A fault early warning method and system for a photovoltaic device based on multi-source data fusion

Abstract

This invention discloses a fault early warning method and system for photovoltaic equipment based on multi-source data fusion, belonging to the field of fault early warning technology. The method includes: acquiring a first multi-source time-series operational dataset and inputting it into an operational monitoring model, outputting operational monitoring results, and concatenating the operational monitoring results with the first multi-source time-series operational dataset to obtain a second multi-source time-series operational dataset; inputting the second multi-source time-series operational dataset into a fault early warning model to obtain a first set of suspected abnormal inverters; determining the group's operational baseline range; calculating a deviation index, comparing the deviation index with a preset deviation threshold to obtain a second set of suspected abnormal inverters; and comparing the first set of suspected abnormal inverters with the second set of suspected abnormal inverters to obtain a set of faulty inverters. This invention employs a dual verification mechanism, significantly reducing the false alarm rate and avoiding the misjudgment of P-V curve changes caused by environmental factors as equipment faults.

Description

Technical Field

[0001] This invention belongs to the field of fault early warning technology, and in particular relates to a fault early warning method and system for photovoltaic equipment based on multi-source data fusion. Background Technology

[0002] In recent years, photovoltaic power generation has shown promising development prospects in my country. In a photovoltaic power generation system, the conversion between solar energy and electrical energy is achieved through an inverter. The inverter converts the direct current (DC) generated by the photovoltaic modules into alternating current (AC), which has a significant impact on the stable operation of the photovoltaic power station.

[0003] In the field of photovoltaic power generation, the most efficient and reliable power conversion control method is maximum power point tracking (MPPT) technology. MPPT technology improves energy conversion efficiency by monitoring the operating status of photovoltaic cells in real time and adjusting their voltage / current parameters to track the maximum power point.

[0004] However, factors such as passing clouds, localized shadows, and rapid temperature changes can cause drastic and frequent fluctuations in the PV curve of a photovoltaic array. Current technology struggles to distinguish between a normal MPPT tracking a changing target and a malfunction in the MPPT itself, leading to a very high false alarm rate and significant difficulties for on-site maintenance. When the MPPT malfunctions, the system cannot operate at its optimal power point, resulting in severe energy loss and directly impacting the overall return on investment of the photovoltaic power plant. Furthermore, it can cause the system to oscillate on the power curve, or even trigger a shutdown. Summary of the Invention

[0005] The purpose of this invention is to provide a fault early warning method and system for photovoltaic equipment based on multi-source data fusion, thereby solving the problems existing in the prior art.

[0006] To achieve the above objectives, this invention provides a fault early warning method for photovoltaic equipment based on multi-source data fusion, comprising the following steps: Step 100: Collect the operating data of multiple inverters in the current photovoltaic power station, and add timestamps to the operating data to obtain the first multi-source time-series operating dataset; The operating data includes: inverter voltage data, current data, power data, temperature data, and environmental data; Step 200: Input the first multi-source time-series running dataset into the pre-trained running monitoring model and output the running monitoring results; concatenate the first multi-source time-series running dataset with the running monitoring results to obtain the second multi-source time-series running dataset; The operation monitoring results include the operation data of each inverter at future points in time; Step 300: Input the second multi-source time-series running dataset into the pre-trained fault warning model, output the fault probability value of each inverter, compare the fault probability value of each inverter with the preset fault warning threshold, and form a set of inverters whose fault probability value is not less than the preset fault warning threshold to generate the first set of suspected abnormal inverters. Step 400: Preprocess the second multi-source time-series running dataset and extract the maximum power point tracking feature parameters of each inverter. Then, concatenate the preprocessed second multi-source time-series running dataset with the maximum power point tracking feature parameters of each inverter to obtain the standardized feature data of each inverter. Maximum power point tracking (MPPT) characteristic parameters include power point tracking speed, power curve fluctuation amplitude, and power conversion efficiency. Step 500: Determine the group operation reference range by calculating the statistical distribution characteristics of the maximum power point tracking characteristic parameters; Statistical distribution characteristics include mean, variance, and quantiles; Step 600: The standardized characteristic data of each inverter is compared horizontally with the group operation benchmark range to calculate the deviation index. The deviation index is compared with the preset deviation threshold. Inverters with deviation indices not less than the preset deviation threshold are grouped into a set to generate a second set of suspected abnormal inverters. Step 700: Compare the first set of suspected abnormal inverters with the second set of suspected abnormal inverters. If the same inverter appears in both the first set of suspected abnormal inverters and the second set of suspected abnormal inverters, it is determined that this inverter has a fault. The set of faulty inverters is obtained and fault warning information is generated and issued.

[0007] Furthermore, the operation monitoring model is based on a deep learning architecture, including a data input layer, a feature extraction layer, and an output layer; The input layer is used to input the first multi-source time-series running dataset; The feature extraction layer is used to extract features from the first multi-source time-series running dataset; The output layer is used to output the operating data of each inverter at a future point in time; The training process for the operation monitoring model is as follows: collect historical operation data and add timestamps, and train the model parameters through supervised learning methods.

[0008] Furthermore, the method for concatenating the first multi-source time-series running dataset with the running monitoring results is as follows: maintain the original data structure of the first multi-source time-series running dataset, add a corresponding field of the running monitoring results after each data record in the first multi-source time-series running dataset, and obtain the second multi-source time-series running dataset.

