Intelligent early warning system for photovoltaic power grid fault based on multi-source data fusion

Abstract

This invention discloses an intelligent early warning system for photovoltaic power grid faults based on multi-source data fusion, relating to the photovoltaic power grid field. It includes a data sensing module, an edge computing module, a preliminary early warning response module, a cloud analysis module, and a secondary early warning response module. By deploying a multimodal sensor network, it collects multi-source sensing data in real time, including electrical, physical, image, environmental, and string IV curve data. The collected data is preprocessed, and feature extraction and fusion are performed using a cross-modal autoencoder. An improved isolated forest model is used to achieve preliminary fault detection and tiered early warning response. The cloud-based system analyzes the root causes of faults and simulates their propagation trends based on digital twins, ultimately generating accurate secondary early warning reports and sending them to the operation and maintenance terminal for response. This invention addresses the problems of delayed alarms, frequent false alarms, and inability to diagnose root causes in existing photovoltaic fault early warning systems. It achieves intelligent early warning throughout the entire process, from rapid response to precise operation and maintenance.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power grids, and specifically to an intelligent early warning system for photovoltaic power grid faults based on multi-source data fusion. Background Technology

[0002] With the acceleration of the global energy transition and the continuous expansion of photovoltaic (PV) power generation, the increasing penetration rate poses a severe challenge to the stability and reliability of the power grid. Traditional PV power plant operation and maintenance heavily relies on manual inspections and post-incident repairs. This model is slow to respond, costly, and unable to meet the management needs of large-scale, distributed PV arrays. More importantly, the inherent randomness and volatility of PV power generation mean that grid connection failures can trigger cascading effects, endangering regional power grid security. Therefore, developing an efficient and accurate intelligent early warning system for PV grid faults has become a key technological requirement for ensuring the healthy development of the PV industry.

[0003] Existing systems largely rely on a single operational data stream and employ simple threshold alarm mechanisms. This approach suffers from severe lag; by the time an alarm is triggered, the fault has often already occurred, rendering the warning meaningless. Furthermore, the lack of cross-validation with multi-dimensional data leads to a high false alarm rate and difficulty in pinpointing the root cause of complex faults. Numerous latent faults remain unidentified, such as minor hotspot effects and slow performance degradation, resulting in significant power generation losses.

[0004] Therefore, a photovoltaic power grid fault intelligent early warning system based on multi-source data fusion is needed to solve the above problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention discloses a photovoltaic power grid fault intelligent early warning system based on multi-source data fusion. Through an intelligent architecture of multi-source data fusion and cloud-edge collaboration, it realizes intelligent early warning throughout the entire process from rapid response to precise operation and maintenance, thereby improving the accuracy of photovoltaic power grid fault detection and the timeliness of response.

[0006] The present invention adopts the following technical solution:

[0007] A photovoltaic power grid fault intelligent early warning system based on multi-source data fusion includes:

[0008] The data sensing module deploys a multimodal sensor network at the distributed monitoring nodes of the photovoltaic power grid to acquire multi-source sensing data of the photovoltaic power grid in real time. The multi-source sensing data includes at least electrical quantity parameters, physical quantity parameters, image parameters, environmental parameters, and the IV curve of the photovoltaic string.

[0009] The edge computing module deploys a preprocessing unit, a feature processing unit, and a preliminary early warning unit at the edge nodes of the photovoltaic grid. The preprocessing unit is used to clean and calibrate the acquired multi-source sensing data. The feature processing unit uses a cross-modal autoencoder to extract and fuse features from the preprocessed multi-source sensing data. The preliminary early warning unit uses an improved isolated forest model to perform preliminary fault detection on the fused features.

[0010] The preliminary early warning response module triggers a preliminary early warning response strategy based on the preliminary fault detection results. When the preliminary fault detection result indicates an emergency fault, it disconnects the faulty equipment from the power grid, isolates the emergency fault, and uploads the fault data to the cloud for analysis. When the preliminary fault detection result indicates a potential fault, it adjusts the sampling rate of the multimodal sensor network and uploads the fault data and incremental data stream to the cloud for analysis. When the preliminary fault detection result indicates no fault, monitoring continues.

[0011] The cloud-based analysis module receives data uploaded from edge nodes and constructs a digital twin based on the received data. The digital twin analyzes the causes of faults based on multi-physics coupling and simulates the trend of fault propagation.

[0012] The secondary early warning response module generates a secondary early warning report based on the cause of the fault and the fault propagation trend, and sends it to the operation and maintenance terminal for response.

[0013] Furthermore, the distributed monitoring nodes are deployed at the locations of photovoltaic strings, combiner boxes, inverters, transformers, and grid connection points in each photovoltaic power station within the photovoltaic grid. The multimodal sensor network includes at least electrical quantity sensors, physical quantity sensors, vision sensors, environmental sensors, and intelligent string controllers. The electrical quantity sensors are used to collect data on voltage, current, power, harmonics, and insulation resistance. The physical quantity sensors are used to collect data on temperature, vibration, stress, and deformation. The vision sensors are used to collect image data on hot spots, damage, and shading of photovoltaic modules. The environmental sensors are used to collect data on light intensity, wind speed, humidity, altitude, and dust concentration. The intelligent string controllers are used to collect IV curve data of the photovoltaic strings.

