Photovoltaic module fault accurate modeling method based on fusion algorithm

Abstract

The invention discloses a photovoltaic module fault accurate modeling method based on a fusion algorithm, and relates to the technical field of photovoltaic module monitoring, fine changes of different fault modes can be better captured by analyzing a time dependence factor Syyz and dynamically adjusting an analysis period through the fusion algorithm, and finally a more accurate fault prediction coefficient Gyxs is output. And the fault prediction precision is obviously improved. According to the method, the memory window and the analysis period of each basic model are dynamically adjusted according to the change of the real-time data, so that the adaptability of the model to the real-time data is enhanced. In a photovoltaic module system, due to continuous changes of environmental conditions and operation states, real-time data always present a relatively strong time-varying characteristic, by dynamically adjusting an analysis period, the method can flexibly adjust model parameters according to data changes, and model deviation caused by data fluctuation or imbalance is reduced, so that the robustness of the system is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic component monitoring, and in particular to a photovoltaic component fault accurate modeling method based on a fusion algorithm. Background Art

[0002] In the context of modern energy transformation, renewable energy, especially photovoltaic power generation technology, is gradually becoming an indispensable part of the global energy structure. With the widespread application of photovoltaic power generation systems, the performance and reliability issues of photovoltaic modules have gradually become key issues in system operation and maintenance. Photovoltaic module failure not only affects power generation efficiency, but may also cause more serious system problems. Therefore, accurate photovoltaic module fault prediction and diagnosis technology is crucial. To achieve this goal, machine learning and data mining technologies are widely used in fault detection and diagnosis of photovoltaic systems. Among them, fusion algorithms, as a method that combines the advantages of multiple models, have achieved remarkable results in many fields.

[0003] Fusion algorithms can provide more robust results when dealing with complex problems by combining the predictive capabilities of multiple individual models. However, in the specific application of accurate modeling of PV module failures, there is a prominent problem: PV system failures are often gradual and occur over a long time scale.

[0004] In the study of accurate modeling of photovoltaic module failures, the lack of long-term continuous data has become a core problem that needs to be solved urgently. Photovoltaic module failures usually show progressive characteristics, and their failure process may gradually appear over a long period of time, which requires the fault prediction model to be able to handle time dependence and long-term data changes. However, many photovoltaic systems currently ignore such problems in actual operation, which makes it difficult for existing models to fully capture the evolution of photovoltaic module failures, thereby affecting the accuracy of fault prediction. However, even fusion algorithms face the problem of unbalanced data sets. Data samples in normal states usually account for the vast majority, while the number of samples in fault states is relatively small, causing the model to over-learn the normal state during training, thereby affecting its ability to identify fault states. Furthermore, the operating data of photovoltaic systems are usually time series data with significant time dependence. Therefore, how to reasonably handle the long-term dependence in time series data and automatically adjust the learning cycle and time window length of each basic model is the key to solving the problem of data imbalance. By dynamically adjusting the memory window, the fusion algorithm can be more flexible in dealing with unbalanced data and improve the sensitivity of fault recognition. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides a method for accurately modeling photovoltaic component faults based on a fusion algorithm, which solves the problems in the above-mentioned background technology.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: A photovoltaic module fault accurate modeling method based on a fusion algorithm, comprising the following steps:

[0007] S1. Use several groups of sensors to monitor the working status information of photovoltaic modules in real time, and combine with historical data sets to obtain the status information of information samples in the historical data sets to determine the balance of information;

[0008] S2. Based on the data balance determined in S1, the dependence of each basic model on the historical time at the current moment in the fusion algorithm is analyzed to calculate the time dependence factor Syyz. Based on the time dependence factor Syyz, the change direction of the analysis cycle involved in each analysis of each basic model is preliminarily identified;

[0009] S3. According to the prediction deviation of each basic model in the fusion algorithm at the current moment, combined with the time dependency factor Syyz, determine the degree to which the corresponding basic model learns the current trend through historical data, so as to update the analysis cycle involved in each analysis of each basic model;

[0010] S4. With the dynamic adjustment of the analysis cycle, the fault assessment of each basic model is updated in real time, and the final fault prediction coefficient Gyxs is fitted and output from the output of the fusion algorithm based on deep learning technology;

[0011] S5. Using historical data, set the fault threshold K. By comparing and analyzing the final fault prediction coefficient Gyxs with the fault threshold K, the fault degree of the current photovoltaic module is determined, and the fault prediction accuracy is evaluated in combination with the actual situation.

[0012] Preferably, the specific steps of S1 include:

[0013] S11, using several groups of sensors to monitor the working status information of the photovoltaic module in real time, wherein the working status information includes the working voltage, current, temperature, power output, ambient temperature, ambient humidity, fault time, fault type, fault prediction value Gyz output by each basic model, and information involved when the fault occurs;

[0014] S12: According to the historical data set of the historical period, the status information of the information samples in the historical data set is obtained, wherein the status information of the information samples in the historical data set includes the number of data samples in the normal state. and the number of data samples in the fault state ,The data set is constructed by collecting statistics of the working status information of the real-time monitoring of the photovoltaic modules and the historical data set;

[0015] S13. According to the status information of the information samples in the historical data set in S12, the balance between the faulty data and the normal data is analyzed to calculate the data set balance Spd. The data set balance Spd is obtained by the following formula: ;

[0016] In the formula, Indicates the balance of the data set at the current time t, represents the number of data samples in the normal state at the current time t, Represents the number of data samples in the fault state at the current time t.

