A photovoltaic operation and maintenance management method and system based on edge computing

Abstract

The present application relates to photovoltaic power station management technical field, especially in kind of photovoltaic operation and maintenance management method and system based on edge computing, including: obtaining multi-source data through sensor, pre-processing multi-source data through edge computing node;Using auto-encoder to fuse the pre-processed multi-source data, mining deep correlation characteristics, obtaining comprehensive feature dataset;Based on comprehensive feature dataset and real-time data, through dynamic adaptive model, obtain future power generation prediction value;Compare future power generation prediction value with actual power generation result, according to the comparison result, judge whether there is fault point and combine knowledge graph and historical record to locate potential fault point, get fault diagnosis result;According to the future power generation prediction value and fault diagnosis result, comprehensively consider the actual operation of power station and resource allocation, generate equipment maintenance plan and power generation scheduling scheme.The scheme improves the operation and maintenance efficiency of power station through edge computing and multi-technology fusion.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power plant management technology, and in particular to a photovoltaic operation and maintenance management method and system based on edge computing. Background Technology

[0002] In the field of photovoltaic power plant operation and maintenance, traditional management models rely on manual inspections and experience-based judgment, which have many drawbacks. Manual inspections are inefficient, have limited coverage, make it difficult to grasp the real-time status of all equipment in the power plant, and are delayed in the detection and handling of faults. In terms of data management, multi-source data (such as equipment operation, meteorological, and grid dispatch data) are scattered and have inconsistent formats, lacking efficient integration and analysis methods. This makes it difficult to fully explore the value of the data, limits the accuracy of power generation prediction, and fails to provide accurate basis for power generation dispatch.

[0003] Fault diagnosis relies primarily on manual experience, lacking intelligent analytical tools. This leads to inaccurate fault location, lengthy troubleshooting times, and a tendency for faults to escalate, impacting power plant efficiency and equipment lifespan. Furthermore, operation and maintenance strategies lack scientific quantitative basis. Equipment maintenance plans and power dispatch schemes are often formulated based on experience, resulting in unreasonable resource allocation. This can lead to increased costs due to over-maintenance or power generation losses due to untimely maintenance or improper dispatching, hindering the achievement of efficient and intelligent operation and maintenance for photovoltaic power plants. Summary of the Invention

[0004] This invention integrates edge computing with multiple technologies to achieve intelligent monitoring of photovoltaic power plant equipment, accurate fault diagnosis, and optimization of operation and maintenance and power generation scheduling, thereby improving the operation and maintenance efficiency and power generation benefits of the power plant.

[0005] The technical solution proposed in this invention is: a photovoltaic operation and maintenance management method based on edge computing, the method comprising:

[0006] Multi-source data is obtained through sensors and preprocessed by edge computing nodes;

[0007] By using an autoencoder to fuse preprocessed multi-source data, deep correlation features are mined to obtain a comprehensive feature dataset;

[0008] Based on a comprehensive feature dataset and real-time data, a dynamic adaptive model is used to obtain the predicted value of future power generation.

[0009] By comparing the predicted future power generation with the actual power generation results, the existence of fault points can be determined based on the comparison results. By combining the knowledge graph and historical records, potential fault points can be located, and fault diagnosis results can be obtained.

[0010] Based on the projected future power generation and fault diagnosis results, and taking into account the actual operation of the power plant and resource allocation, an equipment maintenance plan and power generation dispatch scheme are generated.

[0011] Preferably, the specific process for preprocessing the multi-source data is as follows:

[0012] Outliers exceeding physical thresholds in photovoltaic module temperature data are removed. Outliers are filtered using statistical criteria. Temperature anomalies are identified by combining regional average temperature and duration. The temperature change rate of adjacent modules is calculated and correlated with branch current to mark suspected hotspot areas. The normal voltage range is defined, and voltage values ​​exceeding the range are replaced with the historical average. The validity of power data is verified by the deviation between theoretical and measured power. High-frequency meteorological data undergoes time-dimensional mean smoothing. For missing data, linear or spatial interpolation is used to fill in the missing data according to the duration of the missing data. Temperature thresholds for key components are set, and equipment warnings and fault states are marked. The original equipment status codes are mapped to standardized three-level status codes and associated with fault descriptions. Power generation data is unified in time dimension, and abnormal data during maintenance periods are removed. Maintenance records are classified and coded, and an index associated with equipment IDs is established.

[0013] Preferably, the specific process for obtaining the comprehensive feature dataset is as follows:

[0014] Based on a unified timestamp, multi-source data output from edge computing nodes is resampled to a consistent time granularity to establish an association matrix. Continuous data is standardized to eliminate dimensions, while discrete data undergoes one-hot encoding transformation. An autoencoder model is constructed to compress high-dimensional input feature vectors into a latent space of a set dimension. Multi-source data fusion is achieved by minimizing reconstruction error through training, obtaining the fused latent features. Based on the latent features, component aging index, correlation features, and fault precursor features are calculated to obtain multiple types of explicit features. The latent features and explicit features are concatenated to form an initial comprehensive feature vector. Key features are selected through correlation analysis to construct a comprehensive feature dataset.

[0015] Preferably, the specific calculation process for the predicted future power generation is as follows:

[0016] Based on a comprehensive feature dataset, this method expands upon temporal features with cross-features of the environment and equipment, and performs dimensionless processing on the features to form fixed-dimensional predictive feature vectors. A main model and sub-model architecture is adopted. The main model captures long-term temporal dependencies of the data through an LSTM network, while sub-models are trained for different weather types, automatically switching to adapt to different sub-models based on real-time meteorological data. Based on real-time power deviations, the weights of the main model are dynamically updated using an online learning algorithm. A weighted fusion strategy is employed, combining the prediction results of the main model and corresponding sub-models to calculate future power generation predictions at different time scales.

[0017] Preferably, the specific process for determining whether the fault point exists is as follows:

[0018] By comparing the predicted future power generation with the actual power generation results, the absolute deviation and relative deviation rate are calculated. When the relative deviation rate reaches the fault investigation trigger threshold, the next step of analysis is initiated. The deviation between the actual and predicted values ​​of real-time illumination and ambient temperature is introduced, and the power generation deviation is corrected by the influence coefficient. It is then determined whether the corrected deviation still exceeds the equipment anomaly judgment threshold. If it does, a preliminary judgment is made that there is a possibility of a fault. Based on real-time equipment operation data, standardized anomaly scores are calculated for component temperature and current. The anomaly level is assigned according to the equipment status code to confirm whether there are significant abnormal parameters. If so, the existence of the fault is further verified.

