Photovoltaic panel state detection method and device based on machine learning model and electronic equipment

Abstract

The application discloses a photovoltaic panel state detection method and device based on a machine learning model and an electronic device. The method comprises collecting operation parameter data and outdoor environment data of a target photovoltaic panel, performing parameter type identification on the parameter data and generating parameter vectors through feature conversion; calculating the abnormal similarity of each parameter vector by using a first machine learning model, and calculating the correlation degree of the parameter vectors with other parameter vectors by using a second machine learning model for the parameter vectors with abnormal similarity exceeding a threshold value, to determine abnormal candidate parameters containing strong correlation parameters; obtaining the position information of the sensors corresponding to the abnormal candidate parameters and calculating the position correlation, and generating abnormal candidate vectors from the multiple abnormal candidate parameters with position correlation satisfying a threshold value; and outputting the abnormal state probability of the target photovoltaic panel by using a third machine learning model based on the abnormal candidate vectors. The method can reduce false positives and false negatives under complex working conditions, improve the state detection accuracy and robustness, and reduce the dependence on manual inspection.

Description

Technical Field

[0001] This application relates to the field of photovoltaic technology, and in particular to a photovoltaic panel condition detection method, device, and electronic device based on a machine learning model. Background Technology

[0002] Photovoltaic power generation systems typically include photovoltaic modules (solar panels) and their associated inverter, combiner, and monitoring equipment. Solar panels operate outdoors for extended periods, making them susceptible to damage from dust accumulation, bird droppings, leaf cover, rain, snow, and sandstorms. These factors can lead to decreased light transmittance, increased module temperature, or reduced output power, thus impacting power generation efficiency and operation and maintenance costs. Therefore, continuous monitoring and anomaly identification of solar panel operation is a crucial technical challenge in the field of photovoltaic power plant operation and maintenance.

[0003] In existing technologies, the assessment and maintenance of photovoltaic panels often rely on manual inspections, periodic cleaning, or alarms based on experience thresholds, such as simple judgments based on the degree of power generation decline, abnormal module temperature, or changes in ambient light. These methods typically require manual on-site verification or frequent cleaning operations, resulting in low detection efficiency, delayed response, and high operation and maintenance costs. Furthermore, they are difficult to achieve refined and continuous condition assessment in large-scale module array scenarios.

[0004] Furthermore, photovoltaic panels operate in complex outdoor environments, and their output characteristics are affected by a combination of factors. Weather changes (such as cloud cover, temperature, humidity, wind speed, and rainfall), shading from adjacent modules, localized hot spots, sensor drift, or noise can all cause fluctuations or abnormal performance in operating parameters. Judging based on a single parameter or threshold can easily lead to misinterpreting fluctuations caused by the external environment or shading as module malfunctions or dust accumulation, resulting in false alarms, missed alarms, and unnecessary cleaning and maintenance.

[0005] To improve the reliability of anomaly identification, some solutions attempt to introduce multi-source data fusion or model analysis, but in practical applications, they still face two prominent shortcomings: First, they fail to effectively characterize the correlation between parameters when multiple parameters are abnormal, making it difficult to distinguish between "anomalies caused by other situations" and "anomalies of the component itself"; Second, they do not make full use of the spatial position information of sensors on photovoltaic panels, making it difficult to verify the spatial consistency of local anomalies (such as anomalies caused by local shading or local pollution), affecting the accuracy of anomaly location and status determination.

[0006] Therefore, there is an urgent need for a technical solution for photovoltaic panel condition detection that can collect photovoltaic panel operating parameters and outdoor environmental data, calculate the degree of parameter anomalies based on machine learning models, and further combine the correlation between parameters and sensor location information to form anomaly candidate vectors to output the probability of abnormal state of the target photovoltaic panel. This would improve the accuracy and robustness of condition detection under complex outdoor conditions and reduce reliance on manual intervention and the risk of misjudgment. Summary of the Invention

[0007] This application provides a photovoltaic panel condition detection method, device, and electronic device based on a machine learning model to address the shortcomings of low accuracy in photovoltaic panel condition detection in the prior art.

[0008] To achieve the above objectives, embodiments of this application provide a photovoltaic panel condition detection method based on a machine learning model, comprising: Collect multiple parameter data of the target photovoltaic panel, including at least the operating parameter data of the target photovoltaic panel collected by multiple sensors and environmental data collected by outdoor environment sensors; Parameter type identification is performed on the multiple parameter data respectively to obtain the parameter type corresponding to each parameter data; For each parameter type, feature transformation is performed on the corresponding parameter data to generate the corresponding parameter vector; For multiple parameter vectors, the first machine learning model is used to calculate the anomaly similarity corresponding to each parameter vector. For a first parameter vector whose abnormal similarity is greater than a preset abnormal similarity threshold, a second machine learning model is used to calculate the correlation between each parameter vector and other parameter vectors. Based on the correlation degree, abnormal candidate parameters are determined, wherein the abnormal candidate parameters include a first parameter vector and at least one second parameter vector whose correlation degree with at least one of the first parameter vectors is greater than a preset correlation degree threshold and whose abnormal similarity is less than a preset abnormal similarity threshold. Obtain the position information of the sensor corresponding to the abnormal candidate parameter on the target photovoltaic panel; Based on the location information, the location correlation between the anomaly candidate parameters is calculated; Based on the location correlation, multiple abnormal candidate vectors are generated from the abnormal candidate parameters, wherein each abnormal candidate vector corresponds to two or more abnormal candidate parameters whose location correlation is greater than a preset location threshold. Based on the anomaly candidate vector, a third machine learning model is used to calculate the probability of the target photovoltaic panel's abnormal state.

[0009] According to the photovoltaic panel state detection method based on a machine learning model according to the embodiments of this application, generating multiple anomaly candidate vectors from the anomaly candidate parameters based on the location correlation includes: Based on the location correlation, the abnormal candidate parameters are divided into multiple abnormal candidate parameter groups, and each abnormal candidate parameter group includes two or more abnormal candidate parameters whose location correlation with each other is higher than a preset location threshold. For each group of anomalous candidate parameters, the anomalous weight of the corresponding anomalous candidate parameter is determined by the anomalous similarity of the anomalous candidate parameters in the group. For each abnormal candidate parameter group, an abnormal candidate vector is generated based on the values ​​of each abnormal candidate parameter in the abnormal candidate parameter group, the abnormal weight, and the position information.

[0010] According to the photovoltaic panel status detection method based on machine learning model according to the embodiments of this application, the operating parameter data includes at least one: output voltage, output current, output power, component temperature, inverter operating status, daily power generation or instantaneous power generation; and / or, the environmental data includes at least one: ambient temperature, ambient humidity, wind speed, rainfall or light intensity.

[0011] According to the photovoltaic panel condition detection method based on machine learning model according to the embodiments of this application, parameter type identification includes labeling each parameter data according to at least one of data source identifier, dimensional information and sampling period to obtain parameter type.

