AI-based photovoltaic power generation efficiency real-time monitoring and feedback system

Abstract

The invention provides an AI-based photovoltaic power generation efficiency real-time monitoring and feedback system, and the system comprises an edge end which collects electrical data, meteorological environment data, equipment data and photovoltaic panel image data in real time, carries out the real-time abnormality detection through a lightweight diagnosis model, generates a preliminary diagnosis result, and uploads the preliminary diagnosis result to a cloud end. The cloud constructs a digital twinborn model to calculate the deviation between theoretical output power and actual power, deploys a multi-modal AI analysis engine, integrates a power prediction model based on LSTM, an image analysis model based on OfficientNet-B3 and a root cause analysis model based on LightGBM, and achieves power prediction, visual diagnosis and fault cause analysis. According to the method, time sequence data, image features and multi-source diagnosis information are integrated, fault causes are output, a targeted operation and maintenance scheme is generated, and intelligent monitoring, accurate fault diagnosis and efficiency optimization feedback of the operation state of the photovoltaic power generation system are achieved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of photovoltaic power generation, in particular to an AI-based photovoltaic power generation efficiency real-time monitoring and feedback system. BACKGROUND

[0002] As an important part of clean energy, the stability and reliability of the operation efficiency of photovoltaic power generation have a key influence on energy utilization efficiency and power grid safety. At present, photovoltaic power stations generally use sensor data acquisition and monitoring systems for operation state monitoring, but the existing systems mostly rely on preset threshold alarms and single-dimensional data analysis methods, which are difficult to achieve comprehensive perception and accurate diagnosis of complex operation states.

[0003] In the prior art, photovoltaic monitoring systems often have the problem of insufficient multi-source data fusion, and time series data and visual data are processed independently, which cannot form a comprehensive judgment, resulting in a high rate of missed or false alarms in fault identification. At the same time, the abnormality detection lacks real-time and adaptive capabilities, and cannot combine theoretical generation models with actual outputs for dynamic deviation analysis, making it difficult to accurately locate the root cause of the fault. In addition, the system lacks intelligent predictive maintenance and efficiency optimization feedback mechanisms, and the operation and maintenance decisions rely on human experience, with a lagging response, which affects the continuous improvement of power generation efficiency. SUMMARY

[0004] In order to solve the technical problems mentioned in the background art, the application provides an AI-based photovoltaic power generation efficiency real-time monitoring and feedback system.

[0005] To this end, the technical solution adopted by the application is as follows: The AI-based photovoltaic power generation efficiency real-time monitoring and feedback system comprises an edge end and a cloud end, and comprises: M1: edge module, the edge end collects a data set in real time, and pre-processes the data set, the data set comprising time series data including electrical data, meteorological environmental data, equipment data, and image data of the surface of the photovoltaic panel; performing real-time anomaly detection based on the time series data and a preset lightweight diagnostic model to obtain a preliminary diagnosis result; uploading the data set and the preliminary diagnosis result to the cloud end; M2: cloud module, the cloud end comprising a digital twin model and a multi-modal AI analysis engine; The digital twin model is used to calculate the theoretical output power of the photovoltaic panel; and to calculate the deviation value between the theoretical output power and the actual output power to obtain the performance deviation; The multi-modal AI analysis engine comprises an LSTM-based power prediction model, an EfficientNet-B3-based image analysis model, and a LightGBM-based root cause analysis model; The power prediction model predicts the future power generation of the photovoltaic based on the time series to obtain a predicted output power; and calculates a deviation value between the predicted output power and an actual output power to obtain a prediction deviation; The image analysis model outputs a visual diagnosis result based on the image data; The root cause analysis model outputs a fault cause based on the time series data, the preliminary diagnosis result, the performance deviation, the prediction deviation and the visual diagnosis result; A fault solution is generated based on the fault cause, a preset fault knowledge base and a historical case base.