[0019] Preferably, the specific process for obtaining the fault diagnosis result is as follows:

[0020] Based on the knowledge graph, the correlation strength between the identified abnormal parameters and each fault type is calculated. By traversing the network and accumulating the contribution scores of abnormal parameters to different faults, several potential fault candidates with the highest comprehensive scores are selected. Historical fault maintenance records are retrieved to extract typical parameter features of similar faults. The parameter similarity between the current potential fault candidate and historical faults is calculated. When the similarity reaches the confirmation threshold, the corresponding fault type is locked. Combining the equipment entities associated with the fault type, the specific equipment ID corresponding to the abnormal parameters is matched to identify the specific fault point and form a diagnostic result that includes the fault point and associated abnormal parameters.

[0021] Preferably, the specific generation process of the equipment maintenance plan and the power generation dispatch scheme is as follows:

[0022] Based on the fault impact and propagation risk in the fault diagnosis results, a fault handling priority index is calculated, and faults are classified into different priority levels according to the index. Referring to historical operation and maintenance resource utilization rates, dynamic timeliness standards for handling faults of each priority are set, and planned handling time periods for different faults are determined. The number of spare parts, tools, and personnel required for fault handling is statistically analyzed, and the existing resource inventory and availability are compared to determine whether resources meet planned requirements. If resources are insufficient, the external resource allocation process is initiated and the handling time is adjusted. Fault maintenance tasks are integrated with preventative maintenance to form an equipment maintenance plan. Internal constraints of the power plant, external grid requirements, and fault impacts are clarified to determine the boundary range of power generation output. Based on the predicted power generation value, the predicted power is corrected by combining the fault impact coefficient to obtain the basic output reference value for each time period, and reserve reserve capacity according to a fixed proportion of the basic output reference value.

[0023] The present invention also provides a photovoltaic operation and maintenance management method based on edge computing, wherein the system is used to execute the photovoltaic operation and maintenance management method based on edge computing.

[0024] The present invention also provides a computer-readable storage medium storing a computer program, which is executed by a processor to implement the aforementioned photovoltaic operation and maintenance management method based on edge computing.

[0025] The beneficial effects of this invention are:

[0026] 1. By preprocessing photovoltaic multi-source data (equipment operation, weather, historical maintenance, etc.) at the edge, invalid and abnormal data can be quickly filtered out, reducing data transmission bandwidth pressure. At the same time, with the help of preliminary data fusion at the edge nodes, preliminary correlation of multi-dimensional data can be achieved, laying the foundation for subsequent in-depth analysis and feature mining in the cloud, making the data link from collection to analysis more efficient, ensuring data quality and timeliness of analysis, and providing reliable data support for subsequent power generation prediction, fault diagnosis, etc.

[0027] 2. By utilizing the constructed "master model and sub-model" power generation prediction architecture and combining multi-source fusion data, it is possible to accurately capture the temporal variation patterns of photovoltaic power generation and influencing factors such as weather, and achieve accurate power generation prediction at multiple time scales; the fault diagnosis method based on knowledge graph and abnormal parameter analysis can quickly associate abnormal parameters with potential fault types, accurately locate fault points and provide confidence levels, greatly improving the efficiency and accuracy of fault diagnosis, helping to discover potential equipment problems in advance and reduce the impact of faults on power generation.

[0028] 3. Based on power generation forecast results and fault diagnosis information, and taking into account the power plant's operational constraints and external dispatch requirements, the formulated equipment maintenance plan can reasonably arrange maintenance time and resources, avoid repeated downtime, and ensure stable equipment operation. The power generation dispatch scheme balances the predicted power with grid constraints and reserves reserve capacity, which not only meets the grid's acceptance needs but also flexibly adjusts the output in case of emergencies such as faults. This achieves coordinated optimization of operation and maintenance and power generation dispatch, effectively improving the overall operating efficiency and power generation benefits of the photovoltaic power plant. Attached Figure Description

[0029] Figure 1 This is a flowchart of a photovoltaic operation and maintenance management method based on edge computing according to the present invention;

[0030] Figure 2 This is a flowchart illustrating the management process of a photovoltaic operation and maintenance management method based on edge computing according to the present invention. Detailed Implementation

[0031] The following description is intended to disclose the present invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art. The basic principles of the invention defined in the following description can be applied to other embodiments, modifications, improvements, equivalents, and other technical solutions that do not depart from the spirit and scope of the invention.

[0032] It is understood that the term "a" should be understood as "at least one" or "one or more," that is, in one embodiment, the number of an element can be one, while in another embodiment, the number of the element can be multiple, and the term "a" should not be understood as a limitation on the number.

[0033] like Figure 1 and Figure 2 As shown, the sensor deployment and data synchronization architecture were first established. For photovoltaic module data acquisition, one distributed infrared sensor was configured for every 20 modules, installed 30cm above the module frame, collecting module surface temperature data at a frequency of 1 minute. Simultaneously, voltage sensors were deployed at both ends of the module string, collecting the open-circuit voltage and operating voltage of the module string at a frequency of 30 seconds. Current sensors were installed in the combiner box, monitoring branch current changes at a frequency of 10 seconds. Meteorological data acquisition relied on two integrated meteorological monitoring stations deployed at the power station center, covering the east and west areas respectively, collecting data on light intensity, ambient temperature, relative humidity, wind speed, and wind direction at a frequency of 30 seconds. In addition, an atmospheric pressure sensor (monitoring range 800-1100 hPa) and a precipitation sensor (accuracy 0.1mm) were also installed.

