Unsupervised fusion method for fault data of multiple space-time photovoltaic power stations

Through data alignment and feature space sharing, combined with feature extractor and classifier, the data alignment and sample number imbalance in the fusion of fault data of multi-source domain photovoltaic power stations is solved, and rapid fault analysis and accurate classification are realized, reducing training time .

CN119066609APending Publication Date: 2024-12-03JIANG SU ZHONG DUO LV NENG XIN NENG YUAN KE JI YOU XIAN GONG SI
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411094927.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-11
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The prior art has problems in data alignment, unbalanced sample number and random results in the fusion of fault data of multi-source photovoltaic power stations, resulting in large amount of calculation and poor classification effect.

Method used

Through the combination of data alignment, data quantity balance and feature extractor, multi-class classifier, and feature decoder, a shared feature space is built, and the target domain has labeled data is used to align the target domain with labeled data, and the loss function is optimized to reduce domain offset.

Benefits of technology

It realizes rapid failure analysis on new photovoltaic power stations, reduces the time to accumulate individual training data, and improves the accuracy and efficiency of fault classification.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119066609A_ABST
    Figure CN119066609A_ABST
Patent Text Reader

Abstract

The invention is suitable for the technical field of photovoltaic power station operation and maintenance monitoring, and provides a multi-space-time photovoltaic power station fault data unsupervised fusion method, which comprises the following steps of: data alignment: performing class reconstruction alignment on data of a plurality of photovoltaic power station fault data sources according to a source domain in a learned potential feature space; data expansion: carrying out data volume balance; and data fusion: training a model, constructing a shared feature space, aligning target domain samples and source domain samples in the shared feature space, and realizing fault classification of the samples without labels in the target domain by using the corresponding relationship of the samples with labels in the source domain and the alignment relationship of the samples without labels in the target domain. According to the method, analysis and data volume balance are carried out based on the existing labeled fault data of different sources and different data volumes of the labeled data, and a model capable of carrying out fault analysis on a new photovoltaic power station is obtained, so that the accumulated time required by independent training of the model is shortened, and the method can be quickly used.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of photovoltaic power station operation and maintenance monitoring, and in particular relates to an unsupervised fusion method for multi-temporal and spatial photovoltaic power station fault data. Background Art

[0002] Photovoltaic power stations are the main producers of photovoltaic energy, but photovoltaic power is affected by various factors, and the output varies greatly, including parameters such as region, season, and radiation. At the same time, there are many parameters that affect its normal operation, including weather, dust accumulation, moisture, environment, etc., so its fault analysis is a problem of multi-source data fusion, and usually requires years of data accumulation. Energy groups build multiple photovoltaic power stations in different environments, such as deserts, Gobi, and corporate rooftops. Due to differences in establishment time, location, and environment, the distribution of generated data is offset. Faced with this challenge, industry personnel usually use the labeled data (manually labeled by operation and maintenance personnel) of multiple photovoltaic power stations (source domain) that have been built and operated in different conditions (deserts, Gobi) and different data volumes in the early stage to analyze the operating data of newly built photovoltaic power stations (target domain) with unlabeled data under different conditions (corporate roofs) to obtain their fault types, thereby assisting maintenance work. Theoretically, this is an unsupervised multi-source domain data fusion problem, which aims to find the common similarities between samples, tasks or models in the source domain and the target domain, and use this as knowledge for learning to achieve the purpose of migrating source domain knowledge to the target domain. The core is to reduce the distribution difference between the source domain and the target domain and overcome domain shift. The focus is on fully exploring and utilizing the common domain invariant characteristics between the source domain and the target domain, so that the feature information extracted from the source domain can play a greater role, thereby eliminating or reducing the impact of domain shift, and learning from the labeled source domain to a model that can be well generalized to different but related target domains, thereby achieving accurate prediction of target samples.

