Holographic twin distributed photovoltaic collection control method and system

Abstract

The present invention discloses a holographic twin distributed photovoltaic collection and control method and system, which belongs to the field of photovoltaic collection and control technology, and includes the following steps: deploying a lightweight twin model in the edge node of the photovoltaic device; performing fault detection through the lightweight twin model, locating the source of the fault and predicting the fault spread range; formulating a blocking control strategy based on the prediction results of the fault spread range, and updating the lightweight twin model. By deploying a lightweight twin model at the edge node of the photovoltaic device, the present invention realizes the organic combination of distributed fault detection and centralized decision-making control, so that the lightweight twin model can run efficiently on resource-constrained edge devices. At the same time, the distributed deployment reduces data transmission delay and achieves a millisecond-level fault response speed. This improvement in timeliness far exceeds expectations and provides a strong guarantee for the safe and stable operation of the photovoltaic system.

Description

Technical Field

[0001] The present invention relates to the field of photovoltaic collection and control technology, and in particular to a holographic twin distributed photovoltaic collection and control method and system. Background Art

[0002] Traditional photovoltaic power generation system fault detection relies primarily on centralized monitoring architectures, which suffer from significant data transmission delays, limited processing capabilities, and untimely fault response. Existing technologies typically employ a single anomaly detection algorithm to monitor photovoltaic equipment. However, due to the widespread geographical distribution of photovoltaic sites and the large number of devices, centralized processing approaches struggle to achieve real-time response. This often leads to missed opportunities for optimal control, especially when faults spread rapidly.

[0003] Existing fault detection methods mostly focus on the abnormal state of a single device, lacking in-depth analysis of how faults propagate across devices. When a device in a photovoltaic system fails, the fault can spread to adjacent devices through electrical connections, environmental factors, and other pathways. However, traditional methods are unable to effectively model this propagation mechanism, resulting in inaccurate fault source location and unreliable prediction of the fault's spread. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is that traditional methods may lead to inaccurate fault source positioning and unreliable prediction of the spread range.

[0006] To solve the above technical problems, the present invention provides the following technical solution: a holographic twin distributed photovoltaic collection and control method, which includes the following steps:

[0007] Deploy lightweight twin models within edge nodes of photovoltaic equipment;

[0008] Use lightweight twin models to detect faults, locate the source of faults, and predict the scope of fault spread;

[0009] Based on the prediction results of the fault propagation range, a blocking control strategy is formulated and the lightweight twin model is updated.

[0010] As a preferred solution of the holographic twin distributed photovoltaic collection control method described in the present invention, the step of constructing the lightweight twin model includes:

[0011] Collect data information of photovoltaic equipment and build data sets;

[0012] Use anomaly detection algorithms to filter dirty data in the data set and extract feature data;

[0013] Constructing training data through feature data, and building a lightweight twin model based on the training data;

[0014] The data information includes operation data and fault equipment data.

[0015] As a preferred solution of the holographic twin distributed photovoltaic collection and control method described in the present invention, the step of performing fault detection through a lightweight twin model includes:

[0016] Input the operating data and faulty equipment data into the lightweight twin model to obtain anomaly scores;

[0017] Compare anomaly scores and set anomaly thresholds;

[0018] When the anomaly score is greater than the set anomaly threshold, it is judged as abnormal data and output to the decision center;

[0019] When the abnormality score is not greater than the set abnormality threshold, the abnormal data is spliced, and the abnormal data after data splicing includes data labels and fault information;

[0020] The data labels and fault information are input into the lightweight twin model as new data to re-perform fault detection.

[0021] The beneficial effects of this preferred technical solution are: the threshold comparison and iterative detection mechanism realizes the continuous optimization of fault detection accuracy, and the data labels and fault information are recombined through data splicing to form an enhanced feature vector for re-detection, which can effectively identify potential faults and early abnormal signs that are easily missed by traditional single detection.

[0022] As a preferred solution of the holographic twin distributed photovoltaic collection and control method described in the present invention, the steps of locating the fault source and predicting the fault spread range include:

[0023] Deploy a propagation tracing network on the edge nodes of photovoltaic equipment and define the fault propagation direction and data flow topology in the propagation tracing network;

[0024] Establish a fault propagation association mechanism and fault propagation chain based on the fault propagation direction and data flow topology in the propagation tracing network;

[0025] Performing source tracing analysis on the faulty device data based on the fault propagation chain to determine the location of the fault source in the fault propagation chain;

[0026] The fault propagation range is predicted based on the propagation direction of the fault propagation chain and the data flow topology relationship.

[0027] The beneficial effect of this preferred technical solution is that by constructing a propagation tracing network and a fault propagation chain, the root cause of the fault can be accurately located in a complex multi-device fault scenario, avoiding the limitations of the traditional method of "treating the symptoms rather than the root cause".