[0017] Preferably, the specific steps of S2 include:

[0018] S21. Outputting a corresponding fault prediction value Gyz through various known basic models, wherein the various known basic models include but are not limited to gated recurrent unit, support vector machine, decision tree, random forest and convolutional neural network.

[0019] Preferably, the specific step S2 also includes:

[0020] S22. According to the data set balance Spd obtained in S13, the dependence of each basic model on the historical time at the current moment in the fusion algorithm is analyzed, and after dimensionless processing, a time dependence factor Syyz is calculated and obtained. The time dependence factor Syyz is obtained by the following formula: ;

[0021] In the formula, represents the time-dependent factor at the current time t; Represents the prediction error at the current time t; represents the data change rate at the current time t, and Indicates the historical maximum prediction error and maximum data change rate; and is the weight value, where and The specific value is set by the user according to the situation.

[0022] Preferably, the specific step S2 also includes:

[0023] S23, by adding the time dependence factor at the current time t The time-dependent factor at the previous time t-1 Compare the size to preliminarily identify the change direction of the analysis cycle involved in each analysis of each basic model. If the time-dependent factor at the current time t is >Time-dependent factor at the previous time t-1 , the analysis period of the corresponding basic model involved in the next analysis will tend to be extended; if the time-dependent factor at the current time t <Time-dependent factor at the previous time t-1 , the analysis period involved in the next analysis of the preliminary identification of the corresponding basic model will show a shortened trend; if the time-dependent factor at the current time t = time-dependent factor at the previous time t-1 , the analysis period involved in the preliminary identification of the corresponding basic model will not change in the next analysis.

[0024] Preferably, the specific steps of S3 include:

[0025] S31. According to the prediction deviation of each basic model in the fusion algorithm at the current moment, combined with the time dependency factor Syyz, the degree to which the corresponding basic model learns the current trend through historical data is judged, and after dimensionless processing, the analysis cycle involved in each analysis of each basic model is updated, and marked as the memory window length update value Jcz. The memory window length update value Jcz is obtained by the following formula: ;

[0026] In the formula, Represents the updated value of the memory window length at the current time t, Indicates the updated value of the memory window length at the next time t+1.

[0027] Preferably, the specific steps of S4 include:

[0028] S41, according to the memory window length update value Jcz in S31, update the fault assessment of each basic model in real time to obtain the fault prediction value under the condition of the memory window length update value Jcz at the corresponding moment , and combined with the data set, dynamically update the weight w of the corresponding basic model, which can be obtained in the following way: ;

[0029] In the formula, represents the weight of the i-th basic model at the next time t+1, and They represent the prediction error of the i-th basic model at the current time t and the prediction error of the j-th basic model at the current time t, respectively. and They represent the historical error averages of the i-th basic model and the j-th basic model, respectively. N represents the number of basic models, j=1, 2, ..., N. Represents an exponential function with base e.

[0030] Preferably, the specific step S4 also includes:

[0031] S42. Based on the deep learning technology and in combination with the dynamic update of the weight w of the corresponding basic model, the final fault prediction coefficient Gyxs is fitted and output from the output end of the fusion algorithm, which is specifically obtained by the following formula: ;

[0032] In the formula, represents the weight of the i-th basic model at the next time t+1, Represents the fault prediction value of the i-th basic model at the next time t+1.

[0033] Preferably, the specific steps of S5 include:

[0034] S51, by comparing and analyzing the final fault prediction coefficient Gyxs with the fault threshold K, to determine the fault degree of the current photovoltaic module, if the final fault prediction coefficient Gyxs exceeds the fault threshold K, the current photovoltaic module is in a fault state, and the fault prediction accuracy is evaluated in combination with the actual situation, if the current prediction error If the final fault prediction coefficient Gyxs does not exceed the fault threshold K, the current PV module is not in a fault state.

[0035] The present invention provides a photovoltaic module fault accurate modeling method based on a fusion algorithm, which has the following beneficial effects:

[0036] (1) By using the fusion algorithm to integrate the weighted prediction results of multiple basic models, the advantages of different models can be effectively integrated to avoid the limitations of a single model. In particular, in the prediction of photovoltaic module failures, due to the complexity and progressiveness of the failure mode, a single model is difficult to provide accurate predictions. By analyzing the time-dependent factor Syyz and dynamically adjusting the analysis cycle, the fusion algorithm can better capture the subtle changes in different failure modes, and ultimately output a more accurate fault prediction coefficient Gyxs, significantly improving the accuracy of fault prediction. This method dynamically adjusts the memory window and analysis cycle of each basic model according to the changes in real-time data, thereby enhancing the model's adaptability to real-time data. In photovoltaic module systems, due to the continuous changes in environmental conditions and operating conditions, real-time data often exhibits strong time-varying characteristics. By dynamically adjusting the analysis cycle, the method can flexibly adjust the model parameters according to data changes, reduce model deviations caused by data fluctuations or imbalances, and thus enhance the robustness of the system.