[0003] In this field, the deep cocktail network proposed in the paper "Deep Cocktail Network: Multi-source UnsuperivisedDomain Adaptation with Category Shift" is one of the classic algorithms for solving such problems. However, this algorithm has several limitations: first, it does not consider data alignment, which makes each calculation start from a random way, making the amount of calculation larger; second, it does not consider the number of samples from each source. Different numbers of samples will make the source with more samples occupy a greater weight and have a greater influence during the learning process; third, it does not consider the alignment of target domain samples with source domain samples, which makes the learning end in a random way, thus bringing a certain degree of randomness to the results. In response to the above problems, the present invention proposes an unsupervised fusion method for multi-temporal and spatial photovoltaic power station fault data. Summary of the invention

[0004] The purpose of the present invention is to provide an unsupervised fusion method for multi-temporal and spatial photovoltaic power station fault data, aiming to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] An unsupervised fusion method for multi-temporal and spatial photovoltaic power station fault data includes the following steps:

[0007] Step 1: Data alignment: Align the data from multiple PV power station fault data sources according to the class reconstruction of the source domain in the learned potential feature space; in the learning space, place all samples belonging to different source domains but sharing a common class label closely. The specific process is: count the data samples from each source, use Indicates that from each The samples obtained from the source domains are: Where y is the fault label of the PV power station, bi is the output fault classification number; i Gather them together and sort them in the order of common fault class labels first, then arrange them in the order of sources within the samples with common fault class labels to achieve data alignment;

[0008] Step 2: Data expansion: balance the data volume. The specific process of data volume balance is as follows: for each fault class label, Count the samples from the sources in , count the largest number maxX(bi) among all sources 1 to K, and then expand the number of samples of this fault class from the other K-1 sources to the size of maxX(bi);

[0009] Step 3: Data fusion: Train a model and build a shared feature space. In the shared feature space, align the target domain samples with the source domain samples and use the y s =f(x s ) correspondence f, and the target domain has no labeled samples Align the relationship to achieve fault classification for unlabeled samples in the target domain

[0010] Furthermore, in the data fusion step, the model includes three sub-modules: a feature extractor, a multi-class classifier, and a feature decoder.

[0011] Furthermore, for each of the K source domains S containing the x, y mapping relationship k, each of them designs a feature extractor E, and obtains K feature extractors The feature extractor uses the standard ResNet101 model, distills the first-layer feature output of ResNet101, sorts the features according to the weight, and counts the first 128 features as feature representation; the specific process is: collect inverter-level data of photovoltaic power stations, a total of 18 parameters, and process the data at a time resolution of once every 5 minutes; normalize each parameter according to the historical maximum and minimum values; then align each normalized parameter according to the timing information of each inverter as the resolution of the branch on the hardware to form a piece of data, and arrange the piece of data horizontally into a one-dimensional matrix; then form a 288*18 matrix x of 24 hours of historical data from bottom to top as the input of each learning; finally, mark the 8 types of fault types that are manually marked and cannot be analyzed as 1 to 8 as the output y of each learning.

[0012] Furthermore, the multi-class classifier C that classifies each source domain data k , a B-dimensional vector probability score is generated at the output, indicating the possibility of a specific bi-th type of fault, B=8, that is, bi is between 1 and 8; the model structure of the multi-class classifier includes an MLP model with 1 input layer, 1 hidden layer and 1 output layer. The input layer receives 128 features output by the feature extractor, the hidden layer includes 256 neurons, and the output layer includes 8 neurons, corresponding to the possibility of 8 types of faults.

[0013] Furthermore, for the sample S belonging to source l l and the sample S belonging to source k k , build Feature decoder {D kl}; The model framework of the feature decoder is based on the ResNet101 model framework, including 2 fully connected layers, with 128 neurons in each layer.

[0014] Furthermore, in the data fusion step, the model loss function optimization process includes a source domain alignment loss function, a source-target alignment module and an inter-class maximum module.

[0015] Furthermore, in the source domain alignment loss function, Indicates that from each The source domains share a common class label, namely A sample of Represents the encoded data, represents the lth decoded representation of the kth encoded data. The source domain alignment loss function includes the following three loss functions:

[0016] a. According to the following formula, by minimizing the loss function Minimize ξ i The difference in the encoding feature representation of samples from different source domains; among them, samples belonging to source l and samples belonging to source k Have a common fault class label, i.e., y i =bi:

[0017] Formula 1:

[0018] in express expectations;