[0006] Further, the lightweight diagnosis model adopts a hybrid architecture combining a rule-based expert system and a lightweight machine learning model; The expert system performs abnormality detection on the time series data by defining a set of diagnosis rules containing physical constraints and empirical thresholds; The lightweight machine learning model is an unsupervised abnormality detection model constructed using an isolation forest algorithm; The output of the lightweight machine learning model is an abnormality score; The preliminary diagnosis result includes an abnormality type, an abnormality level, an abnormality occurrence timestamp and related feature values.

[0007] Further, the digital twin model is constructed based on design parameters and actual layout of the photovoltaic power station; The parameters of the digital twin model include electrical parameters of photovoltaic components, component installation parameters, array topology structure and a three-dimensional geometric model of obstructions.

[0008] Further, the calculation steps of the theoretical output power are as follows, The theoretical maximum output power is calculated based on a single diode model, and the calculation formula is as follows: Wherein, Pmax is the theoretical maximum output power, Pstd is the rated power under standard test conditions, Icurr is the current total irradiance of the photovoltaic component, Istd is the standard irradiance, Tcurr is the current temperature of the photovoltaic component, Tstd is the standard temperature, α is the power temperature coefficient; The theoretical maximum output power is corrected based on the loss of the photovoltaic system to obtain the theoretical output power, and the calculation formula is as follows: Wherein, Pth is the theoretical output power, for inverter efficiency, for cable transmission efficiency, for component mismatch loss coefficient, for dust shielding loss coefficient, for power limiting peak clipping loss coefficient.

[0009] Further, the power prediction model adopts an encoder-decoder architecture; The encoder is composed of two layers of LSTM units, responsible for extracting the time sequence feature representation of the input sequence; The decoder is composed of two layers of LSTM units, responsible for generating the predicted sequence; an attention mechanism is introduced between the encoder and the decoder; The power prediction model outputs the predicted power value through a fully connected layer; The power prediction model adopts mean square error as the main loss function, and introduces mean absolute percentage error as an auxiliary evaluation index.

[0010] Further, the image analysis model includes three branches, namely a fault detection branch, a fault classification branch and a fault positioning branch; The fault detection branch outputs a binary classification result through a global average pooling layer and a fully connected classification layer, and the binary classification result is normal or abnormal; The fault classification branch further classifies the fault type for the image detected as abnormal; The fault positioning branch adopts a target detection method to locate the specific position of the fault in the image, and outputs the bounding box coordinates and confidence.

[0011] Further, the loss function of the image analysis model The total loss function is designed as a multi-task learning, which is the weighted sum of the loss of each branch, wherein, The binary cross-entropy loss of the detection branch, The multi-label cross-entropy loss of the classification branch, The bounding box regression loss of the positioning branch adopts Smooth L1 loss, 、 And is the weight coefficient.

[0012] Further, the fault reason includes a fault reason list and a probability corresponding to each fault reason.

[0013] Compared with the prior art, the advantages of the present application are: 1. The application overcomes the limitations of single data dimension and one-sided diagnosis in the prior art by the synergy of edge and cloud architecture, integrating time series data, image features and simulation results of digital twin models, and using a multi-modal AI analysis engine for cross-validation and comprehensive analysis, thereby achieving more comprehensive and accurate identification and root cause positioning of faults.

[0014] 2. The application not only realizes real-time anomaly detection and visual diagnosis, but also combines power prediction and root cause analysis to actively predict performance trends and automatically generate targeted operation and maintenance schemes, transforming traditional passive alarms into proactive predictive maintenance and performance optimization, improving operation efficiency and reliability of the power generation system.

[0015] 3. The application calculates the theoretical deviation by introducing digital twin technology and combines advanced machine learning models such as LightGBM for root cause reasoning. The system can deeply analyze the root cause of performance loss and output probabilistic fault causes, providing high-value decision support for operation and maintenance personnel and reducing over-reliance on human experience. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0017] Fig. 1 is a schematic diagram of the system workflow of the present application; Fig. 2 is a schematic diagram of the data structure set of the present application; Fig. 3 is a schematic diagram of the multi-modal AI analysis engine structure of the present application. DETAILED DESCRIPTION

[0018] To achieve the above purpose, the present application is implemented by the following technical solutions. The present application provides an AI-based photovoltaic power generation efficiency real-time monitoring and feedback system, please refer to Figs. 1 to 3 , including: M1: edge module, the edge end collects data set in real time, and pre-processes the data set, the data set includes time series data such as electrical data, meteorological environment data, equipment data, and image data of the surface of the photovoltaic panel; based on the time series data and a preset lightweight diagnostic model, real-time anomaly detection is performed to obtain preliminary diagnostic results; the data set and the preliminary diagnostic results are uploaded to the cloud end; The edge-end configuration multi-type sensor and data acquisition equipment realize the data acquisition of the key nodes of the photovoltaic power generation system, including photovoltaic array, inverter, bus box, etc., and the specific operation is as follows: The electrical data acquisition is obtained through current sensor, voltage sensor and power analyzer, and the acquisition frequency is set to once per second. The acquired electrical parameters include DC side voltage, DC side current, AC side voltage, AC side current, active power, reactive power, power factor and cumulative value of electric energy; The meteorological environmental data is obtained through the weather station, including solar irradiance, environmental temperature, photovoltaic module temperature, wind speed, wind direction, humidity and atmospheric pressure, and the acquisition frequency is once per minute; The equipment data is obtained through the inverter communication interface and sensor, including inverter running state, inverter temperature, inverter efficiency, DC bus box state and circuit breaker state, and the acquisition frequency is once every ten seconds; The image data acquisition is realized through the deployment of high-definition cameras at key positions of the photovoltaic array. The camera adopts waterproof and dustproof design, and has automatic focusing and light self-adaptive functions; The image acquisition frequency is set according to actual needs. Under normal circumstances, panoramic image is collected once per hour, and when abnormality is detected, it is automatically switched to collect once per minute; The image resolution is set to no less than 1920x1080 pixels, which ensures that the fault features such as dirt, cracks and hot spots on the surface of the photovoltaic panel can be clearly captured.

[0019] For time series data, the preprocessing steps include data cleaning, outlier detection, missing value filling and data standardization, and the specific operation is as follows: Data cleaning identifies and removes obviously incorrect data points by setting a reasonable threshold range, such as negative power values, temperature values exceeding physical limits, etc. Outlier detection uses a statistical method based on sliding window to calculate the deviation of each data point from its neighborhood data. When the deviation exceeds the set three standard deviations, it is marked as an outlier. Missing value filling selects different strategies according to the type and proportion of missing data. For a small amount of missing data in a short time, linear interpolation method is used for filling; for long-term continuous missing, a filling method based on historical same period data is used, that is, the average value of the same period of the previous days is used for filling; Data standardization uses Z-score standardization method to convert data of different dimensions to the same scale range; For image data, the preprocessing includes image denoising, contrast enhancement, size normalization and interested region extraction, and the specific operation is as follows: Image denoising uses Gaussian filter to remove noise generated in the acquisition process; The contrast enhancement improves the visual effect of the image by the histogram equalization method, so that the fault features are more obvious. The size normalization adjusts the images of different resolutions to a fixed size to meet the input requirements of the subsequent neural network model. The region of interest extraction automatically identifies and crops the photovoltaic panel region through the image segmentation algorithm, excluding background interference.

[0020] Considering the computing resource limitations of edge devices, the lightweight diagnostic model adopts a hybrid architecture combining a rule-based expert system and a lightweight machine learning model. The expert system is used to detect explicit threshold overrun anomalies, and the machine learning model is used to identify complex abnormal patterns. The expert system defines a set of diagnostic rules containing physical constraints and empirical thresholds to detect anomalies in time series data. In this embodiment, the specific operation of the expert system is as follows: When the DC side voltage exceeds 1.1 times the rated open-circuit voltage of the photovoltaic module or is lower than 0.7 times the rated maximum power point voltage, it is determined to be a voltage anomaly; when the inverter efficiency is lower than the preset efficiency threshold, it is determined to be an inverter efficiency anomaly, and the efficiency threshold is preferably 95%; when the difference between the photovoltaic module temperature and the ambient temperature exceeds the preset temperature difference threshold, it is determined to be a component overheating anomaly, and the temperature difference threshold is preferably 35℃.

[0021] The lightweight machine learning model is an unsupervised anomaly detection model constructed using the Isolation Forest algorithm. The Isolation Forest algorithm is an efficient unsupervised anomaly detection algorithm, whose core mechanism is based on the assumption that abnormal data points have few and different characteristics, so they are more easily isolated by random partitioning in the feature space.

[0022] The Isolation Forest algorithm belongs to the category of ensemble learning and is implemented by constructing multiple isolation trees. Its key innovation is not to model normal data, but to directly quantify the path length required for each data instance to be isolated by random partitioning rules. The shorter the path length, the higher the likelihood that the instance is judged to be abnormal. Finally, the algorithm calculates a normalized anomaly score by integrating the path lengths of all trees.

[0023] The main advantage of the Isolation Forest algorithm is its linear time complexity and low memory requirements, making it particularly suitable for processing high-dimensional large data sets and resource-constrained edge computing scenarios.

[0024] In this embodiment, the number of isolation trees is preferably 100, and the number of samples for each tree is preferably 256 samples. Samples refer to data in the data set, such as voltage, current, power, power factor, etc. The output of a lightweight machine learning model is an anomaly score, calculated as follows: in, These are abnormal scores. For the first The path length of each sample in the isolation tree. The expected value of the path length for all isolated trees. for Average path length of each sample; When abnormal scores If the value exceeds the preset abnormal threshold, it is judged as abnormal; The typical value for the anomaly threshold is 0.5, based primarily on the following two aspects: Anomalies in photovoltaic systems often develop from minor deviations into serious failures. Setting the threshold to 0.5 can trigger detection in the early stages of anomalies, facilitating timely intervention and preventative maintenance, and preventing small problems from escalating into major failures. Using 0.5 as the starting threshold for anomaly detection leaves ample room for differentiation in subsequent severity levels, enabling operations and maintenance personnel to take differentiated response measures based on different levels and achieve refined management.

[0025] In actual deployment, the abnormal threshold can be fine-tuned and optimized based on the historical operating data and failure cases of the specific photovoltaic power station.

[0026] Preliminary diagnostic results include the type of abnormality, the level of abnormality, the timestamp of the abnormality, and related characteristic values; Abnormality types include voltage abnormality, current abnormality, efficiency abnormality, temperature abnormality, etc. The abnormality level is divided into four levels: normal, mild, moderate, and severe. The determination of anomaly level is mainly based on a comprehensive evaluation of two dimensions: anomaly score and parameter deviation. The specific classification criteria are as follows: Normal: The abnormality score is less than 0.5, and the deviations of all monitoring parameters are within the normal fluctuation range, for example, the power deviation is less than 5% and the temperature deviation is less than 5℃; Minor: Anomaly score between 0.5 and 0.6, or slight deviations in a single parameter, such as power deviation of 5% to 10%, voltage deviation of 3% to 5%, component temperature deviation of 5°C to 15°C, and inverter efficiency decrease of 1% to 3%; Medium: Anomaly score between 0.6 and 0.7, or multiple parameters deviating simultaneously, or a single parameter deviating moderately, such as power deviation of 10% to 20%, voltage deviation of 5% to 10%, component temperature deviation of 15°C to 25°C, and inverter efficiency decrease of 3% to 5%; Severe: abnormal score higher than 0.7, or key parameters deviate significantly, for example, power deviation more than 20%, voltage deviation more than 10%, component temperature deviation more than 25℃, inverter efficiency drop more than 5%, or trigger device protection action; The relevant feature value refers to the actual value of the key monitoring parameter when the abnormality is triggered.

[0027] M2 cloud module, the cloud includes a digital twin model and a multi-modal AI analysis engine; The digital twin model is used to calculate the theoretical output power of the photovoltaic; and calculate the deviation between the theoretical output power and the actual output power to obtain the performance deviation; The multi-modal AI analysis engine includes an LSTM-based power prediction model, a convolutional neural network-based image analysis model, and a LightGBM-based root cause analysis model; The power prediction model predicts the future power generation of the photovoltaic based on the time series to obtain a predicted output power; and calculates the deviation between the predicted output power and the actual output power to obtain a prediction deviation; The image analysis model outputs a visual diagnosis result based on the image data; The root cause analysis model outputs a fault cause based on the time series data, the preliminary diagnosis result, the performance deviation, the prediction deviation, and the visual diagnosis result; Generate an operation and maintenance strategy based on the predicted power, the visual diagnosis result, and the fault cause, the operation and maintenance strategy including a fault solution and an efficiency improvement scheme.

[0028] Based on the design parameters and the actual layout of the photovoltaic power station, a refined photovoltaic array digital twin model is constructed; The parameters of the digital twin model include the electrical parameters of the photovoltaic components, the component installation parameters, the array topology structure, and the three-dimensional geometric model of the obstructions.

[0029] The electrical parameters of the photovoltaic components are the open-circuit voltage, the short-circuit current, the maximum power point voltage, the maximum power point current, and the temperature coefficient under standard test conditions; The component installation parameters include the tilt angle and the azimuth angle; The array topology structure includes the number of strings, the number of components per string, and the number of parallel strings; components, also commonly known as solar panels, are the smallest, indivisible power generation units in a photovoltaic power generation system; connecting multiple components in series forms a string.

[0030] Based on the single-diode model, the theoretical maximum output power of the photovoltaic component under the current working condition is calculated, and the calculation formula is as follows: Wherein, Pmax,th Pn,th Pn Pn,0 Tn Tn,0 α Pn,th Pn,th Pn,th Pn,th ηinv ηc ηm ηd ηp Pn,th

[0031] The power prediction model adopts an encoder-decoder architecture to support multi-step prediction. The encoder consists of two layers of LSTM units, responsible for extracting the temporal feature representation of the input sequence. The decoder also consists of two layers of LSTM units, responsible for generating the predicted sequence. An attention mechanism is introduced between the encoder and the decoder, allowing the decoder to focus on the relevant information in the input sequence for the current prediction step. The power prediction model outputs the predicted power value through a fully connected layer.

[0032] The power prediction model adopts mean squared error as the main loss function, while introducing mean absolute percentage error as an auxiliary evaluation indicator: where N Pn is the predicted power value of the th training sample, and actual power value of the training sample.

[0033] The power prediction model training adopts an Adam optimizer with an initial learning rate of 0.001, adopts a cosine annealing learning rate scheduling strategy, and generates training data in a sliding window manner, with a window size of the input sequence length plus the prediction sequence length. To prevent overfitting, a Dropout regularization is adopted with a dropout rate of 0.2 and an early stopping strategy.

[0034] The image analysis model adopts an EfficientNet-B3-based backbone network for feature extraction, and builds a multi-task learning framework based on it, simultaneously realizing three tasks of fault detection, fault classification and fault positioning; The fault detection branch adds a global average pooling layer and a fully connected classification layer on the last feature map of the backbone network, and outputs a binary classification result, which is normal or abnormal; The fault classification branch further classifies the fault type for the image detected as abnormal, and the fault types include dust pollution, bird droppings obstruction, leaf obstruction, hot spot, crack, snail pattern, PID attenuation, glass breakage, etc. The fault positioning branch adopts a target detection method to locate the specific position of the fault in the image, and outputs the bounding box coordinates and confidence.

[0035] The loss function of the image analysis model is designed as the total loss function of multi-task learning, which is the weighted sum of the loss of each branch: wherein, the binary cross-entropy loss of the detection branch, the multi-label cross-entropy loss of the classification branch, the bounding box regression loss of the positioning branch, adopting a Smooth L1 loss, , and are weight coefficients, and are set as 1.0, 1.0 and 0.5 respectively; The visual diagnosis result includes fault detection flag, fault type, fault position, and fault severity score. The fault detection flag represents the binary determination result of whether there is a fault in the current image, i.e. normal or abnormal. The fault type is the specific fault category identified and its corresponding confidence, supporting the simultaneous identification of multiple faults. The fault position coordinates are the specific position of the fault in the image, represented in the form of a bounding box, including the coordinates of the upper left corner of the bounding box, the width, the height and the positioning confidence. The fault severity score is a quantitative evaluation of the impact of the fault, represented by a value from 0 to 100, with a higher value indicating a more serious fault; it is calculated based on fault type weight and fault area proportion; different types of faults have different basic weights, for example, the weight of hot spot is higher than that of dust pollution, and the proportion of fault area to the total area of the component is also considered, and the final score is calculated by weighting the two.

[0036] The root cause analysis model adopts a multi-classification mode, and the category label is a pre-defined fault cause set, including component aging attenuation, component hot spot fault, component hidden crack fault, junction box poor contact, cable insulation aging, inverter MPPT failure, inverter over-temperature protection, inverter communication fault, serious dust pollution, shading influence, power grid side power cut, weather mutation influence, sensor data anomaly, etc. The hyperparameters of the root cause analysis model are optimized by the Bayesian optimization method, and the key parameters include the number of leaf nodes, the maximum depth, the learning rate, the minimum leaf sample number, and the feature sampling ratio and the sample sampling ratio. The training data uses a historical fault case library, which contains samples with labeled fault causes. To solve the class imbalance problem, the SMOTE oversampling method is used to expand the minority class samples, or the class weight is set to balance. The output of the root cause analysis model is a list of fault causes and the corresponding probability of each fault cause, for example, 85% probability of shadow shading of the 5th string in the 3rd combiner box, 10% probability of connector looseness, and 5% probability of internal component failure.

[0037] Based on the pre-set fault knowledge base and the historical case library, a corresponding solution is recommended for each fault cause.

[0038] The AI-based photovoltaic power generation efficiency real-time monitoring and feedback system proposed in the application constructs a collaborative architecture of edge perception and cloud judgment, deeply integrates time series data, image information and digital twin simulation, uses a multi-modal AI analysis engine to comprehensively diagnose and predict the photovoltaic system, effectively solves the problems of single data dimension, one-sided fault diagnosis and delayed operation and maintenance response in the prior art, and realizes intelligent management of the whole process from real-time monitoring, accurate diagnosis to active operation and maintenance feedback.

[0039] In summary, the application constructs a multi-modal AI analysis system of edge and cloud collaboration, deeply integrates time series data, visual information and digital twin models, realizes intelligent management of the whole link from real-time monitoring, accurate diagnosis to active operation and maintenance of the photovoltaic system, and significantly improves the comprehensiveness of fault diagnosis, the timeliness of operation and maintenance response and the reliability of system efficiency.

[0040] The above merely provides the specific implementation of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An AI-based real-time monitoring and feedback system for photovoltaic power generation efficiency, the system comprising an edge terminal and a cloud terminal, characterized in that, include: M1: Edge module, which collects datasets in real time and preprocesses the datasets. The datasets include time-series data including electrical data, meteorological environmental data, and equipment data, as well as image data of the photovoltaic panel surface. Based on the time-series data and a preset lightweight diagnostic model, real-time anomaly detection is performed to obtain preliminary diagnostic results; the dataset and the preliminary diagnostic results are then uploaded to the cloud. M2: Cloud module, which includes a digital twin model and a multimodal AI analysis engine; The digital twin model is used to calculate the theoretical output power of photovoltaics; and to calculate the deviation between the theoretical output power and the actual output power to obtain the performance deviation; The multimodal AI analysis engine includes an LSTM-based power prediction model, an EfficientNet-B3-based image analysis model, and a LightGBM-based root cause analysis model. The power prediction model predicts the future power generation of photovoltaics based on the time series to obtain the predicted output power; and calculates the deviation between the predicted output power and the actual output power to obtain the prediction deviation; The image analysis model outputs visual diagnostic results based on the image data; The root cause analysis model outputs the cause of the failure based on the time series data, preliminary diagnostic results, performance deviation, prediction deviation, and visual diagnostic results. A fault solution is generated based on the fault cause, a preset fault knowledge base, and a historical case library.

2. The system according to claim 1, characterized in that, The lightweight diagnostic model adopts a hybrid architecture that combines a rule-based expert system with a lightweight machine learning model. The expert system performs anomaly detection on time-series data by defining a set of diagnostic rules that include physical constraints and empirical thresholds; The lightweight machine learning model is an unsupervised anomaly detection model constructed using the isolated forest algorithm; The output of the lightweight machine learning model is an anomaly score; The preliminary diagnostic results include the abnormality type, abnormality level, abnormality occurrence timestamp, and related feature values.

3. The system according to claim 1, characterized in that, The digital twin model is constructed based on the design parameters and actual layout of the photovoltaic power station; The parameters of the digital twin model include the electrical parameters of the photovoltaic module, the module installation parameters, the array topology, and the three-dimensional geometric model of the obstruction.

4. The system according to claim 3, characterized in that, The calculation steps for the theoretical output power are as follows: The theoretical maximum output power is calculated based on a single diode model, using the following formula: in, This is the theoretical maximum output power. Rated power under standard test conditions. This represents the current total irradiance of the photovoltaic module. Standard irradiance, The current temperature of the photovoltaic module. Standard temperature The power temperature coefficient; The theoretical output power is obtained by correcting the theoretical maximum output power based on the losses of the photovoltaic system, and the calculation formula is as follows: in, Theoretical output power, For inverter efficiency, For cable transmission efficiency, This represents the component mismatch loss factor. The dust-induced loss coefficient. This is the peak shaving loss coefficient due to power rationing.

5. The system according to claim 4, characterized in that, The power prediction model adopts an encoder and decoder architecture; The encoder consists of two LSTM units, which are responsible for extracting the temporal feature representation of the input sequence; The decoder consists of two LSTM layers and is responsible for generating the prediction sequence; an attention mechanism is introduced between the encoder and the decoder. The power prediction model outputs predicted power values ​​through a fully connected layer; The power prediction model uses mean squared error as the main loss function, and introduces mean absolute percentage error as an auxiliary evaluation index.

6. The system according to claim 5, characterized in that, The image analysis model includes three branches: fault detection branch, fault classification branch, and fault location branch. The fault detection branch outputs a binary classification result through a global average pooling layer and a fully connected classification layer, and the binary classification result is either normal or abnormal. The fault classification branch further classifies the fault type for images that are detected as abnormal; The fault localization branch uses a target detection method to locate the specific location of the fault in the image and outputs the bounding box coordinates and confidence score.

7. The system according to claim 6, characterized in that, The loss function of the image analysis model The total loss function is designed for multi-task learning, and is a weighted sum of the losses of each branch. in, To detect the binary cross-entropy loss of the branch, For the multi-label cross-entropy loss of the classification branch, To determine the bounding box regression loss for the localization branch, Smooth L1 loss is used. , and These are the weighting coefficients.

8. The system according to claim 7, characterized in that, The causes of failure include a list of causes of failure and the probability corresponding to each cause of failure.

Images