[0037] (2) In the fault monitoring of photovoltaic modules, fault data is usually less, while normal operation data accounts for the vast majority, which causes the problem of imbalanced data sets and affects the model's ability to learn about faults. By utilizing the state information of historical data samples and combining the dynamically adjusted time-dependent factors, the method can effectively improve the ability to identify fault data and reduce overfitting of normal state data, thereby improving the model's fault detection ability under unbalanced data conditions. Photovoltaic system faults usually show gradual changes, and the time scale of fault occurrence is relatively long. By combining the long-term dependency capture of each basic model, especially by analyzing the prediction deviation at the current moment and the trend of historical data learning, the method can continuously update the learning state of the model over a long period of time. In this way, the system can better capture the occurrence and development of progressive faults and improve the foresight and timeliness of fault prediction. Finally, the fault prediction coefficient output by the fusion model is compared and analyzed with the set fault threshold, which can accurately determine the fault degree of the photovoltaic module. This not only improves the accuracy of fault diagnosis, but also enables accurate evaluation by combining the actual situation, further avoiding the misjudgment of the fault state in traditional methods and improving the reliability of the system in practical applications. In summary, this method not only improves the accuracy of fault prediction, but also enhances the robustness and adaptability of the system by integrating the advantages of multiple basic models, dynamically adjusting the analysis cycle, capturing long-term dependencies, and solving data imbalance problems. It has significant practical application value for fault monitoring and prediction of photovoltaic modules.

[0038] (3) In step S22, the learning and memory windows of each basic model can be dynamically adjusted by dimensionless processing and calculating the time-dependent factor Syyz. This calculation process can measure factors such as the prediction error at the current moment, the rate of data change, and the balance of the data set, and accurately reflect the amount of historical information required by the photovoltaic module at a specific moment. The adaptive update of the time-dependent factor enables the model to adjust its reliance on historical information according to the fluctuation of the actual data, making it more flexible and accurate in dealing with different fault modes. This mechanism can effectively deal with progressive faults in photovoltaic systems, especially when fault characteristics gradually emerge under environmental changes or long-term influencing factors.

[0039] (4) By comparing the time-dependent factors of the current moment with the previous moment in real time, the model can dynamically adjust the analysis cycle of the basic model. This dynamic adjustment mechanism enables the basic model to adaptively handle different types of faults. For example, when a PV module fault gradually develops or the external environment changes, the model will automatically extend the analysis cycle to capture more historical data and avoid prediction errors caused by ignoring gradual changes. Conversely, when the fault is more sudden or changes faster, the analysis cycle will automatically shorten to speed up the response and capture the fault in time.

[0040] (5) In step S41, by updating the fault assessment values ​​of each basic model in real time and dynamically adjusting the model weights according to the average value of the current prediction error and the historical error, this method ensures that the model can automatically adjust its importance in the overall fault prediction according to the changes in real-time data. In particular, when the data changes greatly, the weight of the larger memory window model will increase, allowing the model to rely more on the long-term trend of historical data; when the data changes less or sudden faults occur, the adjustment of the weights helps to reduce the impact of excessive reliance on historical data and maintain the agility of the model. Summary: Step S4 dynamically updates the memory window length and the weight of the basic model in real time, combined with deep learning technology, which not only enhances the adaptive ability and real-time response ability of fault prediction, but also effectively avoids the interference of irrelevant historical data, thereby improving the accuracy and reliability of PV module fault detection. In addition, by comparing the fault prediction coefficients, the fault degree of the PV module can be accurately judged, and a scientific basis can be provided for subsequent fault handling and maintenance decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 The present invention is a schematic flow chart of a method for accurately modeling photovoltaic component faults based on a fusion algorithm. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] Example 1

[0044] See also Figure 1 The present invention provides a photovoltaic module fault accurate modeling method based on a fusion algorithm, comprising the following steps:

[0045] S1. Use several groups of sensors to monitor the working status information of photovoltaic modules in real time, and combine with historical data sets to obtain the status information of information samples in the historical data sets to determine the balance of information;

[0046] S2. Based on the data balance determined in S1, the dependence of each basic model on the historical time at the current moment in the fusion algorithm is analyzed to calculate the time dependence factor Syyz. Based on the time dependence factor Syyz, the change direction of the analysis cycle involved in each analysis of each basic model is preliminarily identified;

[0047] S3. According to the prediction deviation of each basic model in the fusion algorithm at the current moment, combined with the time dependency factor Syyz, determine the degree to which the corresponding basic model learns the current trend through historical data, capture long-term dependencies, and update the analysis cycle involved in each analysis of each basic model;

[0048] S4. With the dynamic adjustment of the analysis cycle, the fault assessment of each basic model is updated in real time, and the final fault prediction coefficient Gyxs is fitted and output from the output of the fusion algorithm based on deep learning technology;

[0049] S5. Using historical data, set the fault threshold K. By comparing and analyzing the final fault prediction coefficient Gyxs with the fault threshold K, the fault degree of the current photovoltaic module is determined, and the fault prediction accuracy is evaluated in combination with the actual situation.

