A method and system for inverter health quantification grading and state prediction

Abstract

The application discloses a kind of inverter health degree quantification grading and state prediction method and system thereof, by obtaining the operation data of to be evaluated inverter and carrying out data preprocessing, calculate three main index scores and eight sub-index scores under its jurisdiction, then based on information entropy and / or discrete degree, adaptively determine the weight of each main index and its subordinate sub-index;Then calculate the total score of health degree, and take historical operation data for training model, adopt rolling prediction mode to output the health degree prediction curve in future preset days, finally automatically calculate the current health score of inverter and health degree grade information, and predict the remaining health period from the current health degree grade to next grade.The application can provide future long-time maintenance strategy reference for operation and maintenance personnel.

Description

Technical Field

[0001] This invention belongs to the field of intelligent operation and maintenance technology of photovoltaic power plants, and in particular relates to a method and system for quantitative classification and state prediction of inverter health based on multi-index weighted coupling and long short-term memory network (LSTM) time series modeling. Background Technology

[0002] In the field of photovoltaic power plant operation and maintenance, the inverter, as the core equipment for converting DC power to AC power, directly affects the overall power generation efficiency and economic benefits of the power plant due to its operational stability. Currently, the industry mainly has the following two types of technical solutions regarding inverter operation and maintenance strategies and health status assessments: The first type is the fault-response maintenance solution. Its core logic is to passively wait for a fault to trigger, and then perform maintenance afterward. This involves monitoring the inverter's operating parameters (such as output voltage, current, internal temperature, and fault codes) in real time. When a clear fault occurs in the inverter (such as overcurrent protection, communication interruption, or power slump) and triggers an alarm signal, maintenance personnel then arrange on-site repairs based on the alarm information. This type of solution has the following three major drawbacks: 1. Delayed fault response: The maintenance process can only be initiated after a fault occurs, and potential risks cannot be predicted in advance. If the fault occurs during peak electricity consumption periods or in severe weather such as heavy rain or snow, the failure to maintain the maintenance in time can lead to long-term shutdowns and significant power generation losses. 2. Disorganized operation and maintenance resource scheduling: The uncertainty of sudden failures makes it difficult to plan ahead for maintenance personnel, repair tools and spare parts. There may be insufficient resource allocation when multiple inverters fail at the same time, which will further prolong the failure handling cycle. 3. High labor costs: In order to reduce the risk of sudden failures, some power plants need to adopt a model of inspection one by one, and regularly arrange personnel to check the status of inverters on-site. For large power plants, the increased inspection frequency will lead to a significant increase in labor costs and poor operation and maintenance economy.

[0003] The second category is the inverter health level assessment and maintenance solution. Its core logic is to quantify the current health status and classify the equipment for maintenance. By selecting key operating indicators of the inverter (such as conversion efficiency, failure frequency, and running time), a health scoring model is constructed to calculate the real-time health score of the inverter. Based on the score, the health level is divided. For example, a score greater than 85 is considered healthy, a score greater than 65 but less than 85 is considered sub-healthy, and a score less than 65 is considered a warning. Maintenance personnel can adjust the inspection strategy according to the level. For example, the inspection frequency of healthy equipment is reduced, and equipment with a warning level is prioritized for repair or replacement to reduce labor costs and reduce sudden failures.

[0004] The core problem with the aforementioned second type of solution is the lack of health trend prediction capabilities. It can only assess the current health status of the inverter and cannot predict the trend of health changes over a future period. It is difficult to pinpoint the time point when the health level switches. For example, if an inverter's current health score is 87, which is considered healthy, it is impossible to know when it will drop below 85 and transition to a sub-healthy level. This makes it difficult to achieve forward-looking planning in operation and maintenance strategies. Due to the lack of prediction of future health status, operation and maintenance personnel can only formulate short-term strategies based on the current level. They cannot stockpile spare parts in advance or plan maintenance cycles, and the problem of passive operation and maintenance due to sudden deterioration of health status still exists. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a method and system for quantitative classification and state prediction of inverter health based on multi-index weighted coupling and LSTM time series modeling, which addresses the above-mentioned defects in the existing technology. The method adaptively determines the index weights based on the actual operation data of the power plant, predicts the future long-term health degradation trajectory based on the long short-term memory network model, and automatically calculates the remaining days for level switching. This provides maintenance personnel with a reference for long-term maintenance strategies, thereby achieving cost reduction and efficiency improvement.

[0006] The technical solution adopted by this invention to solve its technical problem is: On the one hand, the present invention provides a method for quantitative classification and state prediction of inverter health, including the following steps: S1: Obtain the operating data of the inverter to be evaluated and perform data preprocessing; S2: Based on the preprocessed data, calculate the scores of three main indicators, including power output, functional status and abnormal risk; then calculate the scores of eight sub-indicators under the three main indicators, including effective utilization hours ranking, output power and irradiance fitting index, inverter conversion efficiency, operating temperature and output power fitting index, three-phase voltage imbalance index, common anomaly index, communication anomaly index and anomaly frequency index. S3: Based on information entropy and / or dispersion, adaptively determine the weights of each main indicator and its subordinate sub-indicators; S4: Calculate the total health score based on the scores of each sub-indicator and their corresponding weights, and use historical running data to train the long short-term memory network model. Using the historical health index scores as input features, output the health prediction curve for the next preset number of days using a rolling prediction method. S5: Based on the health prediction curve and the health level threshold, automatically calculate the inverter's current health score and health level information, and predict the remaining health cycle from the current health level to the next level.