[0028] As a preferred solution of the holographic twin distributed photovoltaic collection and control method described in the present invention, the prediction of the fault propagation range includes:

[0029] Calculating the fault diffusion speed and diffusion path according to the fault propagation chain;

[0030] The degree of output impact based on the fault level of the fault source and the electrical connection relationship between the photovoltaic devices;

[0031] Based on the diffusion speed and impact level, the time range of the fault diffusion and the final impact boundary are predicted.

[0032] The beneficial effects of this preferred technical solution are: through a multi-dimensional diffusion range prediction mechanism, an accurate quantitative assessment of the fault impact is achieved. Not only the diffusion speed and path are calculated, but also the degree of impact is evaluated in combination with the fault level and electrical connection relationship, ultimately forming an impact boundary prediction in both time and space dimensions, which enables fault handling to shift from passive response to active intervention.

[0033] As a preferred solution of the holographic twin distributed photovoltaic collection control method described in the present invention, the step of formulating the blocking control strategy includes:

[0034] Determine the scope of photovoltaic equipment that needs to be isolated based on the fault propagation range;

[0035] When the spread rate exceeds the preset threshold, an emergency blocking instruction is generated; when the spread rate does not exceed the preset threshold, a progressive isolation strategy is formulated;

[0036] Send the control instructions to the corresponding edge node for execution.

[0037] As a preferred solution of the holographic twin distributed photovoltaic collection control method described in the present invention, the steps of executing the control instruction and evaluating the control effect include:

[0038] The edge node receives and executes the control instruction to adjust the specified photovoltaic equipment;

[0039] Real-time monitoring of the operating status and fault propagation of photovoltaic equipment after executing control instructions;

[0040] Compare and analyze the actual fault location information monitored with the predicted fault source;

[0041] The prediction accuracy of the fault location information is judged based on the comparative analysis results, and the judgment results are fed back to the lightweight twin model.

[0042] As a preferred solution of the holographic twin distributed photovoltaic collection control method described in the present invention, the step of updating the lightweight twin model includes:

[0043] When the fault location information is correctly determined, the lightweight twin model is continuously monitored;

[0044] When the fault location information is incorrectly determined, updating the parameters of the lightweight twin model;

[0045] The parameter update of the lightweight twin model includes:

[0046] Calculate the anomaly score and fault level score of the fault source;

[0047] When the product of the abnormality score and the fault level score is greater than a first set value, no parameter update is performed on the lightweight twin model;

[0048] When the product of the abnormality score and the fault level score is not greater than the first set value, a secondary judgment is performed.

[0049] The beneficial effects of this preferred technical solution are: whether to update parameters is determined based on the correctness of the fault location judgment, and the update trigger condition is established by calculating the product of the abnormality score and the fault level score. The intelligent update decision mechanism avoids the impact of frequent invalid updates on the stability of the lightweight twin model.

[0050] As a preferred solution of the holographic twin distributed photovoltaic collection control method described in the present invention, the second-level judgment step includes:

[0051] Determining whether the abnormality score is greater than a second set value;

[0052] When the abnormality score is greater than the second set value, calculating an increment of the abnormality score, and calculating an abnormality score after enhancing the processing capability;

[0053] When the abnormality score is not greater than the second set value, the increment of the abnormality score is set to 0, and the abnormality score after the processing capability is updated is calculated.

[0054] The present invention provides a holographic twin distributed photovoltaic collection and control system.

[0055] To solve the above technical problems, the present invention provides the following technical solutions: a holographic twin distributed photovoltaic collection and control system, comprising a lightweight twin module, a fault detection module, a fault diffusion prediction module, and a control strategy formulation module;

[0056] The lightweight twin module is used to simulate the operating status of the photovoltaic equipment;

[0057] The fault detection module uses a lightweight twin model to perform fault detection;

[0058] The fault spread prediction module analyzes the initial state of the fault based on the lightweight twin model and predicts the fault spread range;

[0059] The control strategy formulation module formulates a blocking control strategy according to the prediction result of the fault diffusion range.

[0060] Beneficial effects of the present invention:

[0061] By deploying lightweight twin models at the edge nodes of photovoltaic equipment, an organic combination of distributed fault detection and centralized decision-making and control is achieved, enabling the lightweight twin models to run efficiently on resource-constrained edge devices. At the same time, distributed deployment reduces data transmission delays and achieves millisecond-level fault response speeds. This improvement in timeliness far exceeds expectations and provides strong guarantees for the safe and stable operation of photovoltaic systems.

