical feature fingerprint as the basis vector. A linear decomposition operation is performed using a swarm optimization algorithm to obtain the contribution component of the fault source information corresponding to each electrical feature fingerprint. This contribution component represents the degree of influence of each fault source on the electrical feature vector under the influence of multiple fault sources. Next, the occurrence frequency of each fault source is obtained. This frequency is obtained through statistical analysis of the occurrence counts of fault sources in the historical records of distributed photovoltaic systems. The occurrence frequency characterizes the probability of each fault source occurring in the corresponding distributed photovoltaic system. The product of the contribution component corresponding to the fault source information and the occurrence frequency is recorded. The probability of a fault source occurrence is analyzed based on the probability of occurrence of all fault source information and the set fault source combination probability threshold. The corresponding fault source combination information is output. The fault source combination information includes several fault sources. The fault source combination probability threshold is set by the operator based on experience, or it can be adjusted according to the false alarm rate of fault source combination judgment. It should be noted that the probability of occurrence of a fault source consists of two parts: contribution component and occurrence frequency. If the contribution component is low but the occurrence frequency is high, then the probability of occurrence of the corresponding fault source will not be too low, indicating that the corresponding fault source also exists. By analyzing the probability of occurrence of fault sources, the accuracy of fault source combination judgment can be improved.

[0032] The analysis is based on the occurrence probability of all fault sources and the set fault source combination probability threshold. Specifically, the occurrence probabilities of all fault sources are added to the judgment set in descending order. The judgment set is initially empty. Each time a fault source probability is added to the judgment set, the sum of the occurrence probabilities of all fault sources in the judgment set is calculated. This process continues until the sum of the occurrence probabilities of all fault sources in the judgment set exceeds the fault source combination probability threshold. Then, the fault source information corresponding to the occurrence probabilities of all fault sources in the judgment set is combined into the corresponding fault source combination information. It should be noted that by adding the occurrence probabilities of multiple fault sources one by one to the judgment set, the contribution of each fault source is gradually accumulated until the sum of these fault sources exceeds the set threshold. This ensures that the fault sources in the combination have sufficient overall support. If the total probability is lower than the threshold, it means that the probability of these fault source combinations is low, and they may just be some random events or noise, which do not have sufficient reliability and are therefore not considered as valid fault source combinations.

[0033] This application matches electrical feature vectors with an electrical feature fingerprint database, calculates similarity, and sets a threshold. In the case of a single fault source, it can directly identify the fault source. In the case of multiple fault sources, it uses a swarm optimization algorithm to perform linear partitioning, combines the contribution components and occurrence frequency of the fault sources, calculates the occurrence probability of the fault sources, and analyzes it with the set fault source combination probability threshold. This ensures that fault source combination information is only output when the contribution of multiple fault sources reaches a certain threshold, thereby optimizing the diagnostic accuracy of fault sources, reducing false alarms and false negatives, and improving the fault location and maintenance efficiency of distributed photovoltaic systems.

[0034] The electrical fingerprint database is constructed in the following way:

[0035] For each fault source information, several training samples labeled by the fault source information are obtained. The training samples include multimodal time series data, which is actually obtained by the operator. The training samples are sent to the distributed photovoltaic diagnostic model for processing, and the electrical feature vectors corresponding to the training samples are output. Then, the electrical feature vectors corresponding to all training samples are clustered. The K-means algorithm can be used, and the cluster center is used as the electrical feature fingerprint corresponding to the fault source information.

[0036] Using electrical feature vectors as target vectors and electrical feature fingerprints as basis vectors, a linear partitioning operation is performed using a swarm optimization algorithm to obtain the contribution components of fault source information corresponding to each electrical feature fingerprint, specifically including the following:

[0037] Set up simulated weight component individuals, each of which includes the simulated weight component corresponding to each electrical feature fingerprint. Then, combine all simulated weight component individuals into a population set and set the maximum number of iterations. Setting up simulated weight component individuals generally involves selecting random values ​​within a set range.

