Photovoltaic module fault diagnosis method based on micro inverter

Abstract

The invention discloses a photovoltaic module fault diagnosis method based on a micro inverter, and relates to the technical field of power system detection, and the method comprises the steps: collecting the real-time operation data of a photovoltaic module, obtaining the actual output power of the photovoltaic module, building a day-ahead power prediction model based on historical data and weather forecast, and carrying out the calculation of the day-ahead power prediction model, calculating the predicted power of the photovoltaic module; based on the actual output power and the predicted power of each photovoltaic module at different time sampling points, calculating the error and the standard deviation of each sampling point, analyzing the change trend of the standard deviation, and identifying and outputting the overall operation state of the photovoltaic module; and identifying an abnormal level based on a preset judgment rule, positioning a fault point and automatically triggering fault early warning. According to the invention, the real-time data of the single photovoltaic module is collected through the micro inverter, the deviation between the actual power and the predicted power can be detected in an extremely short time, and the high-precision real-time detection of a fault or an abnormal state is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system detection, and in particular to a photovoltaic module fault diagnosis method based on micro-inverters. BACKGROUND

[0002] With the wide application of solar photovoltaic power generation in power systems, the health state monitoring and fault diagnosis of individual photovoltaic modules become particularly important. Traditionally, the monitoring of the entire photovoltaic power station is often based on aggregate data, making it difficult to accurately determine the abnormalities or faults of individual modules.

[0003] In recent years, the development of micro-inverter technology enables independent monitoring of each photovoltaic module and sub-module, thereby obtaining more detailed output power data. At the same time, the day-ahead power prediction technology uses data-driven methods such as artificial neural networks (ANN) to predict the output power of the next day, providing a basis for module fault diagnosis.

[0004] Existing research has shown that by comparing the predicted power of the previous day with the actual output, as well as comparing the outputs of adjacent modules, abnormalities, shadows, module degradation or local faults that may exist in photovoltaic modules can be discovered in real time and fault levels can be classified.

[0005] Currently, the related technology mainly uses methods based on physical models, data-driven and hybrid models for power prediction and fault detection. However, in actual application, due to the randomness of the degradation process of the modules, inaccurate prediction models and measurement noise, there are still problems such as large errors and inaccurate alarms.

[0006] Therefore, there is an urgent need for a method that can use micro-inverter data and day-ahead power prediction results to perform real-time and offline fault diagnosis of photovoltaic modules. SUMMARY

[0007] Therefore, it is necessary to provide a photovoltaic module fault diagnosis method based on micro-inverters to solve the above technical problems.

[0008] The present application provides a photovoltaic module fault diagnosis method based on micro-inverters, comprising:

[0009] Collecting real-time operation data of the photovoltaic module, obtaining the actual output power of the photovoltaic module, and constructing a day-ahead power prediction model based on historical data and weather forecasts to calculate the predicted power of the photovoltaic module;

[0010] Based on the actual output power and the predicted power of each photovoltaic module, the error and the standard deviation of the corresponding time sampling points are calculated, and the trend of the standard deviation is analyzed to identify and output the overall operating state of the photovoltaic module.

[0011] When the overall operation state of the photovoltaic module is abnormal, the errors of the individual photovoltaic module at different time sampling points are compared in real time, the abnormal level is identified based on the preset judgment rule, the fault point is located, and the fault warning is automatically triggered;

[0012] Based on the spatial relationship and time evolution characteristics between the photovoltaic modules, a space-time anomaly detection model is constructed, the change trend of the comprehensive diagnostic index of the photovoltaic module is combined, and the fault point is analyzed and classified offline, and the safety response is triggered synchronously.

[0013] Further, a day-ahead power prediction model is constructed combined with historical data and weather forecast, and the predicted power of the photovoltaic module includes:

[0014] According to the preset collection frequency, the historical data of the actual output power in the monitoring process of the photovoltaic module and the weather forecast including the weather observation data, the numerical weather forecast and the component state data are obtained; multi-dimensional features are extracted respectively, and standardized model input data are formed after preprocessing;

[0015] The preprocessed model input data is used to build a training set and a test set, a day-ahead power prediction model is constructed based on the training set, the predicted power of the photovoltaic module is output, and the mapping function is solved through the training sample, and the day-ahead power prediction model is tested by using the test set;

[0016] The day-ahead power prediction model is corrected and smoothed by using weighted moving average and adaptive standard deviation, and the hyperparameters of the Bayesian adjustment neural network are adjusted to optimize the structure and performance of the day-ahead power prediction model.

[0017] Further, based on the actual output power and the predicted power of each photovoltaic module, the error and the standard deviation of the corresponding time sampling point are calculated, and the change trend of the standard deviation is analyzed to identify and output the overall operation state of the photovoltaic module, including:

[0018] The predicted power of the photovoltaic module at the same time sampling point is subtracted from the actual output power to obtain the error of the photovoltaic module at the sampling point, the errors of all photovoltaic modules at the sampling point are counted, and the standard deviation at the sampling point is calculated by using the standard deviation formula;

[0019] The standard deviations of the sampling points are integrated in time sequence to form time series data, if the standard deviation of the current sampling point is less than the preset standard deviation threshold, it is determined that the operation state of all photovoltaic modules is stable, if the standard deviation of the current sampling point is greater than or equal to the preset standard deviation threshold, it is determined that the overall operation state of all photovoltaic modules is abnormal.

[0020] Further, the errors of the single photovoltaic module at different time sampling points are compared in real time, the abnormal level is identified based on a preset judgment rule, the fault point is located, and a fault warning is automatically triggered, including:

[0021] The predicted power of the single photovoltaic module is subtracted from the standard deviation corresponding to the same time sampling point to obtain a first alarm threshold, and if the actual output power of the photovoltaic module at the sampling point is less than the first alarm threshold, it is determined that the photovoltaic module has an abnormality, and a first alarm is triggered.

[0022] The multiple monitored physical quantities of the photovoltaic module participating in abnormality identification are obtained, the diagnostic indicators of the photovoltaic module and adjacent photovoltaic modules are calculated, and the diagnostic indicators are compared with adjacent photovoltaic modules to exclude incidental errors.

[0023] When a single photovoltaic module triggers a first alarm, the error sum of the diagnostic indicators of the photovoltaic module from the start time to the current time is calculated, and if the error sum of the photovoltaic module is greater than a preset cumulative error threshold, a second warning is triggered.

[0024] Further, the multiple monitored physical quantities of the photovoltaic module participating in abnormality identification are obtained, the diagnostic indicators of the photovoltaic module and adjacent photovoltaic modules are calculated, and the diagnostic indicators are compared with adjacent photovoltaic modules to exclude incidental errors, including:

[0025] The multiple monitored physical quantities of the photovoltaic module participating in abnormality identification are obtained, including the actual value and the predicted value of the monitored physical quantity corresponding to different time sampling points.

[0026] The deviation of each monitored physical quantity is calculated by subtracting the predicted value from the actual value, and the absolute values of the deviations of all monitored physical quantities are added to obtain an error accumulation sum as the diagnostic indicator of the photovoltaic module.

[0027] The diagnostic indicators of the photovoltaic module with an abnormality are compared with the diagnostic indicators of adjacent photovoltaic modules at the same time, and based on a preset diagnostic condition of the adjacent photovoltaic modules, incidental errors are excluded.