[0012] According to the photovoltaic panel state detection method based on machine learning model according to the embodiments of this application, feature transformation includes performing at least one preprocessing operation on parameter data and extracting features to generate a parameter vector. The preprocessing operation includes at least one of denoising, missing value imputation, normalization, standardization and sliding window segmentation.

[0013] According to the photovoltaic panel state detection method based on machine learning model according to the embodiments of this application, the calculation of the correlation degree between each parameter vector and other parameter vectors using a second machine learning model includes: for a first parameter vector, calculating the correlation degree between the first parameter vector and each other parameter vector to obtain a correlation degree set corresponding to the first parameter vector.

[0014] In the photovoltaic panel status detection method based on machine learning model according to the embodiments of this application, the location information includes at least one of the following: the installation coordinates of the sensor corresponding to the abnormal candidate parameter on the target photovoltaic panel, the installation area identifier, or the relative distance information.

[0015] In the photovoltaic panel status detection method based on machine learning model according to the embodiments of this application, the position correlation is determined according to the spatial distance between the sensors corresponding to the abnormal candidate parameters, and the position correlation between abnormal candidate parameters whose spatial distance is less than a preset distance threshold is higher than a preset position threshold.

[0016] This application also provides a photovoltaic panel condition detection device based on a machine learning model, characterized in that it includes: The acquisition module is used to acquire multiple parameter data of the target photovoltaic panel. The multiple parameter data includes at least the operating parameter data of the target photovoltaic panel acquired by multiple sensors and the environmental data acquired by outdoor environment sensors. The type recognition module is used to identify the parameter type of multiple parameter data and obtain the parameter type corresponding to each parameter data. The vector generation module is used to perform feature transformation on the corresponding parameter data for each parameter type to generate the corresponding parameter vector. The anomaly calculation module is used to calculate the anomaly similarity of each parameter vector using the first machine learning model for multiple parameter vectors. The correlation calculation module is used to calculate the correlation between each parameter vector and other parameter vectors using a second machine learning model for a first parameter vector whose abnormal similarity is greater than a preset abnormal similarity threshold. The candidate determination module is used to determine abnormal candidate parameters based on the correlation degree, wherein the abnormal candidate parameters include a first parameter vector and at least one second parameter vector whose correlation degree with at least one of the first parameter vectors is greater than a preset correlation degree threshold and whose abnormal similarity is less than a preset abnormal similarity threshold. The location acquisition module is used to acquire the location information of the sensor corresponding to the abnormal candidate parameter on the target photovoltaic panel; The location correlation calculation module is used to calculate the location correlation between abnormal candidate parameters based on location information; The candidate vector generation module is used to generate multiple abnormal candidate vectors from abnormal candidate parameters based on location correlation, wherein each abnormal candidate vector corresponds to two or more abnormal candidate parameters whose location correlation is greater than a preset location threshold. The probability calculation module is used to calculate the probability of abnormal states of a target photovoltaic panel based on anomaly candidate vectors using a third machine learning model.

[0017] This application also provides an electronic device, including: Memory, used to store programs; A processor is configured to run the program stored in the memory, wherein the program executes the photovoltaic panel state detection method based on a machine learning model provided in the embodiments of this application.

[0018] This application also provides a computer-readable storage medium storing a computer program executable by a processor, wherein the program, when executed by the processor, implements the photovoltaic panel state detection method based on a machine learning model as provided in this application.

[0019] The photovoltaic panel condition detection method, device, and electronic equipment based on machine learning models provided in this application collect operating parameter data and outdoor environmental data of the target photovoltaic panel, perform parameter type identification and feature transformation to generate parameter vectors, and use a first machine learning model to calculate the anomaly similarity of each parameter vector to achieve automatic identification of multi-parameter anomalies. For the first parameter vector whose anomaly similarity exceeds a threshold, a second machine learning model is further introduced to calculate its correlation with other parameter vectors, and based on this, the second parameter vectors that are correlated with the first parameter vector are included in the anomaly candidate parameter set. This enables correlation analysis and candidate screening of multi-source anomalies under complex outdoor conditions, reducing the reliance on a single parameter or... This approach addresses false alarms and missed alarms caused by fixed thresholds. Simultaneously, it acquires the location information of sensors corresponding to abnormal candidate parameters on the photovoltaic panel, calculates the positional correlation between abnormal candidate parameters, and combines two or more spatially consistent abnormal candidate parameters to generate an abnormal candidate vector. Finally, a third machine learning model outputs the probability of the abnormal state of the target photovoltaic panel. This integrates three types of information—"abnormality degree," "parameter correlation," and "spatial consistency"—for judgment, effectively distinguishing parameter fluctuations caused by weather changes, shading, or local factors from true abnormal states. This improves the accuracy, robustness, and interpretability of state detection, reduces reliance on manual inspections and unnecessary maintenance interventions, and enhances the efficiency of photovoltaic panel operation monitoring and the level of power plant operation and maintenance management.

[0020] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0021] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart illustrating an embodiment of the photovoltaic panel condition detection method based on a machine learning model provided in this application; Figure 2A schematic diagram of the photovoltaic panel condition detection device based on a machine learning model provided in this application; Figure 3 A schematic diagram of the structure of an embodiment of the electronic device provided in this application. Detailed Implementation

[0022] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0023] Since photovoltaic (PV) panels are typically installed outdoors for extended periods, their surfaces and surrounding environment are susceptible to fluctuations caused by dust accumulation, shading, and weather changes. This fluctuation can affect the PV panel's output voltage, current, power, and other operating parameters, ultimately impacting the power generation efficiency and operational reliability of the PV system. Therefore, continuous and accurate monitoring and anomaly identification of PV panel operation is a crucial aspect of PV power plant operation and maintenance management.

[0024] In existing technologies, photovoltaic panel condition monitoring and maintenance typically rely on manual inspections or periodic cleaning strategies, or trigger alarms and maintenance decisions by setting thresholds for single or a few operating parameters. For example, when a decrease in power generation or abnormal module temperature is detected, maintenance personnel determine whether cleaning or troubleshooting is necessary by conducting on-site inspections. This approach has significant limitations: firstly, in scenarios with a large number of photovoltaic panels, wide distribution of sites, or long operating time spans, manual inspections and experience-based judgments are inefficient, costly, and slow to respond; secondly, judgment methods based on fixed thresholds are difficult to adapt to varying outdoor conditions, and are prone to false alarms or missed alarms under different seasons, weather conditions, installation orientations, or shading conditions.