[0009] Furthermore, the fault warning model adopts a deep learning network structure, including convolutional neural network layers and long short-term memory network layers; Convolutional neural network layers are used to extract spatial features from the second multi-source time-series dataset and identify the relationships between different data dimensions. Long Short-Term Memory (LSTM) network layers are used to extract temporal features, capture trends and patterns in running data over time, and analyze the evolutionary relationship between current actual data and future predicted data. The training process of the fault warning model is as follows: collect historical fault case data and normal operation data of the inverter, label the data, and label the fault type, fault time and fault severity; divide the labeled data into training set, validation set and test set to train the fault warning model.

[0010] Furthermore, the preprocessing operation includes three sub-steps: data cleaning, missing value handling, and data normalization. During the data cleaning phase, the server uses box plots to identify outliers, calculates the quartiles of each data dimension, and determines the outlier criteria. For voltage, current, and power data, if the data value exceeds 1.5 times the interquartile range of the normal range, it is marked as an outlier and removed. During the missing value handling phase, if data is missing at a certain time, linear interpolation is performed based on the valid data at the time before and after that data to fill the missing value; if data is missing for more than 5 consecutive time periods, the data record in that segment is deleted. During the data normalization phase, the server uses the min-max normalization method to map data of different dimensions to the same numerical range.

[0011] Furthermore, the maximum power point tracking characteristic parameters include three parameters: power point tracking speed, power curve fluctuation amplitude, and power conversion efficiency. The power point tracking speed is calculated as follows: The server acquires the inverter's power data sequence within a continuous time window, calculates the power change rate, and the power point tracking speed is defined as the average rate of power change per unit time. The calculation formula is: ; in, For power point tracking speed, The number of data points within the time window. For the first The power value at each moment. For the first The power value at each moment. The sampling time interval; The calculation method for the fluctuation amplitude of the power curve is as follows: The server calculates the standard deviation of the power data sequence. The standard deviation reflects the degree of fluctuation of the power curve, and the calculation formula is: ; in, The amplitude of the power curve fluctuation. The number of data points within the time window. For the first The power value at each moment. This represents the average power. The power conversion efficiency is calculated as follows: The server obtains the input and output power of the inverter and calculates the power conversion efficiency using the following formula: ; in, For power conversion efficiency, For output power, This refers to the input power.

[0012] Furthermore, the steps for determining the group's operational baseline range by calculating the statistical distribution characteristics of the maximum power point tracking characteristic parameters include: Step 401: The server calculates the statistical distribution characteristics of the three maximum power point tracking feature parameters, including the mean, variance, and quantiles. Step 402: The server determines the baseline range for group operation based on the calculated statistical distribution characteristics; The baseline range is determined using the 3σ principle of normal distribution, which assumes that normal data should be distributed within three standard deviations of the mean. For power point tracking speed, the baseline range for group operation is:

[0013] in, This serves as the lower limit of the power point tracking speed baseline. This serves as the baseline upper limit for power point tracking speed. The group mean The group standard deviation; Similarly, the group operating benchmark range for determining the power curve fluctuation amplitude of the server is... The group operating reference range for power conversion efficiency is ; Step 403: The server integrates the group operating reference ranges of the three characteristic parameters to form a complete group operating reference range data structure. This data structure includes the power point tracking speed reference range, the power curve fluctuation amplitude reference range, and the power conversion efficiency reference range.

[0014] Furthermore, regarding power point tracking speed, the server calculates the population mean, population variance, and quantiles, as follows: The formula for calculating the group mean is: ; in, The population mean of the power point tracking speed. This refers to the total number of inverters in the photovoltaic power station. For the first Power point tracking speed of each inverter; The formula for calculating population variance is: ; in, The population variance of the power point tracking speed. For the first Power point tracking speed of each inverter This represents the group mean of the power point tracking speed; The quantile calculation method is as follows: The server sorts the power point tracking speed data of all inverters in ascending order and calculates the 25th quantile, 50th quantile, and 75th quantile; the 25th quantile... This indicates that 25% of the data are less than this value, the 50th quantile. That is, the median, the 75th quantile. This indicates that 75% of the data is less than this value; For both power curve fluctuation amplitude and power conversion efficiency, the server uses the same calculation method to calculate their population mean, population variance, and quantiles, respectively; the population mean for power curve fluctuation amplitude is... The group variance is The group mean of power conversion efficiency is The group variance is .

[0015] Furthermore, the deviation index is calculated as follows: Determine whether the inverter's power point tracking speed is within the group's operating reference range; if it is within the range, the deviation is 0; if it is not within the range, calculate the distance it deviates from the reference range. The formula for calculating the power point tracking speed deviation is: ; in, This is an indicator of the deviation from the power point tracking speed. This refers to the power point tracking speed of the inverter. and These are the lower and upper limits of the power point tracking speed, respectively. The group standard deviation of the power point tracking speed; Using the same calculation method, the deviation index of the power curve fluctuation amplitude is calculated. Deviation index from power conversion efficiency ; After the server calculates the three deviation indicators, it calculates the comprehensive deviation indicator using the following formula: ; in, As a comprehensive deviation index, , and These are the weighting coefficients for power point tracking speed, power curve fluctuation amplitude, and power conversion efficiency, respectively, and the sum of the three weighting coefficients is 1.

[0016] A fault early warning system for photovoltaic equipment based on multi-source data fusion, used to execute any of the aforementioned fault early warning methods for photovoltaic equipment based on multi-source data fusion, comprising: The data acquisition module collects the operating data of multiple inverters in the current photovoltaic power station and adds timestamps to the operating data to obtain the first multi-source time-series operating dataset. The operation monitoring module inputs the first multi-source time-series operation dataset into the pre-trained operation monitoring model and outputs the operation monitoring results; the first multi-source time-series operation dataset and the operation monitoring results are concatenated to obtain the second multi-source time-series operation dataset. The first fault determination module inputs the second multi-source time-series running dataset into the pre-trained fault warning model, outputs the fault probability value of each inverter, compares the fault probability value of each inverter with the preset fault warning threshold, and obtains the first set of suspected abnormal inverters. The group operation reference range determination module preprocesses the second multi-source timing operation dataset and extracts the maximum power point tracking characteristic parameters of each inverter; by calculating the statistical distribution characteristics of the maximum power point tracking characteristic parameters, the group operation reference range is determined. The second fault determination module compares the standardized characteristic data of each inverter with the group's operating benchmark range, calculates the deviation index, and compares the deviation index with the preset deviation threshold to obtain the second set of suspected abnormal inverters. The fault warning module compares the first set of suspected abnormal inverters with the second set of suspected abnormal inverters. If the same inverter appears in both the first set of suspected abnormal inverters and the second set of suspected abnormal inverters, it determines that this inverter has a fault, obtains the set of faulty inverters, and generates and publishes fault warning information.

[0017] Therefore, the present invention employs the above-mentioned fault early warning method and system for photovoltaic equipment based on multi-source data fusion, which has the following beneficial effects: 1. By integrating multi-source operating data such as inverter voltage, current, power, temperature, and environmental data, and using an operating monitoring model to predict operating data at future points in time, the actual operating data is spliced ​​with the predicted operating data, providing rich time-series evolution information for fault early warning. This enables the system to identify potential fault trends in advance, realizing the transformation from post-diagnosis to predictive maintenance and improving the foresight of fault early warning. 2. A dual verification mechanism is adopted, which cross-verifies the identification results of the deep learning fault warning model with the analysis results of the group horizontal comparison. Only inverters that meet both judgment conditions are identified as faulty equipment. This dual verification mechanism effectively solves the technical problem of distinguishing between normal MPPT tracking change targets and MPPT real faults in the existing technology, significantly reduces the false alarm rate, and avoids the situation where PV curve changes caused by environmental factors are misjudged as equipment faults. 3. By extracting maximum power point tracking characteristic parameters, including power point tracking speed, power curve fluctuation amplitude, and power conversion efficiency, and using group data of all inverters in the photovoltaic power station to establish a benchmark range for normal operation, the 3σ principle is used to identify abnormal individuals that deviate from the normal state of the group. This method makes full use of the operational correlation of equipment in the same photovoltaic power station. When environmental factors affect all equipment, a single device will not be judged as abnormal, but when an individual device deviates from the group due to its own fault, it can be accurately identified, thus improving the reliability and accuracy of fault identification. 4. The deep learning architecture that combines convolutional neural networks to extract spatial features and long short-term memory networks to extract temporal features can identify the correlation between different data dimensions, capture the trends and patterns of running data over time, analyze the evolutionary relationship between current actual data and future predicted data, realize intelligent identification of complex fault modes, and improve the system's adaptability to multiple fault types.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] Figure 1 This is a flowchart of a fault early warning method for photovoltaic equipment based on multi-source data fusion according to the present invention. Detailed Implementation

[0020] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0021] Please see Figure 1 A fault early warning method for photovoltaic equipment based on multi-source data fusion includes the following steps: Step 100: Collect the operating data of multiple inverters in the current photovoltaic power station, and add timestamps to the operating data to obtain the first multi-source time-series operating dataset; Specifically, in this embodiment, the executing entity is a fault early warning server, which is connected to multiple inverters in the photovoltaic power station via a communication network. In this embodiment, the photovoltaic power station includes at least 10 inverter devices, and each inverter device is equipped with a data acquisition sensor.

[0022] The server collects real-time operating data from each inverter through a data acquisition interface. The operating data includes five types of data sources: First, inverter voltage data, including DC-side input voltage and AC-side output voltage; second, inverter current data, including DC-side input current and AC-side output current; third, inverter power data, including input power, output power, and real-time power conversion efficiency; fourth, inverter temperature data, including inverter internal temperature, radiator temperature, and power module temperature; and fifth, environmental data, including ambient temperature, humidity, light intensity, and electromagnetic field strength.

[0023] The server collects the aforementioned operational data according to a preset collection frequency. In this embodiment, the collection frequency is set to once per second. When collecting each set of operational data, the server synchronously adds a timestamp identifier, which includes year, month, day, hour, minute, second, and millisecond information to ensure the time sequence of the data.

[0024] The server organizes the collected operating data from each inverter according to timestamp order, constructs a data structure, and forms the first multi-source time-series operating dataset. The data format of the first multi-source time-series operating dataset is as follows: each data record includes inverter identifier, timestamp, voltage data, current data, power data, temperature data, and environmental data.

[0025] Step 200: Input the first multi-source time-series running dataset into the pre-trained running monitoring model and output the running monitoring results; concatenate the first multi-source time-series running dataset with the running monitoring results to obtain the second multi-source time-series running dataset; Specifically, the server pre-builds and trains an operation monitoring model based on a deep learning architecture, including a data input layer, a feature extraction layer, and an output layer. The training process of the operation monitoring model is as follows: collecting historical operation data, predicting operation data at future time points, and training the model parameters through supervised learning methods so that the model can predict the inverter's operation data.

[0026] The server uses the first multi-source time-series operational dataset obtained in step 100 as input data into the operation monitoring model. The operation monitoring model processes the input data, specifically including: first, the data input layer receives the first multi-source time-series operational dataset; second, the feature extraction layer extracts operational status features from the input data, including voltage stability features, current fluctuation features, power output features, temperature change features, and environmental impact features; finally, the output layer generates operation monitoring results based on the extracted features, which include the operational data of each inverter at future time points.

[0027] The specific content of the operation monitoring results is as follows: For each inverter, the model predicts its operation data at future time points. The prediction time window is set to 1 minute in the future. The predicted data includes the predicted voltage, current, power and temperature values ​​at each future time.

[0028] After the server obtains the operation monitoring results output by the operation monitoring model, it concatenates the first multi-source time-series operation dataset with the operation monitoring results. The specific method for concatenation is as follows: Maintaining the original data structure of the first multi-source time-series operation dataset, the corresponding operation monitoring result field is appended to each data record in the first multi-source time-series operation dataset, forming an expanded data record. The concatenated data record simultaneously contains the actual operation data at the current moment and the predicted operation data for future moments, providing time-series evolution information for subsequent fault early warning analysis.

[0029] After the server completes the data splicing, it obtains the second multi-source time-series runtime dataset.

[0030] Step 300: Input the second multi-source time-series running dataset into the pre-trained fault warning model, output the fault probability value of each inverter, compare the fault probability value of each inverter with the preset fault warning threshold, and form a set of inverters whose fault probability value is not less than the preset fault warning threshold to generate the first set of suspected abnormal inverters. Specifically, the server pre-builds and trains a fault warning model, which employs a deep learning network structure, including convolutional neural network layers and long short-term memory network layers. The convolutional neural network layers are used to extract spatial features from the second multi-source time-series running dataset and identify the correlation between different data dimensions; the long short-term memory network layers are used to extract time-series features, capture the trends and patterns of running data changes over time, and analyze the evolutionary relationship between current actual data and future predicted data.

[0031] The training process of the fault early warning model is as follows: collect historical fault case data and normal operation data, and label the data, including fault type, fault time and fault severity; divide the labeled data into training set, validation set and test set in a ratio of 7:2:1; train the model parameters through the training set, adjust the hyperparameters through the validation set, and evaluate the model performance through the test set; after training, the model can comprehensively judge the probability of inverter failure based on the input actual operation data and predicted operation data.

[0032] The server inputs the second multi-source time-series operational dataset obtained in step 200 into the fault warning model. This dataset contains current actual operational data and future predicted operational data, providing the model with rich temporal information. The fault warning model processes the input data in the following steps: First, a convolutional neural network layer performs convolution operations on the second multi-source time-series operational dataset to extract multi-dimensional feature vectors, while also extracting the difference features between the actual and predicted data. Second, a long short-term memory network layer performs temporal analysis on the feature vectors, establishing time dependencies and analyzing the evolution trend from history to the present and into the future. Finally, through a fully connected layer and activation functions, the fault probability value of each inverter is output.

[0033] A fault probability value between 0 and 1 indicates the likelihood of the inverter malfunctioning. The closer the fault probability value is to 1, the higher the probability of the inverter malfunctioning; the closer the fault probability value is to 0, the higher the probability of the inverter operating normally.

[0034] The server sets a preset fault warning threshold. In this embodiment, the fault warning threshold is set to 0.7. The server compares the fault probability value of each inverter with the fault warning threshold. The comparison rule is as follows: if the fault probability value of an inverter is greater than or equal to the fault warning threshold of 0.7, the inverter is marked as a suspected abnormal inverter; if the fault probability value of an inverter is less than the fault warning threshold of 0.7, the inverter is determined to be operating normally.

[0035] The server aggregates all inverters marked as potentially faulty, forming a first set of potentially faulty inverters. This first set of potentially faulty inverters contains a list of inverter identifiers identified based on a deep learning model that may indicate a risk of failure.

[0036] Step 400: Preprocess the second multi-source time-series running dataset and extract the maximum power point tracking feature parameters of each inverter. Then, concatenate the preprocessed second multi-source time-series running dataset with the maximum power point tracking feature parameters of each inverter to obtain the standardized feature data of each inverter. Specifically, the preprocessing operation includes three sub-steps: data cleaning, missing value handling, and data normalization.

[0037] During the data cleaning phase, the server performs outlier detection and removal on the second multi-source time-series dataset. The server uses a box plot method to identify outliers, calculates the quartiles of each data dimension, and determines the outlier judgment boundary. For voltage, current, and power data, if the data value exceeds 1.5 times the interquartile range of the normal range, it is marked as an outlier and removed.

[0038] During the missing value handling phase, the server checks whether there is missing data in the second multi-source time-series dataset. A linear interpolation method is used to handle missing values: if data is missing at a certain time point, linear interpolation is performed based on valid data from the time points before and after that point to fill the missing value; if data is missing for more than 5 consecutive time points, that data record is deleted.

[0039] During the data normalization phase, the server normalizes the cleaned data, mapping data with different dimensions to the same numerical range. The server uses the min-max normalization method.

[0040] After the server completes the preprocessing, it extracts the maximum power point tracking (MPPT) characteristic parameters for each inverter. These MPPT characteristic parameters include three parameters: power point tracking speed, power curve fluctuation amplitude, and power conversion efficiency.

[0041] The power point tracking speed is calculated as follows: The server acquires the inverter's power data sequence within a continuous time window, calculates the power change rate, and the power point tracking speed is defined as the average rate of power change per unit time. The calculation formula is: ; in, For power point tracking speed, The number of data points within the time window. For the first The power value at each moment. For the first The power value at each moment. The sampling time interval is 1 second in this embodiment.

[0042] The calculation method for the fluctuation amplitude of the power curve is as follows: The server calculates the standard deviation of the power data sequence. The standard deviation reflects the degree of fluctuation of the power curve, and the calculation formula is: ; in, The amplitude of the power curve fluctuation. The number of data points within the time window. For the first The power value at each moment. The average power is calculated using the following formula: .

[0043] The power conversion efficiency is calculated as follows: The server obtains the input and output power of the inverter and calculates the power conversion efficiency using the following formula: ; in, For power conversion efficiency, The output power is obtained from the output power data collected in step 100. The input power is obtained from the input power data collected in step 100.

[0044] Step 500: Determine the group operation reference range by calculating the statistical distribution characteristics of the maximum power point tracking characteristic parameters; Statistical distribution characteristics include mean, variance, and quantiles; Specifically, the server uses group data generated by all inverters within the same photovoltaic power plant to determine the group's operational baseline range through statistical analysis. This step establishes a benchmark standard for normal operation using group data, providing a reference for subsequent horizontal comparisons.

[0045] The server calculates the statistical distribution characteristics of the three feature parameters, including the mean, variance, and quantiles.

[0046] For power point tracking speed, the server calculates the population mean, population variance, and quantiles. The formula for calculating the population mean is: ; in, The population mean of the power point tracking speed. This refers to the total number of inverters in the photovoltaic power station. For the first The power point tracking speed of each inverter is calculated in step 400.

[0047] The formula for calculating population variance is: ; in, The population variance of the power point tracking speed. For the first Power point tracking speed of each inverter This represents the group mean of the power point tracking speed.

[0048] The quantiles are calculated as follows: the server sorts the power point tracking speed data of all inverters in ascending order and calculates the 25th, 50th, and 75th quantiles. The 25th quantile... This indicates that 25% of the data are less than this value, the 50th quantile. That is, the median, the 75th quantile. This indicates that 75% of the data is less than this value.

[0049] For both power curve fluctuation amplitude and power conversion efficiency, the server uses the same calculation method to calculate their population mean, population variance, and quantiles, respectively. The population mean for power curve fluctuation amplitude is... The group variance is The group mean of power conversion efficiency is The group variance is .

[0050] The server determines the baseline range for group operation based on the calculated statistical distribution characteristics. The baseline range is determined using the 3σ principle of normal distribution, meaning that normal data should be distributed within three standard deviations of the mean. For power point tracking speed, the baseline range for group operation is: ; in, This serves as the lower limit of the power point tracking speed baseline. This serves as the baseline upper limit for power point tracking speed. The group mean The standard deviation of the group is calculated using the following formula: .

[0051] Similarly, the group operating benchmark range for determining the power curve fluctuation amplitude of the server is... The group operating reference range for power conversion efficiency is .

[0052] The server integrates the group operating baseline ranges of the three characteristic parameters to form a complete group operating baseline range data structure. This data structure includes the power point tracking speed baseline range, the power curve fluctuation amplitude baseline range, and the power conversion efficiency baseline range. The group operating baseline range reflects the normal operating status characteristics of the inverter group within the same photovoltaic power plant, providing a horizontal comparison reference standard for identifying abnormal individuals.

[0053] Step 600: The standardized characteristic data of each inverter is compared horizontally with the group operation benchmark range to calculate the deviation index. The deviation index is compared with the preset deviation threshold. Inverters with deviation indices not less than the preset deviation threshold are grouped into a set to generate a second set of suspected abnormal inverters. Specifically, the server performs a horizontal comparative analysis on each inverter within the photovoltaic power station. This step embodies the core idea of ​​this invention: using group data to discover abnormal individuals. By comparing the operating characteristics of a single inverter with a group benchmark, abnormal inverters that deviate from the normal operating state of the group are identified.

[0054] The server retrieves the maximum power point tracking characteristic parameters of each inverter one by one from step 400, including the inverter's power point tracking speed, power curve fluctuation amplitude, and power conversion efficiency.

[0055] The server performs a horizontal comparison of the characteristic parameters of an individual inverter with the group operating reference range determined in step 500, and calculates the deviation index. The deviation index reflects the degree to which the operating characteristics of an individual inverter deviate from the group reference; the greater the deviation, the greater the difference between the operating state of that inverter and the normal state of the group.

[0056] The calculation method for power point tracking speed deviation is as follows: First, determine whether the inverter's power point tracking speed is within the group's operating reference range; if it is within the range, the deviation is 0; if it is not within the range, calculate its distance from the reference range. The formula for calculating power point tracking speed deviation is: ; in, This is an indicator of the deviation from the power point tracking speed. This refers to the power point tracking speed of the inverter. and These are the lower and upper limits of the power point tracking speed, respectively. This represents the group standard deviation of the power point tracking speed.

[0057] Using the same calculation method, the server calculates the deviation index of the power curve fluctuation amplitude. Deviation index from power conversion efficiency The formula for calculating the deviation of the power curve fluctuation amplitude is: ; in, This is an index of the deviation of the power curve fluctuation amplitude. This represents the fluctuation range of the inverter's power curve. and These are the lower and upper limits of the power curve fluctuation amplitude, respectively. This represents the group standard deviation of the power curve fluctuation amplitude.

[0058] The formula for calculating the deviation of power conversion efficiency is: ; in, This is an indicator of the deviation from power conversion efficiency. This refers to the power conversion efficiency of the inverter. and These are the lower and upper limits of the power conversion efficiency, respectively. This represents the group standard deviation of the power conversion efficiency.

[0059] After the server calculates the three deviation indicators, it performs a comprehensive evaluation to calculate the overall deviation indicator. The overall deviation indicator is calculated using a weighted summation method, and the formula is as follows: ; in, As a comprehensive deviation index, , and These are weighting coefficients for power point tracking speed, power curve fluctuation amplitude, and power conversion efficiency, respectively, and the sum of these three weighting coefficients is 1. In this embodiment, , , .