[0028] Further, based on the spatial relationship and time evolution characteristics between the photovoltaic modules, a spatiotemporal anomaly detection model is constructed, the change trend of the comprehensive diagnostic indicators of the photovoltaic module is combined, and offline statistical analysis and fault level division of the fault point are performed, and a safety response is triggered simultaneously, including:

[0029] Based on the spatial relationship and time evolution characteristics of the photovoltaic module collected by the micro-inverter, a spatial weight matrix is formed, and a spatiotemporal anomaly detection model is constructed with the spatial weight matrix as the core input, so as to map the operating state of the photovoltaic module to a multi-dimensional spatiotemporal data space.

[0030] The diagnostic results of each photovoltaic module output by the spatio-temporal anomaly detection model are aggregated, a single-day comprehensive diagnostic index is calculated, a multi-grade fault is divided through a weighted voting mechanism, and a fault diagnosis result is output.

[0031] Based on the fault statistics of all photovoltaic modules as a whole, the adaptive weight coefficient in the multi-grade fault diagnosis process is dynamically adjusted, and when a severe fault is detected, a safety response is automatically triggered, including reducing the device load, switching to a backup system, or executing a safety shutdown program.

[0032] Further, the diagnostic results of each photovoltaic module output by the spatio-temporal anomaly detection model are aggregated, a single-day comprehensive diagnostic index is calculated, a multi-grade fault is divided through a weighted voting mechanism, and a fault diagnosis result is output, including:

[0033] The sum of the diagnostic indexes of each photovoltaic module at all sampling points within a single day is integrated and calculated to obtain a comprehensive diagnostic index of the photovoltaic module for a single day;

[0034] Based on the mean and standard deviation of the statistics in the current time window, a diagnostic interval for fault grade division is set for the comprehensive diagnostic index, including an upper diagnostic limit and a lower diagnostic limit, and three fault grades are divided according to the diagnostic interval, including a normal state, a mild abnormal state, and a severe abnormal state;

[0035] The comprehensive diagnostic index is compared with the diagnostic interval, and a fault grade diagnosis result corresponding to the comprehensive diagnostic index is output according to the position of the comprehensive diagnostic index.

[0036] Further, a photovoltaic module fault diagnosis method based on a micro-inverter further includes:

[0037] Based on the pre-deployed edge computing nodes, data stream online analysis and early warning tasks are performed, and interaction with cloud data is performed, and the day-ahead power prediction model and the spatio-temporal anomaly detection model are retrained and optimized through dynamic feature decoupling and incremental learning.

[0038] Further, based on the pre-deployed edge computing nodes, data stream online analysis and early warning tasks are performed, and interaction with cloud data is performed, and the day-ahead power prediction model and the spatio-temporal anomaly detection model are retrained and optimized through dynamic feature decoupling and incremental learning, including:

[0039] The edge node separates and analyzes the real-time data stream through a dynamic feature decoupling structure, completes the anomaly analysis and fault diagnosis of the photovoltaic module locally, and interacts with the cloud for the diagnostic results;

[0040] The edge node calculates the fault classification confidence through the multi-modal spatio-temporal attention network, uses a confidence threshold for classification determination, and performs sample processing operations based on the classification result;

[0041] The cloud pushes the trained model parameters to the edge node, replaces the old power prediction model and the spatio-temporal anomaly detection model in the edge node, and the edge node uses the new model to perform real-time diagnosis tasks and synchronously feeds back the early warning effect, and optimizes the weight distribution logic of the Bayesian network.

[0042] Further, the edge node calculates a fault classification confidence through a multi-modal spatio-temporal attention network, uses a confidence threshold for classification determination, and performs a sample processing operation based on a classification result, including:

[0043] A dynamic threshold is set, samples with a confidence greater than or equal to the dynamic threshold are marked as high-confidence samples, and samples with a confidence less than the dynamic threshold are marked as low-confidence samples, and the high-confidence samples are stored in the edge node, and the low-confidence samples are uploaded to the cloud to trigger global reconstruction training;

[0044] The cloud integrates the low-confidence samples uploaded by the edge node, historical fault data and environmental parameters, uses a convolutional neural network and a long short-term memory network for offline deep training, generates new model parameters after training, and optimizes network weights through gradient descent.

[0045] The beneficial effects of the present application are:

[0046] 1. Real-time data of a single photovoltaic module is collected by a micro inverter to realize high-precision real-time monitoring of key parameters such as component-level power, current, voltage and temperature, and combined with a day-ahead power prediction model and a prediction method based on PHANN, the deviation between actual power and predicted power can be detected in a very short time to realize high-precision real-time detection of faults or abnormal states.

[0047] 2. An adaptive fault diagnosis mechanism is provided to optimize operation and maintenance strategies and reduce misjudgments, and the diagnosis system can automatically distinguish fault levels, which helps operation and maintenance personnel to take different levels of maintenance measures.

[0048] 3. An intelligent fault early warning system is deployed on an edge computing node, so that data analysis and anomaly detection are performed on site equipment, thereby reducing data transmission delay, and a Bayesian network (Bayesian Network) is used for multi-feature fusion to optimize fault prediction.

[0049] 4. The LSTM (Long Short-Term Memory) neural network is used to learn historical operation data, and the gradient descent method is used to continuously optimize the prediction model.

[0050] 5. The present application is not limited to monitoring of a single component, but can also be connected with an energy management system (EMS) to realize intelligent photovoltaic power plant operation and maintenance. BRIEF DESCRIPTION OF DRAWINGS

[0051] The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and together with the description serve to explain the application. In the drawings:

[0052] Figure 1 is a flow chart of a micro-inverter-based photovoltaic module fault diagnosis method according to an embodiment of the application. DETAILED DESCRIPTION

[0053] In order to make the objects, technical solutions and advantages of the application clearer, the application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the application and do not limit the application.

[0054] Referring to Figure 1 , a micro-inverter-based photovoltaic module fault diagnosis method is provided, comprising:

[0055] S1, collecting real-time operation data of the photovoltaic module by using the micro-inverter, obtaining actual output power of the photovoltaic module, and constructing a day-ahead power prediction model based on historical data and weather forecasts to calculate the predicted power of the photovoltaic module.

[0056] In the description of the application, constructing a day-ahead power prediction model based on historical data and weather forecasts to calculate the predicted power of the photovoltaic module comprises:

[0057] S11, obtaining historical data of actual output power P actual during the monitoring process of the photovoltaic module according to a preset collection frequency, and weather forecasts, including weather observation data, numerical weather prediction and component state data; respectively extracting multi-dimensional features to form standardized model input data after preprocessing.

[0058] Specifically, the above construction process is an input data system and data types, including historical data (i.e. historical power data), weather observation data, numerical weather prediction, and component state data.

[0059] The corresponding feature dimensions are the actual output power of the component (kW), irradiance (W / m 2 ), ambient temperature (℃), GHI, DNI, cloud cover, wind speed, and temperature coefficient, attenuation rate, and inclination.

[0060] The collection frequency of each dimension feature is 15 minutes, 1 hour, 3 hours, and static collection. Then the input data system is improved by preprocessing methods such as outlier filtering, linear interpolation to 15-minute resolution, spatial and temporal downscaling processing, and normalization processing.

[0061] S12, a training set and a test set are established using the pretreated model input data, a day-ahead power prediction model is constructed based on the training set, and the predicted power P of the photovoltaic module is output pred , and the mapping function is solved by training samples, and the day-ahead power prediction model is tested by using the test set.

[0062] Specifically, according to the construction of the above day-ahead power prediction model, the calculation process of the output prediction result includes effective irradiance correction, and the calculation formula is:

[0063] Geff=Gpoa×cos(θ)×(1-α·ΔT);

[0064] Wherein, Geff is the slope irradiance (W / m 2 ); Gpoa is the solar incident angle; cos(θ) is the glass transmittance coefficient; (1-α·ΔT) is the temperature difference between the module and the environment.

[0065] In the process of establishing the day-ahead power prediction model, the mapping function is solved by training samples:

[0066] P pred =f(x:Θ);

[0067] In the formula, x represents the input vector (temperature, humidity, solar irradiance, clear sky model output); Θ represents the model parameters (weights and biases of neural network).

[0068] S13, the weighted moving average and the adaptive standard deviation are used to correct and smooth the day-ahead power prediction model, and the hyperparameters of the neural network are adjusted by using the Bayesian to optimize the structure and performance of the day-ahead power prediction model.

[0069] Specifically, on the basis of the above day-ahead power prediction model construction process, weighted moving average (WMA) and adaptive standard deviation (ASD) are introduced for error correction and smoothing processing to reduce the fluctuation of prediction error. At the same time, the hyperparameters of the neural network are automatically adjusted by using Bayesian optimization (BO) to optimize the structure and performance of the model, and the accuracy of photovoltaic module power prediction is improved.

[0070] Bayesian optimization uses Gaussian process (GP) and other probabilistic models to model the hyperparameter space, selects the hyperparameters most likely to improve the model performance based on the current evaluation results, and the implementation steps include:

[0071] 1. Define the hyperparameter space: determine the hyperparameters to be optimized, such as learning rate, number of neural network layers, number of neurons per layer, regularization parameter, etc.

[0072] 2. Gaussian Process Modeling: Use the training set and corresponding performance indicators (e.g. MAPE or RMSE) to establish a prediction model through Gaussian Process.

[0073] 3. Select optimization strategy: Use acquisition functions in Bayesian optimization (such as expected improvement function, probability improvement function, etc.) to select the next set of hyperparameters.

[0074] 4. Iterative optimization: Continuously optimize the model until the best performance of the hyperparameter combination is found.

[0075] S2, based on the actual output power and the predicted power of each photovoltaic module, calculate the error and standard deviation of the corresponding time sampling point, and analyze the trend of the standard deviation, identify and output the overall running state of the photovoltaic module.

[0076] In the description of the present application, based on the actual output power and the predicted power of each photovoltaic module, the error and standard deviation of the corresponding time sampling point are calculated, and the trend of the standard deviation is analyzed, the overall running state of the photovoltaic module is identified and output, including:

[0077] S21, subtract the actual output power from the predicted power of the photovoltaic module at the same time sampling point to obtain the error of the photovoltaic module at the sampling point, and calculate the standard deviation at the sampling point using the standard deviation formula.

[0078] Specifically, the error analysis of the predicted power and the actual output power of each time sampling point t and each photovoltaic module i is carried out, and the statistical quantity is calculated to provide a quantitative basis for subsequent diagnosis.

[0079] Wherein, the error calculation formula of the photovoltaic module at the sampling point is:

[0080] ΔP i,t =P pred i,t -P actual i,t ;

[0081] For all N components at time t, the sample standard deviation S t The calculation formula is:

[0082]

[0083] In the formula, ΔP i，t is the difference between the predicted power and the actual output power of photovoltaic module i at time t, i.e. error; S t is the sample standard deviation at time t, reflecting the dispersion degree of the output power of each component; N is the total number of photovoltaic modules sampled at time t; P pred i，t is the predicted power of photovoltaic module i at time t; P actual i，tactual output power of the photovoltaic module i at time t, represents the average error.

[0084] S22, the standard deviation of the sampling points is integrated in time sequence to form time series data, if the standard deviation of the current sampling point is less than the preset standard deviation threshold, it is determined that the running state of all photovoltaic modules is stable, if the standard deviation of the current sampling point is greater than or equal to the preset standard deviation threshold, it is determined that the overall running state of all photovoltaic modules is abnormal.

[0085] Specifically, according to the change trend of the standard deviation, the real-time data stream and the prediction model are combined to analyze the calculation sample, if S t is very small (the standard deviation is less than the preset standard deviation threshold), it indicates that the actual output power of most photovoltaic modules is close to the mean value, indicating that the system works relatively stably, and the error between the prediction and the actual is small. If S t is very large (the standard deviation is greater than or equal to the preset standard deviation threshold), it indicates that the actual output power of some photovoltaic modules is far from the mean value, indicating that there is a large difference in the output power of the system, and further investigation may be needed to determine whether there is a fault or abnormal condition.

[0086] Among them, the standard deviation feature is defined by the application, which is used to indicate that the output power of most photovoltaic modules is close to the mean value, indicating that the system runs stably. The error between the day-ahead power prediction model and the actual output power is small, indicating that the prediction accuracy of the model for the current state is high. The corresponding system environment is relatively stable, and the external conditions such as solar irradiance and temperature do not change dramatically. The system stability judgment is that the system works normally, the prediction error is small, and it belongs to the normal operating state. At this time, routine monitoring can be carried out, and excessive intervention is not necessary.

[0087] For real-time data stream analysis, the current standard deviation can be calculated according to the difference between the actual output power data stream and the predicted power. If the standard deviation is low and stable, it indicates that the accuracy of the model is high, and the system is in a healthy state. Monitoring strategy: continue to monitor to ensure that the system is within the normal working range and no additional troubleshooting is needed. The actual value can be combined to determine whether there is a fault or abnormal phenomenon. The moving window method is used: a sliding window is used to calculate the standard deviation in a certain period of time in the past. The threshold of the standard deviation can be dynamically adjusted according to the data of the past 24 hours, so that it can be adjusted according to the long-term performance and seasonal fluctuations of the system.

[0088] S3, when the overall running state of the photovoltaic module is abnormal, the error of a single photovoltaic module at different time sampling points is compared in real time, the abnormal level is identified based on the preset judgment rule, the fault point is located, and the fault warning is automatically triggered.

[0089] In the description of the present application, the errors of individual photovoltaic modules at different time sampling points are compared in real time, the abnormal level is identified based on preset judgment rules, the fault point is located and the fault warning is automatically triggered, which comprises:

[0090] S31, the predicted power of the individual photovoltaic module is subtracted from the standard deviation corresponding to the same time sampling point to obtain a first alarm threshold, if the actual output power of the photovoltaic module at the sampling point is less than the first alarm threshold, it is determined that the photovoltaic module exists abnormality, and a first alarm is triggered.

[0091] Specifically, the errors calculated at each sampling point for each photovoltaic module are compared in real time, and a fault alarm is automatically issued by using a preset judgment rule, which is mainly used for preliminary screening of the components that may exist abnormality. When a component i at a time t satisfies:

[0092] P actual i,t <P pred i,t -S t ;

[0093] The photovoltaic module is preliminarily considered to exist abnormality, and a first alarm is triggered.

[0094] There are many factors causing false alarms of the first alarm, in order to avoid data errors of a single component, the diagnostic index of the photovoltaic module and its adjacent photovoltaic module k is further compared.

[0095] S32, a plurality of monitored physical quantities of the photovoltaic module participating in abnormality identification are obtained, the diagnostic index of the photovoltaic module and the adjacent photovoltaic module is calculated, and the adjacent photovoltaic module is compared to exclude accidental errors.

[0096] In the description of the present application, a plurality of monitored physical quantities of the photovoltaic module participating in abnormality identification are obtained, the diagnostic index of the photovoltaic module and the adjacent photovoltaic module is calculated, and the adjacent photovoltaic module is compared to exclude accidental errors, which comprises:

[0097] S321, a plurality of monitored physical quantities of the photovoltaic module participating in abnormality identification are obtained, including the actual value and the predicted value of the monitored physical quantity corresponding to different time sampling points.

[0098] S322, the deviation of each monitored physical quantity is calculated by subtracting the predicted value from the actual value, and the absolute values of the deviations of all monitored physical quantities are added to obtain an error accumulation sum as the diagnostic index of the photovoltaic module.

[0099] S323, the diagnostic index of the photovoltaic module with abnormality and the diagnostic index of the adjacent photovoltaic module at the same time are compared, and accidental errors are excluded based on the preset diagnostic determination condition of the adjacent photovoltaic module.

[0100] Specifically, the diagnostic determination condition E is set as: it >Ekt ; wherein, E it is the diagnostic indicator of component i at time t, which is used to comprehensively reflect the cumulative or fluctuation of prediction error; E kt is the diagnostic indicator of component k at time t, which is usually used to compare with the indicator of component i at the same time to determine whether there is a relative anomaly, and to exclude occasional data errors of a single component.

[0101] Diagnostic decision condition operation rules:

[0102] (1) If E it > E kt (abnormal component indicator > adjacent component indicator):

[0103] It indicates that there is a persistent anomaly (non-occasional data error) in component i, which may be caused by device failure (such as hot spot, aging) or local environmental mutation (such as shadow blocking).

[0104] Mark component i as "suspected failure" and start continuous monitoring mechanism; if M consecutive sampling points satisfy E it > E kt (M≥3), trigger failure alarm; generate repair work order and locate the cause of the anomaly (refer to historical decay rate and real-time meteorological data).

[0105] (2) If E it ≤ E kt (abnormal component indicator ≤ adjacent component indicator):

[0106] It indicates that the anomaly of component i is an occasional data error (such as transient cloud blocking, sensor noise).

[0107] The photovoltaic module is classified into "observation queue", and single record does not trigger alarm; clear the historical error count of component i; automatically reset the diagnostic state.

[0108] Exclusion mechanism of occasional error:

[0109] (1) Spatial verification: it needs to be compared with K adjacent components at the same time (K≥3), covering the physical orientation (up / down / left / right) of component i;

[0110] (2) Time verification: only when M consecutive sampling points satisfy the decision condition (M≥3), it is confirmed as a real failure (to avoid temporary interference misjudgment).

[0111] Wherein, the error accumulation can be calculated by the difference between the actual output of the component and the output of the prediction model, that is, the diagnostic indicator of the photovoltaic module, and the calculation formula is:

[0112]

[0113] wherein, represents the actual value of the jth monitored physical quantity of the photovoltaic module i at time t; represents the predicted value or expected value of the jth monitored physical quantity of the photovoltaic module i at time t; n represents the number of monitored physical quantities.

[0114] S33, when a single photovoltaic module triggers a first-level alarm, the error sum of the diagnostic index of the photovoltaic module from the start time to the current time is calculated, and if the error sum of the photovoltaic module is greater than a preset cumulative error threshold, a second-level warning is triggered.

[0115] Specifically, the second-level alarm triggering condition is set by setting a judgment condition, and through the logical chain of preliminary screening, adjacent component comparison and continuous abnormality detection, the false alarm probability is reduced.

[0116] The relevant data triggering the alarm, including time, component number, measured value, predicted value, error, adjacent component state, are stored in the database for subsequent analysis, combined with the error trend and the cumulative error index Ei, to judge the failure probability. If the abnormality may be caused by external factors (temperature, humidity, voltage fluctuation, etc.), check the external environment data and summarize the failure cause.

[0117] The triggering condition of the second-level alarm is that when a certain photovoltaic component C i When the first-level alarm triggering condition is met at a certain time t, the second-level triggering condition will further compare the adjacent components of the component, and the comparison formula is:

[0118]

[0119] Among them, is the cumulative error sum from the start time to the current time t; T cumulative is a preset cumulative error threshold.

[0120] S4, based on the spatial relationship and time evolution characteristics between photovoltaic components, a spatio-temporal anomaly detection model is constructed, combined with the change trend of the comprehensive diagnostic index of the photovoltaic component, the fault point is analyzed and classified, and the safety response is triggered synchronously.

[0121] In the description of the present application, based on the spatial relationship and time evolution characteristics between photovoltaic components, a spatio-temporal anomaly detection model is constructed, combined with the change trend of the comprehensive diagnostic index of the photovoltaic component, the fault point is analyzed and classified, and the safety response is triggered synchronously, including:

[0122] S41, based on the spatial relationship and time evolution characteristics of the photovoltaic component collected by the micro-inverter, a spatial weight matrix is constructed, and a spatio-temporal anomaly detection model is constructed with the spatial weight matrix as the core input, so as to map the running state of the photovoltaic component to a multi-dimensional spatio-temporal data space.

[0123] Specifically, the spatio-temporal anomaly detection model dynamically adjusts the diagnosis strategy through a spatial weight matrix and an intelligent adaptive algorithm, reduces false positives, and achieves accurate classification of fault levels. Its connection with the overall target diagnosis is reflected in:

[0124] 1. Data fusion and intelligent diagnosis: The spatio-temporal anomaly detection model achieves intelligent diagnosis based on multi-dimensional data by combining the spatial and temporal data of components. This process not only diagnoses each component through real-time data, but also optimizes fault diagnosis rules through historical data learning, achieving collaborative analysis of multi-dimensional data.

[0125] 2. Interaction with the overall diagnosis system: The diagnosis results generated by the spatio-temporal model will be aggregated into the overall target diagnosis framework, serving as key diagnostic evidence. The overall target diagnosis system can achieve more accurate fault warning and level classification by aggregating, classifying, and analyzing the diagnosis indicators of each component, thereby providing more comprehensive data support for the maintenance and management of photovoltaic systems.

[0126] This method uses a spatio-temporal anomaly detection model, whose core objective is to identify abnormal behavior of photovoltaic components in a dynamic spatio-temporal environment. This model utilizes the spatial relationships and temporal evolution characteristics between photovoltaic components to map the operation of the photovoltaic system to a multi-dimensional spatio-temporal data space, identifying abnormal phenomena that deviate from normal behavior patterns. Based on the multi-dimensional data of space and time between photovoltaic components, intelligent fault diagnosis is performed.

[0127] In dynamic spatial relationship modeling, a spatial weight matrix between photovoltaic components is defined, converting physical and electrical characteristics such as the geographical position, power output, distance, orientation, etc. of each photovoltaic component into a numerical value for subsequent spatio-temporal analysis. The elements of the spatial weight matrix between photovoltaic components include:

[0128] 1. Spatial relationship: including the physical position between photovoltaic components, the planar coordinates of the components (unit: meters), the Euclidean distance, the distance attenuation coefficient (typical value 50-100 meters), the azimuth angle of the components (relative to the north direction), the azimuth correction coefficient (0.1-0.3), etc. Generally, the power output of adjacent components has strong correlation, so if a component fails, its fault behavior may affect adjacent components.

[0129] 2. Temporal relationship: the power output of photovoltaic components changes over time and is influenced by factors such as weather, environment, and load.

[0130] 3. Electrical characteristics: including the output power, electrical impedance, reference impedance, etc. of the photovoltaic module. These physical characteristics can help identify fault patterns in the comparison between different modules. For example, the relationship between the electrical impedance and the output power of the module can help determine whether the module has a short circuit, open circuit, etc.

[0131] The above elements are combined with the false positive rate in a complex environment, and the cooperative diagnosis of multiple data is linked. The false positive rate (False Positive Rate, FPR) refers to the ratio of the system incorrectly determining normal components as faulty during model diagnosis. According to the normal working fluctuation range of the component, the alarm threshold is dynamically set to avoid false alarms due to data fluctuations. The intelligent adaptive algorithm is not only used to dynamically adjust the alarm threshold in the entire spatio-temporal anomaly detection model, but also models the historical operation data of the device, so that the alarm rule can adapt to different operating loads, environmental conditions and device states. The intelligent adaptive algorithm can optimize the diagnosis strategy according to the running state of the device and the change of the environment through training data.

[0132] S42, aggregate the diagnosis results of each photovoltaic module output by the spatio-temporal anomaly detection model, calculate the single-day comprehensive diagnosis index, divide the multi-grade fault through the weighted voting mechanism, and output the fault diagnosis result.

[0133] In the description of the present application, the diagnosis results of each photovoltaic module output by the spatio-temporal anomaly detection model are aggregated, the single-day comprehensive diagnosis index is calculated, the multi-grade fault is divided through the weighted voting mechanism, and the fault diagnosis result is output, which includes:

[0134] S421, integrate and calculate the sum of the diagnosis indexes of each photovoltaic module at all sampling points within a single day, to obtain the comprehensive diagnosis index of the photovoltaic module in a single day; and the daily diagnosis result, alarm record and related statistical data are encrypted and recorded through a lightweight blockchain node to ensure that the data cannot be tampered with.

[0135] Among them, in this framework, the daily comprehensive application will aggregate the diagnosis data of each photovoltaic module and combine it with the overall index, where the overall index and the daily comprehensive application are based on the diagnosis index, and the comprehensive application further expands this information, providing detailed alarm records, component-level analysis and system health reports.

[0136] S422, based on the mean and standard deviation of the statistics in the current time window, set the diagnosis interval of the comprehensive diagnosis index participating in the fault grade division, including the diagnosis upper limit and the diagnosis lower limit, and divide three fault grades according to the diagnosis interval, including the normal state, the mild abnormal state and the severe abnormal state.

[0137] Specifically, the weighted moving average calculation method and the adaptive standard deviation calculation method are used to calculate the daily comprehensive diagnosis index E of each component ii The intelligent adaptive algorithm models the historical operation data of the device through a machine learning algorithm, dynamically adjusts the alarm threshold, and enables the alarm rule to adapt to different loads, environments and operating states.

[0138] The statistical quantity updated using the intelligent adaptive algorithm defines the fault level, i.e., sets the diagnostic interval participating in the fault level division, including the diagnostic upper limit and the diagnostic lower limit, to be [μ t +σ t , μ t +3σ t ].

[0139] The three types of fault levels are: when in normal state, E i <μ t +σt; when in soft fault state (mild abnormal state), μ t +σ t ≤E i <μ t +3σ t ; and when in hard fault state (severe abnormal state), E i ≥μ t +3σ t .

[0140] In the formula, μ t is the mean value in the statistical quantity updated by the intelligent adaptive algorithm at time t or in the time window t; and σ t represents the standard deviation calculated at time t or in the time window t.

[0141] The statistical quantity is the mean value and the standard deviation calculated by the adaptive algorithm, and the mean value and the standard deviation are updated based on the mean value and the standard deviation in step S2 through the intelligent adaptive algorithm, and are dynamically adjusted based on the historical performance and the current operating state of the entire system, so that the system can more accurately assess the health state of each photovoltaic component.

[0142] S423, compare the comprehensive diagnostic index with the diagnostic interval, and output the fault level diagnostic result corresponding to the comprehensive diagnostic index according to the position of the comprehensive diagnostic index.

[0143] S43, based on the fault statistics result of all photovoltaic components as a whole, dynamically adjust the adaptive weight coefficient in the multi-level fault diagnosis process, i.e., adjust the sensitivity and range of the diagnostic threshold ([μ t +σ t , μ t +3σ t ], for example, if a component is running in a specific environment or load for a long time, the system will adjust μ tSo that it can better reflect the actual working state of the component; if the error fluctuation of the component increases under certain environmental conditions (such as extreme weather), the system will learn the error fluctuation range of the component according to historical data, and dynamically adjust the standard deviation (σ t ), relax the tolerance range of the error. And when detecting a severe fault, automatically trigger a safety response, including reducing device load, switching backup system or executing a safety shutdown procedure.

[0144] Specifically, the spatio-temporal anomaly detection model realizes intelligent diagnosis based on multi-dimensional data by combining the spatial and temporal data of the components. This process not only diagnoses each component through real-time data, but also optimizes the fault diagnosis rules through historical data learning, realizing the collaborative analysis of multi-dimensional data; the diagnosis results generated by the spatio-temporal model will be summarized into the overall target diagnosis framework, becoming the key diagnosis basis.

[0145] The overall target diagnosis system can realize more accurate fault warning and grade division by summarizing, classifying and analyzing the diagnosis indicators of each component, thereby providing more comprehensive data support for the maintenance and management of the photovoltaic system.

[0146] If long-term monitoring finds that the overall abnormal frequency of the system increases, the adaptive adjustment σ t weight is increased to improve the fault detection sensitivity, and the historical fault mode is analyzed in combination with the machine learning model. When a hard fault is detected, an automatic response mechanism can be triggered to reduce device load, switch backup system or execute a safety shutdown procedure.

[0147] S5, based on the pre-deployed edge computing node, perform data stream online analysis and early warning tasks, and interact with cloud data, retrain and optimize the day-ahead power prediction model and spatio-temporal anomaly detection model through dynamic feature decoupling and incremental learning.

[0148] In the description of the present application, based on the pre-deployed edge computing node, perform data stream online analysis and early warning tasks, and interact with cloud data, retrain and optimize the day-ahead power prediction model and spatio-temporal anomaly detection model through dynamic feature decoupling and incremental learning, including:

[0149] S51, the edge node separates and analyzes the real-time data stream through the dynamic feature decoupling structure, and completes the abnormal analysis and fault diagnosis of the photovoltaic component locally, while interacting with the cloud for diagnosis results.