[0014] Furthermore, the preprocessing unit uses adaptive Kalman filtering to denoise the multi-source sensing data and uses a generative adversarial network to complete the missing multi-source sensing data. The cross-modal autoencoder includes an encoder and a decoder. The encoder uses a heterogeneous dual-branch coding structure to independently encode the preprocessed numerical parameters and image parameters in two branches to generate feature vectors. The generated dual-branch feature vectors are weighted and correlated through a self-attention mechanism. At the same time, feature vectors from different modalities are mapped to the same feature space to achieve feature alignment. The decoder reconstructs the data through a multi-task learning framework and outputs fused features.

[0015] Furthermore, the improved isolated forest model employs a two-level detection framework to perform preliminary fault detection on the fused features. The two-level detection framework includes string-level detection and system-level detection.

[0016] The string-level detection constructs an IV curve feature library under normal operating conditions of the photovoltaic power grid based on environmental parameters and the IV curves of the photovoltaic strings. When the features on the IV curves of the photovoltaic strings deviate from the IV curve feature library by more than a preset deviation threshold, it is determined to be a string-level potential fault. The string-level potential fault information and the fused feature vector are then input into the isolated forest model after knowledge distillation.

[0017] The system-level detection uses an isolated tree to score anomalies. When the anomaly score is greater than a first preset threshold, it is determined to be an emergency fault. When the anomaly score is less than or equal to the first preset threshold but greater than a second preset threshold, it is determined to be a potential fault. When the anomaly score is less than or equal to the second preset threshold, it is determined to be a fault-free condition.

[0018] Furthermore, when the initial fault detection result is no fault, the sampling rate of the multimodal sensor network is maintained in the normal mode to acquire multi-source sensing data of the photovoltaic grid in real time; when the initial fault detection result is an emergency fault, a trip command is sent to the actuator of the faulty device through the hard-wired interface, and the actuator disconnects the faulty device from the photovoltaic grid according to the trip command to achieve emergency fault isolation; when the initial fault detection result is a potential fault, a sampling rate adjustment command is sent to the sensor manager through the built-in management interface to adjust the sampling rate of the multimodal sensor network from the normal mode to the abnormal mode to acquire the incremental data stream of the photovoltaic grid.

[0019] Furthermore, the cloud analysis module receives the fusion features of fault data, incremental data streams and multi-source sensing data uploaded by edge nodes through a distributed message queue, and constructs a local digital twin of each photovoltaic power station and a global digital twin of the photovoltaic power grid based on the received data;

[0020] The local digital twin includes a physical model layer, a data-driven layer, and a scene fingerprint. The physical model layer simulates the internal physical field distribution of the photovoltaic power station equipment through finite element analysis and integrates multi-physics coupling equations to simulate the cross-domain parameter interaction of the photovoltaic power station. The data-driven layer uses a long short-term memory network (LSTM) and an attention mechanism to output the real-time operating status parameters of the photovoltaic power station equipment. It also corrects the simulation deviation of the physical model layer by minimizing the mean square error between the predicted and measured values. The scene fingerprint is based on the equipment files of the photovoltaic power station and the historical uploaded data of the edge nodes to build a scene fingerprint library. It also uses cosine similarity matching to locate the current operating scene of the photovoltaic power station.

[0021] The global digital twin integrates the local digital twins of all photovoltaic power plants through federated learning, and realizes state interaction with the local digital twins through a distributed data synchronization protocol.

[0022] Furthermore, for electrical quantity anomalies, the digital twin uses an electric field model to simulate the energy transmission path in the circuit topology and combines it with circuit equations to solve for the voltage and current distribution at the fault point to obtain the cause of the fault; for physical quantity anomalies, the digital twin uses a thermal field model to calculate the thermal balance relationship between the operating losses of the photovoltaic power station equipment and the heat dissipation from the environment, and combines it with a mechanical model to analyze structural deformation and stress distribution to obtain the cause of the fault; for IV curve anomalies, the digital twin uses an equivalent circuit model of the photovoltaic module to simulate the distortion law of the IV curve to obtain the cause of the fault.

[0023] Furthermore, the digital twin uses the equipment of the photovoltaic power station as graph nodes, the electrical connection relationship between the equipment as edges, and the comprehensive coefficient of line transmission efficiency and distance as edge weights to construct a dynamic topology graph of the photovoltaic power station.

[0024] The temporal dynamic features of fault development are extracted by Temporal Convolutional Network (TCN), and the spatial propagation features of faults in the topology are captured by Graph Neural Network (GNN) to simulate the diffusion path, impact range and impact degree of photovoltaic grid faults at different time scales, and output the fault diffusion probability matrix.