[0060] The server sets a preset deviation threshold, which is set to 3.0 in this embodiment. This threshold corresponds to the 3σ principle, meaning that inverters with a deviation exceeding three times the standard deviation are considered abnormal.

[0061] The server compares the overall deviation index of each inverter with the deviation threshold. The comparison rule is: if the overall deviation index of a certain inverter... If the deviation index is greater than or equal to the deviation threshold of 3.0, the inverter is marked as a suspected abnormal inverter; if the comprehensive deviation index of an inverter is... If the deviation is less than the deviation threshold of 3.0, the inverter is considered to be performing normally in the group comparison.

[0062] The server aggregates all inverters marked as potentially abnormal to form a second set of potentially abnormal inverters. This second set of potentially abnormal inverters contains a list of inverter identifiers that deviate from the normal operating state of the group, identified based on a group-wide cross-sectional comparison method.

[0063] Step 700: Compare the first set of suspected abnormal inverters with the second set of suspected abnormal inverters. If the same inverter appears in both the first set of suspected abnormal inverters and the second set of suspected abnormal inverters, it is determined that this inverter has a fault. The set of faulty inverters is obtained and fault warning information is generated and issued.

[0064] Specifically, the server employs a dual-verification mechanism to determine the final faulty inverter. This mechanism combines deep learning model predictions with group cross-validation results, improving the accuracy of fault identification and reducing the false alarm rate through cross-validation of two different technical approaches.

[0065] The server obtains the first set of suspected abnormal inverters obtained in step 300 and the second set of suspected abnormal inverters obtained in step 600. The first set of suspected abnormal inverters is denoted as set [missing information]. The first set includes identifiers of suspected faulty inverters identified based on a deep learning model; the second set of suspected faulty inverters is denoted as set [set name missing]. It includes identifiers of suspected abnormal inverters based on cross-group comparison.

[0066] The server performs an intersection operation on two sets to find the inverter identifiers that appear in both sets. The mathematical expression for the intersection operation is: ; in, The set of faulty inverters represents the set of inverters that are simultaneously identified as faulty by both methods. This is the first set of inverters suspected of being abnormal. This is the second set of inverters suspected of being abnormal. The intersection operation of sets is represented.

[0067] The server determines the result of the intersection operation. The determination rule is as follows: if the identifier of an inverter exists in both the first suspected abnormal inverter set and the second suspected abnormal inverter set, that is, the inverter is determined by the deep learning model to have a high probability of failure and is determined by the group horizontal comparison to deviate from the group benchmark, then the inverter is confirmed as a faulty inverter and added to the faulty inverter set.

[0068] Through a dual verification mechanism, the server effectively filters out false alarms that might arise from a single method. For example, when environmental factors cause changes in the PV curve of a photovoltaic array, the future operating data predicted by the operation monitoring model may deviate from the actual trend. The fault warning model might identify it as an anomaly, but if the inverter performs consistently with other inverters in a group comparison (because the impact of environmental factors is the same on all inverters), it will not be added to the second set of suspected anomaly inverters, thus avoiding false alarms. Conversely, if it is indeed an MPPT fault, the actual operating trend of the inverter will deviate from the predicted trend, and its operating characteristics will also deviate from the group benchmark. It will be identified by both methods simultaneously, ensuring the accuracy of fault identification.

[0069] After the server completes the intersection operation, it obtains the final set of faulty inverters. The set of faulty inverters contains a list of inverter identifiers that have been double-verified and confirmed to be at risk of failure.

[0070] The server generates fault warning information for each inverter in the faulty inverter set. The fault warning information includes: inverter identifier, fault detection time, fault probability value output by deep learning model, comprehensive deviation index calculated by group horizontal comparison, details of abnormal feature parameters, and suggested handling measures.

[0071] The server outputs the generated fault warning information to the monitoring terminal. Output methods include: displaying fault warning prompts on the graphical interface of the monitoring terminal, including the location marker and detailed information of the faulty inverter; sending warning notification messages to maintenance personnel, including SMS notifications and mobile application push notifications; and recording fault warning information in the database to establish a fault warning log for subsequent analysis and statistical use.

[0072] After receiving fault warning information, the monitoring terminal can promptly arrange on-site inspections and repairs based on the warning information, thus avoiding energy loss and system downtime risks caused by the escalation of the fault.

[0073] Through the above steps, this invention achieves fault early warning for photovoltaic equipment based on multi-source data fusion. This method integrates operational information from multiple data sources, predicts future operational trends through an operational monitoring model, and combines deep learning technology with cross-sectional comparison technology of group data, employing a dual verification mechanism to effectively solve the problem of high false alarm rate in MPPT fault identification in existing technologies. By introducing future prediction data, the system can detect potential fault risks in advance, achieving predictive maintenance and improving the accuracy, reliability, and foresight of fault early warning.