[0050] In this embodiment, by using several groups of sensors to monitor the working status information of photovoltaic components in real time, and combining historical data sets to analyze the balance of information samples, it is possible to prepare for the later effective reduction of the negative impact caused by data imbalance. This process ensures that the model can fully consider the various operating states of photovoltaic components and enhance the accuracy of fault detection. At the same time, based on the fusion algorithm, the results of multiple basic models are weighted and fused, which can further improve the model's ability to identify abnormal conditions and fault states and enhance the robustness of the system. Dynamically adjust the memory window of the model to optimize the model learning ability: By calculating the time dependency factor Syyz and dynamically adjusting the analysis cycle of each basic model, the memory window length can be flexibly updated according to the data characteristics of the current moment. This adaptive mechanism can effectively capture the early signs and long-term evolution trends of faults under the progressive characteristics of photovoltaic system faults. Compared with the traditional fixed-cycle analysis method, the dynamically adjusted memory window can better adapt to the complexity of photovoltaic component faults and enhance the model's learning ability for long-term dependency information. Improve the ability to capture and predict progressive faults: The failure of photovoltaic components is usually progressive, and it is difficult to capture its potential risks through short-term data. Through deep learning of historical data by fusion algorithm, the progressive failure and long-term change trend of photovoltaic modules can be more effectively identified. In addition, by correcting the prediction deviation of the basic model based on the time-dependent factor, the learning process of each basic model can be adjusted more accurately, which helps to provide early warning of failures and reduce the failure downtime of the photovoltaic system. Accurately assess the risk of failure and improve the accuracy of fault prediction: In the final fault prediction process, the fusion model based on deep learning combines the prediction results of different basic models and fits the final fault prediction coefficient Gyxs. ​​By comparing with the preset fault threshold K, the degree of failure of the current photovoltaic module can be accurately judged. Combined with the feedback of the actual operating conditions, this process can provide more accurate fault prediction results for operation and maintenance personnel, helping them to take timely measures, thereby effectively reducing the risk of failure and extending the service life of the photovoltaic system.

[0051] Example 2

[0052] Please refer to Figure 1 , specifically: S1 specific steps include:

[0053] S11, using several groups of sensors to monitor the working status information of the photovoltaic module in real time, wherein the working status information includes the working voltage, current, temperature, power output, ambient temperature, ambient humidity, fault time, fault type, fault prediction value Gyz output by each basic model, and information involved when the fault occurs;

[0054] S12: According to the historical data set of the historical period, the status information of the information samples in the historical data set is obtained, wherein the status information of the information samples in the historical data set includes the number of data samples in the normal state. and the number of data samples in the fault state ,The data set is constructed by collecting statistics of the working status information of the real-time monitoring of the photovoltaic modules and the historical data set;

[0055] S13. According to the status information of the information samples in the historical data set in S12, the balance between the faulty data and the normal data is analyzed to calculate the data set balance Spd. The data set balance Spd is obtained by the following formula: ;

[0056] In the formula, Indicates the balance of the data set at the current time t, represents the number of data samples in the normal state at the current time t, Represents the number of data samples in the fault state at the current time t.

[0057] In this embodiment, by using several groups of sensors to monitor the working voltage, current, temperature, power output, ambient temperature, ambient humidity, fault time, fault type and other multi-dimensional working status information of the photovoltaic module in real time, the operating performance of the photovoltaic module in different working environments and states can be understood in multiple dimensions. The collection of such multi-dimensional data can provide a richer information basis for subsequent fault analysis, thereby improving the accuracy and comprehensiveness of fault prediction. Build a high-quality data set to solve the problem of data imbalance: By combining the normal state and fault state data samples in the historical data set, a complete data set containing different state information can be effectively constructed. Quantitatively evaluate the balance of the data set and improve the efficiency of model learning: By calculating the balance of the data set Spd, the degree of balance between fault data and normal data can be accurately measured. By analyzing the balance of the data at the current moment, the weights of various types of data can be effectively adjusted during the model training and prediction process, thereby preventing the model from overfitting the normal state data and improving the sensitivity of the model to the fault data. This process can improve the learning efficiency of the data set and make the fault prediction model more accurate, especially in the case of unbalanced data samples, and can better identify and predict the potential faults of the photovoltaic module.

[0058] Example 3

[0059] Please refer to Figure 1 , specifically: S2 specific steps include:

[0060] S21. Output a corresponding fault prediction value Gyz through various known basic models, wherein the known basic models include but are not limited to a gated recurrent unit (GRU), a support vector machine (SVM), a decision tree (DT), a random forest (RF) and a convolutional neural network (CNN).