[0019] b. For a sample belonging to k sources By the given encoder E k Encode, get Definition D kl is to decode samples from k sources across domains The cross-domain decoding output to the l domain is obtained by minimizing the loss function according to the following formula: Implement the acquisition of class-level shared features between k and l domains;

[0020] Formula 2:

[0021] c. According to the encoding feature representation given by the feature extractor E, for a given data Represents the original true classification of the data, C kb Indicates that the classifier C k The output classification of is based on the following formula, by minimizing the loss function Minimize the cross entropy loss for each classifier;

[0022] Formula 3:

[0023] Furthermore, in the source-target alignment module, for a given target domain sample Through the k-th feature extractor, obtain the k-th feature representation of the target domain sample Then, through the classifier C k The samples were further analyzed using the B-dimensional unstandardized logit scores of the classifiers; assuming yes Feature extractor in k source domain, yes The B-dimensional logit vector defines the source-target alignment consistency loss As follows:

[0024] Formula 4:

[0025] Furthermore, in the maximum inter-class module, for a sample set belonging to class bl Encode it through the given encoder to get its feature center For a source sample set belonging to class bk bl≠bk, encoded by the given encoder, and its feature center is obtained maximize and The distance between the encoded The specific formula is as follows:

[0026] Formula 5:

[0027] Finally, the central loss metric of the source domain is minimized together and minimized by The loss of the source domain and the maximization Given a sample from the target domain, train the parameters of the classifier at the same time:

[0028] Formula 6:

[0029] Then by minimizing Based on the target domain samples, update the feature encoder E k :

[0030] Formula 7:

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] This unsupervised fusion method for multi-temporal and spatial photovoltaic power station fault data is based on the labeled fault data of different data volumes from different sources of existing labeled data, and performs analysis and data volume balance to obtain a model that can perform fault analysis on new photovoltaic power stations, thereby reducing the accumulation time required for its own individual training and enabling rapid use. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 The flowchart of a specific embodiment of the present invention. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0035] The specific implementation of the present invention is described in detail below in conjunction with specific embodiments.

[0036] like Figure 1As shown, an unsupervised fusion method for multi-temporal and spatial photovoltaic power station fault data provided by an embodiment of the present invention can be based on the data of multiple photovoltaic power stations with labeled data (under different temporal and spatial conditions and data volume conditions) to obtain a model and perform fault analysis on unlabeled photovoltaic power stations in new regions.

[0037] The task of the present invention is theoretically defined as: there are multiple source domains (photovoltaic power stations that have been built and operated previously) S = (S 1 ,S 2 ,...), they have accumulated labeled training samples with different data volumes and their own temporal and spatial characteristics (the data distribution is not completely consistent with each other); a target domain T (a photovoltaic power station that has just been built and has no data accumulation) has unlabeled test samples. Training samples can be obtained from multiple source domains. The goal is to merge the labeled data samples of multiple domains, train the model, and then analyze them on the unlabeled target domain samples. That is, there are K source domains S k ,in Each sample consists of a training sample from a specific domain. Composition, of which and x is the status data of the photovoltaic power station; y is the fault label of the photovoltaic power station; R is the real number domain; d is the dimension of the input status data, such as temperature, humidity, etc.; bi is the output fault classification number, from 1 to B, B (Bug). represents the training set of the kth source domain. On the other hand, T represents the sample The only unlabeled target domain.

[0038] Assuming that the distribution of the domains is different, that is, they have their own time and space characteristics, resulting in inconsistency in the data distribution. In many cases, the class conditional probabilities between S and T are also different, that is, the distribution between S is inconsistent, and the sample distribution between T and S is also inconsistent. That is, consider There exists P(S k )≠P(S l )and

[0039] The method specifically comprises the following steps:

[0040] Step 1, data alignment: The data from multiple PV power plant fault data sources are reconstructed and aligned according to the class of the source domain in the learned latent feature space; in the learning space, all samples belonging to different source domains but sharing a common class label are forced to be placed closely, which facilitates the construction of domain-invariant space models for all source domains. This solution is more intuitive and easier to capture their features than simply merging all domains in the original feature space and ignoring their differences. In addition, this also reduces the need to learn a separate mapping for each source domain-target domain.