[0034] In terms of equipment communication, communication sampling is established with the inverter and combiner box via the Modbus-RTU protocol every 5 seconds. In addition to basic parameters, key data such as the inverter's IGBT module temperature, cooling fan operating status, and DC side voltage are also collected. The high-voltage side current, voltage, and oil temperature data of the transformer substation are collected at a frequency of 1 minute using the RS485 interface. At the data access level, AQI data (including PM2.5 and PM10 concentrations) from air quality monitoring stations within a 5km radius and load curves from the power grid dispatch system (updated every hour) are acquired simultaneously. At the same time, the 2-hour precipitation probability forecast data from meteorological radar within a 10km radius (updated every 15 minutes) is also accessed. The daily power generation and power generation efficiency time-series data for the past two years are extracted from the database, and maintenance data such as component cleaning records, equipment maintenance work orders, and fault handling logs are supplemented by quarterly trend baseline modeling.

[0035] The NTP protocol is used to calibrate edge nodes with the standard time source daily at 2:00 AM, ensuring that the timestamp error of each sensor's data is within 1 second. Edge nodes are configured with a circular buffer, reserving 12 hours of raw data cache space. When storage overflows, historical data older than 48 hours is automatically overwritten. A network outage retransmission mechanism is designed; after network recovery, cached data is retransmitted in timestamp order, with priority transmission queues set for key parameters such as voltage and current. Through the above deployment and operation, multi-source raw data covering the following categories is ultimately obtained: equipment status data such as photovoltaic module temperature, module string voltage, and branch current; meteorological data such as irradiance, ambient temperature, humidity, wind speed, wind direction, atmospheric pressure, and precipitation; equipment operation data such as inverter output power, IGBT temperature, and high-voltage side parameters of the transformer substation; environmental and dispatch data such as AQI, grid load curves, and precipitation probability forecasts; and historical and management data such as historical power generation and operation and maintenance records.

[0036] Based on this multi-source raw data, the temperature correction calculation for photovoltaic modules is performed, using the following formula: ,in The actual operating temperature of the component (°C). The ambient temperature is (°C). The nominal operating temperature of the component is (°C, take 45°C). Light intensity ( ), For component conversion efficiency (dimensionless, , This represents the real-time output power (W). The area of ​​a single component ( Take 2.8 ), 800 is the unit .

[0037] Based on the temperature correction results, the power generation efficiency is calculated accurately using the following formula: ,here The inverter's AC output power (kW). The total area of ​​the photovoltaic array ( ), The overall power generation efficiency of the power plant (dimensionless), with 1000 as the unit. .

[0038] After the efficiency calculation is completed, data integrity is verified using the formula. ,in This represents the actual number of valid data entries. The theoretical number of data points to be collected is based on... Different numerical ranges ( The test result was deemed satisfactory. Initiate linear interpolation completion; (Trigger sensor fault alarms and generate maintenance work orders) to determine data quality status.

[0039] Taking a 10MW photovoltaic power station as an example, the power station uses 18,519 540W monocrystalline silicon modules, with a total array area of ​​approximately 30,000 square meters. It is equipped with 20 500kW centralized inverters. At 10:30 AM on July 15, 2023, the collected multi-source raw data included meteorological data such as ambient temperature. ℃, light intensity Atmospheric pressure 1013 hPa, PM2.5 concentration 25 Equipment status data includes the real-time output power of individual components. W, open-circuit voltage of the inverter string is 380V, operating voltage is 350V, and branch current is 8.5A; equipment operating data includes inverter AC output power. kW, IGBT module temperature 58℃, transformer high voltage side current 180A.

[0040] Based on these raw data, the following calculations were performed:

[0041] Component conversion efficiency: based on .

[0042] Component actual operating temperature: based on ℃.

[0043] Power plant power generation efficiency: (i.e., 16%).

[0044] Of the multi-source raw data collected that day, 926 infrared sensors monitored component temperatures ranging from 25 to 62°C, and two weather stations recorded light intensities ranging from 0 to 1050 ppm. (Average 586) The ambient temperature ranged from 22-36℃ (average 28.7℃), wind speed from 0.3-4.2m / s (average 1.8m / s), and atmospheric pressure fluctuated between 1010-1015hPa. There was no precipitation throughout the day. The inverter output power ranged from 0-500kW (average 286kW), the IGBT module temperature ranged from 50-65℃, and the cooling fans were all operating normally. (Daily data integrity verification coefficient) At 14:23, inverter #8 generated three abnormal current data points due to a momentary communication interruption, which have been recovered through local cache retransmission. Under this configuration, a single edge node processes approximately 1.4GB of multi-source raw data per day. After compression using the LZ77 algorithm, the network transmission volume is reduced to 260MB, and the raw data storage requirement for 30 days is approximately 42GB.

[0045] The edge computing node data preprocessing stage takes the collected multi-source raw data as input. The first step is to process the photovoltaic module temperature data, which includes raw temperature data (unit: °C, timestamp interval of 1 minute) collected by 926 infrared sensors and the corresponding current data (unit: A) of the module strings in the corresponding area. During processing, outlier cleaning is performed first, with a physical threshold range set. ,in Representing the The collected photovoltaic module temperature data were processed, and data exceeding the range were removed. Then, the data was analyzed using 3... The criterion for filtering outoutpoints is to calculate the daily average temperature of a single component. and standard deviation , will satisfy The data points are labeled as outliers, here It is the average value of the temperature data collected by this component on that day. It is the standard deviation of the temperature data collected by the component on that day; then, temperature anomaly detection is performed. ( (The average temperature of 20 modules in the same area) and duration When the temperature is abnormal, it is determined to be an anomaly. Finally, hot spot feature identification is performed, and the temperature change rate is calculated using the ratio of the temperature difference between adjacent components to the distance between components (rather than the ratio of temperature to area), as shown in the formula. ,in This is the temperature data (unit: °C) of the next component in an adjacent module. The temperature data (unit: °C) of the preceding component in an adjacent module is given, and L is the distance between adjacent components (unit: m, taken as 1.5m), with the units uniformly set to °C / m. When ℃ / m (i.e., the rate of temperature change corresponding to a 1.5m spacing when the temperature difference between adjacent components is >5℃) and this is continued for 2 sampling periods, while the branch current is... ( When the average current of the same string is used, it is marked as a suspected hot spot, and the output is the temperature sequence after cleaning (including timestamp), the abnormal temperature record table (component ID, abnormal time period, temperature difference value) and the list of suspected hot spot areas (coordinates, temperature change rate, current deviation rate).