[0038] The fitness of each simulated weight component individual is calculated by performing a weighted summation of all electrical feature fingerprints according to the simulated weight components in the simulated weight component individual to obtain a simulated vector. The similarity between the simulated vector and the target vector is used as the fitness of the simulated weight component individual.

[0039] Based on the fitness of the simulated weight component individuals, the population set is iteratively updated using the sparrow search algorithm until the maximum number of iterations is reached. After normalizing the simulated weight component individuals with the highest fitness, the contribution component of the fault source information corresponding to each electrical feature fingerprint is obtained.

[0040] Training a distributed photovoltaic diagnostic model involves the following steps:

[0041] All training samples labeled with fault source information are grouped into a training set. The distributed photovoltaic diagnostic model is trained using the training set. During training, the labeled fault source information is used as the target output. The loss value is calculated based on the target output and the predicted output of the distributed photovoltaic diagnostic model. The MSE method can be used. Then, based on the direction of minimizing the loss value, the parameters of the distributed photovoltaic diagnostic model are updated using the gradient descent method.

[0042] Example 2: A distributed photovoltaic diagnostic system based on multimodal time series data, such as... Figure 1 As shown, it includes:

[0043] The multimodal time-series data acquisition module is used to periodically acquire multimodal time-series data from distributed photovoltaic (PV) systems. This multimodal time-series data refers to current time-series data, equipment temperature time-series data, and illumination time-series data within a preset time window. These data are composed of current values, equipment temperature values, and illumination indices sorted by time, respectively. All of this time-series data is acquired through sensors installed on the distributed PV system. For example, current time-series data is acquired through current sensors, equipment temperature time-series data is acquired through temperature sensors (reflecting the operating temperature of the distributed PV system), and illumination time-series data is acquired through illumination sensors (referring to the ambient light in the environment where the distributed PV system is located). All of this time-series data is marked and aligned using a unified timestamp. Distributed PV refers to photovoltaic power generation devices installed on the user side (e.g., in homes and businesses). The analysis of multimodal time-series data enables the diagnosis of distributed PV systems.

[0044] The electrical feature vector extraction module is used to input the currently acquired multimodal time series data into the distributed photovoltaic diagnostic model for processing and output the corresponding electrical feature vector.

[0045] The single fault source identification module sequentially matches the output electrical feature vector with electrical feature fingerprints in the electrical feature fingerprint database. Specifically, the matching operation calculates the similarity between the electrical feature vector and the electrical feature fingerprint, which can be calculated using a cosine similarity algorithm. It then determines whether the similarity between the electrical feature vector and the electrical feature fingerprint exceeds a similarity threshold. This threshold is set by the operator based on experience and is adjusted in actual operation according to the false alarm rate of the single fault source. A higher false alarm rate for a single fault source necessitates a higher similarity threshold. The false alarm rate for a single fault refers to the analysis of combinations of fault sources into a single fault source, obtained from subsequent feedback from actual maintenance operations of the distributed photovoltaic system. If the similarity between the electrical feature vector and the electrical feature fingerprint exceeds the similarity threshold, the fault corresponding to the electrical feature fingerprint is directly output. Fault source information can be used to construct warning messages and send them to the user side of maintenance personnel, providing information reference for subsequent distributed photovoltaic maintenance operations. If the similarity between the electrical feature vector and the electrical feature fingerprint is not higher than the similarity threshold, a fault source combination judgment operation is performed, and the corresponding fault source combination information is output. The electrical feature fingerprint database includes one-to-one corresponding fault source information and electrical feature fingerprints. It should be noted that since it is impossible to determine whether the currently acquired multimodal time series data is affected by multiple fault sources, the diagnostic operation will not be performed through the distributed photovoltaic diagnostic model. Instead, the acquired electrical feature vector is matched with the electrical feature fingerprint obtained by statistical analysis. The electrical feature fingerprint can reflect the feature information under the influence of a single fault source. If the corresponding similarity during matching meets the expectations, then the judgment of a single fault source can be directly performed.