[0041] Specifically: Count the data samples from each source, and use Indicates that from each The samples obtained from the source domain have a common fault class label, namely Where y is the fault label of the PV power station, bi is the output fault classification number; these ξ i Gather them together and sort them in the order of common fault class labels first, and then arrange them in the order of source within the samples with common fault class labels to achieve data alignment.

[0042] The principle is similar to when people play cards. If the cards in their hands are placed in a disorderly manner, it is difficult to find straights, flushes, and bombs. However, if the cards are placed in order of size and suit, it is relatively easy to find the pattern and find straights, flushes, etc. In this way, the computational efficiency of deep learning networks in finding patterns can be improved, and the accuracy of finding feature patterns can be improved.

[0043] Step 2, data expansion: Considering that the amount of labeled data in the new and old power plants is different, the characteristics of power plants with small data volumes will be weakened during learning, while the characteristics of power plants with large data volumes will be enhanced during learning, thus affecting the final classification effect. Therefore, a data expansion solution based on the consistent data source within the data fault class is adopted to balance the data volume.

[0044] The specific steps for data volume balancing are:

[0045] For each fault class label, The samples from the sources in , count the maximum number maxX(bi) among all sources 1~K, and then expand the number of samples of this fault class (bi) from the other K-1 sources to the size of maxX(bi), according to the scheme in "SSGAN:A Semantic Sililarity-Based GAN for Small-Sample Image Augmentation", expand to the size of maxX(bi).

[0046] Assume that for the second type of fault samples (bi=2), there are a total of 4 sources, namely, 2 rooftop photovoltaic power stations of enterprises in Suzhou, and 1 standard ground photovoltaic power station in Huai'an and Suqian. That is, there are 4 labeled data sets in the source domain, and the number of fault samples of the second type (bi=2) provided by them is {100, 200, 300, 400} respectively. At this moment, maxX(2)=400, and then according to the scheme in "SSGAN: A Semantic Sililarity-Based GAN for Small-Sample Image Augmentation", the second type of fault samples from the other 3 sources are expanded to a size of 400, that is, {400, 400, 400, 400}. In this way, the number of samples is expanded, and the learning weights of the power station features of each source are equal during the learning process.

[0047] Step 3: Data fusion: Train a model and build a shared feature space. In the shared feature space, align the target domain samples with the source domain samples, so that the y of the labeled samples in the source domain can be used. s =f(x s ) correspondence f, and the target domain has no labeled samples Align the relationship to achieve fault classification for unlabeled samples in the target domain

[0048] The sorting principle here is the same as that of data alignment, similar to when people play cards, if people first align the rules according to pairs, and then according to the regular results of suits such as straights, it is easy to find the same type of cards as their friends.

[0049] Considering the implicit differences in the data generation process of different sources, the existing general learning modules cannot fully capture the inherent domain shift between different data sources. In terms of structure, based on the main structure of the deep cocktail network proposed in the paper "Deep Cocktail Network: Multi-source Unsuperivised Domain Adaptation with Category Shift", the model in this invention includes three sub-modules: feature extractor, multi-class classifier, and feature decoder.

[0050] (1) Feature extractor E (Encoder). For each K source domain S containing the mapping relationship between x and y k ,in Each of them designs a feature extractor E, and obtains K feature extractors The feature extractor uses the standard ResNet101 model, distills the first-layer feature output of ResNet101, sorts the features according to the weight, and counts the first 128 features as feature representation; as feature extractor E. The source domain data uses two rooftop photovoltaic power stations in Suzhou, and one standard ground photovoltaic power station in Huai'an and Suqian. That is, there are 4 labeled data sets in the source domain, that is, at this moment, K=4; the target domain is a rooftop photovoltaic power station of a company in Yangzhou. They are owned by the same photovoltaic energy company, with the same electrical structure and photovoltaic panel parameters. The data is centralized in the photovoltaic power station through the 5G network, and the data is sent to the data platform of the photovoltaic energy company once every 5 minutes for storage and analysis.

[0051] In terms of data, the data used is based on the inverter level, including 11 parameters such as current, voltage, rated voltage of solar panels, maximum current, actual output current, actual power, directed and undirected, plus 7 actual parameter data (irradiance, temperature, humidity, wind direction, wind speed, rainfall, air pressure, coordinates) obtained by the multi-factor automatic meteorological station built in the local photovoltaic power station. Weather forecast data is not used, totaling 18 parameters.