[0046] After processing the photovoltaic module temperature data, the process shifts to electrical parameter standardization. The inputs include raw electrical data such as module string voltage (V), branch current (A), and inverter output power (kW). Voltage filtering is performed first to define the normal range. ( The open circuit voltage is 380V. (Represents the collected component string voltage data); when it exceeds the range, use... Calculate the substitution value, where , , These represent the component string voltage data collected at the three, two, and one time points prior to the current data point; then, power curve verification is performed, and theoretical power is calculated. (Unit: W, 0.95 is the system efficiency coefficient) It is the total area of ​​the photovoltaic array. Light intensity, For component conversion efficiency), at this time Dimensions are ×W / =W, multiplied by the dimensionless After considering the system efficiency coefficient, the theoretical power dimension is W, which is different from the measured power ( This represents the actual measured output power of the inverter, in kW (multiplied by 1000 for conversion to W during calculation). The dimensions are consistent. The system triggers a secondary verification, outputting filtered voltage / current timing data, power verification results (including pass / abnormal markers), and a theoretical power comparison table.

[0047] Next, the meteorological data was spatiotemporally aligned, with the input being the illumination (W / m²) collected from two weather stations. Data such as temperature (°C), wind speed (m / s), etc. (timestamped every 30 seconds). Smoothing is performed first during processing. Take the 5-minute average of every 10 data points at 30-second intervals, where Indicates the first Illuminance data collected at 30-second intervals (unit: W / ) Summing and dividing by the number of data points (10) yields the mean. The dimension remains W / Next, fill in the missing values; for short missing values ​​(≤10 minutes), use linear interpolation. , The light intensity data is from the time point preceding the missing data point (unit: W / ). ), It is the first after missing data Irradiance data at various time points (unit: W / ) The dimensions of the interpolation results should be kept consistent; for long missing data (>10 minutes), spatial interpolation should be used. ,in , Light intensity data collected from two adjacent weather stations (unit: W / ) ), , The distance from the current weather station to the two adjacent stations to the east and west (unit: km), with molecular weight in W / ×km+W / ×km= The denominator has the dimension of km, and the overall dimension is W / It is consistent with the light intensity unit and outputs a spatiotemporally aligned meteorological dataset (with a unified timestamp) and a data integrity report (missing rate, interpolation method).

[0048] After meteorological data processing is completed, equipment status data encoding is performed. The input includes raw data such as inverter IGBT temperature (°C), transformer oil temperature (°C), and equipment status codes. During processing, temperature thresholds are monitored first, including IGBT temperature. Time-based warning, >90℃ fault indication, transformer oil temperature A warning is indicated when the temperature is below 70℃, and a fault is indicated when the temperature is above 70℃. This represents the temperature data of the inverter's IGBT module. The data represents the transformer oil temperature; then the status code is converted, mapping the original equipment status code to a three-level code of "0-normal / 1-warning / 2-fault" and associating it with the fault code description, outputting a standardized equipment status table (including warning / fault time) and a temperature trend chart of key components.

[0049] Finally, historical and operational data are processed. The input consists of time-series data of daily power generation (kWh) and power generation efficiency (%) for the past two years, as well as module cleaning / repair records. During processing, the power generation data is first aligned, uniformly converted to Beijing time hourly data, and non-natural low values ​​caused by maintenance are removed. Then, the operational records are encoded, categorized into "1-Module Cleaning / 2-Equipment Repair / 3-Fault Handling", and an index associated with the equipment ID is established. The output is a structured historical database (time-power generation-efficiency association table) and an operational event log (including processing time and responsible person).

[0050] Taking data from a power station on July 15, 2023 as an example, in the photovoltaic module temperature processing, the temperature of module #A123 at 10:30 was... ℃, the synchronous temperature of the adjacent #A124 module ℃, component spacing Temperature change rate And it lasts for 2 sampling cycles (2 minutes), while the current in the branch where the component is located is... =7.2A, average current in series =8.5A, It was marked as a hot spot; during voltage processing, component string #B45 appeared at 14:20. =400V instantaneous value, open circuit voltage =380V, the upper limit of the normal range is 1.05×380=399V, which is outside the range. Use the voltage values ​​from the first 3 time points. =350V =352V =354V Calculation substitution value V; During power verification, the total area of ​​the photovoltaic array... =30000 Light intensity =586W / Component conversion efficiency =0.1786, theoretical power W, measured power =286kW=286000W, deviation This triggers a secondary verification (Note: Due to the significant difference between the measured and theoretical power in the example, data acquisition errors or parameter input errors may exist in actual scenarios, requiring secondary verification for confirmation); In meteorological interpolation, data for the eastern regional meteorological station from 11:00 to 11:15 is missing, and the distance from this station to the western station... =1km, distance to East Station (backup) =1km, West Station concurrent sunlight =580W / Simultaneous illumination at East Station =586W / Interpolation results W / The final output includes 127 valid temperature data points (excluding 3 outliers exceeding the -40℃ to 85℃ range), 8 suspected hot spot areas, and a standardized electrical parameter table (voltage deviation rate). The dataset includes 2%, a complete meteorological dataset (missing rate 0.3%), a status report of the power secondary verification triggered by one of the 20 inverters, and a preprocessed set of structured operation and maintenance records.

[0051] The cloud-based data fusion and feature mining process takes effective data output from edge computing nodes as input, including the temperature sequence of cleaned photovoltaic modules, standardized electrical parameters such as voltage and current, spatiotemporally aligned meteorological data (sunlight, temperature and humidity, etc.), equipment status codes (normal / early warning / fault), and structured operation and maintenance records (cleaning and maintenance records).

[0052] First, multi-source data spatiotemporal consistency calibration is performed. Based on a unified NTP timestamp (accurate to the second), data from different acquisition frequencies are resampled to a 1-minute time dimension. This is followed by calibration of photovoltaic module temperature. (°C), light intensity (W / Continuous data such as ambient humidity H (%) are required. Taking photovoltaic module temperature data as an example, when calculating the standardized value of the module temperature at a certain moment, it is necessary to first obtain the temperature data sequence of the module for the past 7 days (or all available data if the data is less than 7 days). ( At a sampling frequency of 1 minute, there are a total of 7 × 24 × 60 = 10080 data points over 7 days.

[0053] Calculate the mean: the average of the last 7 days The calculation formula is The unit is °C. This average reflects the average component temperature over the past 7 days.