[0046] The fault source combination judgment module uses the electrical feature vector as the target vector and the electrical feature fingerprint as the basis vector. Through a population optimization algorithm, it performs a linear partitioning operation to obtain the contribution component of the fault source information corresponding to each electrical feature fingerprint. This contribution component represents the degree of influence of each fault source on the electrical feature vector under the influence of multiple fault sources. The module then obtains the occurrence frequency of each fault source. The occurrence frequency is obtained by statistically analyzing the occurrence frequency of fault sources in the historical records of distributed photovoltaic systems. The occurrence frequency characterizes the probability of each fault source occurring in the corresponding distributed photovoltaic system. The product of the contribution component corresponding to the fault source information and the occurrence frequency is denoted as... The probability of a fault source occurrence is analyzed based on the probability of occurrence of all fault source information and the set probability threshold for fault source combinations. The corresponding fault source combination information is output. The fault source combination information includes several fault sources. The probability threshold for fault source combinations is set by the operator based on experience, or it can be adjusted according to the false alarm rate of fault source combination judgment. It should be noted that the probability of occurrence of a fault source consists of two parts: contribution component and occurrence frequency. If the contribution component is low but the occurrence frequency is high, then the probability of occurrence of the corresponding fault source will not be too low, indicating that the corresponding fault source also exists. By analyzing the probability of occurrence of fault sources, the accuracy of fault source combination judgment can be improved.

[0047] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims. Parts not described in detail in this specification are prior art known to those skilled in the art.

Claims

1. A distributed photovoltaic diagnostic method based on multimodal time series data, characterized in that, include: Regularly acquire multimodal time-series data from distributed photovoltaic systems; The acquired multimodal time-series data is fed into the distributed photovoltaic diagnostic model for processing, and the corresponding electrical feature vector is output. The output electrical feature vector is sequentially matched with the electrical feature fingerprints in the electrical feature fingerprint database. The electrical feature fingerprints are the cluster centers of the electrical feature vectors corresponding to all training samples after performing cluster analysis. The matching operation specifically calculates the similarity between the electrical feature vector and the electrical feature fingerprint, and determines whether the similarity between the electrical feature vector and the electrical feature fingerprint is higher than the similarity threshold. If the similarity between the electrical feature vector and the electrical feature fingerprint is higher than the similarity threshold, the fault source information corresponding to the electrical feature fingerprint is directly output. If the similarity between the electrical feature vector and the electrical feature fingerprint is not higher than the similarity threshold, the fault source combination judgment operation is performed, and the corresponding fault source combination information is output. The electrical feature fingerprint database contains one-to-one correspondences between fault source information and electrical feature fingerprints. The fault source combination judgment operation is as follows: using the electrical feature vector as the target vector and the electrical feature fingerprint as the basis vector, a linear partitioning operation is performed through a swarm optimization algorithm to obtain the contribution component of the fault source information corresponding to each electrical feature fingerprint. This includes: setting simulated weight component individuals, each of which includes the simulated weight component corresponding to each electrical feature fingerprint; forming a population set from all simulated weight component individuals; setting a maximum number of iterations; calculating the fitness of each simulated weight component individual by performing a weighted summation of all electrical feature fingerprints according to the simulated weight components in the simulated weight component individuals to obtain a simulated vector; using the similarity between the simulated vector and the target vector as the fitness of the simulated weight component individual; iteratively updating the population set based on the fitness of the simulated weight component individuals through a swarm optimization algorithm until the maximum number of iterations is reached; and then normalizing the simulated weight component individual with the highest fitness to obtain the contribution component of the fault source information corresponding to each electrical feature fingerprint. Next, obtain the occurrence frequency corresponding to each fault source. The product of the contribution component corresponding to the fault source information and the occurrence frequency is recorded as the occurrence probability of the fault source. The occurrence probabilities of all fault source information are added to the judgment set in descending order. The judgment set is initially empty. Whenever a fault source occurrence probability is added to the judgment set, the sum of the occurrence probabilities of all fault sources in the judgment set is calculated. This continues until the sum of the occurrence probabilities of all fault sources in the judgment set is higher than the fault source combination probability threshold. Then, the fault source information corresponding to the occurrence probabilities of all fault sources in the judgment set is combined into the corresponding fault source combination information, and the corresponding fault source combination information is output.