[0007] In step S2, the inverter's power output capability is quantified by effectively utilizing the hour ranking index and the output power and irradiance fitting degree index, thus constructing a two-dimensional quantitative model of the inverter's health and power output capability, as detailed below: Effective utilization hours ranking index: Obtain multi-day data of the inverter to be evaluated, divide the output power generation by the installed capacity of the photovoltaic modules to obtain the effective utilization hours of the inverter; within the same power station, rank all inverters to be evaluated according to their effective utilization hours, assign 100 points to the inverter ranked first, and calculate the scores of the remaining inverters according to the following formula: Where valid_hour represents the inverter's own effective utilization hours, and first_valid_hour represents the effective utilization hours of the top-ranked inverter within the same power station. By eliminating environmental interference through horizontal comparison, this indicator can accurately quantify the differences in actual power generation capacity among different inverters within the same power station, providing a core reference dimension for health assessment.

[0008] Output power vs. irradiance fitting index: Data on the output power of a single inverter over several days and the local irradiance were obtained. The data was divided into 3-day intervals, with data less than three days merged into the next unit. A fitting was performed separately for every 3 days of data. The resulting R² value was converted into a score using the following formula: When R² > 0.9, the score = 90 + (R² - 0.9) × 100; When 0.7 < R² ≤ 0.9, the score = 75 + (R² - 0.7) × 75; When 0.5 < R² ≤ 0.7, the score = 50 + (R² - 0.5) × 125; When R²≤0.5, the score = R²×100.

[0009] To improve the timeliness of the results, fitting results closer to the evaluation date are assigned higher weights. The weighting formula is as follows: Where α represents the attenuation coefficient and K is the number of units, the initial weight values ​​are first obtained through the attenuation coefficient. Then normalization is performed, which involves dividing the weight value by the sum of all weight values ​​to obtain the final weight. This calculation method makes the fit results more reflective of the inverter's most recent operating state, providing good timeliness.

[0010] In step S2, the inverter's functional status is quantified through three dimensions: inverter conversion efficiency, inverter operating temperature vs. output power fit, and inverter three-phase voltage imbalance. A multi-parameter evaluation model for inverter health and functional status, comprising these three sub-indicators, is constructed as follows: Inverter conversion efficiency metrics: First, obtain the conversion efficiency data recorded by the inverter itself. For inverters that do not record conversion efficiency, or whose conversion efficiency data contains more than a certain proportion of empty values, obtain the inverter's input power data and output power data. For data where the input power is greater than the rated output power, adjust the input power to be equal to the output power. Based on the processed power data, divide the output power by the input power to obtain the inverter's conversion efficiency at each moment. Finally, calculate the inverter's average conversion efficiency based on the conversion efficiency data of the inverter at each moment.

[0011] Using 95% as the threshold, inverters with a conversion efficiency greater than 95% are assigned a score of 100. The remaining inverters are calculated to have their conversion efficiency deviation score calculated according to the following logic: Wherein, convert_efficiency is the average conversion efficiency of the inverter. The difference between the conversion efficiency threshold (95%) and convert_efficiency is calculated. The deviation between convert_efficiency and the conversion efficiency threshold (95%) is further obtained. The conversion efficiency deviation is multiplied by 100 to get the conversion efficiency index deviation score after conversion. The deviation score is subtracted from 100 to get the conversion efficiency index score.

[0012] Inverter operating temperature vs. output power fit index: First, historical operating data of the inverter to be evaluated was collected. Data on internal temperature and output power during the period of sufficient sunlight from 9:00 AM to 6:00 PM local time were selected. The R² values ​​of both were obtained through regression analysis, and a piecewise calculation method was used to score the results based on the R² values. When R² > 0.9, the score = 90 + (R² - 0.9) × 100; When 0.7 < R² ≤ 0.9, the score = 75 + (R² - 0.7) × 75; When 0.5 < R² ≤ 0.7, the score = 50 + (R² - 0.5) × 125; When R²≤0.5, the score = R²×100.

[0013] By fitting the operating conditions in segments, this model can quantify the matching degree between the inverter's operating temperature and power output, and identify potential performance degradation risks in advance.