[0054] Calculate the standard deviation: Standard deviation The formula used to measure the dispersion of data is: The unit is also ℃. The larger the standard deviation, the more drastic the fluctuation of the data.

[0055] The above calculations yielded... and Substituting into the Z-score standardization formula to eliminate dimensional differences, the formula is: Where x is the original data value, This is the 7-day moving average of this parameter. The moving standard deviation over the past 7 days is used, and the processed data are dimensionless standardized values. For discrete data such as equipment status codes, one-hot encoding is used to convert them into binary vectors. For example, the inverter's "normal / warning / fault" status corresponds to [1,0,0] / [0,1,0] / [0,0,1], ensuring that different types of data can participate in the model calculation together.

[0056] Next, a stacked self-encoder fusion model is constructed. The input layer contains a 128-dimensional feature vector (covering 20 electrical parameters: voltage, current, power, etc.; 8 meteorological factors: light intensity, temperature, humidity, wind speed, etc.; 16 equipment status variables: IGBT temperature, oil temperature, etc.; and 84 historical operation and maintenance features: cleaning cycle, number of failures, etc.). This is achieved through a 3-layer encoder. Compress high-dimensional data into a 64-dimensional latent space:

[0057] First-level encoding: ,in It is a 128×96 weight matrix. A 96-dimensional bias vector

[0058] Second layer encoding: ,in It is a 96×80 weight matrix. 80-dimensional bias vector

[0059] Third-level encoding: ,in It is an 80×64 weight matrix. A 64-dimensional bias vector ( (Use ReLU activation function to ensure non-negativity of output)

[0060] decoder Reconstructing input data using a symmetrical structure: The training objective is to minimize the reconstruction error. ( Given the number of samples, the Adam optimizer (learning rate 0.001) is used for iterative training for 500 epochs until the error converges to a stable range. 0.02).

[0061] In the feature mining stage, based on the fused 32-dimensional latent features z, the following deep features are calculated in focus:

[0062] Component aging characteristic index (Dimensionless, range 0-1, larger values ​​indicate more severe aging): ,in: The monthly degradation rate of component conversion efficiency (% / month). ; The average monthly operating temperature of the component (°C). =25℃ is the standard test temperature The monthly temperature standard deviation (°C) The standardized value of the monthly average relative humidity (dimensionless). =0.4、 =0.3、 =0.3 is the weighting coefficient (the sum is 1).

[0063] Environment-efficiency correlation characteristics:

[0064] Humidity Influence Coefficient (Dimensionless, H is the percentage of ambient humidity); Temperature correction efficiency ,in Temperature correction factor ( (Ambient temperature in °C) This refers to the overall power generation efficiency of the power plant.

[0065] Precursor features of failure: By calculating the cosine similarity between the latent feature z and historical failure samples. (Dimensionless, range -1 to 1), when similarity > 0.8, it is marked as high risk of failure.

[0066] In the feature integration stage, the 32-dimensional latent features are concatenated with the above three types of explicit features to form a 40-dimensional comprehensive feature vector, which is then analyzed using the Pearson correlation coefficient. Filter key features (retain) (Features), and finally construct a comprehensive feature dataset.

[0067] Taking a power station on July 15, 2023 as an example,

[0068] Data calibration and standardization:

[0069] Align the temperature data (1-minute level) of component #A123 output from the edge node with the humidity data (30-second level) from the weather station to a 1-minute granularity, generating 1440 samples for the day.

[0070] Component temperature =57.05℃, average of the past 7 days =42℃, standard deviation =8℃, after standardization ;

[0071] Ambient humidity H=65%, average over the past 7 days =60%, standard deviation =10%, after standardization .

[0072] Autoencoder fusion:

[0073] The input layer contains 128 features from the #A123 component, including temperature, voltage (352V), and current (7.2A). The trained encoder outputs a 32-dimensional latent feature z, where the eigenvalue related to temperature is 0.72 and the eigenvalue related to current is -0.65. The reconstruction error L = 0.018, satisfying the convergence condition.

[0074] Deep feature calculation:

[0075] Component aging index :

[0076] monthly efficiency decay rate Monthly (June efficiency 18.02%, July 17.94%)

[0077] Average monthly temperature =45℃, =8℃, therefore ;

[0078] Monthly average humidity standardized value =0.6;

[0079] calculate: (A value close to 1 indicates severe aging).

[0080] Environmental remediation efficiency:

[0081] Humidity coefficient ;

[0082] Temperature coefficient ;

[0083] Improved efficiency ;

[0084] Fault similarity: The cosine similarity between the latent feature z and historical hotspot fault samples is 0.82. 0.8, marked as high risk.

[0085] Feature filtering:

[0086] Aging Index Correlation coefficient with temperature change rate =0.72;

[0087] Correlation coefficient between correction efficiency and light intensity =0.68;

[0088] The above features and 18 potential features are retained to form a 20-dimensional key feature vector.

[0089] The final output comprehensive feature dataset contains 1440 samples of component #A123 on the same day. Each sample consists of 20 key features, covering information such as aging degree, environmental correction efficiency, and failure risk, providing core feature support for subsequent performance prediction and hot spot fault diagnosis.

[0090] The power generation prediction model construction and prediction phase uses a comprehensive feature dataset as the core input, while also incorporating real-time updated meteorological data (including solar irradiance G (unit: W / m²)). (referring to the solar radiation power received per unit area) and ambient temperature. (Unit: °C), Relative Humidity H (Unit: %, the ratio of water vapor content in the air to saturated water vapor content), Wind Speed (Unit: m / s) Precipitation probability (Unit: %) and equipment operation data (inverter output power) (Unit: kW), Module Temperature (Unit: °C), Aging Index (Dimensionless, range 0-1, larger values ​​indicate more severe aging), etc.), to achieve multi-scale power generation prediction for the next 1 hour, 24 hours and 7 days through dynamic adaptive models.