2. The distributed photovoltaic diagnostic method based on multimodal time series data according to claim 1, characterized in that, The electrical fingerprint database is constructed in the following way: For each fault source information, several training samples labeled by the fault source information are obtained. The training samples include multimodal time series data. The training samples are sent to the distributed photovoltaic diagnostic model for processing, and the electrical feature vectors corresponding to the training samples are output. Then, the electrical feature vectors corresponding to all training samples are clustered and the cluster center is used as the electrical feature fingerprint corresponding to the fault source information.

3. The distributed photovoltaic diagnostic method based on multimodal time series data according to claim 2, characterized in that, The distributed photovoltaic diagnostic model consists of an LSTM layer, a fully connected layer, and an identification layer. The LSTM layer is used to receive multimodal time-series data. In the fully connected layer, the outputs of all LSTM layers are fully connected to achieve feature dimension unification and normalization. The identification layer is used to process the output of the fully connected layer and output a probability vector. The probability vector includes the probability corresponding to each fault source, and the output of the fully connected layer is used as the corresponding electrical feature vector.

4. The distributed photovoltaic diagnostic method based on multimodal time series data according to claim 3, characterized in that, The training of the distributed photovoltaic diagnostic model includes the following steps: forming a training set by combining all training samples labeled with fault source information, training the distributed photovoltaic diagnostic model using the training set, using the labeled fault source information as the target output during training, calculating the loss value based on the target output and the predicted output of the distributed photovoltaic diagnostic model, and then updating the parameters of the distributed photovoltaic diagnostic model using the gradient descent method based on the direction of minimizing the loss value.

5. A distributed photovoltaic diagnostic system based on multimodal time-series data, characterized in that, The system employs a distributed photovoltaic diagnostic method based on multimodal time-series data as described in any one of claims 1-4, comprising: The multimodal time-series data acquisition module is used to periodically acquire multimodal time-series data from distributed photovoltaic power generation. The electrical feature vector extraction module is used to input the currently acquired multimodal time series data into the distributed photovoltaic diagnostic model for processing and output the corresponding electrical feature vector. The single fault source determination module is used to sequentially match the output electrical feature vector with the electrical feature fingerprints in the electrical feature fingerprint database. The matching operation specifically involves calculating the similarity between the electrical feature vector and the electrical feature fingerprint, and determining whether the similarity between the electrical feature vector and the electrical feature fingerprint is higher than the similarity threshold. If the similarity between the electrical feature vector and the electrical feature fingerprint is higher than the similarity threshold, the fault source information corresponding to the electrical feature fingerprint is directly output. If the similarity between the electrical feature vector and the electrical feature fingerprint is not higher than the similarity threshold, the fault source combination determination operation is performed, and the corresponding fault source combination information is output. The electrical feature fingerprint database includes a one-to-one correspondence between fault source information and electrical feature fingerprints. The fault source combination judgment module is used to judge the fault source combination operation. Specifically, it uses the electrical feature vector as the target vector and the electrical feature fingerprint as the basis vector. It performs a linear split operation through a group optimization algorithm to obtain the contribution component of the fault source information corresponding to each electrical feature fingerprint. Then, it obtains the occurrence frequency of each fault source. Based on the occurrence probability of all fault source information and the set fault source combination probability threshold, it analyzes and outputs the corresponding fault source combination information.

Images