[0062] By introducing a propagation traceability network mechanism and establishing a fault propagation chain, a technological leap has been achieved, moving from single-point detection to global propagation analysis. This mechanism not only accurately locates the source of a fault but, more importantly, predicts its propagation path and impact range. This predictive capability forms an intelligent linkage mechanism with real-time blocking control strategies. When the system detects that the fault propagation rate exceeds a preset threshold, it automatically triggers an emergency blocking strategy, achieving proactive fault prevention and control rather than a passive response. This shift in philosophy from treatment to prevention has produced significant protective benefits.

[0063] By comparing and analyzing the execution effect of the control strategy with the accuracy of the fault location prediction in real time, it is possible to automatically identify the changing trend of the lightweight twin model performance and make differentiated parameter adjustments based on multi-level judgment logic. As the operating time increases, the detection accuracy continues to improve, forming a virtuous cycle of "the more it is used, the more accurate it becomes." BRIEF DESCRIPTION OF THE DRAWINGS

[0064] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0065] Figure 1 A structural diagram of a holographic twin distributed photovoltaic collection and control method provided for one embodiment of the present invention. DETAILED DESCRIPTION

[0066] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0067] Example 1, with reference to Figure 1 , is an embodiment of the present invention, which provides a holographic twin distributed photovoltaic collection control method, comprising the following steps:

[0068] S1. Deploy a lightweight twin model in the edge node of the photovoltaic equipment.

[0069] The steps of constructing the lightweight twin model include:

[0070] S1.1. Collect data information of photovoltaic equipment and construct a data set, where the data information includes operating data and fault equipment data;

[0071] In this implementation, edge nodes utilize an ARM Cortex-A78 processor. Each edge node connects to three to eight photovoltaic inverters via an RS485 bus and communicates with the upstream combiner box via an Ethernet interface. The edge nodes also have a built-in Linux operating system and run a lightweight inference engine based on the TensorFlow Lite framework.

[0072] The scope of operational data collection covers the key operational parameters of the photovoltaic power generation system. Electrical parameters include but are not limited to DC side voltage, DC side current, AC side voltage, AC side current, active power, reactive power, and power factor.

[0073] Faulty equipment data mainly includes historical fault records and real-time fault information. Historical fault records cover all fault events that occurred in the past 12 months, including fault type, fault time, fault duration, fault impact range, and fault handling measures.

[0074] S1.2. Use anomaly detection algorithms to filter dirty data in the dataset and extract feature data;

[0075] In this example, an outlier detection method based on the isolation forest algorithm was used to filter dirty data from the photovoltaic equipment dataset. The isolation forest algorithm identifies anomalous data points by constructing a set of randomized decision trees. The algorithm parameters were set as follows: the number of decision trees was set to 100, the subsample size of each tree was set to 256 data points, and the anomaly ratio threshold was set to 5%. This means that the top 5% of the anomaly scores after sorting are marked as anomalous.

[0076] The identification criteria for dirty data are divided into four levels. The first level is physical limit checking, which marks data outside the physical operating range of the device as dirty data. Examples include DC voltage exceeding 1200 volts, DC current exceeding 50 amperes, and ambient temperature exceeding 60 degrees Celsius or falling below -40 degrees Celsius. The second level is logical consistency checking, which identifies data combinations that violate the laws of physics, such as power output under zero irradiance conditions or power output far exceeding the rated power. The third level is temporal continuity checking, which identifies data anomalies caused by sensor failures. When a monitoring parameter maintains the same value for 15 consecutive sampling cycles, it is considered a sensor stuck fault. When a parameter value jumps by more than 10 times the normal range between adjacent sampling cycles, it is considered sensor noise interference. The fourth level is statistical anomaly checking, which establishes a statistical distribution model for each parameter based on six months of historical operating data. Data that deviates from the mean by more than four standard deviations is marked as a statistical anomaly.

[0077] In this embodiment, feature extraction extracts the operating characteristics of the photovoltaic device from the time domain dimension.

[0078] Time domain feature extraction includes first-order and second-order statistical features. First-order statistical features include nine features: arithmetic mean, geometric mean, harmonic mean, median, mode, maximum, minimum, range, and interquartile range. Second-order statistical features include six features: variance, standard deviation, coefficient of variation, skewness, kurtosis, and rate of change.

[0079] Time-domain trend characteristics are obtained through sliding window analysis, with a window length of 60 sampling points corresponding to one minute of data. Trend characteristics include linear trend slope, quadratic trend curvature, trend goodness of fit, and the number of trend change points. Correlation characteristics are obtained by calculating the Pearson correlation coefficient between different monitored parameters, focusing on the correlations between power and irradiance, voltage and current, and temperature and efficiency.

[0080] In another optional implementation, the operating characteristics of the photovoltaic system can also be extracted from the frequency domain dimension.