[0025] Furthermore, photovoltaic (PV) panels operate in complex outdoor environments, where there is a significant coupling relationship between their operating parameters and environmental parameters. Weather changes (such as variations in sunlight intensity, rainfall, and temperature and humidity fluctuations), shading from adjacent PV panels or obstacles, reduced light transmittance due to wind, sand, and dust accumulation, and sensor noise or drift can all cause abnormal performance of operating parameters. Judging solely based on the degree of anomaly in a single parameter makes it difficult to distinguish whether the anomaly is caused by changes in external conditions or by changes in the PV panel's own condition, potentially leading to inaccurate identification of the anomaly's source and affecting the correctness and timeliness of maintenance strategies.

[0026] To address the aforementioned issues, this application proposes a photovoltaic panel condition detection method based on a machine learning model. The basic scheme and principle are as follows: First, multiple parameter data of the target photovoltaic panel are collected. These multiple parameter data include at least operational parameter data collected by multiple sensors and environmental data collected by outdoor environmental sensors. Then, parameter type identification is performed on each parameter data, and feature transformation is applied to the corresponding parameter data for each parameter type to generate parameter vectors. Further, a first machine learning model is used to calculate the anomaly similarity corresponding to each parameter vector to identify potential anomalies. For first parameter vectors with anomaly similarity exceeding a preset anomaly similarity threshold, a second machine learning model is used to calculate the correlation between each parameter vector and other parameter vectors, and an anomaly candidate parameter set is constructed based on the correlation, enabling the introduction of "parameter correlation" information during anomaly determination. Simultaneously, the position information of the sensors corresponding to the anomaly candidate parameters on the target photovoltaic panel is obtained. Based on the position information, the positional correlation between the anomaly candidate parameters is calculated, and multiple anomaly candidate parameters whose spatial positional correlation meets the threshold are combined to generate an anomaly candidate vector. Finally, a third machine learning model is used based on the anomaly candidate vectors to calculate the probability of an abnormal state of the target photovoltaic panel, thereby outputting a condition detection result that better reflects actual working conditions.

[0027] Through the above scheme, this application can perform unified modeling and analysis of multi-source parameters under complex outdoor conditions: on the one hand, it uses anomaly similarity to achieve automatic detection of multi-parameter anomalies, reducing reliance on manual inspection and fixed threshold strategies; on the other hand, it expands anomalies from single parameters to collaborative judgment of "related parameter sets" through correlation analysis, which helps reduce false alarms and missed alarms caused by weather changes, shading, and other factors; further, by combining sensor location information and location correlation, it introduces spatial consistency constraints into anomaly judgment, making it more capable and robust in identifying local and regional anomalies; finally, it outputs the detection results in the form of anomaly state probabilities, which facilitates risk classification, decision ranking, and resource scheduling by the operation and maintenance side, thereby improving the accuracy, stability, and operation and maintenance efficiency of photovoltaic panel status detection.

[0028] Example 1 like Figure 1 As shown, Figure 1 This is a flowchart illustrating a photovoltaic panel state detection method based on a machine learning model according to an embodiment of this application. Figure 1 The method shown may include: S101, collect multiple parameter data of the target photovoltaic panel. The multiple parameter data includes at least the operating parameter data of the target photovoltaic panel collected by multiple sensors and the environmental data collected by outdoor environment sensors.

[0029] In this embodiment, data acquisition can be collaboratively completed by the power plant monitoring system, data acquisition unit (DTU / RTU), inverter communication interface, module-side sensors, and weather station / outdoor environmental sensors. Operating parameter data reflects the electrical and thermal operating status of the target photovoltaic panel, while environmental data characterizes the impact of external operating conditions on the target photovoltaic panel's output. The acquisition process can be performed according to a preset sampling period (e.g., seconds, minutes, or longer), and timestamps and data source identifiers are appended to the acquired data for subsequent synchronization, alignment, and modeling analysis. When data sources with different sampling frequencies exist, interpolation, resampling, or window aggregation (e.g., mean, maximum, quantiles, etc.) can be used to achieve time alignment, thereby forming a multi-parameter data set that can be used for subsequent parameter type identification and feature transformation.

[0030] In addition, the embodiments of this application may also include at least one type of operating parameter data: output voltage, output current, output power, component temperature, inverter operating status, daily power generation or instantaneous power generation; and / or, environmental data may include at least one type of environmental data: ambient temperature, ambient humidity, wind speed, rainfall or light intensity.

[0031] In this embodiment, output voltage, output current, and output power can be acquired from the inverter or combiner box side; component temperature can be acquired from a surface-mount temperature sensor or a backplane temperature probe; inverter operating status can include discrete or semi-discrete information such as grid-connected status, fault codes, and derating status; daily power generation / instantaneous power generation can serve as an operational characterization at the energy and power levels. Ambient temperature, humidity, wind speed, rainfall, and light intensity can be obtained from meteorological station sensors and used to provide explanatory variables or auxiliary features of external influencing factors in subsequent correlation analysis and abnormal state probability calculation, thereby reducing the risk of false alarms caused by relying solely on operating parameters.

[0032] S102, perform parameter type identification on multiple parameter data respectively, and obtain the parameter type corresponding to each parameter data.

[0033] In this embodiment, parameter type identification is used to map parameter data from different sources, with different dimensions, and with different meanings to a unified type space, facilitating the adoption of corresponding feature transformation strategies and machine learning model input formats. Parameter types may include, but are not limited to: voltage, current, power, temperature, status (discrete status / alarm code), irradiance, and rainfall. The identification method can be comprehensively determined based on the data channel configuration table, sensor ID and installation information, protocol field meanings (such as Modbus register definitions), unit and dimension verification, and data statistical characteristics (such as value range and dispersion). When inconsistencies in type or abnormal units are detected, it can be recorded as a data quality problem and trigger a cleaning or error correction process to improve the reliability of subsequent model calculations.

[0034] Furthermore, the embodiments of this application may also include parameter type identification by labeling each parameter data according to at least one of data source identifier, dimensional information and sampling period, so as to obtain the parameter type.

[0035] In this embodiment, each parameter data entry can record its data source identifier (e.g., inverter channel, environmental sensor channel, module-side temperature channel, etc.), dimensional information (e.g., V, A, W, ℃, %RH, etc.), and sampling period information (e.g., 1s, 10s, 60s, etc.), which serve as the basis for type labeling. For example, when the data source is a weather station and the dimension is W / m², it can be labeled as irradiation; when the source is a module temperature probe and the dimension is ℃, it can be labeled as temperature. Type labeling directly determines the window length, normalization method, and model input format used in subsequent feature transformations, avoiding model bias caused by mixing different physical quantities.

[0036] S103: For each parameter type, perform feature transformation on the corresponding parameter data to generate the corresponding parameter vector.