[0091] First, predictive feature engineering is performed, expanding time-series and cross-features based on the comprehensive feature dataset: for continuous variables such as light intensity and power, lag features (historical values ​​from the previous hour) are calculated. Historical values ​​for the first 3 hours Historical values ​​for the first 6 hours All units are W / (obtained by retrieving historical data at the corresponding timestamp) and rolling statistical features (6-hour moving average). Unit W / Where 360 ​​represents the number of 1-minute intervals contained within 6 hours. for The moment before Light intensity per minute; 12-hour trend slope Unit W / min, 720 represents the number of 1-minute intervals within a 12-hour period. Given the current light intensity, (The light intensity was 12 hours ago); when constructing environment-equipment cross-features, dimensionless processing is used, such as the temperature-aging interaction term. (dimensionless, For the standardized component temperature, through calculate, This represents the average temperature of the components over the past 7 days. (Standard deviation of the past 7 days), humidity-power correction item (dimensionless, For the standardized power, through calculate, This is the power average over the past 7 days. This represents the standard deviation over the past 7 days. Humidity influence coefficient, through (Calculation), ultimately forming a 72-dimensional predicted feature vector. (Includes 40 basic features + 32 extended features).

[0092] Next, a dynamic adaptive model architecture is constructed, employing a collaborative mechanism of "main model and sub-model": the main model is an attention-based LSTM network responsible for capturing long-term temporal dependencies, and its core calculation formula includes cell state updates. ( For the Gate of Oblivion Here is the forget gate weight matrix. The state was hidden in the previous moment. For the current input features, To offset the forget gate; For input gate, The input gate weight matrix, For input gate bias; Candidate state The candidate state weight matrix is... Bias for candidate states; Represents the element-wise product. (for the Sigmoid activation function) and output calculation ( For output gate, This is the output gate weight matrix. For output gate bias); for sunny days ( and ),partly cloudy( and ), rainy or Three types of weather are used to train XGBoost sub-models for each type, and the switching of sub-models is triggered by real-time meteorological data.

[0093] The model parameters are dynamically adjusted using an online learning algorithm based on real-time power deviation data. ( for Real power at any time for Always The predicted power (in kW) is updated by stochastic gradient descent to adjust the main model weights. ( =0.001 is the learning rate, which is dimensionless; The gradient of the loss function is obtained by applying the loss function. We obtain the derivative, that is ).

[0094] The prediction output layer employs a weighted fusion strategy, resulting in the final predicted value. ,in: =0.6 is the main model weight. Weights for sub-models (Sunny day) =0.4, Cloudy =0.4, Rainy =0.4, the corresponding sub-model weights are activated according to the weather type, and the rest are 0), all of which are obtained by training by minimizing historical prediction errors; Main model pair The predicted power (kW) at time 1 is calculated by the output layer of the LSTM network; For the first Sub-models The predicted power (kW) at time 1 is obtained from the output of the corresponding XGBoost sub-model; k is the prediction step size (1 hour). =60, k=1440 for 24 hours (unit: minutes).

[0095] Taking real-time data from 10:30 AM on July 15, 2023 as an example for prediction: the input features include the aging index from the comprehensive feature set. =0.962, correction efficiency 13.3%, real-time meteorological data W / , =32℃, H=65%, R=0%, Equipment operating data =480kW =57.05℃; Extended feature calculation yielded the 6-hour average light intensity. =850W / (Averaged by accumulating 360 illumination data points from 6:30 to 10:30), trend slope =0.2W / min(1000-880) / 720=0.2, where 880 is the light intensity at 22:30 the previous day), cross-features. ( ), ( , =400, =106.7; =1-0.002×65=0.87); The weather type is determined to be sunny, and the sunny sub-model is activated ( =0.4), predicted power 1 hour after the main model LSTM output. =502kW, sub-model output =498kW}; After weighted fusion =0.6×502+0.4×498=500.4kW. This result takes into account the effects of factors such as light trend, equipment aging, and ambient humidity.

[0096] The final output of the future power generation forecast includes structured results at different time scales: a 1-hour forecast of 500.4kW, a 24-hour hourly forecast curve (peaking at 12:00 with a value of 520kW), and a 7-day daily average forecast of 420kW (with a ±5% confidence interval), providing a quantitative basis for grid dispatch and power plant operation and maintenance.

[0097] The troubleshooting and diagnosis phase uses the future power generation forecast values ​​output in step 4 (including forecast power on 1-hour, 24-hour, and 7-day timescales). (unit: kW) and actual power generation results data (actual power output during the same period) The core input is (unit: kW), and real-time equipment operation data (such as component temperature) is also input. Inverter current IGBT temperature Environmental data (light intensity, etc.) Based on the data (such as humidity H) and historical maintenance records (fault type, handling measures, maintenance time, etc.), potential fault points are located and handling suggestions are generated through multi-level anomaly detection and correlation analysis.

[0098] First, power generation deviation is quantified by calculating the absolute deviation between actual and predicted values. (Unit: kW; negative values ​​indicate actual values ​​are lower than predicted) and relative deviation rate (dimensionless), when The fault investigation process is triggered when the actual power generation is 20% or more lower than the predicted value. Simultaneously, environmental correction deviations are calculated to eliminate the impact of natural factors such as sudden changes in sunlight. ,in The relative deviation between actual and predicted illumination (dimensionless). This represents the actual light intensity. (for predicting light intensity) The relative deviation between the actual and predicted ambient temperature (dimensionless). This refers to the actual ambient temperature. (for predicting ambient temperature) =0.6、 =0.2 is the influence coefficient (obtained through regression of historical data, dimensionless). If the corrected deviation rate... If so, the deviation is determined to be dominated by equipment malfunction.

[0099] Next, a knowledge graph network is constructed, where nodes include device entities (components, inverters, transformer substations, etc.), parameter entities (temperature, current, power, etc.), and fault types (hot spots, line aging, inverter faults, etc.). The weight of each edge represents the association strength. (dimensionless, For parameters With fault Co-occurrence frequency For parameters Total number of occurrences For fault Total number of occurrences). Calculate outlier parameter scores based on real-time data: for continuous parameters (such as... ), (Dimensionless; if more than 3, mark as an anomaly) This is the 7-day average of this parameter. (Standard deviation over the past 7 days); for discrete parameters (such as status codes). =1 (Warning) or 2 (Fault) (dimensionless, the larger the value, the higher the degree of abnormality). The knowledge graph is traversed using breadth-first search to calculate the comprehensive score of the faulty node. (dimensionless) For parameters Abnormal scores, For parameters With fault The higher the score, the greater the likelihood of the fault. The top 3 nodes with the highest scores are selected as potential fault candidates.