[0061] GRU (Gated Recurrent Unit) Overview: GRU is an improved RNN structure with a relatively simple structure, which only has update gates and reset gates, so the training efficiency is higher. Application: GRU is also suitable for fault prediction of equipment such as photovoltaic modules, especially when dealing with fault signals with strong short-term dependencies, GRU can respond to system changes more quickly. Existing technology: GRU is one of the commonly used models in modern deep learning and is widely used in natural language processing, time series prediction, signal processing and other fields.

[0062] Support Vector Machine (SVM) Overview: SVM is a supervised learning algorithm widely used in classification and regression problems. SVM can effectively solve small sample and nonlinear problems by finding an optimal hyperplane in high-dimensional space for classification. Application: In fault detection, SVM is often used for binary or multi-classification tasks of normal and faulty states of photovoltaic modules. Existing technology: Since its introduction in the late 1990s, SVM has become one of the most classic and mature algorithms in data mining and machine learning.

[0063] Decision Tree (DT) Overview: Decision tree is a supervised learning model that implements classification or regression tasks by recursively splitting data into several subsets. Each node of the tree represents a test of a feature, and the leaf node represents a category or value. Decision trees are easy to understand and interpret, especially when the data is unbalanced, the recognition ability of minority classes can be increased by appropriately adjusting the partitioning criteria. Application: In photovoltaic module fault prediction, decision trees can be used for classification tasks, such as distinguishing between normal, mild fault, severe fault and other states. By dividing the input features (such as temperature, current, voltage, etc.), it helps to make decisions quickly.

[0064] Random Forest (RF) Overview: Random Forest is an ensemble learning method that performs classification or regression through the combination of multiple decision trees. Each tree is trained on different samples and feature subsets, and the final prediction result is obtained by voting or averaging the trees. It has strong generalization ability and can effectively reduce overfitting. Application: In photovoltaic system fault prediction, random forest can classify system faults based on historical data.

[0065] Overview of Convolutional Neural Network (CNN): CNN is a deep learning model that is particularly suitable for processing image data, but can also process one-dimensional time series data. By using convolutional layers to extract local features in the data, CNN can effectively capture the spatial features in the input data. Application: In photovoltaic module failure prediction, CNN can be used to extract useful local features from electrical features (such as current waveforms, power changes) and detect sudden failure modes (such as electrical short circuits, overloads, etc.).

[0066] The specific steps of S2 also include:

[0067] S22. According to the data set balance Spd obtained in S13, the dependence of each basic model on the historical time at the current moment in the fusion algorithm is analyzed, and after dimensionless processing, a time dependence factor Syyz is calculated and obtained. The time dependence factor Syyz is obtained by the following formula: ;

[0068] In the formula, represents the time-dependent factor at the current time t; Represents the prediction error at the current time t; Indicates the rate of change of data at the current time t (the difference of data, measured by standard deviation), and Indicates the historical maximum prediction error and maximum data change rate, which is used for normalization; and is the weight value, where 0< <1,0< <1, and The specific value is set by the user according to the situation.

[0069] It measures the ratio of the error at the current moment to the maximum error. If the prediction error at the current moment is large, the value of this item is small, which means that the model needs more historical information (longer memory window) to "remember" the progressive characteristics of the fault. It reflects the data change rate of the current moment characteristics. The greater the data change rate, the more historical information the model needs to capture complex relationships (for example, environmental changes of PV panels or other long-term influencing factors). It reflects the balance degree between faulty data and normal data in the data set, and the impact on the time-dependent factor.

[0070] The larger the time dependency factor Syyz is, the stronger the time dependency is at the current moment, and more historical data is needed to help the model make accurate predictions.

[0071] In this embodiment, multi-model fusion and fault prediction accuracy are improved: in step S21, fault prediction is performed through a variety of known basic models, such as gated recurrent units, support vector machines, decision trees, random forests, convolutional neural networks, etc., and the corresponding fault prediction value Gyz is output. These basic models each have different learning capabilities and sensitivities to data features, and can predict the faults of photovoltaic modules from different angles. By fusing the prediction results of multiple models, the reliability and accuracy of fault prediction can be improved, thereby enhancing the fault detection capability of the overall system, especially in the case of diversity and complexity of photovoltaic modules. Dynamic adjustment of memory window and optimization of time dependence: in step S22, by calculating the time dependency factor Syyz, the analysis cycle (memory window) of the basic model is dynamically adjusted according to the prediction error of real-time data, the data change rate, and the balance of the data set. When the prediction error is large or the data change rate is high, the model will automatically increase the length of the memory window to enhance the ability to capture long-term dependencies. This dynamic adjustment mechanism enables the model to flexibly respond to the working conditions of different photovoltaic modules, especially under the progressive changes of faults, and can more effectively capture potential fault signals. Dimensionless processing and enhanced data adaptability: Through dimensionless processing, the prediction error and data change rate at the current moment are compared with historical data, making each parameter comparable in scale. This not only eliminates the influence between different dimensions, but also enhances the adaptability of the data, helps to eliminate the influence caused by the difference in measurement units or data sources of different sensors, and improves the stability and reliability of the model under different environmental conditions. Improve the ability to identify progressive faults: Based on the calculation of the time-dependent factor Syyz, it can accurately determine whether more historical data is needed at the current moment to capture the progressive fault characteristics of photovoltaic modules. If the prediction error at the current moment is large or the data change rate is high, it indicates that the photovoltaic module may undergo more complex changes, and more historical data is needed to help the model predict. In this way, the model can adjust the analysis cycle in time according to the dynamic changes of the time-dependent factor, so that the identification of progressive faults is more accurate and potential fault risks are warned in advance.