[0037] In this embodiment, feature transformation is used to convert the original time-series parameter data into a fixed-dimensional, comparable parameter vector to adapt to the input of the first machine learning model. Feature transformation can employ different transformation strategies based on parameter type: for example, for continuous electrical / environmental parameters, a sliding time window can be used to extract statistical features (mean, standard deviation, slope, kurtosis, quantiles, rate of change, etc.), frequency domain features (such as energy spectral density), and deviation features from the baseline; for discrete state parameters, one-hot encoding, frequency statistics, or duration statistics can be performed. The generated parameter vector can contain single-time-point features or multi-time-window features and can be aligned with timestamps, thereby forming multiple parameter vector sets for subsequent anomaly similarity calculations.

[0038] Furthermore, the embodiments of this application may also include feature transformation including performing at least one preprocessing operation on the parameter data and extracting features to generate a parameter vector. The preprocessing operation includes at least one of denoising, missing value imputation, normalization, standardization, and sliding window segmentation.

[0039] In this embodiment, denoising can be achieved using median filtering, moving average, or Kalman filtering to suppress acquisition noise; missing value imputation can be achieved using forward imputation, linear interpolation, or regression imputation based on adjacent parameters; normalization / standardization can be performed separately according to parameter type to avoid the adverse effects of different units on model distance measurement; sliding window segmentation can use a fixed window length (e.g., 5 minutes, 15 minutes) and a fixed step size (e.g., 1 minute) to construct samples, thereby balancing real-time performance and stability. Through the above preprocessing and feature extraction, the quality of parameter vectors and consistency with model input can be improved, enhancing the robustness of anomaly detection.

[0040] S104. For multiple parameter vectors, use the first machine learning model to calculate the anomaly similarity corresponding to each parameter vector.

[0041] In this embodiment, the first machine learning model is used to measure the degree of "deviation from the normal pattern" for each parameter vector, i.e., anomaly similarity. The first machine learning model can be an unsupervised or semi-supervised anomaly detection model, such as a model based on density estimation, distance measurement, reconstruction error, or a single-class classification approach; anomaly similarity can be defined as the distance from the center of the normal cluster, negative log-likelihood, reconstruction error, or a mapped score. During the model training phase, normal distributions or normal representations can be established using normal data selected from historical runs; during the online computation phase, anomaly similarity is output for each parameter vector and compared with a preset anomaly similarity threshold to filter the first parameter vectors that require further association analysis, thereby achieving initial screening and noise reduction for multiple parameters.

[0042] S105, for the first parameter vector whose abnormal similarity is greater than the preset abnormal similarity threshold, the second machine learning model is used to calculate the correlation between each parameter vector and other parameter vectors.

[0043] In this embodiment, the second machine learning model is used to characterize the coupling relationship between parameters to determine whether anomalies exhibit characteristics of "correlation propagation" or "homogeneous change". The correlation degree can reflect the correlation, conditional dependence, or causal consistency between the first parameter vector and other parameter vectors within the same time window. This can be achieved, for example, through multivariate regression residuals, mutual information estimation, graph model edge weights, attention weights, or similarity obtained through contrastive learning. During calculation, the correlation degree calculation can be triggered only for the first parameter vector exceeding a threshold to reduce computational overhead. Different correlation calculation strategies can be set according to parameter types; for example, there may be a stronger physical correlation between operating power and irradiance, or between component temperature and ambient temperature / wind speed, thus making the correlation degree results more consistent with the actual operating conditions of the power plant.

[0044] Furthermore, embodiments of this application may also use a second machine learning model to calculate the correlation between each parameter vector and other parameter vectors, including: for the first parameter vector, calculating the correlation between the first parameter vector and each other parameter vector to obtain a set of correlations corresponding to the first parameter vector.

[0045] In this embodiment, a correlation set can be output for each first parameter vector. This set can be used to construct subsequent anomaly candidate parameters: on the one hand, it can locate parameters that are strongly correlated with the first parameter vector to supplement anomaly evidence; on the other hand, it can also be used to identify external disturbances with "broad correlation" (such as sudden weather causing multiple parameters to change simultaneously), providing input features or weighting basis for determining the probability of subsequent anomalies.

[0046] S106, Based on the correlation degree, determine the abnormal candidate parameters, wherein the abnormal candidate parameters include a first parameter vector and at least one second parameter vector whose correlation degree with at least one of the first parameter vectors is greater than a preset correlation degree threshold and whose abnormal similarity is less than a preset abnormal similarity threshold.

[0047] In this embodiment, the construction of anomaly candidate parameters is used to include a first parameter vector of "significant anomalies" and a second parameter vector of "strong correlation with them but not significant anomalies when viewed alone" in the analysis scope, thereby forming a multi-parameter collaborative set of anomaly evidence. Specifically, the first parameter vector set can be determined first (anomaly similarity exceeding a preset anomaly similarity threshold), and then a second parameter vector satisfying the correlation threshold condition can be found among the remaining parameter vectors; although the anomaly similarity of the second parameter vector is lower than the anomaly threshold, its strong correlation with the first parameter vector can indicate potential homologous changes, local regional problems, or early signs of anomalies. Through this strategy, missed detections caused by relying solely on a single anomaly score can be avoided, and the coverage of complex anomaly patterns can be improved.

[0048] S107, Obtain the position information of the sensor corresponding to the abnormal candidate parameter on the target photovoltaic panel.

[0049] In this embodiment, location information is used to map anomaly candidate parameters to the spatial structure of the photovoltaic panel for subsequent spatial consistency analysis. Location information can come from equipment installation logs, CAD / 3D modeling data, construction acceptance records, QR code / electronic tag binding information, or GIS / digital twin systems. For component-level and cascade-level sensors, a mapping relationship of "sensor—component location—string / array topology" can be established. After obtaining the location information, anomaly candidate parameters can be associated and stored with the corresponding sensor coordinates or area labels, providing a basis for location correlation calculation.

[0050] In addition, the location information in this application embodiment may include at least one of the following: the installation coordinates of the sensor corresponding to the abnormal candidate parameter on the target photovoltaic panel, the installation area identifier, or the relative distance information.

[0051] In this embodiment, the installation coordinates can be two-dimensional planar coordinates or a local coordinate system with the component border as a reference; the installation area identifier can be a discrete area description such as "top left / top right / bottom left / bottom right" or "component in row x and column y"; the relative distance information can represent a distance matrix or adjacency table between sensors. When precise coordinates cannot be obtained, approximate spatial modeling can be achieved using only area identifiers or relative distances, which can still meet the subsequent position correlation calculation requirements for determining "whether they are spatially adjacent / whether they are in the same area".

[0052] S108, based on location information, calculate the location correlation between anomaly candidate parameters.

[0053] In this embodiment, location correlation is used to quantify whether anomalous candidate parameters exhibit spatial clustering or a consistent distribution. Location correlation can be determined by a combination of factors such as the geometric distance between sensors, the topological adjacency of components, and their relative positions within the same string / component plane. For example, distance can be mapped to a correlation score (the closer the distance, the higher the correlation), or sensors that satisfy the adjacency relationship can be assigned a higher correlation. Through location correlation calculation, criteria can be provided for subsequent grouping of anomalous candidate parameters and construction of anomalous candidate vectors, so that anomaly determination considers not only "parameter anomalies and correlations" but also "spatial consistency," thereby improving the ability to identify local anomalies, local occlusion, local contamination, or regional sensor faults.