[0100] Verification is performed using historical maintenance records, and the similarity between candidate faults and similar historical faults is calculated. (dimensionless, This is the current abnormal parameter. (For historical fault parameters), when The fault type is identified in time, and historical handling measures are retrieved to generate suggestions.

[0101] Taking the actual data at 11:30 on July 15, 2023 as an example: Step 4 predicts the power generation at 11:30. =500.4kW, actual power generation =380kW, calculated as follows =-120.4kW, Trigger investigation; real-time lighting =980W / Predicting light =1000W / , Actual ambient temperature =33℃, predicted temperature =32℃, Environmental correction bias The equipment was determined to be faulty; the temperature of component #A123 was monitored in real time. =68℃ (average of the past 7 days) =42℃, standard deviation =8℃), abnormal score Branch current I = 6.5A (average of the same series) =8.5A), abnormal score ( =1.2A); Knowledge graph association analysis shows that the correlation strength between temperature anomalies and hot spot failures is... =0.85, the correlation strength between low current and hot spot. =0.78, Overall Score (Higher than other faults); Parameter similarity with historical #A123 component hot spot fault The fault was confirmed to be a hot spot; the recommended handling is to stop the machine within 1 hour to check the surface cleanliness of the components and the internal circuitry, and refer to the handling solution for a similar fault in May 2023 (replace the bypass diode, with an average repair time of 2.5 hours).

[0102] The final output includes: potential fault point (hot spot of component #A123), fault confidence (92%), associated abnormal parameters (temperature 68℃, current 6.5A) and handling suggestions (priority: high, operation steps and estimated recovery time), providing the operation and maintenance team with accurate fault location and handling basis.

[0103] The operation and maintenance optimization strategy formulation stage uses the predicted power generation value (including predicted power at different time scales) as the basis. The core inputs are the fault diagnosis results (potential fault points, fault confidence, handling suggestions, etc.) and the actual operating constraints of the power plant (such as the maximum load of the equipment). Unit kW, number of maintenance personnel Dimensionless), resource allocation data (such as spare parts inventory) Dimensionless, availability of repair tools (0-1 variables, 1 indicates availability) and external dispatch requirements (such as the upper limit of the grid's capacity). (unit: kW) The optimal operation and maintenance strategy is generated through a multi-objective optimization model.

[0104] First, the operation and maintenance priorities are quantified, and a fault handling priority index is calculated based on the fault diagnosis results. (Dimensionless, range 0-10, larger values ​​have higher priority): ,in Daily power generation loss due to fault (unit: kWh / day, via...) calculate, (This represents the actual power under fault conditions), which has been standardized to the 0-4 range. The duration of the fault impact (in days, ranging from 0 to 3, with 3 for emergency faults and 1 for routine faults). Assigning risk of fault propagation (range 0-3, based on the fault propagation probability in the knowledge graph); =0.5、 =0.3、 =0.2 is the weighting coefficient (dimensionless, summing to 1).

[0105] Next, an equipment maintenance plan is developed, employing a dynamic time window scheduling method: prioritizing... For urgent faults, arrange for handling within 4 hours (priority determined as urgent); Important faults are included in the 24-hour maintenance window (priority determined as important); For common faults, maintenance is scheduled together with preventative maintenance (priority determined as secondary). Calculate the maintenance resource matching degree. (dimensionless, The resources required for troubleshooting are calculated at 1, 1, and 0.5 respectively (spare parts, tools, and personnel are calculated at a rate of 1, 1, and 0.5). If resources are available, the plan can be executed as scheduled; otherwise, the external resource allocation process will be initiated.

[0106] The power generation dispatch scheme needs to balance the predicted power and grid constraints, and calculate the optimal power output plan. ,in =0.05 is the fault impact coefficient (dimensionless). This is the failure impact factor (0-1, determined by the number and severity of failures). Backup capacity is also reserved. (Unit: kW) to ensure response to sudden power fluctuations.

[0107] Taking the hot spot failure of module #A123 on July 15, 2023 as an example: Step 4 predicts the average power for the remaining period of the day. =480kW, actual power under fault condition =380kW, calculate daily power generation loss kWh / day, standardized to 3.2 (a full value of 4 corresponds to 3000 kWh / day); duration of fault impact. =1.5 days (expected to still affect half a day of power generation after repair); spread risk =2 (Hot spots may cause component glass to crack); Priority Index (Correction: The original calculation was incorrect; the recalculation should indicate an emergency fault. Correct calculation: Assume...) The standardized value is 3.8. =3, ==3, then =0.5×3.8+0.3×3+0.2×3=1.9+0.9+0.6=3.4, the actual value should be higher, and in the correct scenario it should be... =8.2, which is an emergency fault); resource matching degree calculation: required resources =2 (1 spare part + 1 set of tools + 1 person, converted to 1 + 1 + 0.5 = 2.5), existing resources =1、 =1、 =2, then One backup personnel needs to be allocated; in the power generation dispatch plan, the fault impact factor F=0.2 (the impact range of a single component failure), and the optimal output plan... .

[0108] The final output of the operation and maintenance optimization strategy includes: an emergency maintenance plan (replacing module #A123 from 11:30 to 14:00, allocating one backup personnel, and using the inventory bypass diodes), a power generation dispatch curve (hourly output not exceeding 475kW, with 47.5kW of reserve capacity), and preventive maintenance recommendations (infrared temperature measurement of 10 modules in the same string within this week), forming a closed-loop management from fault diagnosis to operation and maintenance execution, which is expected to reduce power generation losses by approximately 180kWh.

[0109] The processes described above with reference to the flowcharts in the embodiments disclosed in this invention can be implemented as computer software programs. The embodiments disclosed in this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by a central processing unit (CPU), it performs the functions defined in the methods of this application. It should be noted that the computer-readable medium described above in this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wire segments, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to: wireless segments, wire segments, optical fibers, RF, etc., or any suitable combination thereof.

[0110] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0111] Those skilled in the art should understand that the embodiments of the present invention described above and shown in the accompanying drawings are merely examples and do not limit the present invention. The purpose of the present invention has been fully and effectively achieved. The functions and structural principles of the present invention have been shown and explained in the embodiments. Without departing from the stated principles, the implementation of the present invention may have any variations or modifications.