[0014] Inverter three-phase voltage imbalance index: The three-phase voltage imbalance of the inverter at various times is calculated using the following formula: Summarize the three-phase voltage imbalance at each time point and calculate the average imbalance of the inverter to be evaluated. Determine three imbalance thresholds: 1.3%, 2.0%, and 2.6%, and assign scores based on these thresholds. imbalance ≤ 1.3% 100 points will be awarded. 1.3% <imbalance≤ 2.0%: score = 100-(imbalance-1.3) / (2.0-1.3)×30; 2.0% < imbalance ≤ 2.6% score = 70-(imbalance-2.0) / (2.6-2.0)×40; imbalance > 2.6% score = 30.0 - ((imbalance - 2.6) / 2.6) × 30. In step S2, a three-dimensional anomaly risk assessment model is constructed, comprising scores for common anomaly indicators, communication anomaly indicators, and anomaly frequency indicators. This model quantifies the anomaly risk status of the inverter from different anomaly types, as detailed below: First, acquire the anomaly record data of the inverter to be evaluated during the evaluation period. Calculate the duration of each anomaly based on its start and end times. Only anomalies with a duration exceeding a certain threshold are considered valid anomalies. Based on this, obtain the communication anomaly duration and ordinary anomaly duration of the inverter to be evaluated during the evaluation period, and assign scores using the following calculation method: Common anomaly score = (Total runtime - Common anomaly duration) / Total runtime × 100; Communication anomaly score = (Total runtime - Communication anomaly duration) / Total runtime × 100; Anomaly frequency reflects the number of times anomalies occur in the inverter being evaluated. There's no need to filter valid anomalies based on their duration; a score is assigned solely based on the frequency of off-grid anomalies recorded in the original data. The anomaly frequency score is converted from the anomaly frequency ranking of inverters within the same power station, with lower frequencies receiving higher scores. The lowest frequency inverter receives 100 points, and the remaining inverters follow the following scoring rules: In step S3, the weights of each indicator are calculated based on the information entropy theory to achieve dynamic weight allocation. The steps are as follows: First, calculate the index scores of each inverter in the power station to be evaluated based on historical operating data. Then, based on human experience, divide the health of the inverters into three levels: healthy, sub-healthy, and warning. At least 20% of the inverters need to be classified into the corresponding health level, and ensure that all three levels are covered.

[0015] Then, using the scores of various health indicators of the inverter as input, a KNN model is built. The KNN model calculates the Euclidean distance between each inverter and the inverters whose health levels are manually assigned, and assigns each inverter to the most matching health level, thereby obtaining the health level information of each inverter in the power plant to be evaluated.

[0016] Finally, the information entropy of the health level and the information gain of each health indicator are calculated. The greater the information gain of an indicator, the stronger its ability to distinguish between different health levels, and it should be assigned a greater weight. The weighting formula is as follows: in, As the initial weights, The initial weight is calculated based on the ratio of the information gain of each indicator to the sum of the information gains of all indicators. The final weight value is then adjusted based on human experience. This weight calculation method can be used not only to calculate the weight value of each main indicator, but also to calculate the weight value of each sub-indicator under the main indicator.

[0017] In step S3, the weights of each indicator are calculated based on the degree of dispersion, and a calculation model is constructed, as follows: The greater the dispersion of a certain indicator score, the greater its influence on the health score, and thus a larger weight can be assigned to it. If it is difficult to classify the health level of the inverter, the weight can also be determined based on the dispersion of each indicator to achieve adaptive weight calculation, as shown in the following formula: First, calculate the standard deviation of the scores for each health indicator of each inverter under the power station to be evaluated. and mean The coefficient of variation (CV) is further obtained according to the formula. The initial weight value is determined by the ratio of the coefficient of variation of each indicator to the sum of the coefficients of variation of each indicator. Finally, it is adjusted according to human experience to obtain the final weight value. This weight value calculation method can be used not only to calculate the weight value of each main indicator, but also to calculate the weight value of each sub-indicator under the main indicator.

[0018] In steps S4 and S5, a model-specific Long Short-Term Memory (LSTM) prediction model is constructed to address the temporal correlation of inverter health status. During the data processing phase, the historical operating data of the power plant to be evaluated is categorized and organized by inverter brand and model. Health index scores and levels for corresponding time periods are calculated, and training and testing sets are created. An LSTM model is trained separately for each inverter model, using the index data from the past 30 days as input features. A rolling prediction method is employed, using the prediction result of the current day as the input for the next day's prediction, to predict the health index and total score for the next 300 days. Based on the prediction results, the remaining days from the current level to the next level are automatically calculated. For example, if an inverter currently scores 87 points (healthy level), it is predicted to drop to a sub-healthy level in 276 days and will not drop to a warning level within 300 days. This model enables the prediction of health status from current assessment to future trends, providing data support for proactive operation and maintenance.

[0019] On the other hand, the present invention also provides a system corresponding to the aforementioned method, comprising the following modules: The data acquisition module, including a distributed sensor unit and a device data interface, is used to collect real-time operating data of the inverter, and the device data interface is used to transmit data. The data storage module adopts a dual storage architecture of local storage and cloud database. The local cache module is used to temporarily store the collected real-time data to avoid data loss due to communication interruption. The cloud database is used to store structured data for a long time and supports the creation of indexes by device-time dimension to ensure data query efficiency. The processor module stores the computer's running programs and runs the inverter health assessment algorithm and inverter trend prediction algorithm. The operation and maintenance interaction module includes the operation and maintenance platform's web interface and mobile APP. The web interface allows power plant managers to view the operating status of all equipment, health reports, and health prediction trends, while the mobile APP allows operation and maintenance personnel to receive work orders, upload maintenance data, and provide feedback on processing results.