[0074] A fault early warning system for photovoltaic equipment based on multi-source data fusion includes: The data acquisition module collects the operating data of multiple inverters in the current photovoltaic power station and adds timestamps to the operating data to obtain the first multi-source time-series operating dataset. The operation monitoring module inputs the first multi-source time-series operation dataset into the pre-trained operation monitoring model and outputs the operation monitoring results; the first multi-source time-series operation dataset and the operation monitoring results are concatenated to obtain the second multi-source time-series operation dataset. The first fault determination module inputs the second multi-source time-series running dataset into the pre-trained fault warning model, outputs the fault probability value of each inverter, compares the fault probability value of each inverter with the preset fault warning threshold, and obtains the first set of suspected abnormal inverters. The group operation reference range determination module preprocesses the second multi-source timing operation dataset and extracts the maximum power point tracking characteristic parameters of each inverter; by calculating the statistical distribution characteristics of the maximum power point tracking characteristic parameters, the group operation reference range is determined. The second fault determination module compares the standardized characteristic data of each inverter with the group's operating benchmark range, calculates the deviation index, and compares the deviation index with the preset deviation threshold to obtain the second set of suspected abnormal inverters. The fault warning module compares the first set of suspected abnormal inverters with the second set of suspected abnormal inverters. If the same inverter appears in both the first set of suspected abnormal inverters and the second set of suspected abnormal inverters, it determines that this inverter has a fault, obtains the set of faulty inverters, and generates and publishes fault warning information.

[0075] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for fault early warning of a photovoltaic device based on multi-source data fusion, characterized in that, Includes the following steps: Step 100: Collect the operating data of multiple inverters in the current photovoltaic power station, and add timestamps to the operating data to obtain the first multi-source time-series operating dataset; The operating data includes: inverter voltage data, current data, power data, temperature data, and environmental data; Step 200: Input the first multi-source time-series running dataset into the pre-trained running monitoring model and output the running monitoring results; concatenate the first multi-source time-series running dataset with the running monitoring results to obtain the second multi-source time-series running dataset; The operation monitoring results include the operation data of each inverter at future points in time; Step 300: Input the second multi-source time-series running dataset into the pre-trained fault warning model, output the fault probability value of each inverter, compare the fault probability value of each inverter with the preset fault warning threshold, and form a set of inverters whose fault probability value is not less than the preset fault warning threshold to generate the first set of suspected abnormal inverters. Step 400: Preprocess the second multi-source time-series running dataset and extract the maximum power point tracking feature parameters of each inverter. Then, concatenate the preprocessed second multi-source time-series running dataset with the maximum power point tracking feature parameters of each inverter to obtain the standardized feature data of each inverter. Maximum power point tracking (MPPT) characteristic parameters include power point tracking speed, power curve fluctuation amplitude, and power conversion efficiency. Step 500: Determine the group operation reference range by calculating the statistical distribution characteristics of the maximum power point tracking characteristic parameters; Statistical distribution characteristics include mean, variance, and quantiles; Step 600: The standardized characteristic data of each inverter is compared horizontally with the group operation benchmark range to calculate the deviation index. The deviation index is compared with the preset deviation threshold. Inverters with deviation indices not less than the preset deviation threshold are grouped into a set to generate a second set of suspected abnormal inverters. Step 700: Compare the first set of suspected abnormal inverters with the second set of suspected abnormal inverters. If the same inverter appears in both the first set of suspected abnormal inverters and the second set of suspected abnormal inverters, it is determined that this inverter has a fault. The set of faulty inverters is obtained and fault warning information is generated and issued. 2.The photovoltaic device fault early warning method based on multi-source data fusion of claim 1, wherein, The operation monitoring model is based on a deep learning architecture, including a data input layer, a feature extraction layer, and an output layer; The input layer is used to input the first multi-source time-series running dataset; The feature extraction layer is used to extract features from the first multi-source time-series running dataset; The output layer is used to output the operating data of each inverter at a future point in time; The training process for the operation monitoring model is as follows: collect historical operation data and add timestamps, and train the model parameters through supervised learning methods. 3.The photovoltaic device fault early warning method based on multi-source data fusion of claim 2, characterized in that, The method for concatenating the first multi-source time-series running dataset with the running monitoring results is as follows: maintain the original data structure of the first multi-source time-series running dataset, add the corresponding running monitoring result field after each data record in the first multi-source time-series running dataset, and obtain the second multi-source time-series running dataset. 4.The photovoltaic device fault early warning method based on multi-source data fusion of claim 3, wherein, The fault warning model adopts a deep learning network structure, including convolutional neural network layers and long short-term memory network layers; Convolutional neural network layers are used to extract spatial features from the second multi-source time-series dataset and identify the relationships between different data dimensions. Long Short-Term Memory (LSTM) network layers are used to extract temporal features, capture trends and patterns in running data over time, and analyze the evolutionary relationship between current actual data and future predicted data. The training process of the fault early warning model is as follows: collect historical fault case data and normal operation data of the inverter, and label the data. The labeling content includes fault type, fault time and fault severity. The labeled data is divided into training, validation, and test sets for training the fault warning model. 5.The photovoltaic device fault early warning method based on multi-source data fusion of claim 4, characterized in that: Preprocessing operations include three sub-steps: data cleaning, missing value handling, and data normalization. During the data cleaning phase, the server uses box plots to identify outliers, calculates the quartiles of each data dimension, and determines the outlier criteria. For voltage, current, and power data, if the data value exceeds 1.5 times the interquartile range of the normal range, it is marked as an outlier and removed. During the missing value handling phase, if data is missing at a certain time, linear interpolation is performed based on the valid data at the time before and after that data to fill the missing value; if data is missing for more than 5 consecutive time periods, the data record in that segment is deleted. During the data normalization phase, the server uses the min-max normalization method to map data of different dimensions to the same numerical range. 6.The photovoltaic device fault early warning method based on multi-source data fusion of claim 5, wherein, The maximum power point tracking characteristic parameters include three parameters: power point tracking speed, power curve fluctuation amplitude, and power conversion efficiency. The power point tracking speed is calculated as follows: The server acquires the inverter's power data sequence within a continuous time window, calculates the power change rate, and the power point tracking speed is defined as the average rate of power change per unit time. The calculation formula is: ； in, For power point tracking speed, The number of data points within the time window. For the first The power value at each moment, For the first The power value at each moment, The sampling time interval; The calculation method for the fluctuation amplitude of the power curve is as follows: The server calculates the standard deviation of the power data sequence. The standard deviation reflects the degree of fluctuation of the power curve, and the calculation formula is: ； in, The amplitude of the power curve fluctuation. The number of data points within the time window. For the first The power value at each moment, This represents the average power. The power conversion efficiency is calculated as follows: The server obtains the input and output power of the inverter and calculates the power conversion efficiency using the following formula: ； wherein, PCE is the power conversion efficiency, Pout is the output power, Pin is the input power. 7.The photovoltaic device fault early warning method based on multi-source data fusion of claim 6, wherein, The steps for determining the group operating baseline range by calculating the statistical distribution characteristics of the maximum power point tracking characteristic parameters include: Step 401: The server calculates the statistical distribution characteristics of the three maximum power point tracking feature parameters, including the mean, variance, and quantiles. Step 402: The server determines the baseline range for group operation based on the calculated statistical distribution characteristics; The baseline range is determined using the 3σ principle of normal distribution, which assumes that normal data should be distributed within three standard deviations of the mean. For power point tracking speed, the baseline range for group operation is: in, This serves as the lower limit of the power point tracking speed baseline. This serves as the baseline upper limit for power point tracking speed. The group mean The group standard deviation; Similarly, the server determines the group operation reference range of the fluctuation amplitude of the power curve to be , and the group operation reference range of the power conversion efficiency to be ; Step 403: The server integrates the group operating reference ranges of the three characteristic parameters to form a complete group operating reference range data structure. This data structure includes the power point tracking speed reference range, the power curve fluctuation amplitude reference range, and the power conversion efficiency reference range.