[0072] The impact of enhanced data set balance on prediction effect: In the process of calculating the time-dependent factor, the balance Spd between normal data and fault data in the data set is considered. This balance reflects the proportional relationship between normal data and fault data in the data set, which helps to evaluate whether the model needs to strengthen the learning of fault data at the current moment. In the case of an unbalanced data set, the learning process of the model can be balanced through appropriate adjustments to avoid overfitting normal data, thereby improving the sensitivity to fault states and ultimately improving the accuracy of fault prediction. In summary, step S2 can effectively improve the accuracy and reliability of fault prediction through the fusion of multiple models and the mechanism of dynamically adjusting the memory window. By making full use of the interaction between real-time data and historical data, this method can adaptively capture the progressive fault characteristics of photovoltaic modules, improve the prediction ability of complex fault modes, and thus improve the overall performance and stability of photovoltaic systems.

[0073] Example 4

[0074] Please refer to Figure 1 Specifically: S2 includes the following specific steps:

[0075] S23, by adding the time dependence factor at the current time t The time-dependent factor at the previous time t-1 Compare the size to preliminarily identify the change direction of the analysis cycle involved in each analysis of each basic model. If the time-dependent factor at the current time t is >Time-dependent factor at the previous time t-1 , the analysis period of the corresponding basic model involved in the next analysis will tend to be extended; if the time-dependent factor at the current time t <Time-dependent factor at the previous time t-1 , the analysis period involved in the next analysis of the preliminary identification of the corresponding basic model will show a shortened trend; if the time-dependent factor at the current time t = time-dependent factor at the previous time t-1 , the analysis period involved in the preliminary identification of the corresponding basic model will not change in the next analysis.

[0076] In this embodiment, adaptive analysis cycle adjustment: in step S23, by comparing the time dependency factor at the current moment with the previous moment, the analysis cycle of the basic model can be dynamically adjusted according to the changing trend of the fault. This mechanism enables the system to flexibly adapt to different development stages of photovoltaic module faults. For example, if the time dependency factor at the current moment is greater than that at the previous moment, it means that the fault mode may become more complex and more historical data is needed to assist in judgment, thereby extending the analysis cycle; on the contrary, if the time dependency factor at the current moment is less than that at the previous moment, it means that the fault trend tends to be stable or easy to identify, and the model can shorten the analysis cycle. This adaptive analysis cycle adjustment effectively improves the responsiveness of the system, so that it can detect and predict faults in a timely manner under various fault modes. Enhanced prediction accuracy and robustness: Through the dynamic adjustment of the time dependency factor, the model can capture enough historical information in complex situations, avoid ignoring long-term dependencies, and thus avoid misjudgment and underreporting. For example, some fault features may be relatively weak in the early stage, but as time goes by, the performance of the fault becomes increasingly obvious. The system can ensure the sensitivity and accuracy of the model during the development of the fault by extending the analysis cycle. At the same time, when the fault tends to stabilize, the system will automatically shorten the analysis cycle to avoid excessive historical information affecting the current prediction results, further improving the efficiency of fault detection.

[0077] Example 5

[0078] Please refer to Figure 1 , specifically: S3 specific steps include:

[0079] S31, according to the prediction deviation of each basic model at the current moment in the fusion algorithm (the prediction error at the current moment t ), and combined with the time-dependent factor Syyz, the degree to which the corresponding basic model learns the current trend through historical data is judged, and after dimensionless processing, the analysis cycle involved in each analysis of each basic model is updated, and marked as the memory window length update value Jcz, the memory window length update value Jcz is obtained by the following formula: ;

[0080] In the formula, Represents the updated value of the memory window length at the current time t, Indicates the updated value of the memory window length at the next time t+1.

[0081] In this embodiment, the memory window length is dynamically adjusted to improve the accuracy and adaptability of the prediction: by combining the prediction error at the current moment with the time dependency factor Syyz, the model can dynamically adjust the memory window length of each basic model. This adjustment can flexibly change the analysis cycle according to the changing rules of different fault modes, and improve the model's adaptability to progressive and sudden faults. Specifically, if the prediction error at the current moment is large and the time dependency factor is high, the model will extend the memory window to capture more historical data and help the model better learn the trend of long time scales; on the contrary, when the error is small or the dependency is weak, the memory window will be shortened, thereby speeding up the response speed and identifying faults in time. Optimize the learning process and improve fault identification ability: The dynamic calculation of the memory window length update value Jcz enables the basic model to automatically adjust its learning depth and historical data dependency in different situations. When a photovoltaic module fails, based on the changes in the current error and time dependency factor, the model can determine whether it needs to increase its reliance on historical data to avoid prediction deviations caused by over-reliance on the current state when there is less data. This adaptive adjustment mechanism can significantly improve the accuracy of fault identification, especially when the fault of PV modules has long-term progressive characteristics. The model can capture more subtle changes by extending the memory window.