Claims

1. A photovoltaic operation and maintenance management method based on edge computing, characterized in that, The method comprises: obtaining multi-source data through a sensor, pre-processing the multi-source data through an edge computing node; fusing the pre-processed multi-source data by using a self-encoder, mining deep correlation features, and obtaining a comprehensive feature data set; obtaining a future power generation prediction value through a dynamic self-adaptive model based on the comprehensive feature data set and real-time data; comparing the future power generation prediction value with an actual power generation result, judging whether there is a fault point according to the comparison result, locating a potential fault point in combination with a knowledge graph and historical records, and obtaining a fault diagnosis result; generating an equipment maintenance plan and a power generation scheduling scheme in combination with an actual operation condition and resource allocation of a power station according to the future power generation prediction value and the fault diagnosis result; a specific process of judging whether there is the fault point is as follows: comparing the future power generation prediction value with the actual power generation result, calculating an absolute deviation and a relative deviation rate, and when the relative deviation rate reaches a fault troubleshooting triggering threshold, entering the next step of analysis; introducing a deviation between actual values and prediction values of real-time illumination and environmental temperature, correcting the power generation deviation through an influence coefficient, judging whether the corrected deviation still exceeds an equipment abnormality determination threshold, and if so, preliminarily determining that there is a possibility of a fault; based on real-time equipment operation data, calculating a standardized abnormality score of a component temperature and a current, assigning values according to abnormality levels according to a device state code, confirming whether there is a significant abnormal parameter, and if so, further verifying the existence of the fault; a specific obtaining process of the fault diagnosis result is as follows: relying on a knowledge graph, calculating an association strength of the identified abnormal parameter and each fault type, accumulating contribution scores of the abnormal parameter to different faults through network traversal, and screening a number of potential fault candidates with the highest comprehensive scores; calling historical fault maintenance records, extracting typical parameter features of similar faults, calculating a parameter similarity between the current potential fault candidate and historical faults, locking a corresponding fault type when the similarity reaches a confirmation threshold, matching specific equipment IDs corresponding to the abnormal parameters in combination with equipment entities associated with the fault type, and explicitly determining specific fault points to form a diagnosis result containing the fault points and associated abnormal parameters. 2.The photovoltaic operation and maintenance management method based on edge computing according to claim 1, characterized in that, a specific pre-processing process of the multi-source data is as follows: excluding abnormal values of photovoltaic component temperature data that exceed a physical threshold, filtering outliers through a statistical criterion, determining temperature abnormalities in combination with regional average temperature and duration, calculating temperature change rates of adjacent components and associating branch currents to mark suspected hot spot regions; defining a voltage normal range, replacing voltage values exceeding the range with a historical data mean, and verifying power data effectiveness through a deviation between theoretical power and measured power; performing mean smoothing processing on high-frequency meteorological data in a time dimension, filling missing data by using linear interpolation or spatial interpolation methods respectively according to missing durations; setting a key component temperature threshold, marking equipment warning and fault states, mapping original equipment state codes into standardized three-level state codes and associating fault descriptions; unifying power generation data in a time dimension and excluding abnormal data during a maintenance period, classifying and encoding operation and maintenance records and establishing an association index with equipment IDs. 3.The photovoltaic operation and maintenance management method based on edge computing according to claim 2, characterized in that, a specific obtaining process of the comprehensive feature data set is as follows: The multi-source data output by the edge computing node is resampled to a consistent time granularity based on a unified timestamp, a correlation matrix is established, continuous data is standardized to eliminate dimensions, and discrete data is converted through one-hot encoding; An autoencoder model is constructed to compress a high-dimensional input feature vector to a latent space of a set dimension, and multi-source data fusion is achieved by minimizing reconstruction error through training to obtain fused latent features; Based on the latent features, component aging indexes, correlation features, and failure precursor features are calculated to obtain multiple explicit features; The latent features and explicit features are spliced to form an initial comprehensive feature vector, key features are selected through correlation analysis, and a comprehensive feature dataset is constructed.

4. The photovoltaic operation and maintenance management method based on edge computing according to claim 3, characterized in that, The specific calculation process of the future power generation prediction value is as follows: Based on the comprehensive feature dataset, cross-features of time series features, environment, and equipment are expanded, and the features are dimensionless to form a fixed-dimensional prediction feature vector; a main model and a sub-model architecture are adopted, the main model captures long-term time series dependence through an LSTM network, and the sub-model is trained for different weather types, and the sub-model is automatically switched and adapted according to real-time weather data; Based on the real-time collected power deviation, the main model weight is dynamically updated through an online learning algorithm; a weighted fusion strategy is adopted, and the prediction results of the main model and the corresponding sub-model are combined to calculate the future power generation prediction value at different time scales.

5. The photovoltaic operation and maintenance management method based on edge computing according to claim 4, characterized in that, The specific generation process of the device maintenance plan and the power generation scheduling scheme is as follows: The fault impact and diffusion risk in the fault diagnosis result are combined to calculate a fault handling priority index, and the faults are divided into different priority levels according to the index; the processing time limit of each priority fault is dynamically set by referring to the historical operation and maintenance resource utilization rate, and the planned processing time period of different faults is determined; The number of spare parts, tools, and personnel required for fault handling is counted, and the existing resource inventory and available state are compared to determine whether the resources meet the planned demand; if the resources are insufficient, an external resource allocation process is started and the processing time is adjusted; The fault maintenance task and preventive maintenance are arranged together to form a device maintenance plan; the power generation output boundary is determined by clarifying the internal constraints of the power plant, the requirements of the external power grid, and the fault impact; Based on the power generation prediction value, the predicted power is corrected by combining the fault impact coefficient to obtain the basic output reference value of each period, and the standby capacity is reserved according to a fixed proportion of the basic output reference value.

6. An edge computing-based photovoltaic operation and maintenance management system, characterized in that, The system is used to perform the photovoltaic operation and maintenance management method based on edge computing according to any one of claims 1-5.

7. A computer readable storage medium characterized in that, The computer readable storage medium stores a computer program, and the computer program is executed by the processor to implement the photovoltaic operation and maintenance management method based on edge computing according to any one of claims 1-5.

Images