[0025] The beneficial effects of this invention are as follows:

[0026] 1. This invention, by deploying a multimodal sensor network at distributed monitoring nodes in the photovoltaic power grid, can acquire various data in real time, including electrical quantities, physical quantities, environmental parameters, image data, and the IV curves of photovoltaic strings. This combination of multi-source data enables the system to more comprehensively perceive the state of the power grid, greatly improving the accuracy and reliability of monitoring.

[0027] 2. This invention deploys preprocessing units, feature processing units, and preliminary early warning units at the edge nodes of the photovoltaic power grid, enabling localized data processing and rapid response. Through edge computing, the system can clean, calibrate, and extract features from the collected data, reducing data transmission latency and improving the real-time performance of fault detection. Furthermore, the improved isolated forest model for preliminary fault detection of features can effectively identify potential faults and reduce the false alarm rate.

[0028] 3. After initial fault detection, this invention employs different early warning and response strategies based on different types of faults. In the event of an emergency fault, the faulty equipment is immediately disconnected and isolated to ensure the safe operation of the power grid; in the event of a potential fault, the sensor sampling rate is adjusted to reduce data traffic while uploading the data to the cloud for in-depth analysis. This multi-layered early warning and response strategy effectively ensures the stability and security of the system.

[0029] 4. This invention receives uploaded data from edge nodes and establishes a digital twin model. It analyzes the root cause of the fault through multiphysics coupling and simulates the fault's propagation trend. Through digital twin simulation analysis, more accurate fault diagnosis can be achieved, and potential risks can be predicted in advance, allowing for preventative measures to reduce losses. After initial fault location, a secondary early warning response is implemented based on the fault cause and propagation trend, further improving grid protection and ensuring the stable operation of the photovoltaic grid. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the system architecture of the present invention;

[0031] Figure 2 This is an architecture diagram of the edge computing module in this invention;

[0032] Figure 3 This is a diagram illustrating the architecture of the digital twin in this invention. Detailed Implementation

[0033] The following will refer to the appendices in the embodiments of the present invention. Figure 1 To be continued Figure 3 The technical solutions in the embodiments of the present invention are clearly and completely described herein. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0034] This invention discloses an intelligent early warning system for photovoltaic power grid faults based on multi-source data fusion, as shown in the attached figure. Figure 1 As shown, it includes:

[0035] The data sensing module deploys a multimodal sensor network at the distributed monitoring nodes of the photovoltaic power grid to acquire multi-source sensing data of the photovoltaic power grid in real time. The multi-source sensing data includes at least electrical parameters, physical parameters, image parameters, environmental parameters, and the IV curve of the photovoltaic string.

[0036] Multimodal sensor networks are deployed as distributed monitoring nodes within photovoltaic power grids. These monitoring nodes are deployed at key equipment and locations in the photovoltaic power plant, differentiated according to their function and importance, including but not limited to:

[0037] Temperature sensors are deployed on the photovoltaic strings, utilizing intelligent string controllers; multiple current sensors and internal temperature sensors are deployed in the combiner boxes. High-precision electrical quantity sensors are deployed on the DC input and AC output sides of the inverter, and vibration and temperature sensors are deployed on the inverter housing. Electrical quantity sensors and environmental sensors are deployed at transformers, switchgear, and other step-up substation and grid connection points.

[0038] This layered and hierarchical deployment strategy enables comprehensive and full-state perception of photovoltaic power plants, from components to grid connection points.

[0039] A multimodal sensor network consists of the following types of sensors, used to acquire the multi-source sensing data in real time:

[0040] Electrical quantity sensors: including but not limited to high-precision current transformers (CTs), voltage transformers (PTs), power analyzers, power quality analyzers, and insulation resistance testers. These are used to collect key electrical parameters such as voltage, current, power, harmonic content, and system-to-ground insulation resistance at the photovoltaic DC side and AC grid connection point.

[0041] Physical quantity sensors: including but not limited to thermocouples, infrared temperature sensors, vibration acceleration sensors, and stress strain gauges. Used to collect physical state parameters of key equipment such as temperature distribution (e.g., component backplanes, inverter heat sinks), mechanical vibration (e.g., transformers, inverters), internal mechanical stress, and deformation.

[0042] Visual sensors: These include high-definition visible light cameras and infrared thermal imagers deployed at fixed locations or mounted on drones. They are used to periodically or on demand acquire images of hot spots on photovoltaic modules, images of surface damage, and images of snow, dust, or vegetation obstruction, enabling visual diagnosis of faults.

[0043] Environmental sensors: Small weather stations, either integrated or deployed independently, are used to collect environmental parameters such as light intensity (irradiance), ambient temperature, wind speed, relative humidity, altitude, and dust concentration at the power plant site, providing an environmental benchmark for power prediction and fault analysis.

[0044] Intelligent string controller: As a dedicated data acquisition unit, it can scan photovoltaic strings through a precision electronic load in online or offline mode to obtain complete IV characteristic curve data, thereby accurately diagnosing problems such as string-level performance degradation, series mismatch, and bypass diode failure.

[0045] The edge computing module deploys preprocessing units, feature processing units, and preliminary early warning units at the edge nodes of the photovoltaic grid, as shown in the attached diagram. Figure 2 As shown, the preprocessing unit is used to clean and calibrate the acquired multi-source sensing data, the feature processing unit uses a cross-modal autoencoder to extract and fuse features from the preprocessed multi-source sensing data, and the preliminary early warning unit uses an improved isolated forest model to perform preliminary fault detection on the fused features.

[0046] The preprocessing unit is responsible for performing preliminary processing on the raw multi-source sensing data uploaded by the data sensing module, providing a high-quality data foundation for subsequent feature fusion.

[0047] For time-series data such as electrical and physical quantities, an adaptive Kalman filter algorithm is used for noise reduction. This algorithm can dynamically adjust the filter gain according to the real-time changes in measurement noise, effectively smoothing random interference while preserving the true abrupt changes in the data (such as sudden changes caused by faults), which is superior to traditional fixed-parameter filters. The algorithm dynamically adjusts the Kalman gain by estimating the covariance matrix of the measurement noise in real time. Specifically, when the difference between the observed and predicted values ​​(innovation) continuously increases, the algorithm judges that a fault or strong interference may have occurred, thus automatically reducing the confidence in the current observation (gain reduction), relying more on the prediction of the system model, and avoiding over-smoothing of true fault characteristics. Compared with fixed-parameter filtering algorithms, this is more sensitive to capturing sudden faults.

[0048] To address data loss caused by brief communication interruptions or momentary sensor malfunctions, a data completion model based on generative adversarial networks (GANs) is introduced. This model is trained using historical complete data. The generator can generate the most probable value that conforms to the true data distribution based on valid data before and after the missing data point, achieving higher accuracy than simple interpolation methods (such as linear interpolation). First, the GAN is pre-trained using historical complete data. The generator learns to generate reasonable filler values ​​based on valid data within the time window before and after the missing data point; the discriminator learns to distinguish between "true complete data" and "data completed by the generator." In practical applications, when data loss is detected, valid data before and after the missing time period is input into the pre-trained generator, which outputs the completed data sequence.

[0049] The feature processing unit employs an advanced cross-modal autoencoder to achieve deep feature extraction and fusion of multi-source heterogeneous data. The cross-modal autoencoder includes an encoder and a decoder.

[0050] The encoder employs a heterogeneous dual-branch coding structure, including a numerical parameter branch and an image parameter branch. For preprocessed numerical data such as electrical quantities, physical quantities, and environmental parameters, feature encoding is performed using a fully connected layer or a one-dimensional convolutional neural network. The input consists of denoised and completed numerical data such as electrical, environmental, and temperature data (with a dimension of n). First, a fully connected layer performs an initial transformation, and then two layers of one-dimensional convolutional neural networks are used to extract the local correlation and trend features of the data in the time dimension. For infrared and visible light images, a lightweight convolutional neural network is used for feature extraction. The input consists of preprocessed infrared and visible light images (with a size regularized to 224x224). A lightweight convolutional neural network (such as a variant of MobileNetV2) is used for feature extraction, and the final output is a high-dimensional feature vector.

[0051] By using a self-attention mechanism to interactively compute the feature vectors generated by the two branches, the importance weights of different modal features under specific operating conditions are dynamically learned and assigned, achieving weighted fusion. Essentially, this process maps heterogeneous features to a unified, comparable latent feature space, completing feature alignment. The feature vectors output from the two branches are then input into a self-attention mechanism layer, which calculates the correlation weights between different dimensions of numerical and image features. For example, it might learn that a slight decrease in current in a branch is highly correlated with a hot spot appearing in the image of a component in that branch—two cross-modal signals—and assign them higher weights during fusion. Through weighted summation, a unified, deep fused feature vector is obtained. This process maps heterogeneous data to the same semantic feature space, achieving feature alignment.

[0052] The decoder employs a multi-task learning framework, simultaneously reconstructing the fused feature vector into both numerical and image data. This reconstruction task forces the encoder to retain the most essential and critical information from each modality during the fusion process, ensuring the generated fused feature vector has high representativeness and integrity. The decoder has two output heads: one for reconstructing numerical data (through a fully connected layer), and the other for reconstructing image data (through a deconvolutional network). By minimizing the loss (e.g., mean squared error MSE) between the reconstructed data and the original input data, the encoder is forced to learn fused features that contain sufficient information, thus guaranteeing feature quality.

[0053] The preliminary early warning unit adopts an improved isolated forest model and performs fault detection based on a two-level detection framework, which balances detection efficiency and accuracy. The two-level detection framework includes string-level detection and system-level detection.

[0054] String-level testing is based on historical normal data, constructing a standard IV curve feature library for each photovoltaic string under different environmental parameters (mainly irradiance and temperature). This feature library stores the mapping relationship between key feature values ​​such as maximum power point, open-circuit voltage, and short-circuit current and the environment.

[0055] In real-time monitoring, the intelligent string controller periodically collects IV curves. Based on the currently measured irradiance and temperature, the system interpolates the corresponding standard feature values ​​from a feature library. It then calculates the relative deviation between the measured values ​​and the standard values ​​(e.g., Pmpp deviation rate). If the deviation rate of any key feature exceeds a preset deviation threshold (e.g., Pmpp deviation > 5%), the string is determined to have a potential string-level fault.

[0056] The system-level detection concatenates the "string-level potential fault" label (as an important Boolean feature) output from the first-level detection with the fused feature vector generated by the feature processing unit, using both as input. A lightweight isolated forest model optimized with knowledge distillation is then employed for computation. During training, a complex deep neural network (teacher model) guides the isolated forest (student model), enabling it to learn more complex anomaly patterns. The model calculates an anomaly score for each input sample, with higher scores indicating greater anomalies.

[0057] The model operates on the input fused feature vector and outputs an anomaly score.

[0058] If the abnormal score is greater than the first preset threshold (e.g., 0.75), it is determined to be an emergency fault (such as DC arcing, insulation breakdown, etc., which require immediate handling).

[0059] If the abnormal score is less than or equal to the first preset threshold but greater than the second preset threshold, it is judged as a potential fault (such as performance degradation, minor abnormalities, or other phenomena that require attention).

[0060] If the anomaly score is less than or equal to the second preset threshold (e.g., 0.5), it is determined to be fault-free.

[0061] The preset thresholds are based on a parameter system determined by combining historical data, theoretical calculations, and business requirements.

[0062] The preliminary early warning response module triggers the preliminary early warning response strategy based on the preliminary fault detection results.

[0063] When the initial fault detection result indicates no fault, monitoring continues, maintaining the sampling rate of the multimodal sensor network in normal mode. This allows for real-time acquisition of multi-source sensing data from the photovoltaic grid without sending adjustment commands to any external devices. The multimodal sensor network periodically collects data at a default, lower sampling rate (i.e., normal mode, e.g., electrical parameters once per second, image parameters once per minute) to maintain basic monitoring and conserve edge-side computing and communication resources. Data is uploaded to the cloud according to the normal process for long-term trend analysis.

[0064] When the initial fault detection result indicates an emergency fault, the connection between the faulty equipment and the power grid is disconnected to isolate the emergency fault, and the fault data is uploaded to the cloud for analysis. Immediately, a trip command is sent to the actuator of the nearest intelligent circuit breaker or contactor via the hardwired digital output interface or high-speed industrial real-time Ethernet interface of the edge computing gateway. This command is a highest-priority switching signal to ensure minimal transmission delay. Upon receiving the command, the actuator drives the circuit breaker to trip, physically disconnecting the faulty equipment from the power grid and achieving emergency fault isolation. Simultaneously with sending the trip command, the module marks and caches high-frequency raw data packets for a period before and after the fault (e.g., 10 seconds before and after), and immediately uploads these packets to the cloud analysis module via the communication network as a highest-priority task for post-fault deep root cause analysis.

[0065] When the initial fault detection indicates a potential fault, the sampling rate of the multimodal sensor network is adjusted, and the fault data and incremental data streams are uploaded to the cloud for analysis. A structured sampling rate adjustment command is sent to the unified sensor manager via the edge computing gateway's built-in management interface. This command explicitly specifies the sensor types requiring adjustment and their target sampling rates. The sensor manager parses the command and increases the sampling rate of the relevant sensors from normal mode to abnormal mode. For example, the sampling rate of the infrared thermal imager associated with the potential fault is increased from 1 Hz to 1 Hz for 10 seconds; the sampling rate of the current sensor in a specific branch is increased from 1 Hz to 100 Hz. Only the incremental data stream resulting from the increased sampling rate and the initial diagnostic results are uploaded, not all the raw data. This significantly optimizes the efficiency of uplink bandwidth utilization, ensuring that the cloud can obtain the necessary detailed data for in-depth analysis.

[0066] The cloud-based analysis module receives data uploaded from edge nodes and constructs a digital twin based on the received data. The digital twin analyzes the causes of faults based on multi-physics coupling and simulates the trend of fault propagation.

[0067] The cloud analytics module asynchronously and reliably receives data uploaded from a massive number of edge nodes via a high-throughput distributed message queue. This data includes high-frequency data packets for urgent faults, incremental data streams for potential faults, and characteristic data from routine monitoring. The digital twin built in the cloud, as shown in the attached diagram... Figure 3 As shown, it includes a local digital twin for each photovoltaic power station and a global digital twin for the photovoltaic power grid. The local digital twin is a dedicated digital twin for each independent photovoltaic power station; the global digital twin is a global model that can reflect the operating status of the entire regional photovoltaic power grid.

[0068] The local digital twin adopts an architecture that integrates physical models and data-driven models. The local digital twin includes a physical model layer, a data-driven layer, and a scene fingerprint.

[0069] Physical Model Layer: Based on the precise physical parameters and geometry of photovoltaic modules, cables, inverters, and other equipment, a finite element analysis model is established. This model is used to simulate the physical field distribution inside the equipment, such as solving the balance between Joule heating and heat dissipation in electro-thermal coupled fields, and simulating the temperature distribution of the equipment under different loads. Structural mechanics field analysis is performed on the deformation and stress distribution under wind loads and thermal stresses. A simulation benchmark conforming to physical laws is provided to ensure the extrapolation and interpretability of the model.

[0070] The data-driven layer employs a deep learning model combining a Long Short-Term Memory (LSTM) network with an attention mechanism. This model takes edge-up time-series data as input and outputs more refined, difficult-to-measure operational status parameters of devices (such as inverters) (e.g., health status, efficiency degradation factor). A key function of this layer is to correct biases in the physical model layer. By minimizing the mean square error between the data-driven layer's predictions and actual measurements, a "bias field" or "correction factor" is generated and fed back to the physical model layer to dynamically calibrate its simulation parameters (e.g., adjusting material properties and boundary conditions), enabling the digital twin to approximate the real power plant almost perfectly.

[0071] Scene fingerprinting is based on the power plant's equipment files (model, capacity, installation angle) and historically uploaded operational data (average irradiance, temperature range). A high-dimensional vector is constructed as the scene fingerprint for the power plant and stored in a fingerprint database. During analysis, the cosine similarity between the current operational data and historical scenes in the fingerprint database is calculated to quickly locate the typical operational scenario currently in which the power plant is located (such as "high irradiance at noon in summer" or "low irradiance on a cloudy day in winter"), providing context for subsequent analysis.

[0072] The global digital twin is itself a meta-model or model factory, integrating the experience of local digital twins from all photovoltaic power plants through a federated learning framework. After each power plant's local model is trained on local data, only the updated model parameters (not the original data) are encrypted and uploaded to the cloud. The cloud aggregates these updates to generate a more powerful and generalizable global model, which is then distributed to each local digital twin for collaborative evolution. State synchronization and command exchange between the local and global twins are achieved through a distributed data synchronization protocol (such as gRPC).

[0073] For different types of anomalies, the digital twin initiates the corresponding physical model for root cause analysis:

[0074] In response to abnormal electrical quantities (such as voltage drops or excessive harmonics), an electric field model is activated. In the simulated circuit topology, by solving circuit equations (such as Kirchhoff's laws), the energy transmission path is simulated, the impedance anomaly point is located, and the precise distribution of voltage and current at the fault point is calculated, thereby determining whether it is a loose connection in the line, a sudden change in load, or internal damage to the equipment.

[0075] For abnormal physical quantities (such as excessively high temperature or abnormal vibration), activate the thermal field model and the mechanical model. The thermal field model calculates the balance between operating losses (copper losses, iron losses) and environmental heat dissipation to determine the cause of overheating. The mechanical model analyzes vibration frequencies and modes to determine whether mechanical failure is caused by loose bolts, internal eccentricity, etc.

[0076] To address abnormal IV curves, an equivalent circuit model of the photovoltaic module (such as a single diode model) is activated. By adjusting parameters such as series resistance, parallel resistance, and photocurrent in the model, the fault mode that best matches the measured distorted IV curve is simulated, thereby accurately diagnosing whether it is a hot spot, aging, or shading.

[0077] The digital twin uses each device (module, combiner box, inverter) of a photovoltaic power plant as a graph node, electrical connections as edges, and the inverse of the combined line transmission efficiency and physical distance as edge weights to construct a dynamic graph model describing the power plant's topology. A temporal convolutional network (TCN) is used to extract the evolution of fault signals across multiple time scales, and a graph neural network (GNN) is used to simulate how the fault propagates along topological connections (edges) between nodes. By fusing spatiotemporal features, Monte Carlo simulations or physical rule-based deductions are performed within the digital twin, ultimately outputting a fault propagation probability matrix. This matrix clearly shows the probability, scope, and severity of the fault's propagation to other devices at different future time points (e.g., 1 hour later, 3 hours later), providing a quantitative basis for operation and maintenance decisions.

[0078] The secondary early warning response module generates a secondary early warning report based on the cause of the fault and the fault propagation trend, and sends it to the operation and maintenance terminal for response.

[0079] The secondary early warning response module is an intelligent decision support system deployed in the cloud. It receives analysis results from the digital twin and generates and issues secondary early warning reports through the following process:

[0080] The module has a built-in risk assessment engine. This engine comprehensively considers the following factors to quantify the risk of early warning events: the severity of the root cause of the fault (e.g., internal short circuit is critical, minor hot spot is a warning); the predicted scope and degree of impact in the fault propagation probability matrix (e.g., impact on the core inverter is high, impact on only a single string is low); and the estimated potential power generation loss and equipment repair costs.

[0081] Based on the above quantitative results, the warning level is automatically divided into multiple levels such as critical, serious, general, and caution.

[0082] Based on a predefined template, the module automatically generates a structured secondary early warning report. This report includes at least the following sections: Early Warning Summary: Event overview, risk level, and main conclusions; Fault Location and Diagnosis: Clearly identifies the faulty equipment / location and elaborates on the root cause of the fault derived from digital twin analysis; Impact Prediction: Displays the fault propagation trend in the form of visual charts (such as topology coloring diagrams and trend graphs), including potentially affected equipment, timeline estimates, and potential consequences; Maintenance Decision Recommendations: This is the core value of the report.

[0083] The system will match and recommend the optimal handling strategy library from the knowledge base based on the fault type and risk level, for example:

[0084] Example of a disposal strategy library:

[0085] For critical levels: It is recommended to shut down the machine immediately and arrange on-site repairs within 24 hours.

[0086] For critical levels: It is recommended to reduce the load during the next 72 hours and to plan maintenance within this week.

[0087] For general levels: It is recommended to focus on inspecting this area during the next planned inspection or to increase the frequency of drone inspections of this area.

[0088] For the level of concern: It is recommended to automatically list the spare parts (such as fuses, damaged components) and special tools that may be needed based on the cause of the failure, so that they can be used.

[0089] After generating the report, the module intelligently distributes it to the relevant maintenance personnel through multiple channels: automatically converting the early warning report into a standardized maintenance work order and pushing it to the power plant's asset management system or maintenance management platform. The work order clearly includes the fault location, recommended measures, risk level, and processing time limit. Through a dedicated APP or SMS / email gateway, the summary and key information of the early warning report are instantly pushed to the mobile terminals (such as mobile phones and tablets) of relevant maintenance personnel. For critical and major early warnings, sound and vibration alerts are provided. On the large visualization screen in the central control center, the status icon of the corresponding faulty equipment changes color according to the risk level (e.g., red represents critical), and simple handling suggestions are displayed, providing dispatchers with a global situational awareness.

[0090] The system provides an interface for operations and maintenance personnel to report the actual cause of the fault, the measures taken, and the time taken after handling the fault on-site. This valuable on-site feedback data will be recorded. The actual results of the on-site feedback will be compared with the previous predictive analysis of the digital twin. Through this comparison, the analytical model of the digital twin and the decision knowledge base of the secondary early warning response module can be continuously verified and optimized, forming a continuous optimization closed loop from analysis to action and then to verification.

[0091] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, any equivalent modifications or substitutions made by those skilled in the art to the relevant technical features will fall within the scope of protection of the present invention.

Claims

1. A photovoltaic power grid fault intelligent early warning system based on multi-source data fusion, characterized in that, include: The data sensing module deploys a multimodal sensor network at the distributed monitoring nodes of the photovoltaic power grid to acquire multi-source sensing data of the photovoltaic power grid in real time. The multi-source sensing data includes at least electrical quantity parameters, physical quantity parameters, image parameters, environmental parameters, and the IV curve of the photovoltaic string. The edge computing module deploys a preprocessing unit, a feature processing unit, and a preliminary early warning unit at the edge nodes of the photovoltaic grid. The preprocessing unit is used to clean and calibrate the acquired multi-source sensing data. The feature processing unit uses a cross-modal autoencoder to extract and fuse features from the preprocessed multi-source sensing data. The preliminary early warning unit uses an improved isolated forest model to perform preliminary fault detection on the fused features. The preliminary early warning response module triggers a preliminary early warning response strategy based on the preliminary fault detection results. When the preliminary fault detection result indicates an emergency fault, it disconnects the faulty equipment from the power grid, isolates the emergency fault, and uploads the fault data to the cloud for analysis. When the preliminary fault detection result indicates a potential fault, it adjusts the sampling rate of the multimodal sensor network and uploads the fault data and incremental data stream to the cloud for analysis. When the preliminary fault detection result indicates no fault, monitoring continues. The cloud-based analysis module receives data uploaded from edge nodes and constructs a digital twin based on the received data. The digital twin analyzes the causes of faults based on multi-physics coupling and simulates the trend of fault propagation. The secondary early warning response module generates a secondary early warning report based on the cause of the fault and the fault propagation trend, and sends it to the operation and maintenance terminal for response.

2. The photovoltaic power grid fault intelligent early warning system based on multi-source data fusion according to claim 1, characterized in that, The distributed monitoring nodes are deployed at the locations of photovoltaic strings, combiner boxes, inverters, transformers, and grid connection points in each photovoltaic power station in the photovoltaic grid. The multimodal sensor network includes at least electrical quantity sensors, physical quantity sensors, vision sensors, environmental sensors, and intelligent string controllers. The electrical quantity sensors are used to collect data on voltage, current, power, harmonics, and insulation resistance. The physical quantity sensors are used to collect data on temperature, vibration, stress, and deformation. The vision sensors are used to collect image data of hot spots, damage, and shading of photovoltaic modules. The environmental sensors are used to collect data on light intensity, wind speed, humidity, altitude, and dust concentration. The intelligent string controllers are used to collect IV curve data of the photovoltaic strings.