[0081] Frequency domain feature extraction uses the Fast Fourier Transform (FFT) method to convert the time domain signal into the frequency domain for analysis. The sampling frequency is 1 Hz, and the analysis frequency range is 0 to 0.5 Hz. Frequency domain features include eight features: dominant frequency, subdominant frequency, spectral center of gravity, spectral variance, spectral skewness, spectral kurtosis, frequency domain energy, and spectral entropy.

[0082] S1.3. Constructing training data using feature data, and building a lightweight twin model based on the training data;

[0083] The lightweight twin model uses a two-branch fully connected neural network architecture with shared weights. The network structure consists of three main parts: feature encoding branch, similarity calculation branch, and output branch.

[0084] The feature encoding branch uses a four-layer fully connected network structure. The input layer receives a 35-dimensional feature vector. The first hidden layer contains 128 neurons, the second hidden layer contains 64 neurons, and the third hidden layer contains 32 neurons. The encoding output layer contains 16 neurons to generate an encoded representation of the device's operating mode. ReLU activation functions are used between each layer to provide nonlinear transformation capabilities. A dropout layer is added after the first and second hidden layers, with a dropout ratio of 0.3 to prevent overfitting.

[0085] The similarity calculation branch measures the similarity of device operating modes by calculating the Euclidean distance between two code vectors. The distance calculation formula is the square root of the sum of the squared differences between the corresponding elements of the two 16-dimensional code vectors. To enhance the robustness of the distance calculation, the code vectors are L2-normalized before the distance calculation.

[0086] The output branch converts the Euclidean distance into a similarity score between 0 and 1. This conversion uses a negative exponential function, where the similarity equals the negative distance. This design ensures that smaller distances result in a similarity closer to 1, while larger distances result in a similarity closer to 0, which aligns with human intuition about similarity.

[0087] Model training is optimized using a contrastive loss function. The contrastive loss function is designed to minimize the distance between positive pairs while maximizing the distance between negative pairs. The loss function consists of two terms: a similarity term and a dissimilarity term. The similarity term is the square of the distance between the positive pairs, while the dissimilarity term is the square of the difference between the maximum boundary and the negative pairs. The boundary parameter is set to 1.0.

[0088] S2. Perform fault detection through lightweight twin models to locate the source of the fault and predict the scope of fault spread.

[0089] The steps for fault detection using lightweight twin models include:

[0090] Input the operating data and faulty equipment data into the lightweight twin model to obtain anomaly scores, compare the anomaly scores, and set anomaly thresholds;

[0091] In this implementation, anomaly score calculation is performed through forward reasoning within a lightweight twin model. Real-time operating data and each reference fault sample are fed into two branches of the model. After encoding through a four-layer fully connected network, a 16-dimensional feature vector is generated. The two 16-dimensional vectors are then Euclidean distance calculated to obtain a distance value, which is then converted into a similarity score using a negative exponential function.

[0092] The specific calculation process is as follows: First, the sum of the squared differences between the elements of the two encoded vectors is calculated, then the square root is taken to obtain the Euclidean distance. Finally, the similarity score between 0 and 1 is calculated using the formula: similarity = exp(-distance). The closer the similarity score is to 1, the more similar the current operating state is to the reference fault mode, and the higher the degree of abnormality.

[0093] The anomaly threshold is determined based on statistical analysis of historical data. By analyzing the past six months of normal operation data and calculating the distribution characteristics of anomaly scores under normal conditions, the 95th percentile is used as the baseline threshold for anomaly detection, with a specific value set at 0.75.

[0094] When the anomaly score is greater than the set anomaly threshold, it is judged as abnormal data and output to the decision center;

[0095] When the abnormality score is not greater than the set abnormality threshold, the abnormal data is spliced, and the abnormal data after data splicing includes data labels and fault information;

[0096] The data labels and fault information are input into the lightweight twin model as new data to re-perform fault detection.

[0097] In this implementation, when the anomaly score exceeds the anomaly threshold of 0.75, the current operating state is determined to be abnormal, and the abnormal data output process is initiated. The abnormal data includes the following information: device identification code, anomaly detection timestamp, anomaly score value, matching fault type, current device operating parameters, anomaly level indicator, and recommended treatment measures. The abnormal data is transmitted to the decision center via the edge node's communication interface. Upon receiving the abnormal data, the decision center analyzes and stores the data, triggering the corresponding alarm and processing processes. Alarm information is notified to operations and maintenance personnel through various means, including display on the monitoring interface, SMS notifications, and email push notifications.

[0098] When the anomaly score is no greater than the anomaly threshold of 0.75, the system initiates data concatenation and enhanced detection to further improve the accuracy and sensitivity of fault detection. The data concatenation process combines current operating data with historical data and contextual information to form a richer feature representation. Data labels are generated based on a multi-dimensional analysis of the current operating status. Time labels, including hour, date, season, and weather status, characterize time-related operating patterns. Operational labels, including load level, efficiency range, and stability, characterize equipment operating characteristics. Environmental labels, including irradiance level, temperature range, and wind speed level, characterize environmental factors. Fault information is constructed based on the analysis of the current anomaly score. Although the anomaly score does not exceed the alarm threshold, it still contains valuable fault propensity information. Fault information includes the most similar fault type and its similarity, the second most similar fault type and its similarity, fault probability distribution, and predicted fault development trends.

[0099] Steps A1 to A4 are included to locate the fault source and predict the fault spread range:

[0100] A1. Deploy a propagation and tracing network on the edge nodes of photovoltaic equipment and define the fault propagation direction and data flow topology in the propagation and tracing network.

[0101] In this example, a 25MW photovoltaic power plant was selected as the application scenario. The plant consists of 100 250kW inverters, grouped into subarrays of 10 inverters. The transmission and traceability network is deployed using a layered architecture, with an edge node deployed in each subarray area, for a total of 10 edge nodes. Each edge node is responsible for monitoring the operating status of the 10 inverters within its subarray. Edge nodes are connected via Ethernet to form a mesh topology, ensuring communication reachability between any two nodes.

[0102] The definition of fault propagation direction is based on the electrical connection characteristics and physical layout features of the PV power plant. At the electrical level, fault propagation is defined based on the current flow from the PV module to the inverter, from the inverter to the combiner box, and from the combiner box to the transformer. At the physical level, fault propagation between adjacent devices takes into account environmental factors, including physical phenomena such as temperature conduction, electromagnetic interference, and vibration transmission. Data flow topology relationships are stored using an adjacency matrix. Each element in the matrix indicates whether a direct fault propagation path exists between two devices. When there is an electrical connection or physical proximity between the devices, the corresponding matrix element is set to 1; otherwise, it is set to 0.

[0103] A2. Establish a fault propagation association mechanism and fault propagation chain based on the fault propagation direction and data flow topology in the propagation tracing network.

[0104] The establishment of the association propagation mechanism is based on the analysis of multiple associations between devices, primarily encompassing three dimensions: electrical association, spatial association, and functional association. Electrical association is determined by analyzing the connections between devices within an electrical circuit. Devices on the same circuit have strong associations, while weak associations exist between different circuits through combiner boxes or transformers. Spatial association is determined by calculating the physical distance between devices. The closer the devices, the stronger the association. A strong association is established when the distance between devices is less than 50 meters, a moderate association is established between 50 and 100 meters, and no association is established when the distance is greater than 100 meters.

[0105] Fault propagation chains are constructed using a graph traversal algorithm, with each device as a node in the graph and the relationships between devices as edges. When the system needs to analyze the propagation impact of a fault source, it performs a breadth-first search starting from that source node, building chains in layers according to the length of the propagation path. The first-level propagation chain includes devices directly associated with the fault source, the second-level propagation chain includes devices associated with the first-level devices, and so on until the third-level propagation chain. Each propagation chain records the identities and propagation directions of all devices along the propagation path, forming a complete description of the fault propagation trajectory.

[0106] A3. Based on the fault propagation chain, perform source tracing analysis on the faulty device data to determine the location of the fault source in the fault propagation chain;

[0107] Fault tracing analysis begins with the device currently experiencing an anomaly and traces back along the established fault propagation chain. When an edge node detects an anomaly on inverter INV-025, the system first queries the device's position within the propagation chain, identifying possible upstream influencing factors as other inverters INV-021 through INV-024 connected to the same combiner box, as well as the physically adjacent inverters INV-026 and INV-030. The system then retrieves the operating data for these suspected devices over the past hour and determines the fault's origin by comparing the time series and numerical values ​​of each device's anomaly scores.

[0108] The decision logic for source tracing analysis is based on the dual principles of time priority and score priority. The time priority principle requires that the device that first experiences an anomaly is more likely to be the source of the fault. The system determines the order in which anomalies occurred by comparing the timestamps at which the anomaly scores of each device exceeded the threshold. The score priority principle requires that devices with higher anomaly scores have stronger fault characteristics. When multiple devices experience anomalies within a similar timeframe, the device with the highest anomaly score is selected as the primary source of the fault. Through a comprehensive analysis of the time series and score distribution, inverter INV-023 was ultimately determined to be the source device of this fault. This device is upstream in the fault propagation chain and has the highest anomaly score among all related devices.

[0109] A4. Predict the fault propagation range based on the propagation direction of the fault propagation chain and the data flow topology relationship.