8. A fault early warning method for photovoltaic equipment based on multi-source data fusion according to claim 7, characterized in that, For power point tracking speed, the server calculates the population mean, population variance, and quantiles, as follows: The formula for calculating the group mean is: ； in, The population mean of the power point tracking speed. This refers to the total number of inverters in the photovoltaic power station. For the first Power point tracking speed of each inverter; The formula for calculating population variance is: ； in, The population variance of the power point tracking speed. For the first Power point tracking speed of each inverter This represents the group mean of the power point tracking speed; The quantile calculation method is as follows: The server sorts the power point tracking speed data of all inverters in ascending order and calculates the 25th quantile, 50th quantile, and 75th quantile; the 25th quantile... This indicates that 25% of the data are less than this value, the 50th quantile. That is, the median, the 75th quantile. This indicates that 75% of the data is less than this value; For both power curve fluctuation amplitude and power conversion efficiency, the server uses the same calculation method to calculate their population mean, population variance, and quantiles, respectively; the population mean for power curve fluctuation amplitude is... The group variance is The group mean of power conversion efficiency is The group variance is .

9. A fault early warning method for photovoltaic equipment based on multi-source data fusion according to claim 8, characterized in that, The deviation index is calculated as follows: Determine whether the inverter's power point tracking speed is within the group's operating reference range; if it is within the range, the deviation is 0; if it is not within the range, calculate the distance it deviates from the reference range. The formula for calculating the power point tracking speed deviation is: ； in, This is an indicator of the deviation from the power point tracking speed. This refers to the power point tracking speed of the inverter. and These are the lower and upper limits of the power point tracking speed, respectively. The group standard deviation of the power point tracking speed; Using the same calculation method, the deviation index of the fluctuation amplitude of the power curve is calculated and the deviation index of the power conversion efficiency ; After the server calculates the three deviation indicators, it calculates the comprehensive deviation indicator using the following formula: ； in, As a comprehensive deviation index, , and These are the weighting coefficients for power point tracking speed, power curve fluctuation amplitude, and power conversion efficiency, respectively, and the sum of the three weighting coefficients is 1.

10. A fault early warning system for photovoltaic equipment based on multi-source data fusion, characterized in that, A fault early warning method for photovoltaic equipment based on multi-source data fusion as described in any one of claims 1-9 includes: The data acquisition module collects the operating data of multiple inverters in the current photovoltaic power station and adds timestamps to the operating data to obtain the first multi-source time-series operating dataset. The operation monitoring module inputs the first multi-source time-series operation dataset into the pre-trained operation monitoring model and outputs the operation monitoring results; the first multi-source time-series operation dataset and the operation monitoring results are concatenated to obtain the second multi-source time-series operation dataset. The first fault determination module inputs the second multi-source time-series running dataset into the pre-trained fault warning model, outputs the fault probability value of each inverter, compares the fault probability value of each inverter with the preset fault warning threshold, and obtains the first set of suspected abnormal inverters. The group operation reference range determination module preprocesses the second multi-source timing operation dataset and extracts the maximum power point tracking characteristic parameters of each inverter; by calculating the statistical distribution characteristics of the maximum power point tracking characteristic parameters, the group operation reference range is determined. The second fault determination module compares the standardized characteristic data of each inverter with the group's operating benchmark range, calculates the deviation index, and compares the deviation index with the preset deviation threshold to obtain the second set of suspected abnormal inverters. The fault warning module compares the first set of suspected abnormal inverters with the second set of suspected abnormal inverters. If the same inverter appears in both the first set of suspected abnormal inverters and the second set of suspected abnormal inverters, it determines that this inverter has a fault, obtains the set of faulty inverters, and generates and publishes fault warning information.

Images