3. The intelligent early warning system for photovoltaic power grid faults based on multi-source data fusion according to claim 1, characterized in that, The preprocessing unit uses adaptive Kalman filtering to denoise the multi-source sensing data and uses a generative adversarial network to complete the missing multi-source sensing data. The cross-modal autoencoder includes an encoder and a decoder. The encoder uses a heterogeneous dual-branch coding structure to independently encode the preprocessed numerical parameters and image parameters in two branches to generate feature vectors. The generated dual-branch feature vectors are weighted and correlated through a self-attention mechanism. At the same time, the feature vectors of different modalities are mapped to the same feature space to achieve feature alignment. The decoder reconstructs the data through a multi-task learning framework and outputs fused features.

4. The photovoltaic power grid fault intelligent early warning system based on multi-source data fusion according to claim 1, characterized in that, The improved isolated forest model uses a two-level detection framework to perform preliminary fault detection on the fused features. The two-level detection framework includes string-level detection and system-level detection. The string-level detection constructs an IV curve feature library under normal operating conditions of the photovoltaic power grid based on environmental parameters and the IV curves of the photovoltaic strings. When the features on the IV curves of the photovoltaic strings deviate from the IV curve feature library by more than a preset deviation threshold, it is determined to be a string-level potential fault. The string-level potential fault information and the fused feature vector are then input into the isolated forest model after knowledge distillation. The system-level detection uses an isolated tree to score anomalies. When the anomaly score is greater than a first preset threshold, it is determined to be an emergency fault. When the anomaly score is less than or equal to the first preset threshold but greater than a second preset threshold, it is determined to be a potential fault. When the anomaly score is less than or equal to the second preset threshold, it is determined to be a fault-free condition.

5. The intelligent early warning system for photovoltaic power grid faults based on multi-source data fusion according to claim 1, characterized in that, When the initial fault detection result is no fault, the sampling rate of the multimodal sensor network is maintained in the normal mode to acquire multi-source sensing data of the photovoltaic grid in real time. When the initial fault detection result is an emergency fault, a trip command is sent to the actuator of the faulty device through the hard-wired interface. The actuator disconnects the faulty device from the photovoltaic grid according to the trip command to achieve emergency fault isolation. When the initial fault detection result is a potential fault, a sampling rate adjustment command is sent to the sensor manager through the built-in management interface to adjust the sampling rate of the multimodal sensor network from the normal mode to the abnormal mode to acquire the incremental data stream of the photovoltaic grid.

6. The photovoltaic power grid fault intelligent early warning system based on multi-source data fusion according to claim 1, characterized in that, The cloud-based analysis module receives the fusion features of fault data, incremental data streams, and multi-source sensing data uploaded by edge nodes through a distributed message queue, and constructs a local digital twin of each photovoltaic power station and a global digital twin of the photovoltaic power grid based on the received data. The local digital twin includes a physical model layer, a data-driven layer, and a scene fingerprint. The physical model layer simulates the internal physical field distribution of the photovoltaic power station equipment through finite element analysis and integrates multi-physics coupling equations to simulate the cross-domain parameter interaction of the photovoltaic power station. The data-driven layer uses a long short-term memory network (LSTM) and an attention mechanism to output the real-time operating status parameters of the photovoltaic power station equipment. It also corrects the simulation deviation of the physical model layer by minimizing the mean square error between the predicted and measured values. The scene fingerprint is based on the equipment files of the photovoltaic power station and the historical uploaded data of the edge nodes to build a scene fingerprint library. It also uses cosine similarity matching to locate the current operating scene of the photovoltaic power station. The global digital twin integrates the local digital twins of all photovoltaic power plants through federated learning, and realizes state interaction with the local digital twins through a distributed data synchronization protocol.

7. The photovoltaic power grid fault intelligent early warning system based on multi-source data fusion according to claim 1, characterized in that, In response to electrical anomalies, the digital twin uses an electric field model to simulate the energy transmission path in the circuit topology, and combines the circuit equations to solve for the voltage and current distribution at the fault point to obtain the cause of the fault. For anomalies in physical quantities, the digital twin calculates the thermal balance relationship between the operating losses of photovoltaic power station equipment and environmental heat dissipation through a thermal field model, and analyzes structural deformation and stress distribution through a mechanical model to obtain the cause of the fault. For anomalies in the IV curve, the digital twin simulates the distortion law of the IV curve through an equivalent circuit model of the photovoltaic module to obtain the cause of the fault.

8. The photovoltaic power grid fault intelligent early warning system based on multi-source data fusion according to claim 1, characterized in that, The digital twin uses the equipment of the photovoltaic power station as graph nodes, the electrical connection relationship between the equipment as edges, and the comprehensive coefficient of line transmission efficiency and distance as edge weights to construct a dynamic topology graph of the photovoltaic power station. The temporal dynamic features of fault development are extracted by Temporal Convolutional Network (TCN), and the spatial propagation features of faults in the topology are captured by Graph Neural Network (GNN) to simulate the diffusion path, impact range and impact degree of photovoltaic grid faults at different time scales, and output the fault diffusion probability matrix.

Images