[0110] In this implementation, inverter INV-023 is used as the fault source, and diffusion prediction is performed along the forward path of the propagation chain. The first-level diffusion range includes inverters INV-021, INV-022, INV-024, and INV-025 connected to the same combiner box. The second-level diffusion range includes inverters connected to adjacent combiner boxes and other physically adjacent devices. Diffusion parameters are set based on the fault type and severity. If the fault is electrical and the anomaly score is greater than 0.8, the predicted diffusion will affect the first-level devices within 15 minutes and the second-level devices within 45 minutes.

[0111] The spread range prediction also considers the impact of the equipment's operating status and environmental conditions on the spread process. Equipment in good operating condition has a certain resistance to fault propagation, slowing the spread or mitigating its impact. Equipment in poor operating condition, however, may accelerate the spread of the fault. Adverse environmental conditions, such as high temperatures and strong winds, can increase the likelihood and severity of fault propagation. The system uses real-time environmental data to refine the prediction results. The final prediction results show that under current conditions, the fault may affect nine inverters within a 3×3 area centered on INV-023, with a total affected capacity of 2.25 MW. The estimated time for spread is one hour.

[0112] The prediction of the fault spread range includes B1~B2:

[0113] B1. Calculating the fault diffusion speed and diffusion path according to the fault propagation chain;

[0114] The fault propagation rate is calculated based on a comprehensive analysis of three primary factors: fault type, device characteristics, and propagation distance. For an electrical fault originating from inverter INV-023, the baseline propagation rate is determined based on historical data to be one adjacent device every 15 minutes. The propagation distances between devices are obtained through the PV power plant's geographic information system. The physical distance between INV-023 and its four directly adjacent devices, INV-021, INV-022, INV-024, and INV-025, is 25 meters. The propagation time calculated using the baseline rate is 15 minutes. Taking into account environmental factors, the current ambient temperature is 35 degrees Celsius, 30 degrees Celsius higher than the standard temperature, which increases the propagation rate by 20%, shortening the actual propagation time to 12 minutes.

[0115] The diffusion path calculation uses a multi-path parallel analysis method. Four main diffusion paths exist from the fault source, INV-023. The first path propagates along the electrical connection to the combiner box DCB-03, passing through INV-021 and INV-022 to reach the combiner box, potentially affecting other inverters connected to the same combiner box. The second path propagates northward along the physical proximity, passing through INV-024 and INV-034 in sequence. The third path propagates southward, passing through INV-025 and INV-015. The fourth path propagates eastward and westward, passing through INV-013 and INV-033, respectively. The system evaluates the propagation probability of each path. The highest probability is 0.9 for the electrical connection path, 0.7 for the physical proximity path, and 0.4 for the other paths.

[0116] B2. Output the impact level based on the fault level of the fault source and the electrical connection relationship between the photovoltaic devices;

[0117] The fault level assessment is based on an analysis of the anomaly score and fault characteristics of the fault-causing device, INV-023. The device's current anomaly score is 0.87. According to the pre-set grading standard, an anomaly score of 0.8 or higher is considered a severe fault level, a score between 0.6 and 0.8 is considered a moderate fault level, and a score between 0.4 and 0.6 is considered a minor fault level. Combined with an analysis of the device's specific fault characteristics, INV-023 exhibited an abnormal drop in output power and fluctuations in DC current. These characteristics closely matched the typical fault patterns of power devices within the inverter, further confirming the severe fault level.

[0118] The analysis of the electrical connections between PV devices is based on the power plant's electrical system diagram. INV-023 has a strong electrical coupling with the other four inverters connected to the same combiner box, DCB-03. A fault in INV-023 would directly impact the combiner box's current distribution and voltage stability, further impacting INV-021, INV-022, INV-024, and INV-025 to varying degrees. The degree of impact is quantified using an impact coefficient: 0.8 for devices directly electrically connected to the fault source, 0.4 for physically adjacent but electrically independent devices, and 0.1 for devices more distant. Based on this assessment method, the impact on system output is shown to be severe for the combiner box, DCB-03, and the four directly connected inverters, moderate for the six adjacent inverters, and minor or no impact for the remaining devices.

[0119] B3. Based on the diffusion speed and impact level, predict the time range of the fault diffusion and the final impact boundary.

[0120] The timeframe for fault propagation is predicted based on a time series analysis of the previously calculated propagation speed and path. Starting from the fault source, INV-023, it is predicted that the fault will spread to the four directly adjacent devices on the first tier within 12 minutes, to adjacent devices on the second tier within 36 minutes, and reach its maximum spread within 60 minutes. This timeframe also takes into account possible control intervention. If emergency isolation measures are implemented within 15 minutes of the fault occurrence, the fault can be effectively prevented from spreading to devices on the second tier, limiting the impact to less than 30 minutes. If no control measures are implemented, the fault propagation process will stabilize after 90 minutes, at which point the system will reach a new equilibrium state.

[0121] The prediction of the final impact boundary is defined by comprehensively considering three dimensions: geographical scope, equipment scope, and capacity scope. The geographical scope is centered on the fault source INV-023, and all potentially affected equipment is included in a circular area with a radius of 100 meters. This area covers three adjacent sub-array areas. The equipment scope includes 9 inverters, 3 junction boxes, and related monitoring equipment, totaling 15 major equipment units. The capacity scope involves an installed capacity of 2.25MW, accounting for 9% of the total capacity of the entire power station, and has an observable impact on the overall power generation capacity of the power station. The predicted results of the impact boundary provide important decision-making basis for subsequent emergency response and control strategy formulation, enabling operation and maintenance personnel to accurately assess the impact scope and severity of the fault.

[0122] S3. Develop a blocking control strategy based on the prediction results of the fault propagation range and update the lightweight twin model.

[0123] The steps of formulating the blocking control strategy include C1 to C3:

[0124] C1. Determine the scope of photovoltaic equipment that needs to be isolated based on the fault propagation range;

[0125] In this example, based on the predicted spread of the fault on inverter INV-023 in the previous steps, the system begins to determine the appropriate device isolation range. Based on the predicted analysis, the fault's final impact boundary includes nine inverters within a 3×3 area centered on INV-023. Specifically, the nine inverters listed are INV-013, INV-021, INV-022, INV-023, INV-024, INV-025, INV-033, INV-034, and INV-015. The system classifies these nine devices according to the degree of impact. Among them, INV-023, as the source device of the fault, is listed as a first-level isolation object and must be removed immediately; INV-021, INV-022, INV-024 and INV-025, which are directly electrically connected to INV-023, are listed as second-level isolation objects and require preventive isolation before the fault spreads; INV-013, INV-033, INV-034 and INV-015, which are physically adjacent, are listed as third-level isolation objects as a backup isolation plan.

[0126] Determining the isolation scope also requires consideration of electrical system continuity and the economic viability of power plant operations. System analysis revealed that isolating all nine devices would completely shut down the entire subarray, significantly impacting the power plant's revenue. Through optimization analysis, the system ultimately determined a minimum isolation scope encompassing the fault source, INV-023, and the other four devices connected to the same combiner box. Isolating a total of five devices effectively prevented the spread of the fault while keeping outage losses within acceptable limits. The remaining four physically adjacent devices will not be included in the isolation scope for now, but require enhanced monitoring, and additional isolation measures will be implemented immediately if any anomalies are detected.

[0127] C2: When the spread rate exceeds the preset threshold, an emergency blocking instruction is generated. When the spread rate does not exceed the preset threshold, a progressive isolation strategy is formulated;

[0128] The threshold determination of the spread rate is based on the system's preset hierarchical response mechanism. When the fault spread rate exceeds the preset threshold of affecting one device every 10 minutes, the system determines that it is in rapid spread mode and needs to initiate an emergency blocking instruction. The current spread rate of the INV-023 fault is affecting one device every 12 minutes, which is slightly lower than the emergency threshold. However, considering the severity of the fault and the potential scope of impact, the system has decided to adopt an emergency blocking strategy after a comprehensive assessment to ensure safety. The content of the emergency blocking instruction includes immediately cutting off all electrical connections at the source of the fault, INV-023, and at the same time, preventively disconnecting the other four inverters in the same combiner box DCB-03. The entire operation process must be completed within 5 minutes.

[0129] If the spread rate does not exceed the preset threshold, the system will develop a progressive isolation strategy as a backup plan. This progressive isolation strategy is implemented in a step-by-step manner. First, the device at the source of the fault is isolated and the spread of the fault is observed. If the spread is controlled, further operations are stopped. If the spread continues, other devices are gradually isolated according to a predetermined sequence. This strategy has the advantage of maximizing the power plant's generating capacity, but requires longer observation and decision-making time. It is suitable for faults with slower spread and relatively controllable impacts. The progressive isolation interval is set at 15 minutes per step, and the overall completion time is controlled within 45 minutes.

[0130] C3. Send the control instructions to the corresponding edge node for execution.

[0131] Control instructions are generated using a standardized format, encompassing five essential elements: instruction type, target device, execution time, operation details, and confirmation requirements. For the current fault, the system generated an emergency blocking instruction, targeting five inverters (INV-021 through INV-025). The execution time requirement is three minutes after receiving the instruction, and the operation involves disconnecting all electrical connections on both the AC and DC sides of the device. Confirmation requires that each device transmit a confirmation signal after isolation is complete. Control instructions are encrypted using digital signature technology to ensure the security and integrity of instruction transmission and prevent tampering or forgery.