[0020] The beneficial effects of this invention are as follows: 1. Three major evaluation indicators were established, including power output indicators, abnormal risk indicators, and functional status indicators. This enabled a multi-indicator weighted evaluation system for inverter health, which accurately classifies inverter health levels and reflects the current working status of the inverter, providing maintenance personnel with short-term maintenance strategy references.

[0021] 2. By constructing an LSTM prediction model based on historical data, dynamic prediction of the daily health score of the inverter over a period of time, such as 300 days, can be achieved, clearly presenting the trajectory of the health score change and helping maintenance personnel to grasp the direction of the health status evolution in advance.

[0022] 3. Based on the prediction results of future health scores, the system automatically calculates the number of days it will take for an inverter to switch from its current level to the next level. For example, if an inverter currently scores 87 points, which is at the healthy level, it is expected to drop to the sub-healthy level in 276 days and will not drop to the warning level within 300 days. This provides maintenance personnel with a reference for long-term maintenance strategies, thereby reducing costs and increasing efficiency. Attached Figure Description

[0023] Figure 1 The flowchart shows the overall process of the inverter health quantification and state prediction method provided in the embodiments of the present invention.

[0024] Figure 2 This is a schematic diagram of the hierarchical structure of the three main indicators and eight sub-indicators provided in the embodiments of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0026] Example 1: Method Example like Figure 1 As shown, this embodiment provides a method for quantitative classification and state prediction of inverter health, including steps S1 to S5.

[0027] Step S1: Data Acquisition and Preprocessing Taking a photovoltaic power station as an example, the station has deployed 64 inverters of the same model. The raw operating data of each inverter was obtained, including: internal temperature time-series data, input power time-series data, output power time-series data, operational anomaly records, three-phase voltage time-series data, conversion efficiency time-series data, daily power generation time-series data, irradiance data, and the photovoltaic module installed capacity and rated output power corresponding to the inverter being evaluated.

[0028] Step S2: Calculation of multi-dimensional indicator scores Based on the preprocessed data, scores for eight sub-indicators under the three main indicators are calculated, with all sub-indicators using a percentage system. Calculating each indicator requires obtaining time-series data of the inverter over a specific period; in this embodiment, all indicators are calculated based on data from the past seven days.

[0029] (a) Power output indicators 1. Ranking indicators for effective utilization hours The effective utilization hours of the inverter to be evaluated within a certain period are calculated by dividing the daily power generation of the inverter to be evaluated during that period by the corresponding photovoltaic module installed capacity. Based on this method, the effective utilization hours of all inverters to be evaluated in the power plant under evaluation are calculated and ranked according to their values. The inverter ranked first is assigned 100 points, and the scores of the remaining inverters are converted using the following formula to obtain the final score for this indicator: Where valid_hour is the effective utilization hours of the inverter itself, and first_valid_hour is the effective utilization hours of the inverter ranked first in the same power station.

[0030] By eliminating the interference of environmental factors through horizontal comparison, this indicator can accurately quantify the differences in the actual power generation capacity of different inverters under the same power station, providing a core reference dimension for health assessment.

[0031] 2. Fit index between output power and irradiance The output power and irradiance data of the inverter to be evaluated are obtained at various times over 7 days. The data is divided into 3-day units, and data with less than 3 days is merged into the next unit. In this embodiment, the 7-day data is divided into a first unit of 3 days and a second unit of 4 days, as shown in the table below: After obtaining the cell division, the R² values ​​of the two time series data within each cell are calculated and then converted into corresponding cell fractions using the following formula: When R² > 0.9, the score = 90 + (R² - 0.9) × 100; When 0.7 < R² ≤ 0.9, the score = 75 + (R² - 0.7) × 75; When 0.5 < R² ≤ 0.7, the score = 50 + (R² - 0.5) × 125; When R²≤0.5, the score = R²×100.

[0032] To improve the timeliness of the results, fitting results closer to the evaluation date are assigned higher weights. The weighting formula is as follows: Where α represents the attenuation coefficient, and k is the number of units. The initial weight value is obtained through the attenuation coefficient. The initial weight values ​​are normalized by dividing each weight value by the sum of all weight values ​​to obtain the final weights. This calculation method makes the fit results more reflective of the inverter's most recent operating state, providing good timeliness.

[0033] In this embodiment, let the attenuation coefficient α be 0.6, then the weight values ​​of the two units are: The following formula can then be used to obtain the goodness-of-fit index score between the output power and irradiance of the inverter to be evaluated: Unit 1 R² score × 0.625 + Unit 2 R² score × 0.375 = score (ii) Functional status indicators 1. Inverter conversion efficiency indicators Obtain the past 7 days' conversion efficiency, input power, and output power data for the inverter to be evaluated. If the percentage of null values ​​in the conversion efficiency data exceeds 30%, it is considered invalid, and the conversion efficiency must be recalculated based on the input and output power. If high solar irradiance causes some input power to exceed the rated output power of the inverter to be evaluated, direct calculation may result in an underestimation of the conversion efficiency. Therefore, this portion of the input power is adjusted to be equal to the output power, i.e., the conversion efficiency of this portion of the data is set to 100%. After this preprocessing, the conversion efficiency is calculated again. If the percentage of null values ​​in the final conversion efficiency data is less than 30%, linear interpolation is used to fill the null values ​​to improve data quality.

[0034] Finally, based on the conversion efficiency data, the average conversion efficiency value of the inverter to be evaluated is obtained. Using 95% conversion efficiency as a threshold, the average conversion efficiency value is converted into a score in the following way to obtain the inverter conversion efficiency index score: Wherein, convert_efficiency is the average conversion efficiency of the inverter. The difference between the conversion efficiency threshold (95%) and convert_efficiency is calculated. The deviation between convert_efficiency and the conversion efficiency threshold (95%) is further obtained. The conversion efficiency deviation is multiplied by 100 to get the conversion efficiency index deviation score after conversion. The deviation score is subtracted from 100 to get the conversion efficiency index score.

[0035] 2. Operating temperature vs. output power fit index Obtain the past 7 days' operating temperature and output power time-series data of the inverter to be evaluated, and filter the data for the period from 9:00 AM to 6:00 PM local time, when solar radiation is sufficient and output power is effective. Fit the two time-series data points within this period to obtain the R² value. Assign scores to each segment according to the following formula to obtain the segmented fitting index score for the inverter's operating temperature and output power: When R² > 0.9, the score = 90 + (R² - 0.9) × 100; When 0.7 < R² ≤ 0.9, the score = 75 + (R² - 0.7) × 75; When 0.5 < R² ≤ 0.7, the score = 50 + (R² - 0.5) × 125; When R²≤0.5, the score = R²×100.

[0036] By fitting the operating conditions in segments, this model can quantify the matching degree between the inverter's operating temperature and power output, and identify potential performance degradation risks in advance.

[0037] 3. Three-phase voltage imbalance index Obtain the three-phase voltage data of the inverter to be evaluated for the past 7 days, and calculate the three-phase voltage imbalance at each time point using the following formula: Finally, the average unbalance of the inverter under evaluation is calculated based on the three-phase voltage unbalance at each time point. Three unbalance thresholds are determined: 1.3%, 2.0%, and 2.6%. Scores are then assigned based on these three thresholds. Imbalance ≤ 1.3% Assign a score of 100 points; 1.3% < imbalance ≤ 2.0% score = 100-(imbalance-1.3) / (2.0-1.3)×30; 2.0% < imbalance ≤ 2.6%: score = 70-(imbalance-2.0) / (2.6-2.0)×40; imbalance > 2.6% score = 30.0 - ((imbalance - 2.6) / 2.6) ×30.

[0038] (III) Abnormal Risk Indicators 1. Common anomaly indicators and communication anomaly indicators Communication anomalies refer to anomalies caused by interruptions in data communication transmission of the inverter under evaluation. General anomalies refer to anomalies caused by problems inherent to the inverter under evaluation, excluding grid anomalies and communication anomalies. For general anomaly and communication anomaly indicators, only valid anomalies lasting longer than 15 minutes are counted. For example, if an inverter under evaluation has the following 5 anomaly records within 7 days: According to the table above, the total duration of communication anomalies for this inverter is 20m 33s, and the duration of normal anomalies is 3h 53m 20s. Therefore, the scores for normal anomaly indicators and communication anomaly indicators of this inverter are: Common anomaly score = (Total runtime - Common anomaly duration) / Total runtime × 100 Communication anomaly score = (Total runtime - Communication anomaly duration) / Total runtime × 100 2. Abnormal frequency index The anomaly frequency score aims to count the total number of non-grid anomalies of all inverters under evaluation in the power plant under evaluation over the past 7 days. No valid anomalies need to be filtered. Each inverter under evaluation is ranked according to its anomaly frequency. The inverter with the lowest anomaly frequency in the power plant under evaluation is assigned 100 points. The remaining inverters are assigned the following scores: Divide 100 by the maximum rank (max_rank) and the abnormal frequency score can be calculated using the formula.

[0039] Here are some examples of scoring: A 0 1 100 B 2 2 83.33 C 3 3 66.67 D 3 3 66.67 E 6 4 50 F 7 5 33.33 G 10 6 16.67 Step S3: Adaptive weight allocation After calculating the scores of each indicator for all inverters in the power station according to the above calculation rules, the weights of the three main indicators and the weights of each sub-indicator under the main indicators can be determined based on the two calculation methods.

[0040] Method A: Calculate the weight of the main indicator based on information entropy Based on human experience, at least 20% of the inverters to be evaluated in the power station are divided into three levels: healthy, sub-healthy, and warning. The healthy level is characterized by excellent working condition and low inspection frequency, the sub-healthy level is characterized by good working condition and moderate inspection frequency, and the warning level is characterized by average working condition and slightly higher inspection frequency.