[0052] The specific process is as follows: collect inverter-level data of the photovoltaic power station, a total of 18 parameters, and process the data at a time resolution of once every 5 minutes; normalize each parameter according to the historical maximum and minimum values; then use each inverter as the resolution of the branch on the hardware to align each normalized parameter according to the timing information to form a piece of data, and arrange the piece of data horizontally into a one-dimensional matrix; then from bottom to top, 24 hours, a total of 24*12=288 pieces of historical data are formed into a 288*18 matrix x as the input for each learning; finally, the 7 types of fault types that are manually marked and the type that cannot be analyzed are marked as 1, and the total of 8 types of fault types are marked as 1 to 8, which are used as the output y of each learning.

[0053] (2) Multi-class classifier C (Classifier). Multi-class classifier C that classifies each source domain data k , a probability score of a B-dimensional (in this embodiment, B=8) vector is generated at the output, indicating the possibility of a specific bi-th type of fault, that is, bi is between 1 and 8;

[0054] The model structure of the multi-class classifier includes an MLP model with 1 input layer, 1 hidden layer and 1 output layer. Since the distilled features of the ResNet101 model used as feature extractor E total 128, the input (first layer) of this classifier is also 128 neurons, its hidden layer (second layer) is 256 neurons, and its output is 8 neurons, corresponding to the possibility of 8 types of faults.

[0055] (3) Feature decoder D (Decoder). This embodiment hopes to model a common feature space to achieve the goal of combining the feature encoding of source domain samples with all E after training. k The first is that the target domain data is basically the same as the source domain data. The second is that the target domain data is difficult to distinguish from the source domain data. Therefore, for the sample S belonging to source l l , samples S belonging to source k k Here, we construct Decoder kl}, and find their decoded relationships.

[0056] The model framework of the feature decoder is based on the ResNet101 model framework, which includes 2 fully connected layers with 128 neurons in each layer.

[0057] Similar to the adversarial generation framework of the classic Deep Cocktail Network, the present invention needs to control the loss function. In principle, this module uses adversarial training driven by classifier integration (source domain) to project the target domain samples into the same shared feature space. From a global perspective, it reduces the differences in feature distribution between multiple source domains, realizes the alignment of target domain samples with the source domain in the shared feature space, and increases the distance between the average centers of each fault class, thereby improving the overall classification accuracy as much as possible. Source domain alignment is mainly controlled by potential loss and decoder loss. Potential loss attempts to make class-level feature embeddings of different source domains similar; decoder loss performs cross-domain sample reconstruction at the class level. The latent space can learn shared feature representations at the class scale while ensuring class-level compactness. In addition, source domain-specific classifiers make the latent space more distinctive in nature. On the other hand, when predicting the labels of target domain data under the adversarial training strategy, the differences between these classifiers are reasonably utilized. Based on the classic Deep Cocktail Network framework, the loss function in the present invention is as follows:

[0058] 1) Source domain alignment loss function: Indicates that from each source domains share a common class label A sample of Represents the encoded data, represents the lth decoded representation of the kth encoded data. The source domain alignment loss function includes the following three loss functions:

[0059] a. Loss Function The purpose is to make i The encoding feature representation of all samples in is consistent;

[0060] According to Formula 1, by minimizing Minimize ξi The difference in the encoding feature representation of samples from different source domains; among them, samples belonging to source l and samples belonging to source k Have a common fault class label, i.e., y i =bi:

[0061] Formula 1:

[0062] where ξ i is obtained by random sampling, express expectations;

[0063] b. Loss function The goal is to achieve cross-domain reconstruction as much as possible on the decoder side, so as to ensure that the representation can effectively capture class-level shared features;

[0064] For a sample belonging to k sources By the given encoder E k Encode, get Definition D kl is to decode samples from k sources across domains The cross-domain decoding output to the l domain is obtained by minimizing This cross-domain class sample reconstruction difference can be minimized and can be used to improve the encoder E k The performance of the proposed method is to obtain the class-level shared features between the k and l domains. By considering minimizing all possible pairwise reconstruction loss functions It can make the encoded representation of all source domains domain-invariant features;

[0065] Formula 2:

[0066] c. Loss function The goal is to learn a separate classifier for source domain-specific data;

[0067] According to the encoding feature representation given by E, for a given data Represents the original true classification of the data (binary), C kb Indicates that the classifier C k The output classification (binary) is obtained by minimizing Minimize the cross entropy loss for each classifier;

[0068] Formula 3:

[0069] In this way, by minimizing Equations 1 to 3, the class-level feature embeddings of different source domains are guaranteed to be similar.