[0082] Example 6

[0083] Please refer to Figure 1 , specifically: S4 specific steps include:

[0084] S41, according to the memory window length update value Jcz in S31, update the fault assessment of each basic model in real time to obtain the fault prediction value under the condition of the memory window length update value Jcz at the corresponding moment , and combined with the data set, dynamically update the weight w of the corresponding basic model, which can be obtained in the following way: ;

[0085] In the formula, represents the weight of the i-th basic model at the next time t+1, and They represent the prediction error of the i-th basic model at the current time t and the prediction error of the j-th basic model at the current time t, respectively. and They represent the historical error averages of the i-th basic model and the j-th basic model, respectively. N represents the number of basic models, j=1, 2, ..., N. It represents an exponential function with base e, where e is the Euler number and its value is approximately 2.71828.

[0086] The length of the memory window directly affects the amount of training data and memory range of the model at a specific moment. Therefore, the updated memory window length determines the time dependency range of each model, and the model weight will be dynamically adjusted with the length of the memory window. The larger the window, the stronger the model's reliance on historical data. Accordingly, the model's prediction results may be more dependent on long-term historical trends, and the weight will increase accordingly.

[0087] The specific steps of S4 also include:

[0088] S42. Based on the deep learning technology and in combination with the dynamic update of the weight w of the corresponding basic model, the final fault prediction coefficient Gyxs is fitted and output from the output end of the fusion algorithm, which is specifically obtained by the following formula: ;

[0089] In the formula, represents the weight of the i-th basic model at the next time t+1, Represents the fault prediction value of the i-th basic model at the next time t+1.

[0090] The specific steps of S5 include:

[0091] S51, by comparing and analyzing the final fault prediction coefficient Gyxs with the fault threshold K, to determine the fault degree of the current photovoltaic module, if the final fault prediction coefficient Gyxs exceeds the fault threshold K, the current photovoltaic module is in a fault state, and the fault prediction accuracy is evaluated in combination with the actual situation, if the current prediction error If the final fault prediction coefficient Gyxs does not exceed the fault threshold K, the current PV module is not in a fault state.

[0092] In this embodiment, the flexibility and accuracy of the prediction are enhanced: by adjusting the weight of each basic model according to the updated value Jcz of the memory window length, the model can adapt to different failure modes of photovoltaic components on different time scales. When the memory window is extended, the model can absorb more historical data, which is crucial for the prediction of progressive faults (such as long-term power attenuation). On the contrary, when the memory window is shortened, the model will pay more attention to the impact of current data and react more sensitively to short-term faults (such as sudden circuit problems). The dynamically updated weights enable the fusion algorithm to intelligently fuse the prediction results of each basic model to improve the accuracy of the overall prediction. Improve the real-time and accuracy of fault prediction: The change in the length of the memory window not only affects the prediction output of the basic model, but also directly affects the adjustment of the model weight. A longer memory window allows the model to rely more on historical data when predicting, ensuring that the changing trend can be accurately captured for long-term gradually changing faults, while a shorter memory window allows the model to quickly respond to instantaneous or short-term changing faults. Through this dynamic weight update mechanism, the "influence" of each model will be automatically adjusted according to its current performance and time dependency, thereby improving the real-time and accuracy of the overall fault prediction.

[0093] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A photovoltaic module fault accurate modeling method based on a fusion algorithm, characterized by: The following steps are involved: S1. Use several groups of sensors to monitor the working status information of photovoltaic modules in real time, and combine with historical data sets to obtain the status information of information samples in the historical data sets to determine the balance of information; S2. Based on the data balance determined in S1, the dependence of each basic model on the historical time at the current moment in the fusion algorithm is analyzed to calculate the time dependence factor Syyz. Based on the time dependence factor Syyz, the change direction of the analysis cycle involved in each analysis of each basic model is preliminarily identified; S3. According to the prediction deviation of each basic model in the fusion algorithm at the current moment, combined with the time dependency factor Syyz, determine the degree to which the corresponding basic model learns the current trend through historical data, so as to update the analysis cycle involved in each analysis of each basic model; S4. With the dynamic adjustment of the analysis cycle, the fault assessment of each basic model is updated in real time, and the final fault prediction coefficient Gyxs is fitted and output from the output of the fusion algorithm based on deep learning technology; S5. Using historical data, set the fault threshold K. By comparing and analyzing the final fault prediction coefficient Gyxs with the fault threshold K, the fault degree of the current photovoltaic module is determined, and the fault prediction accuracy is evaluated in combination with the actual situation.