[0041] Next, using the scores of various indicators of the inverters as input and the health level as output, a KNN model is built. The KNN model calculates the Euclidean distance between other inverters in the power plant to be evaluated and the inverters classified by human experience, and classifies all inverters to be evaluated into the corresponding health level.

[0042] After classifying the inverters into health levels, the information entropy of each health level is calculated. For example, if a power plant has 64 inverters, after classification using the KNN model, 42 inverters are classified as healthy, 16 as sub-healthy, and 6 as warning. The formula for calculating the information entropy is as follows: For any indicator, the difference between its maximum and minimum scores is taken, and this difference is divided into three levels. The number of inverters to be evaluated at each health level within each level is counted, and the information entropy of each level of the indicator is calculated. For example, if the maximum score for the functional status indicator is 99.82 and the minimum score is 60.37, then there are three levels within the range of 99.82-60.37. Based on the above results, the information gain value of the functional status index is further obtained: Using the same calculation method, the information gain value of the power output indicator is 0.4278, and the information gain value of the abnormal risk indicator is 0.5267. The information gain ratio of different indicators can then be calculated using the following formula: in, As the initial weights, The initial weight is calculated based on the ratio of the information gain of each indicator to the sum of the information gains of all indicators. The final weight value is then adjusted based on human experience. This weight calculation method can be used not only to calculate the weight value of each main indicator, but also to calculate the weight value of each sub-indicator under the main indicator.

[0043] Functional status indicators 0.2079 17.9% 10% Power output index 0.4278 36.8% 40% Abnormal risk indicators 0.5267 45.3% 50% sum 1.1624 100% 100% Method B: Calculating sub-index weights based on dispersion The greater the dispersion of a certain calculation indicator, the greater its impact on the health score, and therefore it should be assigned a larger weight. If it is difficult to classify the health level of the inverter to be evaluated, the weight can be determined by the dispersion of the scores of each indicator, as shown in the following formula: For example, when calculating the weights of each sub-indicator under the functional status index, the following calculation results can be obtained: In this embodiment, the weights of the three main indicators are calculated using an information entropy-based method, and the weights of the sub-indicators are calculated using a dispersion-based method. The final results are shown in the table below: Step S4: Calculation of total health score and training of LSTM model After completing the weighting of the indicators, the total score can be obtained by substituting the weights and scores of each indicator. Historical operating data of the same model of inverters within the power plant to be evaluated are compiled, and health scores for each time period are calculated. Historical operating data from a past period is used to train the LSTM model. In this embodiment, the power plant to be evaluated has a total of 64 inverters of the same model. Data from the past two years of these 64 inverters is used to train the LSTM model. The health scores of each indicator over the past 30 days are used as input features, and the health scores of each indicator for the next day are used as output labels. A rolling prediction method is adopted, using the prediction result of the current day as the input for the prediction of the next day, to achieve health score prediction for the next 300 days and generate a future health prediction curve. Based on the predicted health scores, the number of days required for the inverter to switch from the current level to the next level can be automatically calculated.

[0044] Step S5: Results Output and Operation and Maintenance Decisions Output complete information about the inverter to be evaluated, including its current health score and health level. For example, an inverter currently scores 87, which is considered healthy. It is expected to decline to a sub-healthy level within 276 days and will not decline to a warning level within 300 days. If the weight calculation uses a discrete weighting method, and this method does not build a KNN health level classification model, then the health level can be classified based on a score threshold: greater than 85 is healthy, greater than 65 but less than 85 is sub-healthy, and less than 65 is a warning.

[0045] The health assessment information of the inverters to be evaluated can be used as a reference for the operation and maintenance strategies of photovoltaic power plant operation and maintenance personnel, such as inspection strategies and spare parts storage strategies. If the health score of the inverter is low, a reminder will be sent to the platform. For inverters that are about to enter the warning level within 30 days, a maintenance work order will be sent to the platform.

[0046] Example 2: System Example This embodiment provides a system corresponding to the method described in Embodiment 1 above, including the following modules: The data acquisition module, including a distributed sensor unit and a device data interface, is used to collect real-time operating data of the inverter, and the device data interface is used to transmit data. The data storage module adopts a dual storage architecture of local storage and cloud database. The local cache module is used to temporarily store the collected real-time data to avoid data loss due to communication interruption. The cloud database is used to store structured data for a long time and supports the creation of indexes by device-time dimension to ensure data query efficiency. The processor module stores the computer's running programs and runs the inverter health assessment algorithm and inverter trend prediction algorithm. The operation and maintenance interaction module includes the operation and maintenance platform's web interface and mobile APP. The web interface allows power plant managers to view the operating status of all equipment, health reports, and health prediction trends, while the mobile APP allows operation and maintenance personnel to receive work orders, upload maintenance data, and provide feedback on processing results.