[0070] 2) Source-target alignment module: In an ideal situation, given a target domain sample, Encoder E k , you will get The encoding representation of the target domain sample. That is, for a given target domain sample Through the kth feature extractor, obtain its kth feature representation Then, through the classifier C k , and further analyze the samples using the B-dimensional unstandardized logit scores of each classifier. In this way, by improving the consistency between the logit scores of these K classifiers, we can achieve the following goals:

[0071] a. Each feature extractor shares information between its source domain and target domain, making the feature encoder (E k ) can capture the correlation between two domains.

[0072] b. In the feature space, by constraining the classifier’s consistency expectation, the target domain samples can be brought to the common source domain distribution. Assume that yes Feature extractor in k source domain, yes The B-dimensional logit vector defines the source-target alignment consistency loss Formula 4:

[0073] Formula 4:

[0074] 3) Maximum module between classes;

[0075] Loss Function The goal is to make the difference between two samples of different classes as large as possible after being decoded at the decoder, so as to avoid misclassification. Encode it through the given encoder to get its feature center For a Source sample set of classes Encode it through the given encoder to get its feature center It is hoped that after these two feature centers are encoded, the distance between the two sides will increase;

[0076] Formula 5:

[0077] Finally, the central loss metric of the source domain is minimized together and minimized by The loss of the source domain and the maximization Given a sample from the target domain, train the parameters of the classifier at the same time:

[0078] Formula 6:

[0079] Then by minimizing Based on the target domain samples, update the feature encoder (E k ):

[0080] Formula 7:

[0081] In this way, by implementing the present invention, users can obtain a fault data fusion algorithm for multiple photovoltaic power stations built at different times within the group, and analyze and balance the labeled fault data of different data volumes from different sources of existing labeled data, and obtain a model that can perform fault analysis on new photovoltaic power stations, thereby reducing the time required for its own individual training and enabling quick use.

[0082] The above are only preferred embodiments of the present invention. It should be pointed out that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention. These should also be regarded as the protection scope of the present invention. These will not affect the effect of the implementation of the present invention and the practicality of the patent.

Claims

1. An unsupervised fusion method for multi-temporal and spatial photovoltaic power station fault data, characterized in that: The following steps are involved: Step 1: Data alignment: Align the data from multiple PV power plant fault data sources according to the class reconstruction of the source domain in the learned latent feature space; In the learning space, all samples belonging to different source domains but sharing a common class label are placed closely. The specific process is: count the data samples from each source, use Indicates that from each The samples obtained from the source domains are: Where y is the fault label of the PV power station, bi is the output fault classification number; i Gather them together and sort them in the order of common fault class labels first, then arrange them in the order of sources within the samples with common fault class labels to achieve data alignment; Step 2: Data expansion: balance the data volume. The specific process of data volume balance is as follows: for each fault class label, Count the samples from the sources in , count the largest number maxX(bi) among all sources 1 to K, and then expand the number of samples of this fault class from the other K-1 sources to the size of maxX(bi); Step 3: Data fusion: Train a model and build a shared feature space. In the shared feature space, align the target domain samples with the source domain samples and use the y s =f(x s ) correspondence f, and the target domain has no labeled samples Align the relationship to achieve fault classification for unlabeled samples in the target domain 2. The unsupervised fusion method for multi-temporal and spatial photovoltaic power station fault data according to claim 1 is characterized in that: In the data fusion step, the model includes three submodules: feature extractor, multi-class classifier, and feature decoder.