[0150] Specifically, part of the data analysis and early warning decision is deployed on the edge node to realize a low-delay and fast-response early warning mechanism. The edge node uses real-time data streams for online analysis, and the fusion algorithm is configured in the cloud. At the same time, it interacts with the cloud data, adopts the fusion algorithm and integrated learning, Bayesian network, and weighted correction on the early warning result to realize more accurate early warning judgment. The historical faults include wear and tear, cracks and fractures, corrosion and rust, fouling and shielding, and the detected abnormal and fault data are retrained offline by using deep learning and convolutional neural network and long short-term memory network, and the updated model parameters are fed back to the edge device to realize self-optimization of the prediction model.

[0151] Through the dynamic feature decoupling structure, the edge feature decoupling module containing the environment sensitive branch and the device body branch. In modern photovoltaic fault diagnosis and prediction, the dynamic feature decoupling structure is a structure for decomposing and optimizing complex input features to improve the prediction ability and accuracy of the model. It mainly acts on the edge node and separates and analyzes data in real time through the dynamic feature decoupling structure to make preliminary fault diagnosis locally. This is a local information-based instant decision-making process that ensures low latency and fast response. The cloud is responsible for optimizing global data and feeding back updated model parameters to the edge node, which makes more accurate diagnosis according to the latest model. The cloud feeds back the updated parameters to the edge node to ensure the accuracy of local fault diagnosis. The core is to decouple different types of input data (such as environmental data and device data) so that the system can process these data specifically and improve the sensitivity of the model to different types of features.

[0152] In the present application, the dynamic feature decoupling structure is divided into an environment sensitive branch and a device body branch. These two branches process environmental factors (such as temperature, humidity, irradiance) and electrical states (such as current, voltage) of the device itself in the photovoltaic system. These two branches are calculated separately through feature decoupling technology to avoid interference from different data types, so that each branch can accurately capture and process features in its specific field.

[0153] In this way, the edge node can quickly and accurately process data streams from various dimensions while reducing computational burden and improving processing efficiency. The edge node separates and analyzes data in real time through the dynamic feature decoupling structure to make preliminary fault diagnosis locally. This is a local information-based instant decision-making process that ensures low latency and fast response.

[0154] The cloud is responsible for the optimization of global data, and the updated model parameters are fed back to the edge node, and the edge node makes more accurate diagnosis according to the latest model. This is the cooperation mechanism between the edge node and the cloud, which ensures that the local diagnosis model of the edge node is always in the latest and optimized state.

[0155] The dynamic feature decoupling structure enables the edge node to make a quick response based on real-time collected data, limited to local data processing and preliminary diagnosis, mainly to improve the local response speed and accuracy of the edge node to faults. The cloud is responsible for the optimization of global data, and the updated model parameters are fed back to the edge node, and the updated model parameters are the optimal parameters generated by the cloud based on the deep learning algorithm.

[0156] Among them, the input dimension of the environment branch is temperature, irradiance, humidity, and the input dimension of the device branch is current, voltage.

[0157] S52, the edge node calculates the fault classification confidence through the multi-modal spatio-temporal attention network, uses the confidence threshold for classification judgment, and executes a sample processing operation based on the classification result.

[0158] Among them, the multi-modal spatio-temporal attention network based on confidence criterion level discrimination is a deep learning architecture, which aims to classify the input samples by combining the spatio-temporal data features of different modalities, and judge the reliability of each sample through the confidence criterion, further improve the diagnostic accuracy and early warning ability of the model.

[0159] Confidence is a measure used in machine learning and deep learning to quantify the "confidence" or "reliability" of a model's prediction of a certain result. In the context of fault diagnosis, confidence is used to represent the network's self-confidence in classifying or predicting the current sample. The method of calculating confidence in this patent is to use the softmax function, whose output represents the probability distribution of the prediction belonging to each class.

[0160] Among them, the confidence calculation formula is as follows:

[0161]

[0162] Among them, is the confidence of the sample belonging to the ith class; z i is the original score output by the neural network, the numerator is the exponential score of the class, and the denominator is the sum of the scores of all classes, z j is the original score of the jth class output by the neural network (j traverses all classes); e is the natural constant (≈2.71828), and ∑j represents the sum of j from 1 to K (K=total number of classes); this formula uses exponentialization to convert the negative original score of the model output into a positive number, amplifying the difference between high scores.

[0163] In the multi-modal spatio-temporal attention network, the confidence is represented by calculating the probability of each predicted category, and the reliability of the sample can be judged according to the confidence. The method adopts a soft threshold method to determine the threshold for dividing high confidence and low confidence, that is, the distribution of the model output probability is observed to select the threshold. If the model output prediction probability is obviously biased towards a certain category in most cases, a higher threshold is selected. For example, if the model often gives a probability value of 0.8 or more, 0.75 is selected as the threshold at this time.

[0164] In the description of the application, the edge node calculates the fault classification confidence through the multi-modal spatio-temporal attention network, classifies and judges by using the confidence threshold, and performs a sample processing operation based on the classification result, which includes:

[0165] S521, set a dynamic threshold, mark the sample with a confidence greater than or equal to the dynamic threshold as a high confidence sample, and mark the sample with a confidence less than the dynamic threshold as a low confidence sample. And store the high confidence sample to the edge node, upload the low confidence sample to the cloud, and trigger the global reconstruction training.

[0166] Specifically, for high confidence samples, the classification result of the model already has high reliability, so it will not be modified too much. But it can further optimize the accuracy of its classification. For these samples, the method further trains the model based on the previously trained parameters:

[0167]

[0168] Where, θ is the model parameter, θ new and θ old and are the new model parameters and the old model parameters respectively, η is the learning rate, L is the loss function, X and Y are the input data and the label respectively, and the goal of fine-tuning is to minimize the loss function, thereby improving the prediction accuracy of high confidence samples.

[0169] For low confidence samples, the model is not confident in its prediction, so reconstruction training is needed. These samples will be uploaded to the cloud for more comprehensive reconstruction training, and the weights of the entire network will be updated so that the model can better cope with the complexity of low confidence samples.

[0170] The reconstruction training method adopted by the application is a gradient descent-based optimization algorithm, and the update rule is:

[0171]

[0172] Where, θ t and θ t+1and are the current and updated model parameters, X' and Y' are the low-confidence sample data returned by the cloud, and the goal is to optimize the overall performance of the model.

[0173] S522, the cloud integrates the low-confidence samples uploaded by the edge nodes, historical fault data and environmental parameters, performs offline deep training using a convolutional neural network and a long short-term memory network, generates new model parameters after training, and optimizes the network weights through gradient descent.

[0174] Specifically, by combining the confidence criterion, the model can distinguish between high-confidence samples and low-confidence samples locally, and decide whether to fine-tune or upload the cloud for reconstruction training according to their reliability. This mechanism can realize self-optimization of the model, so that the system can adapt to changing environments and different types of faults.

[0175] Based on the confidence criterion, the level discrimination multi-modal spatio-temporal attention network is used to fine-tune high-confidence samples and upload feature vectors to trigger cloud reconstruction training. A hierarchical response mechanism, learning response optimization, and model update using gradient descent method are required, and the parameter update formula is:

[0176]

[0177] In the formula, θ represents the model parameters; η is the learning rate; and L is the loss function (such as mean square error MSE).

[0178] S53, the cloud pushes the trained model parameters to the edge nodes to replace the old power prediction model and spatio-temporal anomaly detection model in the edge nodes, and the edge nodes use the new model to perform real-time diagnosis tasks and synchronize the warning effect, and optimize the weight distribution logic of the Bayesian network.

[0179] Specifically, after training is completed, the cloud transmits the new model parameters to the edge device to update the edge model, and then uses a convolutional neural network to recognize abnormal patterns and improve the edge node's ability to perceive complex features. The detected abnormal data and fault data are stored in the cloud database to form a training data set. The model performance is evaluated regularly, and if the new model is better than the old model, it is pushed to the edge node for updating, otherwise the original model is maintained. Using this method, the system can continuously iterate to improve the accuracy and intelligence level of the warning, and realize a self-optimizing closed loop.

[0180] A photovoltaic module fault diagnosis method based on a micro inverter is described in detail below in conjunction with specific embodiments.

[0181] The application utilizes micro-inverters to realize fine monitoring of the output power of each photovoltaic component, obtains the predicted power of each component by combining a database based on day-ahead power prediction and a method combining ANN and a clear-sky solar radiation model (CSRM), compares the actual power with the predicted power in real time, and accurately identifies failure or abnormal trends by combining the output data of adjacent components, thereby improving the accuracy and immediacy of failure diagnosis, and providing a basis for system decision-making by simultaneously performing real-time online diagnosis and offline data statistical analysis.

[0182] The real-time online diagnosis method calculates the predicted power and the actual power of each photovoltaic component and micro-inverter at each time sampling point, and compares them in real time to determine whether to generate a first-level or second-level alarm. Offline data analysis is performed, i.e., after a day ends, the daily diagnosis index of each component is calculated, and statistical methods (mean and standard deviation) are used to classify the state of each module, thereby determining the state of "normal", "soft failure" or "hard failure". The day-ahead power prediction uses a hybrid method (PHANN, Physics-Hybrid Artificial Neural Network) based on ANN and data-driven physical model, the input of which includes historical weather parameters and CSRM data, and the output is the predicted direct current output power. The error analysis in the prediction process uses normalized root mean square error to define and quantify common errors.

[0183] Taking a simplified photovoltaic power station system as an example, 10 photovoltaic components are monitored and diagnosed in real time in the laboratory, one of which is simulated to be shaded. The actual monitoring data shows that the actual output power of the component numbered Pv-001 deviates significantly from the predicted value, the calculated diagnosis threshold index exceeds the set value, the system displays the corresponding diagnosis alarm in the real-time monitoring instrument, and determines the cause of the accident in offline statistics. The data of the photovoltaic component during abnormal diagnosis is shown in Table 1.

[0184] Table 1: Data of photovoltaic component during abnormal diagnosis

[0185]

[0186] The calculation formula of the power drop rate is as follows:

[0187]

[0188] Through field experiments, the module causes a power drop of about 20% under partial shading conditions and a power drop of 80% under complete shading conditions. In high-temperature weather, the power drops by 12%, and in weather mutations, the power drops by 44%. Data analysis shows that sensor and acquisition errors are relatively stable, while environmental influences and modeling errors fluctuate greatly in certain time periods, greatly affecting the overall accuracy of the system.

[0189] In the intelligent photovoltaic fault diagnosis system, errors mainly come from sensor measurement, data transmission, environmental influence, modeling deviation, and insufficient data acquisition accuracy. Data analysis shows that sensor and acquisition errors are relatively stable, while environmental influences and modeling errors fluctuate greatly in certain time periods, affecting the overall accuracy of the system. For different error types, optimization measures include using high-precision sensors, improving communication protocols, applying deep learning to optimize prediction models, and improving ADC sampling accuracy and filtering algorithms. Through these optimization strategies, errors can be effectively reduced, and the monitoring accuracy and operating efficiency of the photovoltaic system can be improved, thereby optimizing power plant operation and maintenance and increasing power generation revenue.

[0190] In summary, with the above technical solutions of the present application, real-time data of individual photovoltaic modules are collected by micro-inverters to achieve high-precision real-time monitoring of key parameters such as module-level power, current, voltage, and temperature. Combined with day-ahead power prediction models and PHANN-based prediction methods, the deviation between actual power and predicted power can be detected in a very short time, enabling high-precision real-time detection of faults or abnormal states. An adaptive fault diagnosis mechanism is provided to optimize maintenance strategies and reduce misjudgments. The diagnosis system can automatically distinguish fault levels, which helps maintenance personnel take different levels of maintenance measures. An intelligent fault warning system is deployed on edge computing nodes, enabling data analysis and anomaly detection on-site equipment, thereby reducing data transmission delays. Bayesian Network is used for multi-feature fusion to optimize fault prediction. LSTM neural network is used to learn historical operating data, and the prediction model is continuously optimized through gradient descent method. The present application is not limited to single-module monitoring, but can also be connected with an energy management system (EMS) to realize intelligent photovoltaic power plant operation and maintenance.

[0191] It should be understood that although the steps in the flowcharts of the drawings are shown in sequential order, such that each step depends on completion of the previous step before execution of the next step, the steps are not necessarily performed in the order indicated by the arrows. Unless specifically stated otherwise, the steps can be performed in other orders. Moreover, at least some of the steps in the flowcharts of the drawings can include multiple sub-steps or stages, which are not necessarily performed at the same time, but can be performed at different times, and which are not necessarily performed sequentially, but can be performed in rotation or alternation with other steps or sub-steps or stages of other steps.

Claims

1. A method for fault diagnosis of photovoltaic modules based on microinverters, characterized in that, include: Collect real-time operating data of photovoltaic modules to obtain the actual output power of photovoltaic modules, and build a day-ahead power prediction model based on historical data and weather forecasts to calculate the predicted power of photovoltaic modules; Based on the actual output power and predicted power of each photovoltaic module, the error and standard deviation of the corresponding time sampling points are calculated, the trend of standard deviation is analyzed, and the overall operating status of the photovoltaic module is identified and output. When the overall operating status of photovoltaic modules is abnormal, the error of a single photovoltaic module at different sampling points at different times is compared in real time, the abnormality level is identified based on preset judgment rules, the fault point is located and the fault warning is automatically triggered. Based on the spatial relationships and temporal evolution characteristics among photovoltaic modules, a spatiotemporal anomaly detection model is constructed. Combining the changing trends of the comprehensive diagnostic indicators of photovoltaic modules, offline statistical analysis and fault level classification are performed on the fault points, and a safety response is triggered simultaneously.

2. The photovoltaic module fault diagnosis method based on a micro-inverter according to claim 1, characterized in that, The method of constructing a day-ahead power prediction model by combining historical data and weather forecasts to calculate the predicted power of photovoltaic modules includes: According to the preset acquisition frequency, historical data of actual output power during photovoltaic module monitoring, as well as weather forecasts, including meteorological observation data, numerical weather forecasts, and module status data, are acquired; multi-dimensional features are extracted and preprocessed to form standardized model input data; A training set and a test set are constructed using the preprocessed model input data. A day-ahead power prediction model is built based on the training set to output the predicted power of photovoltaic modules. The mapping function is solved using the training samples, and the day-ahead power prediction model is tested using the test set. We used weighted moving average and adaptive standard deviation to correct and smooth the day-ahead power prediction model, and used Bayesian adjustment of the hyperparameters of the neural network to optimize the structure and performance of the day-ahead power prediction model.

3. The photovoltaic module fault diagnosis method based on a micro-inverter according to claim 1, characterized in that, The process of calculating the error and standard deviation of corresponding time sampling points based on the actual output power and predicted power of each photovoltaic module, analyzing the trend of standard deviation changes, and identifying and outputting the overall operating status of the photovoltaic modules includes: The error of the photovoltaic module at that sampling point is obtained by subtracting the actual output power from the predicted power of the photovoltaic module at the same sampling point. The errors of all photovoltaic modules within that sampling point are statistically analyzed, and the standard deviation at that sampling point is calculated using the standard deviation formula. The standard deviations of the sampling points are integrated in chronological order to form time series data. If the standard deviation of the current sampling point is less than the preset standard deviation threshold, the operating status of all photovoltaic modules is determined to be stable. If the standard deviation of the current sampling point is greater than or equal to the preset standard deviation threshold, the overall operating status of all photovoltaic modules is determined to be abnormal.

4. The photovoltaic module fault diagnosis method based on a micro-inverter according to claim 1, characterized in that, The step of comparing the errors of a single photovoltaic module at different sampling points in real time, identifying the anomaly level based on preset judgment rules, locating the fault point, and automatically triggering a fault warning includes: The predicted power of a single photovoltaic module is subtracted from the standard deviation corresponding to the sampling point at the same time to obtain the first-level alarm threshold. If the actual output power of the photovoltaic module at the sampling point is less than the first-level alarm threshold, the photovoltaic module is determined to be abnormal and a first-level alarm is triggered. Multiple monitoring physical quantities of photovoltaic modules involved in anomaly identification are obtained, diagnostic indicators of photovoltaic modules and adjacent photovoltaic modules are calculated, and compared with adjacent photovoltaic modules to eliminate occasional errors. When a single photovoltaic module triggers a Level 1 alarm, the total error of the diagnostic indicators of that photovoltaic module from the start time to the current time is calculated. If the total error of the photovoltaic module is greater than the preset cumulative error threshold, a Level 2 warning is triggered.

5. A photovoltaic module fault diagnosis method based on a micro-inverter according to claim 4, characterized in that, The process of acquiring multiple monitoring physical quantities of photovoltaic modules for anomaly identification, calculating diagnostic indicators for the photovoltaic module and adjacent photovoltaic modules, and comparing them with adjacent photovoltaic modules to exclude occasional errors includes: Acquire multiple monitoring physical quantities of photovoltaic modules involved in anomaly identification, including the actual and predicted values ​​of the monitoring physical quantities at different time sampling points; The deviation between the actual value and the predicted value of each monitored physical quantity is calculated, and the absolute values ​​of the deviations of all monitored physical quantities are added together to obtain the cumulative error, which is used as a diagnostic indicator for photovoltaic modules. The diagnostic indicators of abnormal photovoltaic modules are compared with the diagnostic indicators of adjacent photovoltaic modules at the same time. Based on the preset diagnostic judgment conditions of adjacent photovoltaic modules, occasional errors are eliminated.

6. The photovoltaic module fault diagnosis method based on a micro-inverter according to claim 1, characterized in that, Based on the spatial relationships and temporal evolution characteristics between photovoltaic modules, a spatiotemporal anomaly detection model is constructed. Combined with the changing trends of comprehensive diagnostic indicators of photovoltaic modules, offline statistical analysis and fault level classification are performed on fault points, and a safety response is triggered simultaneously, including: Based on the spatial relationship and temporal evolution characteristics of photovoltaic modules collected by micro-inverters, a spatial weight matrix is ​​constructed, and a spatiotemporal anomaly detection model is built with the spatial weight matrix as the core input to map the operating status of photovoltaic modules to a multidimensional spatiotemporal data space. The diagnostic results of each photovoltaic module output by the spatiotemporal anomaly detection model are summarized, the daily comprehensive diagnostic index is calculated, and the fault diagnosis results are output by classifying multiple levels of faults through a weighted voting mechanism. Based on the overall fault statistics of all photovoltaic modules, the adaptive weighting coefficients in the multi-level fault diagnosis process are dynamically adjusted, and when a severe fault is detected, a safety response is automatically triggered, including reducing equipment load, switching to backup systems, or executing a safety shutdown procedure.

7. A photovoltaic module fault diagnosis method based on a micro-inverter according to claim 6, characterized in that, The diagnostic results of each photovoltaic module output by the spatiotemporal anomaly detection model are aggregated, and a daily comprehensive diagnostic index is calculated. A weighted voting mechanism is used to classify faults into multiple levels, and the output fault diagnosis results include: The sum of diagnostic indicators for each photovoltaic module at all sampling points within a single day is integrated and calculated to obtain the comprehensive diagnostic indicator for the photovoltaic module on a single day. Based on the mean and standard deviation of the statistics within the current time window, a diagnostic interval for the comprehensive diagnostic indicators to participate in the fault level classification is set, including the upper and lower diagnostic limits. Based on the diagnostic interval, three fault levels are divided, including normal state, mild abnormal state and severe abnormal state. The comprehensive diagnostic index is compared with the diagnostic interval, and the fault level diagnosis result corresponding to the comprehensive diagnostic index is output according to the position of the comprehensive diagnostic index.

8. A method for fault diagnosis of photovoltaic modules based on microinverters according to claim 1, characterized in that, The method also includes: Based on pre-deployed edge computing nodes, online data flow analysis and early warning tasks are performed, and data is interacted with in the cloud. The daytime power prediction model and the spatiotemporal anomaly detection model are retrained and optimized through dynamic feature decoupling and incremental learning.

9. A photovoltaic module fault diagnosis method based on a micro-inverter according to claim 8, characterized in that, The aforementioned task, based on pre-deployed edge computing nodes, involves performing online data flow analysis and early warning, interacting with cloud data, and retraining and optimizing the day-ahead power prediction model and spatiotemporal anomaly detection model through dynamic feature decoupling and incremental learning, including: Edge nodes use a dynamic feature decoupling structure to separate and analyze real-time data streams, perform anomaly analysis and fault diagnosis of photovoltaic modules locally, and exchange the diagnosis results with the cloud. Edge nodes calculate fault classification confidence through a multimodal spatiotemporal attention network, use confidence thresholds to make classification decisions, and perform sample processing operations based on the classification results; The cloud pushes the trained model parameters to the edge nodes, replacing the old power prediction model and spatiotemporal anomaly detection model in the edge nodes. The edge nodes use the new model to perform real-time diagnostic tasks and synchronously provide feedback on the early warning effect, thus optimizing the weight allocation logic of the Bayesian network.

10. A method for fault diagnosis of photovoltaic modules based on microinverters according to claim 9, characterized in that, The edge nodes calculate fault classification confidence through a multimodal spatiotemporal attention network, perform classification determination using a confidence threshold, and execute sample processing operations based on the classification results, including: Set a dynamic threshold, mark samples with confidence scores greater than or equal to the dynamic threshold as high-confidence samples, and mark samples with confidence scores less than the dynamic threshold as low-confidence samples; store high-confidence samples to edge nodes, and upload low-confidence samples to the cloud to trigger global reconstruction training; The cloud integrates low-confidence samples, historical fault data, and environmental parameters uploaded from edge nodes, and performs offline deep training using convolutional neural networks and long short-term memory networks. After training, new model parameters are generated, and the network weights are optimized through gradient descent.

Images