[0047] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for quantitative classification and state prediction of inverter health, characterized in that, Includes the following steps: S1: Obtain the operating data of the inverter to be evaluated and perform data preprocessing; S2: Based on the preprocessed data, calculate the scores of three main indicators, including power output, functional status and abnormal risk; then calculate the scores of eight sub-indicators under the three main indicators, including effective utilization hours ranking, output power and irradiance fitting index, inverter conversion efficiency, operating temperature and output power fitting index, three-phase voltage imbalance index, common anomaly index, communication anomaly index and anomaly frequency index. S3: Based on information entropy and / or dispersion, adaptively determine the weights of each main indicator and its subordinate sub-indicators; S4: Calculate the total health score based on the scores of each sub-indicator and their corresponding weights, and use historical running data to train the long short-term memory network model. Using the historical health index scores as input features, output the health prediction curve for the next preset number of days using a rolling prediction method. S5: Based on the health prediction curve and the health level threshold, automatically calculate the inverter's current health score and health level information, and predict the remaining health cycle from the current health level to the next level.

2. The inverter health quantification classification and state prediction method according to claim 1, characterized in that, In step S2, the inverter's power output capability is quantified by effectively utilizing the hour ranking index and the output power and irradiance fitting index, thus constructing a two-dimensional quantitative model for the inverter's health and power output capability, as detailed below: Effective utilization hours ranking index: Obtain data from several days of the inverter to be evaluated, divide the output power generation by the installed capacity of the photovoltaic modules to obtain the effective utilization hours of the inverter; within the same power plant area, rank all inverters to be evaluated by their effective utilization hours, assign 100 points to the inverter ranked first, and calculate the scores of the remaining inverters according to the following formula: Where valid_hour is the effective utilization hours of the inverter itself, and first_valid_hour is the effective utilization hours of the inverter ranked first in the same power station. Output power vs. irradiance fitting index: Data on the output power of a single inverter over several days and the local irradiance are obtained. The data is divided into n-day intervals, with data from less than n days merged into the next unit. A fitting is performed separately for each n-day interval. The resulting R² value is converted into a score using the following formula: When R² > 0.9, the score = 90 + (R² - 0.9) × 100; When 0.7 < R² ≤ 0.9, the score = 75 + (R² - 0.7) × 75; When 0.5 < R² ≤ 0.7, the score = 50 + (R² - 0.5) × 125; When R²≤0.5, the score = R²×100; To improve the timeliness of the results, fitting results closer to the evaluation date are assigned higher weights. The weighting formula is as follows: Where α represents the attenuation coefficient and K is the number of units, the initial weight values ​​are first obtained through the attenuation coefficient. Then normalization is performed, which involves dividing the weight value by the sum of all weight values ​​to obtain the final weight. .

3. The inverter health quantification classification and state prediction method according to claim 1, characterized in that, In step S2, the inverter's functional status is quantified through three dimensions: inverter conversion efficiency, inverter operating temperature vs. output power fit, and inverter three-phase voltage imbalance. A multi-parameter evaluation model for inverter health and functional status, comprising these three sub-indicators, is constructed as follows: Inverter conversion efficiency metrics: First, the conversion efficiency data recorded by the inverter under evaluation is obtained. For inverters that do not record conversion efficiency or whose conversion efficiency data contains more than a certain proportion of empty values, the input power data and output power data of the inverter are obtained. For data where the input power is greater than the rated output power, the input power is adjusted to be equal to the output power. Based on the processed power data, the output power is divided by the input power to obtain the conversion efficiency of the inverter at each time. Finally, the average conversion efficiency of the inverter is further calculated based on the conversion efficiency data of the inverter under evaluation at each time. Using 95% as the threshold, inverters with a conversion efficiency greater than 95% are assigned a score of 100. The remaining inverters are calculated with the following logic to determine their conversion efficiency deviation score: Wherein, convert_efficiency is the average conversion efficiency of the inverter. The difference between the conversion efficiency threshold (95%) and convert_efficiency is calculated. The deviation between convert_efficiency and the conversion efficiency threshold (95%) is further obtained. The conversion efficiency deviation is multiplied by 100 to get the conversion efficiency index deviation score after conversion. The deviation score is subtracted from 100 to get the conversion efficiency index score. Inverter operating temperature vs. output power fit index: First, historical operating data of the inverter to be evaluated was collected. Data on internal temperature and output power during the period of sufficient sunlight from 9:00 AM to 6:00 PM local time were selected. The R² values ​​of both were obtained through regression analysis, and a piecewise calculation method was used to score the results based on the R² values. When R² > 0.9, the score = 90 + (R² - 0.9) × 100; When 0.7 < R² ≤ 0.9, the score = 75 + (R² - 0.7) × 75; When 0.5 < R² ≤ 0.7, the score = 50 + (R² - 0.5) × 125; When R²≤0.5, the score = R²×100; Inverter three-phase voltage imbalance index: The three-phase voltage imbalance of the inverter at various times is calculated using the following formula: Summarize the three-phase voltage imbalance at each time point and calculate the average imbalance of the inverter to be evaluated; determine three imbalance thresholds of 1.3%, 2.0%, and 2.6%, and assign scores based on these three imbalance thresholds: imbalance ≤ 1.3% 100 points will be awarded. 1.3% <imbalance≤ 2.0%: score = 100-(imbalance-1.3) / (2.0-1.3)×30; 2.0% < imbalance ≤ 2.6% score = 70-(imbalance-2.0) / (2.6-2.0)×40; imbalance > 2.6% score = 30.0 - ((imbalance - 2.6) / 2.6) × 30.