3. The unsupervised fusion method for multi-temporal and spatial photovoltaic power station fault data according to claim 2 is characterized in that: For each K source domain S containing the mapping relationship between x and y k , each of them designs a feature extractor E, and obtains K feature extractors The feature extractor uses the standard ResNet101 model, distills the first-layer feature output of ResNet101, sorts the features according to their weights, and counts the first 128 features as feature representation; The specific process is as follows: collect inverter-level data of the photovoltaic power station, a total of 18 parameters, and process the data at a time resolution of once every 5 minutes; normalize each parameter according to the historical maximum and minimum values; then use each inverter as the resolution of the branch on the hardware to align each inverter according to the timing information to form a piece of data, and arrange the piece of data horizontally into a one-dimensional matrix; then form a 288*18 matrix x of 24 hours of historical data from bottom to top as the input for each learning; finally, mark the 8 types of fault types that are manually marked and cannot be analyzed as 1 to 8, which are used as the output y of each learning.

4. The unsupervised fusion method for multi-temporal and spatial photovoltaic power station fault data according to claim 3 is characterized in that: The multi-class classifier C that classifies each source domain data k , a B-dimensional vector probability score is generated at the output, indicating the possibility of a specific bi-th type of fault, B=8, that is, bi is between 1 and 8; the model structure of the multi-class classifier includes an MLP model with 1 input layer, 1 hidden layer and 1 output layer. The input layer receives 128 features output by the feature extractor, the hidden layer includes 256 neurons, and the output layer includes 8 neurons, corresponding to the possibility of 8 types of faults.

5. The unsupervised fusion method for multi-temporal and spatial photovoltaic power station fault data according to claim 4 is characterized in that: For a sample S belonging to source l l and the sample S belonging to source k k , build Feature decoder {D kl }; The model framework of the feature decoder is based on the ResNet101 model framework, including 2 fully connected layers, with 128 neurons in each layer.

6. The unsupervised fusion method for multi-temporal and spatial photovoltaic power station fault data according to claim 5 is characterized in that: In the data fusion step, the model loss function optimization process includes a source domain alignment loss function, a source-target alignment module, and an inter-class maximum module.

7. The unsupervised fusion method for multi-temporal and spatial photovoltaic power station fault data according to claim 6 is characterized in that: In the source domain alignment loss function, Indicates that from each The source domains share a common class label, namely A sample of Represents the encoded data, represents the lth decoded representation of the kth encoded data. The source domain alignment loss function includes the following three loss functions: a. According to the following formula, by minimizing the loss function Minimize ξ i The difference in the encoding feature representation of samples from different source domains; among them, samples belonging to source l and samples belonging to source k Have a common fault class label, i.e., y i =bi: Formula 1: in express expectations; b. For a sample belonging to k sources By the given encoder E k Encode, get Definition D kl is to decode samples from k sources across domains The cross-domain decoding output to the l domain is obtained by minimizing the loss function according to the following formula: Implement the acquisition of class-level shared features between k and l domains; Formula 2: c. According to the encoding feature representation given by the feature extractor E, for a given data Represents the original true classification of the data, C kb Indicates that the classifier C k The output classification of is based on the following formula, by minimizing the loss function Minimize the cross entropy loss for each classifier; Formula 3:

8. The unsupervised fusion method for multi-temporal and spatial photovoltaic power station fault data according to claim 7 is characterized in that: In the source-target alignment module, for a given target domain sample Through the k-th feature extractor, obtain the k-th feature representation of the target domain sample Then, through the classifier C k The samples were further analyzed using the B-dimensional unstandardized logit scores of the classifiers; assuming yes Feature extractor in k source domain, yes The B-dimensional logit vector defines the source-target alignment consistency loss As follows: Formula 4:

9. The unsupervised fusion method for multi-temporal and spatial photovoltaic power station fault data according to claim 8 is characterized in that: In the maximum module between classes, for a sample set belonging to class bl Encode it through the given encoder to get its feature center For a source sample set belonging to class bk bl≠bk, encoded by the given encoder, and its feature center is obtained maximize and The distance between the encoded The specific formula is as follows: Formula 5: Finally, the central loss metric of the source domain is minimized together and minimized by The loss of the source domain and the maximization Given a target domain sample, train the classifier parameters at the same time: Formula 6: Then by minimizing Based on the target domain samples, update the feature encoder E k : Formula 7: