Intelligent operation and maintenance methods and systems for power generation equipment based on model-driven autonomous selection.

Abstract

The application relates to the technical field of power station operation and maintenance, and discloses a power generation equipment intelligent operation and maintenance method and system based on model autonomous selection, which comprises the following steps: acquiring photovoltaic power station multi-modal operation and maintenance data, and generating a multi-modal operation and maintenance data set; inputting each mode data of the multi-modal operation and maintenance data set into a corresponding module for feature extraction; inputting the extracted feature vectors into a comparative learning network for cross-modal alignment, outputting an executable decision result by combining a knowledge graph through retrieval enhancement generation technology; constructing a privacy protection training framework through federated learning, inputting the executable decision result into a digital twin system for strategy verification, and generating a trained multi-modal large model, so as to select a corresponding module in the trained multi-modal large model based on feature data of the photovoltaic power station for real-time monitoring. The present application solves the problems of poor accuracy and untimely reaction in the traditional photovoltaic power station operation and maintenance process, and can timely and accurately perform power station operation and maintenance.

Description

Technical Field

[0001] This application relates to the field of power plant operation and maintenance technology, such as a model-based intelligent operation and maintenance method and system for power generation equipment. Background Technology

[0002] The power generation industry is the core of global energy supply and is undergoing a transformation from traditional fossil fuels to renewable energy. It faces challenges such as carbon neutrality goals, technological upgrades, operation and maintenance optimization, and grid stability. At the same time, it also needs to promote the large-scale application of digital and intelligent technologies.

[0003] As a crucial component of clean energy, photovoltaic (PV) power generation plays a vital role in promoting energy structure transformation through its efficient and stable operation. However, the current operation and maintenance (O&M) of PV power plants faces numerous systemic challenges. The performance degradation of components due to changes in the natural environment and electrical faults caused by equipment aging, intertwined with external factors such as grid dispatch requirements and market trading mechanisms, make traditional O&M models inadequate for meeting the demands of precise management. Especially with the continuous expansion of PV power plant construction and the increasing diversification of application scenarios, O&M methods relying solely on manual inspections and experience-based judgment suffer from low efficiency, poor accuracy, and delayed response, making them unsuitable for the industry's development requirements.

[0004] Chinese invention patent application number CN202411400293.2 discloses a photovoltaic (PV) operation and maintenance (O&M) AI assistant system and application method based on a deep learning large-scale model, relating to the field of PV power plant technology. The system includes acquiring user-input text, images, voice information, questions, and historical O&M data. Using a multimodal fusion model, this information is integrated and features are extracted, and semantic classification technology is used to categorize it into four types: training, access, analysis, and reporting data. Furthermore, this categorized data, along with historical data, is input into a comprehensive data processing model. This model integrates pre-training, database access, big data analysis, and report generation functions to process and generate detailed O&M reports. Finally, a natural language processing model summarizes the information, forming comprehensive information, which is then sent to the user's device to guide O&M decisions. This invention optimizes the O&M management process of PV power plants through intelligent processing.

[0005] Chinese invention patent application number CN202510459081.X discloses a cloud-based intelligent operation and maintenance method and system for photovoltaic power plants, relating to the field of photovoltaic power plant operation and maintenance management technology. This cloud-based intelligent operation and maintenance method for photovoltaic power plants includes the following steps: obtaining characteristic data of operating status data and parameter data and storing them in a cloud platform; using a solar energy conversion efficiency model to evaluate the power generation performance of photovoltaic equipment in real time; determining the optimal cleaning cycle for photovoltaic equipment; and inputting the detection results into a photovoltaic equipment operation fault anomaly database for fault identification. This invention uses a photoelectric conversion model to evaluate power generation efficiency in real time, enabling operation and maintenance personnel to understand the power generation status of the photovoltaic power plant, avoiding frequent cleaning or excessively long cleaning intervals, thereby improving the ability to detect and identify fault data, and ultimately extending the service life and operating efficiency of the equipment, thus improving the operation and maintenance efficiency of the photovoltaic power plant.

[0006] However, the above method has the following limitations:

[0007] 1) Insufficient depth of multimodal data fusion. Existing technologies mostly use simple feature splicing or rule matching, failing to fully explore the deep correlations between different modalities such as electrical parameters, thermal infrared imaging, and visual data, resulting in low accuracy in identifying complex faults.

[0008] 2) Lack of real-time performance and edge computing capabilities. Relying on centralized cloud computing makes it difficult to meet the low-latency fault response requirements of photovoltaic power plants, especially in scenarios with poor network conditions such as distributed photovoltaic or remote power plants.

[0009] 3) Insufficient support for knowledge reasoning and intelligent decision-making. Existing systems are mostly based on static rules or historical data matching, lacking dynamic knowledge graphs and causal reasoning capabilities, and thus unable to achieve closed-loop decision-making from "fault detection - root cause analysis - optimization suggestions".

[0010] 4) Lack of data security and privacy protection. The lack of effective federated learning or differential privacy protection mechanisms during cross-power plant data sharing poses a risk of sensitive operational data leakage.

[0011] 5) Limited adaptive optimization capability. The lack of a continuous learning mechanism makes it difficult to adapt to changing operation and maintenance needs under new photovoltaic modules or extreme weather conditions, resulting in insufficient model generalization ability.

[0012] In summary, existing technical solutions suffer from poor accuracy and slow response.

[0013] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this application. Summary of the Invention

[0014] To provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. This summary is not intended as a general commentary, nor is it intended to identify key / important components or describe the scope of protection of these embodiments, but rather as a prelude to the detailed description that follows.

[0015] The intelligent operation and maintenance method and system for power generation equipment based on model autonomous selection provided in this disclosure solves the problems of poor accuracy and untimely response in the traditional operation and maintenance process of photovoltaic power plants.

[0016] The intelligent operation and maintenance method for power generation equipment based on model-driven autonomous selection in this embodiment includes:

[0017] Acquire multimodal operation and maintenance data of photovoltaic power plants, classify them by combining historical fault records, and extract features of each modality using signal processing and computer vision techniques to generate a multimodal operation and maintenance dataset;

[0018] An infrared thermal imaging feature extraction module is constructed using the ViT model, a visual data processing module is constructed based on U-Net and YOLOv8, and an environmental data analysis module is constructed using linear regression. The modal data of the multimodal operation and maintenance dataset are input into the corresponding modules for feature extraction.

[0019] A knowledge graph module is constructed using the Neo4j graph database, and a multi-agent collaborative decision-making module is constructed based on the attention mechanism. The extracted feature vectors are input into a contrastive learning network for cross-modal alignment, and executable decision results are output by combining retrieval enhancement generation technology with the knowledge graph.

[0020] A power plant simulation module is constructed, a policy optimization module is built using multi-objective reinforcement learning, a decision evaluation module is built based on causal reasoning, and a privacy-preserving training framework is built through federated learning. The executable decision results are input into the digital twin system for policy verification, generating a trained multimodal large model, which is used by photovoltaic power plants to select the corresponding module in the trained multimodal large model based on feature data for real-time monitoring.

[0021] The intelligent operation and maintenance system for power generation equipment based on model-driven autonomous selection in this embodiment includes:

[0022] The acquisition module is used to acquire multimodal operation and maintenance data of photovoltaic power plants, classify them by combining historical fault records, and extract the features of each mode using signal processing and computer vision technology to generate a multimodal operation and maintenance dataset.

[0023] The processing module is used to construct an infrared thermal image feature extraction module using the ViT model, a visual data processing module based on U-Net and YOLOv8, and an environmental data analysis module using linear regression. The module inputs the modal data of the multimodal operation and maintenance dataset into the corresponding module for feature extraction.

[0024] The processing module is also used to construct a knowledge graph module using the Neo4j graph database, construct a multi-agent collaborative decision-making module based on the attention mechanism, input the extracted feature vectors into the contrastive learning network for cross-modal alignment, and output executable decision results by combining retrieval enhancement generation technology with the knowledge graph.

[0025] The processing module is also used to construct a power plant simulation module, construct a strategy optimization module using multi-objective reinforcement learning, construct a decision evaluation module based on causal reasoning, construct a privacy-preserving training framework through federated learning, input the executable decision results into the digital twin system for strategy verification, and generate a trained multimodal large model for photovoltaic power plants to select the corresponding module in the trained multimodal large model based on feature data for real-time monitoring.

[0026] The intelligent operation and maintenance method and system for power generation equipment based on model autonomous selection provided in this disclosure can achieve the following technical effects:

[0027] This disclosure can enhance the intelligence and multimodal fusion capabilities of photovoltaic power plant fault detection, strengthen the knowledge-driven and interpretable nature of operation and maintenance decisions, optimize adaptive operation and maintenance strategies in complex environments, improve the continuous learning and generalization capabilities of models, achieve real-time response and privacy protection through edge-cloud collaboration, and thus enable accurate and timely monitoring and operation and maintenance of photovoltaic power plants.

[0028] The above general description and the description below are exemplary and illustrative only and are not intended to limit this application. Attached Figure Description

[0029] One or more embodiments are illustrated by way of example with reference to the accompanying drawings. These illustrations and drawings do not constitute a limitation on the embodiments. Elements having the same reference numerals in the drawings are shown as similar elements. The drawings are not to be scaled. And wherein:

[0030] Figure 1 This is a flowchart illustrating a model-based intelligent operation and maintenance method for power generation equipment, provided in an embodiment of this disclosure.

[0031] Figure 2 This is a flowchart of an overall technical route provided by an embodiment of the present disclosure;

[0032] Figure 3 This is a network structure diagram of a multimodal data preprocessing and feature extraction module provided in an embodiment of this disclosure;

[0033] Figure 4 This is an overall framework diagram provided in an embodiment of the present disclosure;

[0034] Figure 5 This is a flowchart illustrating the actual deployment of a model according to an embodiment of this disclosure;

[0035] Figure 6 This is a schematic diagram of the structure of a model-based intelligent operation and maintenance system for power generation equipment, provided in an embodiment of this disclosure.

[0036] Figure 7 This is a schematic diagram of the structure of a model-based intelligent operation and maintenance device for power generation equipment, provided in an embodiment of this disclosure. Detailed Implementation

[0037] To provide a more detailed understanding of the features and technical content of the embodiments of this disclosure, the implementation of the embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. The accompanying drawings are for illustrative purposes only and are not intended to limit the embodiments of this disclosure. In the following technical description, for ease of explanation, several details are used to provide a full understanding of the disclosed embodiments. However, one or more embodiments may still be implemented without these details. In other cases, well-known structures and devices may be simplified in their depiction to simplify the drawings.

[0038] The terms "first," "second," etc., used in the embodiments of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this disclosure described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion.

[0039] Unless otherwise stated, the term "multiple" means two or more.

[0040] In this embodiment of the disclosure, the character " / " indicates that the objects before and after it are in an "or" relationship. For example, A / B means: A or B.

[0041] The term "and / or" describes an association between objects, indicating that three relationships can exist. For example, A and / or B means: A or B, or A and B.

[0042] The term "correspondence" can refer to an association or binding relationship. The correspondence between A and B means that there is an association or binding relationship between A and B.

[0043] The following description, in conjunction with the accompanying drawings, illustrates the intelligent operation and maintenance method, system, equipment, and storage medium for power generation equipment based on model-driven autonomous selection, as provided in the embodiments of this disclosure.

[0044] Figure 1 This is a flowchart illustrating a model-based intelligent operation and maintenance method for power generation equipment, as provided in this embodiment.

[0045] like Figure 1 As shown, the intelligent operation and maintenance method for power generation equipment based on model-driven autonomous selection can include:

[0046] S101: Obtain multimodal operation and maintenance data of photovoltaic power plants, classify them by combining historical fault records, and extract the features of each mode using signal processing and computer vision technology to generate a multimodal operation and maintenance dataset.

[0047] S102 uses the ViT model to build an infrared thermal imaging feature extraction module, builds a visual data processing module based on U-Net and YOLOv8, and builds an environmental data analysis module through linear regression. The modal data of the multimodal operation and maintenance dataset are input into the corresponding modules for feature extraction.

[0048] S103 uses the Neo4j graph database to build a knowledge graph module and an attention mechanism to build a multi-agent collaborative decision-making module. The extracted feature vectors are input into a contrastive learning network for cross-modal alignment. The executable decision results are output by combining retrieval enhancement generation technology with the knowledge graph.

[0049] S104. Construct a power plant simulation module, use multi-objective reinforcement learning to construct a policy optimization module, construct a decision evaluation module based on causal reasoning, construct a privacy-preserving training framework through federated learning, input the executable decision results into the digital twin system for policy verification, and generate a trained multimodal large model for real-time monitoring of photovoltaic power plants by selecting the corresponding module in the trained multimodal large model based on feature data.

[0050] In the process of generating the trained multimodal large model based on the model-driven intelligent operation and maintenance method for power generation equipment in S101-S104, this disclosure is generated by analyzing and training different feature data. Therefore, in the specific use process, different analysis and processing modules can be used to process different feature data of photovoltaic power plants, so that the modules in the subsequently trained multimodal large model can monitor different feature data of photovoltaic power plants in real time, thereby enabling accurate and timely monitoring and operation and maintenance of photovoltaic power plants.

[0051] In some embodiments, the acquisition of multimodal operation and maintenance data of photovoltaic power plants, combined with historical fault records for category labeling, and the extraction of modal features using signal processing and computer vision techniques to generate a multimodal operation and maintenance dataset, includes:

[0052] Four types of heterogeneous data were collected using electrical parameter sensors, infrared thermal imagers, environmental sensors, and drone inspection equipment.

[0053] Based on historical fault records, four types of heterogeneous data are categorized and labeled, with the label set being Y = {"normal", "electrical fault", "hot spot", "fouling", "occlusion"}.

[0054] In some embodiments, feature extraction is performed on electrical signals in a multimodal operation and maintenance dataset, including:

[0055] For electrical signals, Kalman filtering is used to denoise the voltage V(t) and current I(t), satisfying the formula:

[0056]

[0057] in, Let A be the state vector, B be the state transition matrix, and u be the control matrix. t To control the input, w t Let Q be the process noise, and let Q be the noise covariance matrix.

[0058] By utilizing the multi-resolution analysis properties of wavelet transform, the denoised electrical signal is decomposed into approximate components A at different scales. j and detail component D j Extract energy features E at various scales j =∑∣D j | 2 And input the LSTM network modeling time dependencies, and update the hidden state as follows:

[0059] h t =LSTM(E j ,h t-1 ),

[0060] Final output electrical feature vector

[0061] In some embodiments, feature extraction of infrared thermal imaging data in a multimodal operation and maintenance dataset includes:

[0062] By calculating the current frame T img Mean of background image The difference is used to obtain the foreground temperature matrix T. fore :

[0063]

[0064] Using a pre-trained ViT-B / 16 model, the foreground temperature matrix after background removal is input into the ViT-B / 16 model, and the output is a 768-dimensional thermal anomaly feature vector f. ir =ViT(T) fore ).

[0065] In some embodiments, feature extraction is performed on environmental data in a multimodal operations and maintenance dataset, including:

[0066] Data collection includes light intensity S and ambient temperature T. a Environmental data such as humidity H and wind speed v were collected and normalized to the range of [-1,1] to eliminate the difference in dimensionality between different data dimensions.

[0067] By constructing environmental feature vectors Quantify environmental information into feature vector form;

[0068] Using a linear regression model Predicted theoretical power generation P pred Where w is the weight vector and b is the bias term, the theoretical power generation P is... pred With actual power P real By comparison, the efficiency deviation ΔP = P is calculated. pred -P real This is used to reflect the difference between the operating efficiency of a photovoltaic power plant under current environmental conditions and its expected efficiency.

[0069] In some embodiments, feature extraction is performed on drone data in a multimodal operation and maintenance dataset, including:

[0070] Image enhancement is performed using nonlocal mean filtering, and the filtered pixel value I′ i,j The calculation is as follows:

[0071]

[0072] Where Ω is the search window, used to find regions similar to the current pixel; Z is the normalization factor, ensuring that the brightness of the filtered image remains unchanged; σ is the standard deviation of the Gaussian kernel, controlling the weight of the similarity measure; and I... k,l This represents the pixel value at coordinates (k, l) in the image;

[0073] The U-Net model is used for semantic segmentation of RGB images, and the defect mask M∈{0,1} is output by the U-Net model. H×W Pixels with a value of 1 represent detected defect areas, and the defect area percentage is further calculated. Used to quantify the extent of damage to photovoltaic modules;

[0074] The YOLOv8 model is used for object detection, and the model outputs the detection results. Where c m Category confidence represents the probability that a detected target belongs to a certain category; bbox m= (x1, y1, x2, y2) represents the bounding box coordinates, which determine the position of the target in the image. Finally, the annotated image is used as the visualization result f. vis Output.

[0075] In some embodiments, the above-mentioned module for constructing a knowledge graph using the Neo4j graph database and a multi-agent collaborative decision-making module based on an attention mechanism are used. The extracted feature vectors are input into a contrastive learning network for cross-modal alignment. The executable decision result is output through retrieval enhancement generation techniques combined with the knowledge graph, including:

[0076] Electrical encoders E were built using the Transformer architecture respectively. elec Infrared encoder E ir Visual encoder E vis and environmental encoder E env Finally, the modal feature vectors z are output. elec ,z ir ,z vis ,z env ;

[0077] The InfoNCE loss function is used to distinguish between positive and negative sample pairs:

[0078]

[0079] Where z i ,z j For positive sample features, z k For negative sample features, τ is the temperature parameter, and K is the number of negative samples;

[0080] In the cross-modal knowledge distillation process, the teacher model T is trained on a large amount of labeled data to obtain a high-confidence output p. T The student model S minimizes the distillation loss. During the learning process, the distillation loss is:

[0081]

[0082] Where C is the number of fault categories, p S (c) Predict probabilities for the student model;

[0083] The knowledge graph is built using the Neo4j graph database, and includes fault types. Component Model Maintenance measures A knowledge graph of photovoltaic operation and maintenance at each node;

[0084] The retrieval enhancement generation mechanism is triggered after the model detects the fault feature x. By performing vector similarity retrieval between the feature vector and the entity in the knowledge graph, the relevant knowledge triple (f,x,a) is quickly located.

[0085] Electrical Intelligent Agent A elec Long Short-Term Memory (LSTM) networks based on time-series data are used to capture abnormal fluctuation patterns in electrical parameters; Infrared intelligent agent A ir Rapid localization and severity assessment of hotspot regions using YOLO series algorithms; Visual agent A vis Component appearance defects are identified using semantic segmentation technology, and each agent independently outputs a fault probability distribution p. elec ,p ir ,p vis ;

[0086] The weights α of each agent are calculated using a cross-modal attention mechanism. elec ,α ir ,α vis The calculation formula is:

[0087]

[0088] Where w m Given the modal weight vector, the final decision probability is:

[0089] p final =α elec p elec +α ir p ir +α vis p vis .

[0090] In some embodiments, the aforementioned power plant simulation module employs multi-objective reinforcement learning to construct a policy optimization module, causal reasoning to construct a decision evaluation module, and federated learning to construct a privacy-preserving training framework. The executable decision results are input into a digital twin system for policy verification, generating a trained multimodal large model for real-time monitoring of photovoltaic power plants by selecting the corresponding module from the trained multimodal large model based on feature data. This includes:

[0091] A twin model is constructed based on physical equations, with input environmental parameters f. env And the operation and maintenance action 'a', through refined mathematical modeling, outputs the predicted power generation P. sim and failure rate r sim ;

[0092] Multi-objective reinforcement learning is employed, with the state space being... Among them, f elec Represents electrical parameters, f ir For infrared thermal imaging data, fvis It is the result of visual detection, f env For environmental parameters, the action space is... The reward function is designed as follows:

[0093] r=λ1(P real -P sim )+λ2(1-r sim )-λ3c a ,

[0094] Wherein, λ1, λ2, and λ3 are weighting coefficients, optimized using cross-validation to balance power generation improvement, failure rate reduction, and operation and maintenance cost control; c a For the cost of action and the consumption of resources;

[0095] The PC algorithm is used to mine the causal relationship between fault variable F and maintenance action A from historical data, and a causal graph G(V,E) is constructed, where V is the variable node, including fault type, environmental factors, and operation and maintenance operation; E is the directed edge, which clarifies the causal dependency between variables.

[0096] For an action 'a' that has already been performed, the potential outcome 'Y' when the action has not been performed is simulated using the intervention model do(A=a). a＝0 Based on the causal graph structure, and keeping other variables constant, the causal effect τ = Y is calculated by modifying the intervention conditions in the causal model. a＝1 -Y a＝0 It is used to help assess the necessity of maintenance decisions and avoid over-maintenance or under-maintenance.

[0097] Multiple photovoltaic power plants are used as clients, each training its own modal encoder locally. Each client only uploads the model update gradient. The data is sent to a central server to prevent the leakage of raw data; the central server aggregates the gradients from each client, updates the global model, and distributes it, forming a closed-loop mechanism of training-update-feedback.

[0098] Adding Laplace noise ∈ ~Laplace(b) to the gradient, the perturbed gradient is:

[0099]

[0100] Wherein, ∈ controls the privacy budget, and b is the noise scale parameter. By setting the parameter, privacy attacks can be resisted while ensuring model performance.

[0101] Based on local data diversity and model improvement Calculate the power plant weight:

[0102]

[0103] Among them, D k The local data distribution is measured using KL divergence. k With global data distribution Based on the differences in weights, the gradient aggregation ratio of each client is dynamically adjusted, and power plants whose incentive data and training effects meet the preset conditions are included in federated learning.

[0104] In some embodiments, the above Figure 1 The methods may also include:

[0105] A multimodal large-scale model, trained locally at a photovoltaic power plant, is deployed to process electrical signals and drone images in real time, outputting real-time fault warnings.

[0106] The electrical feature vector f extracted at the edge elec Infrared thermal image feature vector f ir Image feature vector f vis Environmental feature vector f env Uploaded to the cloud, the trained multimodal large model deployed in the cloud is based on the Transformer architecture and integrates the attention mechanism to perform in-depth analysis of multi-source heterogeneous data. Through cross-modal alignment and knowledge distillation technology, it generates a detailed decision report containing fault type, maintenance suggestions, and causal analysis: report = {fault type, maintenance suggestions, causal analysis}.

[0107] A closed-loop feedback optimization mechanism is introduced. Based on actual operation and maintenance experience, the decision report is scored s∈[0,1]. When the score s is less than a preset threshold, it indicates that there is a deviation between the model decision and the actual situation, triggering the incremental learning process.

[0108] Use the newly labeled data {(x i ,y i The trained multimodal large model is updated using an elastic weight consolidation algorithm, which retains old knowledge while learning new data. The loss function is:

[0109]

[0110] Among them, cross-entropy loss Regularization terms used to optimize learning for new tasks Used to constrain changes in model parameters, where F(θ) is the Fisher information matrix of the old task parameters, θ0 is the old parameters, and λ is the balance coefficient;

[0111] The model, optimized through incremental learning, is redeployed to both the edge and the cloud. This enables the edge to acquire more accurate real-time analysis capabilities and improves the quality of cloud-based decision-making, thus forming a complete closed loop of monitoring, decision-making, and feedback. This ensures the system continues to evolve during long-term operation and guarantees operational decision-making capabilities.

[0112] Figure 2 This is a flowchart of an overall technical route provided by an embodiment of the present disclosure. Figure 3 This is a network structure diagram of a multimodal data preprocessing and feature extraction module provided in an embodiment of this disclosure. Figure 4 This is an overall framework diagram provided in an embodiment of the present disclosure. Figure 5 This is a flowchart illustrating the actual deployment of a model according to an embodiment of this disclosure, combined with... Figures 2 to 5 ,right Figure 1 The intelligent operation and maintenance method for power generation equipment based on model-driven autonomous selection is further described.

[0113] Specifically, S1, Construction of a Multimodal Operation and Maintenance Dataset for Photovoltaic Power Plants: Addressing the multi-source and heterogeneous nature of photovoltaic power plant operation and maintenance data, this invention constructs a four-modal dataset based on multi-sensor fusion technology, encompassing electrical parameters, infrared thermal imaging, environmental data, and UAV vision. Voltage and current signals are collected by electrical parameter sensors, temperature distribution matrices are obtained by an infrared thermal imager, meteorological parameters are recorded by environmental sensors, and images of the component surfaces are acquired by a dual-spectrum camera mounted on a UAV. The raw data obtained in S1 is input into an expert annotation system, categorized by combining historical fault records, and features for each modality are extracted using signal processing and computer vision techniques.

[0114] S2, Multimodal Data Preprocessing and Feature Extraction Module Design: Addressing the challenges of noise interference and feature extraction in photovoltaic data, this invention constructs an electrical signal processing module based on Kalman filtering and wavelet transform theory, an infrared thermal imaging feature extraction module using the ViT model, a visual data processing module based on U-Net and YOLOv8, and an environmental data analysis module using linear regression. The modal data obtained in S1 are input into the corresponding modules for feature extraction, yielding discriminative high-dimensional feature vectors.

[0115] S3, Construction of a Multimodal Large-Scale Model Fusion and Decision Support System: Addressing the problems of multimodal feature fusion and intelligent decision-making, this invention constructs a feature alignment module based on contrastive learning and knowledge distillation theories, a knowledge graph module using the Neo4j graph database, and a multi-agent collaborative decision-making module based on an attention mechanism. The feature vectors obtained in S2 are input into a contrastive learning network for cross-modal alignment, and executable decision suggestions are output through retrieval-enhanced generation techniques combined with the knowledge graph.

[0116] S4, Adaptive Optimization and Continuous Learning: Addressing long-term operation and maintenance optimization needs, this invention constructs a power plant simulation module based on digital twin technology, a strategy optimization module using multi-objective reinforcement learning, a decision evaluation module based on causal reasoning, and a privacy-preserving training framework through federated learning. The decision results obtained in S3 are input into the digital twin system for strategy verification, and reinforcement learning is used to optimize the long-term operation and maintenance plan.

[0117] S5, System Deployment and Intelligent Operation and Maintenance Applications: Addressing the needs for real-time monitoring and continuous improvement, this invention constructs a lightweight inference module based on an edge computing architecture and employs incremental learning to build a model optimization module. The trained multimodal large model is deployed to edge devices for real-time monitoring, and model performance is continuously optimized through a closed-loop feedback mechanism.

[0118] The construction of the multimodal operation and maintenance dataset for photovoltaic power plants in S1 may include the following process:

[0119] S1-1, Multi-source data acquisition: Four types of heterogeneous data are collected through electrical parameter sensors, infrared thermal imagers, environmental sensors, and UAV inspection equipment:

[0120] (1) Electrical parameters: Voltage V, current I, and power P signals are acquired in real time from sensors connected to the photovoltaic modules, with a sampling frequency of f. s =10Hz, used to detect electrical faults such as short circuits, open circuits, and PID effects.

[0121] (2) Infrared thermal imaging: An infrared camera with a resolution of 640×480 pixels was used to acquire the component temperature distribution matrix T at a frequency of 1 frame / s. img To identify abnormal states within them.

[0122] (3) Environmental parameters: Light intensity S and ambient temperature T were obtained from the weather station. a Data such as humidity (H) and wind speed (v) are recorded every minute to analyze the impact of the environment on power generation efficiency.

[0123] (4) UAV inspection images: The UAV is equipped with an RGB camera (4096×3072 resolution) and an infrared camera, and acquires surface images of the components at a flight speed of 5m / s. rgb and infrared image I ir It covers visual features such as dirt, cracks, and shadows.

[0124] S1-2, Data Labeling and Feature Extraction: Photovoltaic operation and maintenance experts, based on historical fault records, categorize the data, resulting in a label set.

[0125] The frequency domain features F are extracted from the electrical signal using Short Time Fourier Transform (STFT). elec=STFT(V,I,P), identifies anomalous harmonic components.

[0126] For infrared thermal images, the pixel proportion of temperature anomaly regions is extracted by threshold segmentation. Where N abn N represents the number of pixels with abnormal temperatures. total This represents the total number of pixels.

[0127] For RGB images of UAVs, a pre-trained YOLO model is used to detect the coordinates (x1, y1, x2, y2) of defect target boxes and the class confidence c.

[0128] Design of multimodal data preprocessing and feature extraction module in S2, multimodal data preprocessing and

[0129] The network structure diagram of the feature extraction module is as follows: Figure 3 As shown, the process may include:

[0130] S2-1, Electrical Signal Preprocessing and Feature Extraction: During the operation of a photovoltaic power station, electrical signals are susceptible to electromagnetic interference and equipment noise, resulting in a large amount of redundant information in the data. To obtain accurate operating characteristics, this system uses Kalman filtering to denoise the voltage V(t) and current I(t). Its state transition equation is:

[0131]

[0132] in Let A be the state vector, B be the state transition matrix, and u be the control matrix. t To control the input, w t Let Q be the process noise and Q be the noise covariance matrix.

[0133] In the feature extraction stage, the multi-resolution analysis characteristics of wavelet transform are utilized to decompose the denoised electrical signal into approximate components A at different scales. j and detail component D j Extracting energy features at various scales.

[0134] E j =∑∣D j | 2 And input the LSTM network modeling time dependencies, and update the hidden state as follows:

[0135] h t =LSTM(E j ,h t-1 ),

[0136] Final output electrical feature vector

[0137] S2-2, Infrared Thermal Imaging Data Processing: Infrared thermal imaging data can intuitively reflect the temperature distribution of photovoltaic power station equipment, which is of great significance for early fault warning. However, the pixel values ​​in the raw thermal imaging data do not directly represent the actual temperature and temperature calibration is required. According to the blackbody radiation law, there is the following relationship between the thermal imaging pixel value p and the actual temperature T:

[0138]

[0139] Where K1 and K2 are camera calibration constants. The above formula converts thermal imaging pixel values ​​into actual temperatures.

[0140] In practical applications, ambient background temperature can interfere with thermal imaging data, thus requiring background removal processing. A Gaussian mixture model (GMM) is used to model the reference background image B. This is achieved by calculating the current frame T. img Mean of background image The difference is used to obtain the foreground temperature matrix:

[0141]

[0142] To extract thermal anomaly features, this system employs a pre-trained ViT-B / 16 model. The foreground temperature matrix, after background removal, is input into the ViT-B / 16 model, which outputs a 768-dimensional thermal anomaly feature vector f. ir =ViT(T) fore ).

[0143] S2-3, UAV Visual Data Processing: During the acquisition process, RGB images captured by the UAV may become blurred due to lighting conditions and flight vibrations, affecting the subsequent defect detection results. To improve image quality, non-local mean filtering (NLM) is used for image enhancement.

[0144] Filtered pixel value I′ i,j The calculation is as follows:

[0145]

[0146] Where Ω is the search window, used to find regions similar to the current pixel; Z is the normalization factor, ensuring that the brightness of the filtered image remains unchanged; σ is the standard deviation of the Gaussian kernel, controlling the weight of the similarity measure; and I... k,l This represents the pixel value at coordinates (k, l) in the image.

[0147] Based on image enhancement, the U-Net model is used for semantic segmentation of RGB images. The U-Net model outputs a defect mask M∈{0,1}. H×WPixels with a value of 1 represent detected defect areas. Further calculation of the defect area percentage is then performed. Used to quantify the degree of damage to photovoltaic modules.

[0148] The YOLOv8 model is used for object detection, and the model outputs the detection results. Where c m Category confidence represents the probability that a detected target belongs to a certain category; bbox m = (x1, y1, x2, y2) represents the bounding box coordinates, which determine the position of the target in the image. Finally, the annotated image is used as the visualization result f. vis Output.

[0149] S2-4, Environmental Data Fusion Analysis: Data collection of light intensity S and ambient temperature T a Humidity (H), wind speed (v), and other relevant environmental data were collected and normalized to the [-1,1] interval to eliminate dimensional differences between different data dimensions. An environmental feature vector was then constructed. Environmental information is quantified into feature vector form.

[0150] Using a linear regression model Predicted theoretical power generation P pred Where w is the weight vector and b is the bias term. The theoretical power generation P... pred With actual power P real By comparison, the efficiency deviation ΔP = P is calculated. pred -P real This deviation value reflects the difference between the operating efficiency of a photovoltaic power plant under current environmental conditions and its expected efficiency.

[0151] The construction of multimodal large model fusion and decision support systems in S3 can include:

[0152] S3-1, Multimodal contrastive learning and feature alignment: Electrical encoders E are constructed using the Transformer architecture. elec Infrared encoder E ir Visual encoder E vis and environmental encoder E env Finally, the modal feature vectors z are output. elec ,z ir ,z vis ,z env .

[0153] The contrastive learning loss module uses the InfoNCE loss function to distinguish between positive sample pairs (different modes of the same fault) and negative sample pairs (modes of different faults):

[0154]

[0155] Where z i ,z j For positive sample features, z k τ represents the negative sample feature, K represents the temperature parameter, and K represents the number of negative samples.

[0156] In the cross-modal knowledge distillation process, the teacher model T is trained on a large amount of labeled data to obtain a high-confidence output p. T The student model S minimizes the distillation loss. During the learning process, the distillation loss is:

[0157]

[0158] Where C is the number of fault categories, p S (c) Predict probabilities for the student model.

[0159] By using different analysis and processing modules to process different feature data, the modules in the subsequent trained multimodal large model can perform real-time monitoring of different feature data of photovoltaic power plants.

[0160] S3-2, Knowledge-Enhanced Decision Generation: Knowledge graph construction relies on the Neo4j graph database to build a system containing fault types. Component Model Maintenance measures A photovoltaic (PV) operation and maintenance knowledge graph for each node. Module models are associated with their technical parameters and historical failure probabilities; maintenance measures are detailed down to operational specifications such as cleaning frequency and replacement procedures. A knowledge network is constructed by associating "fault-feature" relationships R(f,x) and "fault-measure" relationships R(f,a).

[0161] The retrieval enhancement generation mechanism is triggered after the model detects a fault feature x. By performing vector similarity retrieval between the feature vector and entities in the knowledge graph, it quickly locates the relevant knowledge triple (f, x, a). Large language models, such as Llama3, receive the retrieval results and the original fault features, and combined with natural language generation capabilities, output decision suggestions adv = LLM(x, retrieval results) that include fault cause analysis, processing priorities, and operational steps.

[0162] S3-3, Multi-Agent Collaborative Decision-Making: Modal agents are designed specifically for electrical, infrared, and visual modal data. Electrical Agent A elec Long Short-Term Memory (LSTM) networks based on time-series data are used to capture abnormal fluctuation patterns in electrical parameters; Infrared intelligent agent A ir Rapid localization and severity assessment of hotspot regions using YOLO series algorithms; Visual agent A visComponent appearance defects are identified using semantic segmentation technology. Each agent independently outputs a fault probability distribution p. elec ,p ir ,p vis .

[0163] The weights α of each agent are calculated using a cross-modal attention mechanism. elec ,α ir ,α vis The calculation formula is:

[0164]

[0165] Where w m Let be the modal weight vector. The final decision probability is:

[0166] p final =α elec p elec +α ir p ir +α vis p vis ,

[0167] p final By integrating multimodal information, misjudgments based on single modality can be effectively avoided, providing a more robust basis for operation and maintenance decisions.

[0168] Adaptive optimization and continuous learning in S4 can include:

[0169] S4-1, Digital Twin-Driven Reinforcement Learning: Constructing a twin model based on physical equations, with input environmental parameters f env And the operation and maintenance action 'a', through refined mathematical modeling, outputs the predicted power generation P. sim and failure rate r sim .

[0170] Multi-objective reinforcement learning (MORL) is employed, with the state space being... Among them, f elec Represents electrical parameters, f ir For infrared thermal imaging data, f vis It is the result of visual detection, f env These are environmental parameters. The motion space is... The reward function is designed as follows:

[0171] r=λ1(P real -P sim )+λ2(1-r sim )-λ3c a ,

[0172] Wherein, λ1, λ2, and λ3 are weighting coefficients, optimized using cross-validation to balance power generation improvement, failure rate reduction, and operation and maintenance cost control; c a This refers to the cost of the action and the consumption of resources.

[0173] S4-2, Causal Reasoning and Counterfactual Analysis: The PC algorithm is used to mine the causal relationship between fault variable F and maintenance action A from historical data, and a causal graph G(V,E) is constructed. Here, V is the variable node, covering fault type, environmental factors, and operation and maintenance operation; E is the directed edge, which clarifies the causal dependency between variables.

[0174] For an action 'a' that has already been performed, the potential outcome 'Y' when the action has not been performed is simulated using the intervention model do(A=a). a＝0 Specifically, based on the causal graph structure, and keeping other variables constant, the causal effect τ = Y is calculated by modifying the intervention conditions in the causal model. a＝1 -Y a＝0 It is used to help assess the necessity of maintenance decisions and avoid over-maintenance or under-maintenance.

[0175] S4-3, Federated Learning Framework Construction: Multiple photovoltaic power plants act as clients, each training its own modal encoder locally. (k is the power station index). Each client only uploads the model update gradient. The data is routed to a central server to prevent leakage of raw data. The central server aggregates gradients from each client, updates the global model, and distributes it, forming a closed-loop mechanism of "training-update-feedback".

[0176] To further enhance privacy protection, Laplace noise ∈ ~Laplace(b) is added to the gradient. The perturbed gradient is:

[0177]

[0178] Here, ∈ controls the privacy budget, and b is the noise scaling parameter. By setting the parameters appropriately, privacy attacks can be effectively resisted while ensuring model performance.

[0179] Based on local data diversity and model improvement Calculate the power plant weight:

[0180]

[0181] Among them, D k The local data distribution is measured using KL divergence. k With global data distribution The differences are analyzed. Based on the calculated weights, the gradient aggregation ratio of each client is dynamically adjusted to incentivize power stations with abundant data and good training results to actively participate in federated learning.

[0182] System deployment and intelligent operation and maintenance applications in S5 can include:

[0183] S5-1, Edge-Cloud Collaborative Computing: Deploys a lightweight model locally at the photovoltaic power plant to process electrical signals and drone images in real time, and outputs real-time fault warnings.

[0184] The feature vector f extracted at the edge elec (Electrical characteristics), f ir (Infrared thermal imaging characteristics), f vis (image features), f env (Environmental characteristics) are uploaded to the cloud. The cloud-based multimodal large model is based on the Transformer architecture and incorporates an attention mechanism to perform in-depth analysis of multi-source heterogeneous data. The model generates a detailed decision report, report = {fault type, maintenance recommendation, causal analysis}, containing fault type, maintenance recommendation, and causal analysis, through cross-modal alignment and knowledge distillation techniques.

[0185] S5-2, Closed-Loop Feedback Optimization: To continuously improve the accuracy and reliability of system decisions, a closed-loop feedback optimization mechanism is introduced. Maintenance personnel score the decision reports based on their practical maintenance experience.

[0186] s∈[0,1]. When the score s is less than the preset threshold, it indicates that there is a deviation between the model's decision and the actual situation, at which point the incremental learning process is triggered:

[0187] Use the newly labeled data {(x i ,y i The model is updated. To prevent catastrophic forgetting, the Elastic Weights Consolidation (EWC) algorithm is used to retain old knowledge while learning new data. The loss function is:

[0188]

[0189] Cross-entropy loss Regularization terms used to optimize learning for new tasks Used to constrain changes in model parameters, where F(θ) is the Fisher information matrix of the old task parameters, θ0 is the old parameter, and λ is the balance coefficient.

[0190] The model, optimized through incremental learning, is redeployed to both the edge and the cloud. The edge gains more accurate real-time analytics capabilities, while cloud-based decision-making quality is also improved, thus forming a complete closed loop of "monitoring-decision-feedback." This ensures the system continues to evolve over long-term operation, maintaining a consistently high level of operational decision-making capabilities.

[0191] In a specific example, based on a prototype of an intelligent operation and maintenance system for a photovoltaic power station, a 30-day field test was conducted at three typical photovoltaic power stations (A, B, and C) in a certain region. The test covered 12 environmental conditions, including sunny, cloudy, rainy, high temperature, and high humidity, accumulating 50GB of data collection. This included 12,000 normal state samples and 3,800 fault samples (850 electrical faults, 1,200 hot spots, 1,100 soiling, and 650 shading). The hardware configuration was as follows: the edge device used a combination of a quad-core ARM architecture processor and a 256-core GPU, while the cloud deployment consisted of a cloud server instance with a 16-core CPU and 64GB of memory.

[0192] 1. Multimodal data processing performance

[0193] The results were verified from two aspects: data preprocessing time and feature extraction accuracy.

[0194] Evaluation indicators:

[0195] Preprocessing latency: Processing time for single-sample multimodal data (ms)

[0196] Feature recall: Defect feature detection accuracy (%)

[0197] Cross-modal consistency: The correlation (cosine similarity) of features of the same type of fault across different modalities.

[0198] Table 1. Experimental Results of Multimodal Data Processing Performance S

[0199]

[0200] 2. Decision performance of multimodal large models

[0201] Validation was conducted on fault detection accuracy, decision generation efficiency, and modality missing robustness:

[0202] Evaluation indicators:

[0203] Fault Detection F1-score: Overall Precision and Recall

[0204] Decision generation time: Time elapsed from data input to maintenance recommendation output (s)

[0205] Modal Missing Tolerance: Detection accuracy retention rate (%) when a single modality fails.

[0206] Table 2 Experimental results of multimodal large model decision performance

[0207]

[0208] 3. Adaptive Optimization and Federated Learning Results

[0209] Verify the performance improvement of continuous learning and the privacy protection capabilities of federated learning:

[0210] Evaluation indicators:

[0211] Incremental learning efficiency: Model update time (min) when adding 100 new samples.

[0212] Federated learning convergence speed: the number of iterations of the global model across 3 clients.

[0213] Privacy Breach Risk: Assessing the Probability (%) of Original Data Leakage Through Reconstruction Attacks

[0214] Table 3. Experimental Results of Adaptive Optimization and Federated Learning

[0215]

[0216] Experimental results demonstrate that this multimodal large-model decision support system exhibits significant advantages in photovoltaic power plant operation and maintenance scenarios: Regarding the effectiveness of multimodal fusion, the cross-modal feature alignment error is below 0.87, and the decision accuracy is 12-18% higher than that of single-modal solutions, maintaining a detection capability of over 93% even in scenarios with missing modalities. In terms of real-time performance and scalability, the edge computing layer meets industrial-grade real-time early warning requirements, while the cloud-based large model supports complex fault reasoning. The federated learning architecture achieves fast convergence with manageable privacy risks across three nodes. Regarding continuous evolution capabilities, incremental learning and closed-loop feedback mechanisms enable rapid system optimization when new fault samples are added, and the federated learning contribution evaluation incentivizes data diversity, effectively improving long-term operation and maintenance efficiency. This system can be further promoted in photovoltaic power plant clusters, continuously improving the generalization capability of fault diagnosis through large-scale data iteration.

[0217] The intelligent operation and maintenance method for power generation equipment based on model autonomous selection provided in this disclosure involves the construction process of a multimodal large-scale model decision support system for intelligent operation and maintenance of photovoltaic power plants. First, multimodal data (including electrical signals, thermal imaging, environmental parameters, and visual images) of the photovoltaic power plant are collected using electrical parameter sensors, infrared thermal imagers, environmental sensors, and UAV inspection equipment. Data annotation and feature extraction are then performed to construct a high-quality photovoltaic operation and maintenance dataset. Next, a multimodal data preprocessing and feature extraction module based on signal processing, computer vision, and deep learning technologies is designed to extract features from electrical signals, thermal imaging, visual images, and environmental data, and feature alignment is achieved through cross-modal comparative learning. Subsequently, a decision support system based on a multimodal large-scale model is constructed, integrating knowledge graph-enhanced retrieval generation (RAG) technology and a multi-agent collaborative decision-making mechanism to achieve fault diagnosis, performance evaluation, and maintenance strategy optimization. During training, the model's decision-making capabilities are optimized through digital twin-driven reinforcement learning and causal reasoning. Finally, the system is deployed using an edge-cloud collaborative computing architecture to achieve real-time monitoring and closed-loop feedback for intelligent operation and maintenance.

[0218] This disclosure utilizes electrical parameter sensors, infrared thermal imagers, environmental sensors, and drone inspections to collect multi-source data and construct a multimodal operation and maintenance dataset; it employs contrastive learning and knowledge distillation to achieve cross-modal feature alignment; it combines knowledge graphs and retrieval enhancement generation techniques to enable large language models to provide interpretable decision suggestions; it optimizes long-term operation and maintenance strategies through digital twins and reinforcement learning; and it adopts an edge-cloud collaborative architecture to ensure system real-time performance.

[0219] Compared with the prior art, this disclosure has the following beneficial effects:

[0220] (1) It can improve the intelligence and multimodal fusion capabilities of photovoltaic power plant fault detection. Existing technologies mostly rely on a single data source or shallow multimodal fusion, which makes it difficult to accurately identify complex faults. This invention deeply integrates electrical signal, thermal imaging, visual and environmental data features through multimodal comparative learning and cross-modal knowledge distillation, which significantly improves the accuracy of fault detection.

[0221] (2) It can enhance the knowledge-driven and interpretable nature of operation and maintenance decisions. Existing systems are mostly based on rule matching or statistical models, lacking dynamic knowledge reasoning capabilities. This invention introduces photovoltaic operation and maintenance knowledge graph and retrieval enhancement generation technology, enabling large models to associate fault characteristics, historical maintenance records and industry standards in real time to generate actionable decision suggestions.

[0222] (3) It can optimize adaptive operation and maintenance strategies in complex environments. Existing methods have poor adaptability to environmental changes and cannot dynamically adjust maintenance plans. This invention uses digital twin-driven multi-objective reinforcement learning to simulate power plant performance under different environments, optimize cleaning cycles and inverter parameters, and balance power generation efficiency and operation and maintenance costs.

[0223] (4) It can improve the model's continuous learning and generalization capabilities. Existing system model updates rely on manual intervention and are difficult to adapt to new components or unknown failure modes. This invention continuously optimizes the model through incremental learning and counterfactual analysis modules.

[0224] (5) It enables real-time response and privacy protection in edge-cloud collaboration. Existing centralized cloud architectures suffer from high latency and data security risks. This invention uses a lightweight edge model to monitor key indicators in real time, while a large cloud model handles complex analysis tasks. Simultaneously, through federated learning and differential privacy technology, it supports cross-power station data sharing without leaking sensitive information, solving the privacy and collaborative optimization challenges of distributed photovoltaic systems.

[0225] and Figure 1 Corresponding to the model-based autonomous selection method for intelligent operation and maintenance of power generation equipment, this disclosure also provides a model-based autonomous selection system for intelligent operation and maintenance of power generation equipment, such as... Figure 6 As shown, the system may specifically include:

[0226] The acquisition module 601 is used to acquire multimodal operation and maintenance data of photovoltaic power plants, classify them by combining historical fault records, and extract the features of each mode using signal processing and computer vision technology to generate a multimodal operation and maintenance dataset.

[0227] The processing module 602 is used to construct an infrared thermal image feature extraction module using the ViT model, construct a visual data processing module based on U-Net and YOLOv8, construct an environmental data analysis module through linear regression, and input the modal data of the multimodal operation and maintenance dataset into the corresponding module for feature extraction.

[0228] The processing module 602 is also used to construct a knowledge graph module using the Neo4j graph database, construct a multi-agent collaborative decision-making module based on the attention mechanism, input the extracted feature vectors into the contrastive learning network for cross-modal alignment, and output executable decision results by combining retrieval enhancement generation technology with the knowledge graph.

[0229] The processing module 602 is also used to construct a power plant simulation module, construct a policy optimization module using multi-objective reinforcement learning, construct a decision evaluation module based on causal reasoning, construct a privacy-preserving training framework through federated learning, input the executable decision results into the digital twin system for policy verification, and generate a trained multimodal large model for photovoltaic power plants to select the corresponding module in the trained multimodal large model based on feature data for real-time monitoring.

[0230] In some embodiments, the acquisition of multimodal operation and maintenance data of photovoltaic power plants, combined with historical fault records for category labeling, and the extraction of modal features using signal processing and computer vision techniques to generate a multimodal operation and maintenance dataset, includes:

[0231] Four types of heterogeneous data were collected using electrical parameter sensors, infrared thermal imagers, environmental sensors, and drone inspection equipment.

[0232] Based on historical fault records, four types of heterogeneous data are categorized and labeled, with the label set being Y = {"normal", "electrical fault", "hot spot", "fouling", "occlusion"}.

[0233] In some embodiments, feature extraction is performed on electrical signals in a multimodal operation and maintenance dataset, including:

[0234] For electrical signals, Kalman filtering is used to denoise the voltage V(t) and current I(t), satisfying the formula:

[0235]

[0236] in, Let A be the state vector, B be the state transition matrix, and u be the control matrix. t To control the input, w t Let Q be the process noise, and let Q be the noise covariance matrix.

[0237] By utilizing the multi-resolution analysis properties of wavelet transform, the denoised electrical signal is decomposed into approximate components A at different scales. j and detail component D j Extract energy features E at various scales j =∑∣D j | 2 And input the LSTM network modeling time dependencies, and update the hidden state as follows:

[0238] h t =LSTM(E j ,h t-1 ),

[0239] Final output electrical feature vector

[0240] In some embodiments, feature extraction of infrared thermal imaging data in a multimodal operation and maintenance dataset includes:

[0241] By calculating the current frame T img Mean of background image The difference is used to obtain the foreground temperature matrix T. fore :

[0242]

[0243] Using a pre-trained ViT-B / 16 model, the foreground temperature matrix after background removal is input into the ViT-B / 16 model, and the output is a 768-dimensional thermal anomaly feature vector f. ir =ViT(T) fore ).

[0244] In some embodiments, feature extraction is performed on environmental data in a multimodal operations and maintenance dataset, including:

[0245] Data collection includes light intensity S and ambient temperature T. a Environmental data such as humidity H and wind speed v were collected and normalized to the range of [-1,1] to eliminate the difference in dimensionality between different data dimensions.

[0246] By constructing environmental feature vectors Quantify environmental information into feature vector form;

[0247] Using a linear regression model Predicted theoretical power generation P pred Where w is the weight vector and b is the bias term, the theoretical power generation P is... pred With actual power P real By comparison, the efficiency deviation ΔP = P is calculated. pred -P real This is used to reflect the difference between the operating efficiency of a photovoltaic power plant under current environmental conditions and its expected efficiency.

[0248] In some embodiments, feature extraction is performed on drone data in a multimodal operation and maintenance dataset, including:

[0249] Image enhancement is performed using nonlocal mean filtering, and the filtered pixel value I′ i,j The calculation is as follows:

[0250]

[0251] Where Ω is the search window, used to find regions similar to the current pixel; Z is the normalization factor, ensuring that the brightness of the filtered image remains unchanged; σ is the standard deviation of the Gaussian kernel, controlling the weight of the similarity measure; and I...k,l This represents the pixel value at coordinates (k, l) in the image;

[0252] The U-Net model is used for semantic segmentation of RGB images, and the defect mask M∈{0,1} is output by the U-Net model. H×W Pixels with a value of 1 represent detected defect areas, and the defect area percentage is further calculated. Used to quantify the extent of damage to photovoltaic modules;

[0253] The YOLOv8 model is used for object detection, and the model outputs the detection results. Where c m Category confidence represents the probability that a detected target belongs to a certain category; bbox m = (x1, y1, x2, y2) represents the bounding box coordinates, which determine the position of the target in the image. Finally, the annotated image is used as the visualization result f. vis Output.

[0254] In some embodiments, the above-mentioned module for constructing a knowledge graph using the Neo4j graph database and a multi-agent collaborative decision-making module based on an attention mechanism are used. The extracted feature vectors are input into a contrastive learning network for cross-modal alignment. The executable decision result is output through retrieval enhancement generation techniques combined with the knowledge graph, including:

[0255] Electrical encoders E were built using the Transformer architecture respectively. elec Infrared encoder E ir Visual encoder E vis and environmental encoder E env Finally, the modal feature vectors z are output. elec ,z ir ,z vis ,z env ;

[0256] The InfoNCE loss function is used to distinguish between positive and negative sample pairs:

[0257]

[0258] Where z i ,z j For positive sample features, z k For negative sample features, τ is the temperature parameter, and K is the number of negative samples;

[0259] In the cross-modal knowledge distillation process, the teacher model T is trained on a large amount of labeled data to obtain a high-confidence output p. T The student model S minimizes the distillation loss. During the learning process, the distillation loss is:

[0260]

[0261] Where C is the number of fault categories, p S (c) Predict probabilities for the student model;

[0262] The knowledge graph is built using the Neo4j graph database, and includes fault types. Component Model Maintenance measures A knowledge graph of photovoltaic operation and maintenance at each node;

[0263] The retrieval enhancement generation mechanism is triggered after the model detects the fault feature x. By performing vector similarity retrieval between the feature vector and the entity in the knowledge graph, the relevant knowledge triple (f,x,a) is quickly located.

[0264] Electrical Intelligent Agent A elec Long Short-Term Memory (LSTM) networks based on time-series data are used to capture abnormal fluctuation patterns in electrical parameters; Infrared intelligent agent A ir Rapid localization and severity assessment of hotspot regions using YOLO series algorithms; Visual agent A vis Component appearance defects are identified using semantic segmentation technology, and each agent independently outputs a fault probability distribution p. elec ,p ir ,p vis ;

[0265] The weights α of each agent are calculated using a cross-modal attention mechanism. elec ,α ir ,α vis The calculation formula is:

[0266]

[0267] Where w m Given the modal weight vector, the final decision probability is:

[0268] p final =α elec p elec +α ir p ir +α vis p vis .

[0269] In some embodiments, the aforementioned power plant simulation module employs multi-objective reinforcement learning to construct a policy optimization module, causal reasoning to construct a decision evaluation module, and federated learning to construct a privacy-preserving training framework. The executable decision results are input into a digital twin system for policy verification, generating a trained multimodal large model for real-time monitoring of photovoltaic power plants by selecting the corresponding module from the trained multimodal large model based on feature data. This includes:

[0270] A twin model is constructed based on physical equations, with input environmental parameters f. env And the operation and maintenance action 'a', through refined mathematical modeling, outputs the predicted power generation P. sim and failure rate r sim ;

[0271] Multi-objective reinforcement learning is employed, with the state space being... Among them, f elec Represents electrical parameters, f ir For infrared thermal imaging data, f vis It is the result of visual detection, f env For environmental parameters, the action space is... The reward function is designed as follows:

[0272] r=λ1(P real -P sim )+λ2(1-r sim )-λ3c a ,

[0273] Wherein, λ1, λ2, and λ3 are weighting coefficients, optimized using cross-validation to balance power generation improvement, failure rate reduction, and operation and maintenance cost control; c a For the cost of action and the consumption of resources;

[0274] The PC algorithm is used to mine the causal relationship between fault variable F and maintenance action A from historical data, and a causal graph G(V,E) is constructed, where V is the variable node, including fault type, environmental factors, and operation and maintenance operation; E is the directed edge, which clarifies the causal dependency between variables.

[0275] For an action 'a' that has already been performed, the potential outcome 'Y' when the action has not been performed is simulated using the intervention model do(A=a). a＝0 Based on the causal graph structure, and keeping other variables constant, the causal effect τ = Y is calculated by modifying the intervention conditions in the causal model. a＝1 -Y a＝0 It is used to help assess the necessity of maintenance decisions and avoid over-maintenance or under-maintenance.

[0276] Multiple photovoltaic power plants are used as clients, each training its own modal encoder locally. Each client only uploads the model update gradient. The data is sent to a central server to prevent the leakage of raw data; the central server aggregates the gradients from each client, updates the global model, and distributes it, forming a closed-loop mechanism of training-update-feedback.

[0277] Adding Laplace noise ∈ ~Laplace(b) to the gradient, the perturbed gradient is:

[0278]

[0279] Wherein, ∈ controls the privacy budget, and b is the noise scale parameter. By setting the parameter, privacy attacks can be resisted while ensuring model performance.

[0280] Based on local data diversity and model improvement Calculate the power plant weight:

[0281]

[0282] Among them, D k The local data distribution is measured using KL divergence. k With global data distribution Based on the differences in weights, the gradient aggregation ratio of each client is dynamically adjusted, and power plants whose incentive data and training effects meet the preset conditions are included in federated learning.

[0283] In some embodiments, the above method further includes:

[0284] A multimodal large-scale model, trained locally at a photovoltaic power plant, is deployed to process electrical signals and drone images in real time, outputting real-time fault warnings.

[0285] The electrical feature vector f extracted at the edge elec Infrared thermal image feature vector f ir Image feature vector f vis Environmental feature vector f env Uploaded to the cloud, the trained multimodal large model deployed in the cloud is based on the Transformer architecture and integrates the attention mechanism to perform in-depth analysis of multi-source heterogeneous data. Through cross-modal alignment and knowledge distillation technology, it generates a detailed decision report containing fault type, maintenance suggestions, and causal analysis: report = {fault type, maintenance suggestions, causal analysis}.

[0286] A closed-loop feedback optimization mechanism is introduced. Based on actual operation and maintenance experience, the decision report is scored s∈[0,1]. When the score s is less than a preset threshold, it indicates that there is a deviation between the model decision and the actual situation, triggering the incremental learning process.

[0287] Use the newly labeled data {(x i ,y i The trained multimodal large model is updated using an elastic weight consolidation algorithm, which retains old knowledge while learning new data. The loss function is:

[0288]

[0289] Among them, cross-entropy loss Regularization terms used to optimize learning for new tasks Used to constrain changes in model parameters, where F(θ) is the Fisher information matrix of the old task parameters, θ0 is the old parameters, and λ is the balance coefficient;

[0290] The model, optimized through incremental learning, is redeployed to both the edge and the cloud. This enables the edge to acquire more accurate real-time analysis capabilities and improves the quality of cloud-based decision-making, thus forming a complete closed loop of monitoring, decision-making, and feedback. This ensures the system continues to evolve during long-term operation and guarantees operational decision-making capabilities.

[0291] Combination Figure 7 As shown in the illustration, this disclosure also provides a model-based autonomous selection intelligent operation and maintenance device 700 for power generation equipment, including a processor 704 and a memory 701. Optionally, the system may further include a communication interface 702 and a bus 703. The processor 704, communication interface 702, and memory 701 can communicate with each other via the bus 703. The communication interface 702 can be used for information transmission. The processor 704 can call logical instructions in the memory 701 to execute the model-based autonomous selection intelligent operation and maintenance method for power generation equipment described in the above embodiments.

[0292] Furthermore, the logic instructions in the aforementioned memory 701 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.

[0293] The memory 701, as a computer-readable storage medium, can be used to store software programs and computer-executable programs, such as program instructions / modules corresponding to the methods in the embodiments of this disclosure. The processor 704 executes functional applications and data processing by running the program instructions / modules stored in the memory 701, thereby realizing the intelligent operation and maintenance method for power generation equipment based on model autonomous selection in the above embodiments.

[0294] The memory 701 may include a program storage area and a data storage area. The program storage area may store the operating system and application programs required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory 701 may include high-speed random access memory and may also include non-volatile memory.

[0295] This disclosure provides a computer-readable storage medium storing computer-executable instructions configured to execute a model-based intelligent operation and maintenance method for power generation equipment.

[0296] The aforementioned computer-readable storage medium may be a transient computer-readable storage medium or a non-transitory computer-readable storage medium.

[0297] The technical solutions of this disclosure can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes one or more instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of this disclosure. The aforementioned storage medium can be a non-transitory storage medium, including: a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, and other media capable of storing program code; it can also be a transient storage medium.

[0298] The foregoing description and accompanying drawings fully illustrate embodiments of the present disclosure to enable those skilled in the art to practice them. Other embodiments may include structural, logical, electrical, procedural, and other changes. The embodiments represent only possible variations. Individual components and functions are optional unless explicitly required, and the order of operation may vary. Parts and features of some embodiments may be included in or replace parts and features of other embodiments. As used in the description of the embodiments, the singular forms “a,” “an,” and “(the)” are intended to equally include the plural forms unless the context clearly indicates otherwise. Similarly, the term “and / or” as used in this application means including one or more of the associated listed items and all possible combinations thereof. Additionally, when used in this application, the terms “comprise” and its variations “comprises” and / or “comprising” refer to the presence of stated features, integrals, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof. Unless otherwise specified, an element defined by the phrase "comprising a..." does not exclude the presence of other identical elements in the process, method, or apparatus that includes the element. In this document, each embodiment may focus on describing the differences from other embodiments, and similar or identical parts between embodiments can be referred to mutually. For methods, products, etc., disclosed in the embodiments, if they correspond to the method section disclosed in the embodiments, then the relevant parts can be referred to the description of the method section.

[0299] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the embodiments of this disclosure. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0300] The methods and products (including but not limited to devices and equipment) disclosed in the embodiments herein can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of units may be merely a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection of devices or units may be electrical, mechanical, or other forms. Units described as separate components may or may not be physically separate, and components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to implement this embodiment according to actual needs. Furthermore, the functional units in the embodiments of this disclosure may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0301] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. In the descriptions corresponding to the flowcharts and block diagrams in the accompanying drawings, the operations or steps corresponding to different blocks may also occur in a different order than disclosed in the description, and sometimes there is no specific order between different operations or steps. For example, two consecutive operations or steps may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. Each block in a block diagram and / or flowchart, and combinations of blocks in a block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

Claims

1. A power generation equipment intelligent operation and maintenance method based on model autonomous selection, characterized in that, The method comprises: acquiring multi-modal operation and maintenance data of a photovoltaic power station, combining historical fault records to perform category labeling, and extracting features of each mode by using signal processing and computer vision technology to generate a multi-modal operation and maintenance dataset; an infrared thermal image feature extraction module is constructed by using a ViT model, a visual data processing module is constructed based on U-Net and YOLOv8, an environmental data analysis module is constructed by linear regression, and each mode of data of the multi-modal operation and maintenance dataset is input into the corresponding module for feature extraction; a knowledge graph module is constructed by using a Neo4j graph database, a multi-agent collaborative decision-making module is constructed based on an attention mechanism, a feature vector extracted is input into a contrastive learning network for cross-modal alignment, and an executable decision result is output by using retrieval enhancement generation technology in combination with the knowledge graph; a power station simulation module is constructed, a strategy optimization module is constructed by using multi-objective reinforcement learning, a decision evaluation module is constructed based on causal reasoning, a privacy protection training framework is constructed by federated learning, an executable decision result is input into a digital twin system for strategy verification, and a trained multi-modal large model is generated for real-time monitoring of a photovoltaic power station based on feature data.

2. The method of claim 1, wherein, The acquisition of multi-modal operation and maintenance data of a photovoltaic power station, the combination of historical fault records for category labeling, and the extraction of features of each mode by using signal processing and computer vision technology to generate a multi-modal operation and maintenance dataset comprises: Four types of heterogeneous data are collected through electrical parameter sensors, infrared thermal imagers, environmental sensors, and unmanned aerial vehicle inspection equipment. The four types of heterogeneous data are labeled in categories in combination with historical fault records, and the label set is Y = {"normal", "electrical fault", "hot spot", "dirt", "shading"}.

3. The method of claim 1, wherein, Feature extraction is performed on the electrical signals in the multi-modal operation and maintenance dataset, comprising: For electrical signals, Kalman filtering is used to denoise voltage V(t) and current I(t) to satisfy the formula: wherein, is a state vector, A is a state transition matrix, B is a control matrix, u t is a control input, w t is a process noise, Q is a noise covariance matrix; Using the multi-resolution analysis characteristic of wavelet transform, the de-noised electrical signal is decomposed into different scale approximate components A j and detail components D j , and the energy features under each scale are extracted E j =∑∣D j ∣ 2 and input the LSTM network modeling timing dependence, the hidden state update is: h t = LSTM(E j ,h t-1 ), Final output electrical feature vector 4. The method of claim 1, wherein, Feature extraction is performed on the infrared thermal imaging data in the multi-modal operation and maintenance dataset, comprising: By calculating the difference between the current frame T img and the background image mean , the foreground temperature matrix T fore is obtained. The pre-trained ViT-B / 16 model is adopted, the foreground temperature matrix after background removal processing is input into the ViT-B / 16 model, and a thermal anomaly feature vector f with an output dimension of 768 is output ir = ViT(T fore ).

5. The method of claim 1, wherein, Feature extraction is performed on the environmental data in the multi-modal operation and maintenance dataset, comprising: The environmental data including illumination intensity S, environmental temperature T a , humidity H, and wind speed v are collected and normalized to the interval [-1, 1] to eliminate the dimensional differences between different data dimensions; by constructing an environmental feature vector quantifying the environmental information into a feature vector form; A linear regression model is employed The predicted theoretical power generation P pred where w is a weight vector and b is a bias term, compares the theoretical power generation P pred with the actual power P real to calculate an efficiency deviation ΔP = P pred - P real for reflecting the difference between the operating efficiency of the photovoltaic power station under the current environmental conditions and the expected efficiency.

6. The method of claim 1, wherein, Feature extraction is performed on the unmanned aerial vehicle data in the multi-modal operation and maintenance dataset, comprising: The image is enhanced by non-local mean filtering, and the pixel value I' of the filtered image is i,j The calculation is as follows: where Ω is a search window used to find a region similar to the current pixel; Z is a normalization factor to ensure the brightness of the filtered image remains unchanged; σ is a standard deviation of the Gaussian kernel to control the weight of the similarity measure; I k,l represents the pixel value at coordinate (k, l) in the image; The U-Net model is used for semantic segmentation of the RGB image, and a defect mask M∈{0,1} is output by the U-Net model H×W wherein the pixel with the value of 1 represents a detected defect area, and a defect area ratio is further calculated for quantifying the damage degree of the photovoltaic module; The YOLOv8 model is used for target detection, and the model outputs a detection result where c m is a class confidence, indicating the probability that the detected target belongs to a certain class; bbox m =(x1,y1,x2,y2) is a bounding box coordinate, which determines the position of the target in the image, and finally the labeled image is taken as a visualization result f vis output.

7. The method of claim 1, wherein, The construction of a knowledge graph module by using a Neo4j graph database, the construction of a multi-agent collaborative decision-making module based on an attention mechanism, the input of a feature vector extracted into a contrastive learning network for cross-modal alignment, and the output of an executable decision result by using retrieval enhancement generation technology in combination with the knowledge graph comprise: The electrical encoder E is built using a Transformer architecture elec , the infrared encoder E ir , the visual encoder E vis and the environmental encoder E env , respectively, with the final output of the respective modal feature vectors z elec , z ir , z vis , z env ; An InfoNCE loss function is used to distinguish positive sample pairs and negative sample pairs: Where z i ,z j For positive sample features, z k For negative sample features, τ is the temperature parameter, and K is the number of negative samples; In the cross-modal knowledge distillation process, a teacher model T is trained based on a large amount of labeled data to obtain a high confidence output p T A student model S learns by minimizing a distillation loss The distillation loss is: where C is the number of failure categories, p S (c) is the predicted probability for the student model. The knowledge graph is constructed based on a Neo4j graph database, and contains fault types Component model Maintenance measures A photovoltaic operation and maintenance knowledge graph of nodes The retrieval enhancement generation mechanism is triggered after the model detects fault features x, vector similarity retrieval is performed between the feature vector and entities in the knowledge graph, and relevant knowledge triples (f, x, a) are quickly located. Electrical intelligent agent A elec Long short-term memory network based on time series data to capture abnormal fluctuation patterns of electrical parameters; infrared intelligent agent A ir Using YOLO series algorithm to realize rapid positioning and severity evaluation of hot spot area; visual intelligent agent A vis Identify component appearance defects through semantic segmentation technology, and each intelligent agent independently outputs fault probability distribution p elec ,p ir ,p vis ; The weight of each agent is calculated through a cross-modal attention mechanism elec ,α ir ,α vis , and the calculation formula is: where w m is the modal weight vector, and the final decision probability is: p final = a elec p elec + a ir p ir + a vis p vis .

8. The method of claim 1, wherein, The power station simulation module adopts a multi-objective reinforcement learning strategy optimization module, a decision evaluation module based on causal reasoning, and a privacy protection training framework constructed through federated learning, inputs executable decision results into a digital twin system for strategy verification, generates a trained multi-modal large model, and uses the trained multi-modal large model for real-time monitoring of photovoltaic power stations based on feature data to select corresponding modules in the trained multi-modal large model, including: Based on physical equations, construct twin models, input environmental parameters f env and operation actions a, through fine mathematical modeling, output predicted power generation P sim and failure rate r sim ; Multi-objective reinforcement learning is adopted, and the state space is where f elec represents electrical parameters, f ir is infrared thermal image data, f vis is visual inspection results, and f env is environmental parameters, and the action space is The reward function is designed as: r = λ1(P real - P sim ) + λ2(1 - r sim ) - λ3c a , wherein λ1, λ2, λ3 are weight coefficients, optimized by a cross-validation method, for balancing the power generation increase, the failure rate reduction, and the operation and maintenance cost control; c a are the action cost and resource consumption; The PC algorithm is used to mine the causal relationship between the fault variable F and the maintenance action A from historical data to construct a causal graph G(V, E), where V is a variable node including fault types, environmental factors, and operation and maintenance operations; E is a directed edge that clearly defines the causal dependence between variables; For the performed action a, the potential outcome Y when the action is not performed is simulated by the intervention model do(A=a) a＝0 Based on the causal graph structure, the causal effect τ=Y a＝1 -Y a＝0 , for helping to assess the necessity of maintenance decisions, to avoid over-maintenance or under-maintenance; A plurality of photovoltaic power stations are taken as clients, each of which trains a modal encoder locally Each client only uploads a model update gradient to a central server, avoiding leakage of original data; the central server aggregates the gradients of each client, updates the global model and issues it, forming a closed-loop mechanism of training-update-feedback; Laplace noise ∈ is added to the gradient, and the perturbed gradient is: Where ∈ controls the privacy budget, and b is the noise scale parameter. By setting the parameters, the model performance is guaranteed while resisting privacy attacks; According to local data diversity And model lift Compute power station weights: wherein D k The local data distribution p k is measured by KL divergence from the global data distribution Based on the calculated weights, dynamically adjust the gradient aggregation proportion of each client, and encourage power plants that meet the preset conditions of data and training effect to participate in federated learning.

9. The method of claim 1, wherein, The method further includes: Deploying the trained multi-modal large model locally in the photovoltaic power station to process electrical signals and UAV images in real time and output real-time fault early warning The edge end will upload the extracted electrical feature vector f elec , infrared thermal image feature vector f ir , image feature vector f vis , environment feature vector f env to the cloud, and the trained multi-modal large model deployed in the cloud is based on the Transformer architecture and fuses the attention mechanism to perform deep analysis on multi-source heterogeneous data. Through cross-modal alignment and knowledge distillation technology, a detailed decision report report={fault type, maintenance suggestion, causal analysis} containing fault type, maintenance suggestion, and causal analysis is generated. A closed-loop feedback optimization mechanism is introduced, and the decision report is scored s∈[0, 1] based on actual operation and maintenance experience. When the score s is less than the preset threshold, it indicates that the model decision deviates from the actual situation, triggering an incremental learning process: The trained multi-modal large model is updated using new labeled data {(x i ,y i )}, and an elastic weight consolidation algorithm is used to retain old knowledge while learning new data, and the loss function is: where the cross-entropy loss for optimizing new task learning, a regularization term for constraining the change of model parameters, where F(θ) is the Fisher information matrix of the old task parameters, θ0 is the old parameter, and λ is a balance coefficient; The model optimized through incremental learning is redeployed to the edge and cloud to enable the edge to obtain more accurate real-time analysis capabilities, and the cloud decision quality is also improved, thereby forming a complete closed loop of monitoring-decision-feedback, ensuring the continuous evolution of the system in long-term operation, and ensuring the operation and maintenance decision-making capability.

10. An intelligent operation and maintenance system of a power generation device based on model autonomous selection, configured to perform the method of any one of claims 1-9. The system includes: An acquisition module is configured to acquire multi-modal operation and maintenance data of a photovoltaic power station, label categories in combination with historical fault records, and extract features of each modality using signal processing and computer vision techniques to generate a multi-modal operation and maintenance dataset; A processing module is configured to construct an infrared thermal image feature extraction module using a ViT model, construct a visual data processing module based on U-Net and YOLOv8, and construct an environmental data analysis module through linear regression, and input each modality data of the multi-modal operation and maintenance dataset into the corresponding module for feature extraction; The processing module is further configured to construct a knowledge graph module using a Neo4j graph database, construct a multi-agent collaborative decision-making module based on an attention mechanism, input the extracted feature vectors into a contrastive learning network for cross-modal alignment, and output executable decision results through retrieval enhancement generation technology combined with the knowledge graph; The processing module is further configured to construct a power station simulation module, adopt a multi-objective reinforcement learning strategy optimization module, construct a decision evaluation module based on causal reasoning, construct a privacy protection training framework through federated learning, input executable decision results into a digital twin system for strategy verification, generate a trained multi-modal large model, and use the trained multi-modal large model for real-time monitoring of photovoltaic power stations based on feature data to select corresponding modules in the trained multi-modal large model.

Images