[0054] Furthermore, in this application embodiment, the positional correlation can also be determined based on the spatial distance between the sensors corresponding to the abnormal candidate parameters, and the positional correlation between abnormal candidate parameters whose spatial distance is less than a preset distance threshold is higher than a preset position threshold.

[0055] In this embodiment, spatial distance can be mapped to positional correlation, for example, using a piecewise function or a monotonically decreasing function: when the spatial distance is less than a threshold, a strong correlation is considered to be satisfied, and the positional correlation is made greater than a preset positional threshold; when the spatial distance is greater than the threshold, the positional correlation is made lower than the threshold, thereby achieving rapid screening based on the distance threshold. The distance threshold can be set according to engineering parameters such as component size, sensor deployment density, and inter-string spacing to adapt to the layout structure of different power plants.

[0056] S109, Based on the location correlation, generate multiple abnormal candidate vectors from the abnormal candidate parameters, wherein each abnormal candidate vector corresponds to two or more abnormal candidate parameters whose location correlation is greater than a preset location threshold.

[0057] In this embodiment, the anomaly candidate vector is used to jointly represent a set of spatially related anomaly candidate parameters, so that the third machine learning model can distinguish "region / cluster-level anomaly patterns". The generation method may include: constructing an undirected graph or adjacency matrix based on a location relevance threshold, connecting anomaly candidate parameters with location relevance higher than a preset location threshold into a connected subgraph or cluster; each cluster corresponds to one anomaly candidate vector. The anomaly candidate vector may include the current value of the parameter within the cluster, anomaly similarity, a summary of its relevance to the first parameter vector, and location features (such as cluster center coordinates, cluster scale / radius, coverage area, etc.), thereby reducing dimensionality and enhancing spatial consistency features while maintaining sufficient information.

[0058] Furthermore, embodiments of this application can also generate multiple abnormal candidate vectors from abnormal candidate parameters based on location correlation, including: dividing the abnormal candidate parameters into multiple abnormal candidate parameter groups based on location correlation, each abnormal candidate parameter group including two or more abnormal candidate parameters whose location correlation with each other is higher than a preset location threshold; for each abnormal candidate parameter group, determining the abnormal weight of the corresponding abnormal candidate parameter using the abnormal similarity of the abnormal candidate parameters in the abnormal candidate parameter group; for each abnormal candidate parameter group, generating an abnormal candidate vector for the abnormal candidate parameter group based on the value, abnormal weight, and location information of each abnormal candidate parameter in the abnormal candidate parameter group.

[0059] In this embodiment, the anomaly candidate parameter groups can be implemented using clustering methods such as connected component decomposition, hierarchical clustering, and DBSCAN, where the clustering distance metric is derived from location correlation. For each anomaly candidate parameter group, anomaly weights can be assigned to each parameter within the group based on anomaly similarity. For example, the anomaly similarity can be normalized and used as the weight, or parameters exceeding anomaly thresholds can be given higher weights to highlight the main anomaly source. Subsequently, the observed values ​​of the parameters within the group are weighted and fused with the anomaly weights to form numerical features, and location information (such as the group center point, group coverage area, maximum / minimum distance within the group, etc.) is concatenated as spatial features to form the anomaly candidate vector for that group. This method can complete the information integration of "spatial grouping + strong / weak anomaly weighting" before inputting into the third machine learning model, improving the sensitivity and stability of anomaly state probability calculation to local anomaly morphology.

[0060] S110, based on the anomaly candidate vector, uses a third machine learning model to calculate the probability of anomalies in the target photovoltaic panel.

[0061] In this embodiment, a third machine learning model is used to comprehensively judge the anomaly candidate vectors and output the probability value of the target photovoltaic panel being in an abnormal state, which can be used for subsequent alarm, dispatching, or maintenance decisions. The third machine learning model can be a binary or multi-class probabilistic model, such as logistic regression, gradient boosting tree, neural network, or sequence-based model; its input is one or more anomaly candidate vectors. When there are multiple anomaly candidate vectors, the output probability at the photovoltaic panel level can be obtained by using methods such as maximizing, weighted summation, or attention convergence. Model training can use historically confirmed abnormal samples and normal samples. Labels can come from manual review, maintenance records, or system alarm closed-loop results. By learning the mapping relationship between anomaly candidate vectors and abnormal states, the output probability can simultaneously reflect the degree of anomaly, parameter correlation, and spatial consistency, thereby improving the accuracy of state detection under complex operating conditions and reducing false alarms and missed alarms.

[0062] The photovoltaic panel condition detection method based on a machine learning model provided in this application collects operating parameter data and outdoor environmental data of the target photovoltaic panel, performs parameter type identification and feature transformation to generate parameter vectors, and uses a first machine learning model to calculate the anomaly similarity of each parameter vector to achieve automatic identification of multi-parameter anomalies. For the first parameter vector whose anomaly similarity exceeds a threshold, a second machine learning model is further introduced to calculate its correlation with other parameter vectors, and based on this, the second parameter vectors that are correlated with the first parameter vector are included in the anomaly candidate parameter set. This enables correlation analysis and candidate screening of multi-source anomalies under complex outdoor conditions, reducing the reliance on a single parameter or fixed threshold. This process addresses false alarms and missed alarms. Simultaneously, it acquires the location information of the sensors corresponding to the abnormal candidate parameters on the photovoltaic panel, calculates the positional correlation between the abnormal candidate parameters, and combines two or more spatially consistent abnormal candidate parameters to generate an abnormal candidate vector. Finally, a third machine learning model outputs the probability of the abnormal state of the target photovoltaic panel. This integrates and determines the three types of information—"abnormality degree," "parameter correlation," and "spatial consistency"—effectively distinguishing between parameter fluctuations caused by weather changes, shading, or local factors and true abnormal states. This improves the accuracy, robustness, and interpretability of state detection, reduces reliance on manual inspections and unnecessary maintenance interventions, and enhances the efficiency of photovoltaic panel operation monitoring and the level of power plant operation and maintenance management.

[0063] Example 2 Figure 2 This is a schematic diagram of the photovoltaic panel condition detection device based on a machine learning model provided in this application. This machine learning model-based photovoltaic panel condition detection device can be used to implement, for example, [reference needed]. Figure 1 The present application describes a photovoltaic panel state detection method based on a machine learning model. The photovoltaic panel state detection device based on the machine learning model may include: a data acquisition module 21, a type identification module 22, a vector generation module 23, an anomaly calculation module 24, a correlation calculation module 25, a candidate determination module 26, a location acquisition module 27, a location correlation calculation module 28, a candidate vector generation module 29, and a probability calculation module 30.