2. The photovoltaic module fault accurate modeling method based on fusion algorithm according to claim 1 is characterized by: The specific steps of S1 include: S11, using several groups of sensors to monitor the working status information of the photovoltaic module in real time, wherein the working status information includes the working voltage, current, temperature, power output, ambient temperature, ambient humidity, fault time, fault type, fault prediction value Gyz output by each basic model, and information involved when the fault occurs; S12: According to the historical data set of the historical period, the status information of the information samples in the historical data set is obtained, wherein the status information of the information samples in the historical data set includes the number of data samples in the normal state. and the number of data samples in the fault state ,The data set is constructed by collecting statistics of the working status information of the real-time monitoring of the photovoltaic modules and the historical data set; S13. According to the status information of the information samples in the historical data set in S12, the balance between the faulty data and the normal data is analyzed to calculate the data set balance Spd. The data set balance Spd is obtained by the following formula: ； In the formula, Indicates the balance of the data set at the current time t, represents the number of data samples in the normal state at the current time t, Represents the number of data samples in the fault state at the current time t.

3. The photovoltaic module fault accurate modeling method based on fusion algorithm according to claim 2 is characterized by: The specific steps of S2 include: S21. Outputting a corresponding fault prediction value Gyz through various known basic models, wherein the various known basic models include but are not limited to gated recurrent unit, support vector machine, decision tree, random forest and convolutional neural network.

4. The photovoltaic module fault accurate modeling method based on fusion algorithm according to claim 3 is characterized by: The specific steps of S2 also include: S22. According to the data set balance Spd obtained in S13, the dependence of each basic model on the historical time at the current moment in the fusion algorithm is analyzed, and after dimensionless processing, a time dependence factor Syyz is calculated and obtained. The time dependence factor Syyz is obtained by the following formula: ； In the formula, represents the time-dependent factor at the current time t; Represents the prediction error at the current time t; represents the data change rate at the current time t, and Indicates the historical maximum prediction error and maximum data change rate; and is the weight value, where and The specific value is set by the user according to the situation.

5. The photovoltaic module fault accurate modeling method based on fusion algorithm according to claim 4 is characterized by: The specific steps of S2 also include: S23, by adding the time dependence factor at the current time t The time-dependent factor at the previous time t-1 Compare the size to preliminarily identify the change direction of the analysis cycle involved in each analysis of each basic model. If the time-dependent factor at the current time t is >Time-dependent factor at the previous time t-1 When the initial identification of the corresponding basic model is carried out, the analysis period involved in the next analysis will tend to be extended; If the time-dependent factor at the current time t is <Time-dependent factor at the previous time t-1 , the analysis period involved in the next analysis of the preliminary identification of the corresponding basic model will show a shortened trend; if the time-dependent factor at the current time t = time-dependent factor at the previous time t-1 , the analysis period involved in the preliminary identification of the corresponding basic model will not change in the next analysis.

6. The photovoltaic module fault accurate modeling method based on fusion algorithm according to claim 5 is characterized by: The specific steps of S3 include: S31. According to the prediction deviation of each basic model in the fusion algorithm at the current moment, combined with the time dependency factor Syyz, the degree to which the corresponding basic model learns the current trend through historical data is judged, and after dimensionless processing, the analysis cycle involved in each analysis of each basic model is updated, and marked as the memory window length update value Jcz. The memory window length update value Jcz is obtained by the following formula: ； In the formula, Represents the updated value of the memory window length at the current time t, Indicates the updated value of the memory window length at the next time t+1.

7. The photovoltaic module fault accurate modeling method based on fusion algorithm according to claim 6 is characterized by: The specific steps of S4 include: S41, according to the memory window length update value Jcz in S31, update the fault assessment of each basic model in real time to obtain the fault prediction value under the condition of the memory window length update value Jcz at the corresponding moment , and combined with the data set, dynamically update the weight w of the corresponding basic model, which can be obtained in the following way: ； In the formula, represents the weight of the i-th basic model at the next time t+1, and They represent the prediction error of the i-th basic model at the current time t and the prediction error of the j-th basic model at the current time t, respectively. and They represent the historical error averages of the i-th basic model and the j-th basic model, respectively. N represents the number of basic models, j=1, 2, ..., N. Represents an exponential function with base e.

8. The photovoltaic module fault accurate modeling method based on fusion algorithm according to claim 7 is characterized by: The specific steps of S4 also include: S42. Based on the deep learning technology and in combination with the dynamic update of the weight w of the corresponding basic model, the final fault prediction coefficient Gyxs is fitted and output from the output end of the fusion algorithm, which is specifically obtained by the following formula: ； In the formula, represents the weight of the i-th basic model at the next time t+1, Represents the fault prediction value of the i-th basic model at the next time t+1.

9. The photovoltaic module fault accurate modeling method based on fusion algorithm according to claim 1 is characterized by: The specific steps of S5 include: S51, by comparing and analyzing the final fault prediction coefficient Gyxs with the fault threshold K, to determine the fault degree of the current photovoltaic module, if the final fault prediction coefficient Gyxs exceeds the fault threshold K, the current photovoltaic module is in a fault state, and the fault prediction accuracy is evaluated in combination with the actual situation, if the current prediction error If the final fault prediction coefficient Gyxs does not exceed the fault threshold K, the current PV module is not in a fault state.

Images