4. The inverter health quantification classification and state prediction method according to claim 1, characterized in that, In step S2, a three-dimensional anomaly risk assessment model is constructed, comprising scores for common anomaly indicators, communication anomaly indicators, and anomaly frequency indicators. This model quantifies the anomaly risk status of the inverter from different anomaly types, as detailed below: First, acquire the anomaly record data of the inverter to be evaluated during the evaluation period. Calculate the duration of each anomaly based on its start and end times. Anomalies with a duration exceeding a certain threshold are considered valid anomalies. Based on this, obtain the communication anomaly duration and ordinary anomaly duration of the inverter to be evaluated during the evaluation period, and assign scores using the following calculation method: Common anomaly score = (Total runtime - Common anomaly duration) / Total runtime × 100; Communication anomaly score = (Total runtime - Communication anomaly duration) / Total runtime × 100; The abnormal frequency index score is converted from the abnormal frequency ranking of inverters within the same power station. The lower the frequency, the higher the score, with the lowest frequency assigned 100 points. The remaining inverters are assigned scores according to the following rules: 。 5. The inverter health quantification classification and state prediction method according to claim 1, characterized in that, In step S3, the weights of each indicator are calculated based on the information entropy theory to achieve dynamic weight allocation. The steps are as follows: First, calculate the index scores of each inverter in the power station to be evaluated based on historical operating data. Then, based on human experience, divide the health of the inverters into three levels: healthy, sub-healthy, and warning. At least 20% of the inverters will be assigned to the corresponding health level, and all three levels will be covered. Then, using the scores of each health indicator of the inverter as input, a KNN model is built. The KNN model calculates the Euclidean distance between each inverter and the inverter with manually assigned health levels, and assigns each inverter to the most matching health level, thereby obtaining the health level information of each inverter in the power station to be evaluated. Finally, the information entropy of the health level and the information gain of each health indicator are calculated. The greater the information gain of an indicator, the stronger its ability to distinguish between different health levels, and it should be assigned a greater weight. The weighting formula is as follows: in, As the initial weights, The information gain of a certain indicator is calculated by first determining the initial weight based on the ratio of the information gain of each indicator to the sum of the information gains of all indicators, and then adjusting the weight value based on human experience to obtain the final weight value.

6. The inverter health quantification classification and state prediction method according to claim 1, characterized in that, In step S3, the weights of each indicator are calculated based on the degree of dispersion, and a calculation model is constructed, as follows: The weights are determined based on the dispersion of each indicator to achieve adaptive weight calculation, as shown in the following formula: First, calculate the standard deviation of the scores for each health indicator of each inverter under the power station to be evaluated. and mean Then, the coefficient of variation (CV) is obtained according to the formula. The initial weight value is determined based on the ratio of the coefficient of variation of each indicator to the sum of the coefficients of variation of each indicator. Finally, the weight value is adjusted based on human experience to obtain the final weight value.

7. The inverter health quantification classification and state prediction method according to claim 1, characterized in that, Steps S4 and S5 are used to construct model-specific long short-term memory network prediction models based on the temporal correlation of inverter health status, as detailed below: During the data processing phase, the historical operating data of the power plants to be evaluated are sorted and organized according to the inverter brand and model. The health index scores and levels for the corresponding time periods are calculated, and training sets and test sets are divided. LSTM models are trained separately for each inverter model. The index data of the past m days are used as input features. A rolling prediction method is adopted, and the prediction results of the current day are used as the prediction input for the next day to achieve the prediction of the health index and total score for the next x days. Based on the prediction results, the remaining number of days from the current level to the next level is automatically calculated.

8. The system corresponding to the method described in any one of claims 1-7, characterized in that, The system includes the following modules: The data acquisition module, including a distributed sensor unit and a device data interface, is used to collect real-time operating data of the inverter, and the device data interface is used to transmit data. The data storage module adopts a dual storage architecture of local storage and cloud database. The local cache module is used to temporarily store the collected real-time data to avoid data loss due to communication interruption. The cloud database is used to store structured data for a long time and supports the creation of indexes by device-time dimension to ensure data query efficiency. The processor module stores the computer's running programs and runs the inverter health assessment algorithm and inverter trend prediction algorithm. The operation and maintenance interaction module includes the operation and maintenance platform's web interface and mobile APP. The web interface allows power plant managers to view the operating status of all equipment, health reports, and health prediction trends, while the mobile APP allows operation and maintenance personnel to receive work orders, upload maintenance data, and provide feedback on processing results.

Images