[0064] The acquisition module 21 can be used to acquire multiple parameter data of the target photovoltaic panel. The multiple parameter data includes at least the operating parameter data of the target photovoltaic panel acquired by multiple sensors and the environmental data acquired by outdoor environment sensors.

[0065] In this embodiment, the acquisition module 21 can be implemented by a data acquisition terminal, an edge computing gateway, or a server-side acquisition component. It supports communication with devices such as inverters, combiner boxes, component-side temperature sensors, and weather stations to obtain operating parameter data and environmental data. The acquisition module 21 can timestamp the acquired data, bind data source identifiers, and record the sampling frequency. When multiple sources with different frequencies exist, it performs alignment processing (e.g., resampling, interpolation, or aggregation by time window) to form a unified dataset that can be used for subsequent identification and modeling.

[0066] In addition, the operating parameter data collected by the acquisition module 21 in this embodiment may include at least one of the following: output voltage, output current, output power, component temperature, inverter operating status, daily power generation or instantaneous power generation; and / or, the environmental data collected by the acquisition module 21 may include at least one of the following: ambient temperature, ambient humidity, wind speed, rainfall or light intensity.

[0067] In this embodiment, the aforementioned data can originate from the inverter interface, DC-side measurement unit, module backplane temperature probe, and outdoor environmental sensor, respectively. By using electrical, thermal, and environmental quantities as input data sources, a more complete operating context can be provided for subsequent anomaly calculations and correlation analyses, thereby improving the stability and reliability of the detection results.

[0068] The type recognition module 22 can be used to identify the parameter type of multiple parameter data separately and obtain the parameter type corresponding to each parameter data.

[0069] In this embodiment, the type identification module 22 is used to map parameter data with different physical meanings, dimensions, and sampling characteristics to a predefined set of parameter types, such as voltage, current, power, temperature, state, irradiance, and rainfall. The type identification module 22 can determine the type based on a pre-configured channel dictionary, sensor ID and installation information, protocol field meanings, unit / dimension verification, and data statistical characteristics (range, dispersion, whether discrete values ​​are taken, etc.), and use the parameter type as the basis for subsequent feature conversion strategy selection.

[0070] In addition, in this embodiment of the application, the type identification module 22 can also label the type of each parameter data according to at least one of the data source identifier, dimensional information and sampling period to obtain the parameter type.

[0071] In this embodiment, the type identification module 22 can use "source identifier (e.g., inverter channel / weather station channel) + dimensions (e.g., V, A, W, ℃, W / m²) + sampling period (e.g., 1s, 60s)" as input fields for type labeling and output a traceable type label. This method facilitates maintaining type consistency when data channels change, equipment is replaced, or multiple vendor protocols coexist, reducing the risk of confusion in subsequent model inputs.

[0072] The vector generation module 23 can be used to perform feature transformation on the corresponding parameter data for each parameter type to generate the corresponding parameter vector.

[0073] In this embodiment, the vector generation module 23 converts the original time-series parameter data into parameter vectors of fixed dimensions to adapt to the input of the machine learning model. For continuous variables, statistical features (mean, standard deviation, maximum / minimum, quantiles, slope, rate of change, etc.) and deviation features can be extracted using a sliding window; for discrete / state variables, encoding and statistics (such as one-hot encoding, frequency of occurrence, duration) can be performed. The vector generation module 23 can be configured with different window lengths, feature sets, and normalization strategies according to parameter types to improve the comparability and discriminability of parameter vectors of different types.

[0074] In addition, in this embodiment of the application, the vector generation module 23 may perform at least one preprocessing operation on the parameter data and extract features to generate the parameter vector before generating the parameter vector. The preprocessing operation includes at least one of denoising, missing value imputation, normalization, standardization and sliding window segmentation.

[0075] In this embodiment, the vector generation module 23 can apply median filtering / moving average to channels with high noise, forward padding / linear interpolation to missing values, normalization or standardization to parameters with different dimensions, and segment the data into samples according to a fixed window. Through preprocessing and feature generation, the impact of sensor jitter, packet loss, and other issues on the model output can be reduced, enhancing the robustness of anomaly detection.

[0076] The anomaly calculation module 24 can be used to calculate the anomaly similarity corresponding to each parameter vector using the first machine learning model for multiple parameter vectors.

[0077] In this embodiment, the anomaly calculation module 24 is used to output anomaly similarity for each parameter vector to characterize the degree to which the parameter vector deviates from the normal pattern. The first machine learning model can be an unsupervised or semi-supervised anomaly detection model (e.g., a model based on density estimation, distance metric, reconstruction error, or one-class classification). The anomaly similarity can be defined as distance / error / negative log-likelihood, etc., and compared with a preset anomaly similarity threshold to filter the first parameter vectors that need further correlation analysis, thereby reducing the scale and noise interference of subsequent calculations.

[0078] The association calculation module 25 can be used to calculate the association degree between each parameter vector and other parameter vectors using a second machine learning model for a first parameter vector whose abnormal similarity is greater than a preset abnormal similarity threshold.

[0079] In this embodiment, the correlation calculation module 25 is used to characterize the coupling relationship between parameters and output the correlation degree to assist in anomaly attribution and candidate construction. The second machine learning model can be a correlation learning model, a graphical model, an attention mechanism model, a mutual information estimation model, or a dependency measurement model based on regression residuals. The correlation calculation module 25 can trigger the calculation only for entries determined as the first parameter vector and form a correlation degree result of "first parameter vector - other parameter vectors", providing input for the subsequent candidate determination module 26.

[0080] In addition, the embodiments of this application can also associate the calculation module 25 to calculate the correlation degree between the first parameter vector and each other parameter vector for the first parameter vector, so as to obtain the correlation degree set corresponding to the first parameter vector.

[0081] In this embodiment, the correlation calculation module 25 can generate a correlation degree set for each first parameter vector, and sort them from high to low correlation degree or perform threshold filtering to support the rapid selection of strongly correlated parameters in the future. This set can also be used as one of the input features of the subsequent probability model to express the pattern feature of "whether the anomaly is accompanied by changes in related parameters".

[0082] The candidate determination module 26 can be used to determine abnormal candidate parameters based on the degree of correlation. The abnormal candidate parameters include a first parameter vector and at least one second parameter vector whose degree of correlation with at least one of the first parameter vectors is greater than a preset degree of correlation threshold and whose abnormal similarity is less than a preset abnormal similarity threshold.

[0083] In this embodiment, the candidate determination module 26 integrates the first parameter vector of "significant anomalies" and the second parameter vector of "strongly correlated but not exceeding the anomaly threshold" into the anomaly candidate parameter set. This strategy can supplement evidence through correlation in the early stages of anomalies or when anomalies are scattered, reducing missed detections caused by relying solely on the anomaly score of a single parameter; it also helps to identify widespread synchronous changes caused by environmental disturbances, thereby providing more complete candidate inputs for subsequent spatial consistency analysis and anomaly state probability calculation.

[0084] The location acquisition module 27 can be used to acquire the location information of the sensor corresponding to the abnormal candidate parameter on the target photovoltaic panel.

[0085] In this embodiment, the location acquisition module 27 can read sensor location information from the installation log, sensor binding configuration, CAD / GIS / digital twin system, or database mapping table, and establish a correspondence between "abnormal candidate parameters - sensor - location". The location information can be used to express the component, string, row, or column position or coordinate point of the sensor, providing basic data for subsequent location correlation calculations.

[0086] In addition, the location information obtained by the location acquisition module 27 in this embodiment may include at least one of the following: the installation coordinates of the sensor corresponding to the abnormal candidate parameter on the target photovoltaic panel, the installation area identifier, or the relative distance information.

[0087] In this embodiment, if accurate surveying conditions are available, installation coordinates can be used; if only engineering annotation information is available, installation area identifiers (such as row x, column y) can be used; if the system maintains topological relationships or an adjacency list, relative distance information or a distance matrix can be used. Any of these methods can support subsequent location correlation calculations and spatial grouping.

[0088] The location correlation calculation module 28 can be used to calculate the location correlation between abnormal candidate parameters based on location information.

[0089] In this embodiment, the location correlation calculation module 28 is used to calculate the strength of the spatial relationship between each pair of anomaly candidate parameters and output a location correlation index. Location correlation can be obtained by geometric distance mapping or generated by rules such as topological adjacency, same string / same region determination. Through location correlation, it can be determined whether anomalies are spatially clustered, thereby providing spatial consistency constraints for the subsequent candidate vector generation module 29.

[0090] In addition, in this embodiment, the position correlation calculation module 28 can determine the position correlation based on the spatial distance between the sensors corresponding to the abnormal candidate parameters, and the position correlation between abnormal candidate parameters whose spatial distance is less than a preset distance threshold is higher than a preset position threshold.

[0091] In this embodiment, the location correlation calculation module 28 can use a distance threshold as a rapid discrimination condition: when the distance between two sensors is less than the threshold, it is determined to be a strong spatial correlation, and a location correlation higher than the preset location threshold is assigned; otherwise, a lower correlation is assigned. The distance threshold can be configured according to the component size, sensor density, and array spacing to adapt to different power plant structures.

[0092] The candidate vector generation module 29 can be used to generate multiple abnormal candidate vectors from abnormal candidate parameters based on positional correlation, wherein each abnormal candidate vector corresponds to two or more abnormal candidate parameters whose positional correlation is greater than a preset positional threshold.

[0093] In this embodiment, the candidate vector generation module 29 aggregates anomalous candidate parameters that are spatially close to each other or clustered in the same region into anomalous candidate vectors. Specifically, an adjacency graph can be constructed based on a location correlation threshold, and clustering / connected component decomposition can be performed on nodes that meet the threshold to obtain multiple clusters of anomalous candidate parameters; each cluster outputs an anomalous candidate vector. The anomalous candidate vector can be composed of the current value of the parameters within the cluster, anomalous similarity, correlation summary, and spatial features (cluster center, coverage area, etc.), so that the third machine learning model can learn anomalous patterns at a higher level.

[0094] Furthermore, in this embodiment, the candidate vector generation module 29 can divide the abnormal candidate parameters into multiple abnormal candidate parameter groups based on positional relevance. Each abnormal candidate parameter group includes two or more abnormal candidate parameters whose positional relevance to each other is higher than a preset position threshold. For each abnormal candidate parameter group, the abnormal weight of the corresponding abnormal candidate parameter is determined using the abnormal similarity of the abnormal candidate parameters in the abnormal candidate parameter group. For each abnormal candidate parameter group, an abnormal candidate vector of the abnormal candidate parameter group is generated based on the value, abnormal weight, and position information of each abnormal candidate parameter in the abnormal candidate parameter group.

[0095] In this embodiment, the candidate vector generation module 29 first completes grouping, then calculates anomaly weights within each group based on anomaly similarity (e.g., using normalized values ​​as weights, or assigning greater weights to those exceeding anomaly thresholds), and weights and aggregates the parameter values ​​within each group to form fusion features. Simultaneously, it encodes location information into group-level spatial features and concatenates them to obtain the final anomaly candidate vector. This approach highlights the "main anomaly source" before inputting it into the probability model while preserving spatial structure information, thereby improving the discriminativeness and stability of the anomaly state probability output.

[0096] The probability calculation module 30 can be used to calculate the probability of an abnormal state of a target photovoltaic panel based on anomaly candidate vectors using a third machine learning model.

[0097] In this embodiment, the probability calculation module 30 infers from one or more abnormal candidate vectors and outputs the probability value of the target photovoltaic panel being in an abnormal state. The third machine learning model can be a probabilistic classification model (such as logistic regression, gradient boosting tree, neural network, etc.). When there are multiple abnormal candidate vectors, the maximum value, weighted aggregation, or attention aggregation can be used to obtain the photovoltaic panel-level probability output. The abnormal state probability output by this module can be used for alarm classification, priority ranking of operation and maintenance dispatch orders, or combined with historical trends for continuous monitoring, thereby improving the real-time performance and accuracy of power plant status detection.

[0098] Example 3 The above describes the internal functions and structure of a photovoltaic panel condition detection device based on a machine learning model, which can be implemented as an electronic device. Figure 3 A schematic diagram illustrating the structure of an embodiment of the electronic device provided in this application. (See attached diagram.) Figure 3 As shown, the electronic device includes a memory 31 and a processor 32.

[0099] Memory 31 is used to store programs. In addition to the programs described above, memory 31 can also be configured to store various other data to support operation on the electronic device. Examples of this data include instructions for any application or method used to operate on the electronic device, contact data, phonebook data, messages, pictures, videos, etc.

[0100] The memory 31 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0101] The processor 32 is not limited to a processor (CPU), but may also be a graphics processing unit (GPU), a field-programmable gate array (FPGA), an embedded neural network processor (NPU), or an artificial intelligence (AI) chip. The processor 32 is coupled to the memory 31 and executes the program stored in the memory 31 to perform the photovoltaic panel state detection method based on the machine learning model described in Embodiment 1 above.

[0102] Furthermore, such as Figure 3 As shown, the electronic device may also include other components such as a communication component 33, a power supply component 34, an audio component 35, and a display 36. Figure 3 The diagram only shows some components and does not mean that the electronic device includes only these components. Figure 3 The components shown.

[0103] Communication component 33 is configured to facilitate wired or wireless communication between electronic devices and other devices. The electronic devices can access wireless networks based on communication standards, such as WiFi, 3G, 4G, or 5G, or combinations thereof. In one exemplary embodiment, communication component 33 receives broadcast signals or broadcast-related information from an external broadcast management system via a broadcast channel. In one exemplary embodiment, communication component 33 also includes a near-field communication (NFC) module to facilitate short-range communication. For example, the NFC module may be implemented based on radio frequency identification (RFID) technology, Infrared Data Association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology, and other technologies.

[0104] Power supply component 34 provides power to various components of the electronic device. Power supply component 34 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to the electronic device.

[0105] Audio component 35 is configured to output and / or input audio signals. For example, audio component 35 includes a microphone (MIC) configured to receive external audio signals when the electronic device is in an operating mode, such as call mode, recording mode, and voice recognition mode. The received audio signals may be further stored in memory 31 or transmitted via communication component 33. In some embodiments, audio component 35 also includes a speaker for outputting audio signals.

[0106] Display 36 includes a screen, which may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen can be implemented as a touchscreen to receive input signals from a user. The touch panel includes one or more touch sensors to sense touches, swipes, and gestures on the touch panel. The touch sensors can sense not only the boundaries of the touch or swipe action but also the duration and pressure associated with the touch or swipe operation.

[0107] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0108] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A photovoltaic panel condition detection method based on a machine learning model, characterized in that, include: Collect multiple parameter data of the target photovoltaic panel, including at least the operating parameter data of the target photovoltaic panel collected by multiple sensors and environmental data collected by outdoor environment sensors; Parameter type identification is performed on the multiple parameter data to obtain the parameter type corresponding to each parameter data; For each parameter type, feature transformation is performed on the corresponding parameter data to generate the corresponding parameter vector; For multiple parameter vectors, the first machine learning model is used to calculate the anomaly similarity corresponding to each parameter vector. For a first parameter vector whose abnormal similarity is greater than a preset abnormal similarity threshold, a second machine learning model is used to calculate the correlation between each parameter vector and other parameter vectors. Based on the correlation degree, abnormal candidate parameters are determined, wherein the abnormal candidate parameters include a first parameter vector and at least one second parameter vector whose correlation degree with at least one of the first parameter vectors is greater than a preset correlation degree threshold and whose abnormal similarity is less than a preset abnormal similarity threshold. Obtain the position information of the sensor corresponding to the abnormal candidate parameter on the target photovoltaic panel; Based on the location information, the location correlation between the anomaly candidate parameters is calculated; Based on the location correlation, multiple abnormal candidate vectors are generated from the abnormal candidate parameters, wherein each abnormal candidate vector corresponds to two or more abnormal candidate parameters whose location correlation is greater than a preset location threshold. Based on the anomaly candidate vector, a third machine learning model is used to calculate the probability of the target photovoltaic panel's abnormal state.

2. The method according to claim 1, characterized in that, The step of generating multiple anomaly candidate vectors from the anomaly candidate parameters based on the location correlation includes: Based on the location correlation, the abnormal candidate parameters are divided into multiple abnormal candidate parameter groups, and each abnormal candidate parameter group includes two or more abnormal candidate parameters whose location correlation with each other is higher than a preset location threshold. For each group of anomalous candidate parameters, the anomalous weight of the corresponding anomalous candidate parameter is determined by the anomalous similarity of the anomalous candidate parameters in the group. For each abnormal candidate parameter group, an abnormal candidate vector is generated based on the values ​​of each abnormal candidate parameter in the abnormal candidate parameter group, the abnormal weight, and the position information.

3. The method according to claim 1, characterized in that, The operating parameter data includes at least one of the following: output voltage, output current, output power, component temperature, inverter operating status, daily power generation or instantaneous power generation; and / or, the environmental data includes at least one of the following: ambient temperature, ambient humidity, wind speed, rainfall or light intensity.

4. The method according to claim 1, characterized in that, The parameter type identification includes labeling each parameter data according to at least one of the data source identifier, dimensional information, and sampling period to obtain the parameter type.

5. The method according to claim 1, characterized in that, The feature transformation includes performing at least one preprocessing operation on the parameter data and extracting features to generate a parameter vector. The preprocessing operation includes at least one of denoising, missing value imputation, normalization, standardization, and sliding window segmentation.

6. The method according to claim 1, characterized in that, The step of using the second machine learning model to calculate the correlation between each parameter vector and other parameter vectors includes: for the first parameter vector, calculating the correlation between the first parameter vector and each of the other parameter vectors to obtain a set of correlations corresponding to the first parameter vector.

7. The method according to claim 1, characterized in that, The location information includes at least one of the following: the sensor's installation coordinates on the target photovoltaic panel corresponding to the abnormal candidate parameters, the installation area identifier, or the relative distance information.

8. The method according to claim 1, characterized in that, The location correlation is determined based on the spatial distance between the sensors corresponding to the abnormal candidate parameters, and the location correlation between abnormal candidate parameters whose spatial distance is less than a preset distance threshold is higher than a preset location threshold.

9. A photovoltaic panel condition detection device based on a machine learning model, characterized in that, include: The acquisition module is used to acquire multiple parameter data of the target photovoltaic panel. The multiple parameter data includes at least the operating parameter data of the target photovoltaic panel acquired by multiple sensors and the environmental data acquired by outdoor environment sensors. The type recognition module is used to identify the parameter type of multiple parameter data and obtain the parameter type corresponding to each parameter data. The vector generation module is used to perform feature transformation on the corresponding parameter data for each parameter type to generate the corresponding parameter vector. The anomaly calculation module is used to calculate the anomaly similarity of each parameter vector using the first machine learning model for multiple parameter vectors. The correlation calculation module is used to calculate the correlation between each parameter vector and other parameter vectors using a second machine learning model for a first parameter vector whose abnormal similarity is greater than a preset abnormal similarity threshold. The candidate determination module is used to determine abnormal candidate parameters based on the correlation degree, wherein the abnormal candidate parameters include a first parameter vector and at least one second parameter vector whose correlation degree with at least one of the first parameter vectors is greater than a preset correlation degree threshold and whose abnormal similarity is less than a preset abnormal similarity threshold. The location acquisition module is used to acquire the location information of the sensor corresponding to the abnormal candidate parameter on the target photovoltaic panel; The location correlation calculation module is used to calculate the location correlation between abnormal candidate parameters based on location information; The candidate vector generation module is used to generate multiple abnormal candidate vectors from abnormal candidate parameters based on location correlation, wherein each abnormal candidate vector corresponds to two or more abnormal candidate parameters whose location correlation is greater than a preset location threshold. The probability calculation module is used to calculate the probability of abnormal states of a target photovoltaic panel based on anomaly candidate vectors using a third machine learning model.

10. An electronic device, characterized in that, include: Memory, used to store programs; A processor is configured to run the program stored in the memory, wherein the program executes the photovoltaic panel state detection method based on a machine learning model as described in any one of claims 1 to 8.

Images