[0132] The steps of executing the control instruction and evaluating the control effect include D1 to D4:

[0133] D1. The edge node receives and executes the control instruction to adjust the specified photovoltaic equipment;

[0134] In this embodiment, upon receiving the emergency blocking control command, the edge node immediately initiates the command parsing and execution process. The edge node first performs format verification and permission checks on the received command to confirm the legitimacy of the command source and the validity of the operation permissions. After verification, the edge node identifies the target devices to be operated as INV-021 through INV-025 based on the command content and checks their current operating status and operability. Once the system confirms that all target devices are controllable, it begins executing the isolation operations in a predetermined order.

[0135] D2. Real-time monitoring of the operating status and fault propagation of photovoltaic equipment after executing control instructions;

[0136] After the control command is executed, the edge node immediately initiates enhanced monitoring mode, intensively monitoring all relevant equipment within and outside the isolated area. The monitoring frequency is increased from once per minute during normal operation to once every 10 seconds, and the monitoring parameters are expanded to a comprehensive set of parameters including voltage, current, power, temperature, vibration, and communication status. For the five isolated devices, the focus is on monitoring the maintenance of their disconnection status and insulation status to ensure the effectiveness and safety of the isolation measures. For non-isolated devices but at the edge of the impact range, the focus is on monitoring the stability of their operating parameters and the changing trends of their anomaly scores to promptly detect possible signs of fault spread.

[0137] D3. Compare and analyze the actual fault location information monitored with the predicted fault source;

[0138] The comparative analysis process includes two aspects: location accuracy analysis and type matching analysis. Location accuracy analysis evaluates the consistency between the predicted fault source and the actual fault source. In this fault event, the predicted location and the actual location completely matched, with a location prediction accuracy of 100%. Type matching analysis evaluates the consistency between the predicted fault type and the actual fault type. The predicted electrical fault and the actual power device fault belong to the same fault category, with a type prediction accuracy of 90%. The impact range prediction was verified by observing the changes in equipment status after the isolation measures were implemented. The actual impact range was successfully controlled within the predicted range, and no unexpected spread occurred. The impact range prediction accuracy was 95%.

[0139] D4. Determine the prediction accuracy of the fault location information based on the comparative analysis results, and feed the judgment results back to the lightweight twin model.

[0140] Based on the comparative analysis results, a comprehensive evaluation of the lightweight twin model's prediction accuracy was conducted. The evaluation metrics included fault source location accuracy, fault type identification accuracy, impact range prediction accuracy, and diffusion time prediction accuracy. In this fault incident, the fault source location accuracy was 100%, the fault type identification accuracy was 90%, the impact range prediction accuracy was 95%, and the diffusion time prediction accuracy was 85%. The overall prediction accuracy reached 92.5%, exceeding the system's set accuracy threshold of 90%, and the prediction was deemed correct.

[0141] The steps to update the lightweight twin model include:

[0142] When the fault location information is correctly determined, the lightweight twin model is continuously monitored;

[0143] When the fault location information is incorrectly determined, updating the parameters of the lightweight twin model;

[0144] In this embodiment, the analysis is continued based on the aforementioned inverter INV-023 fault event. After completing the fault handling and effect evaluation, the system obtained a judgment result on the prediction accuracy. The comprehensive prediction accuracy was 92.5%, which exceeded the 90% accuracy threshold set by the system, and was therefore determined to be correct in the fault location information. In this case, the system starts the continuous monitoring mode of the lightweight twin model without adjusting the parameters. In the continuous monitoring mode, the system maintains the current model parameter configuration unchanged, but will strengthen the tracking and analysis of the model's operating status, and increase the monitoring frequency from once a day to once an hour, focusing on observing the model's prediction stability and consistency performance.

[0145] In another embodiment, assume that during another fault event, the system's prediction of the fault location for inverter INV-087 deviates. The actual fault source is confirmed on-site as INV-089, while the system predicts INV-087. The location prediction accuracy is only 60%, below the 90% accuracy threshold, and therefore the fault location information is determined to be incorrect. In this case, the system initiates a parameter update process for the lightweight twin model. By analyzing the causes and characteristics of the prediction deviation, the model parameters are optimized and adjusted to improve the accuracy of subsequent predictions.

[0146] The parameter update of the lightweight twin model includes:

[0147] Calculate the abnormality score and fault level score of the fault source; when the product of the abnormality score and the fault level score is greater than the first set value, do not update the parameters of the lightweight twin model; when the product of the abnormality score and the fault level score is not greater than the first set value, perform a secondary judgment.

[0148] Importantly, the anomaly score and fault severity score for the actual fault source, INV-089, were calculated. The anomaly score was calculated by inputting the operating data of INV-089 at the time of the fault into the current lightweight twin model, resulting in a score of 0.73. The fault severity score was manually assessed based on the actual fault conditions confirmed on-site. INV-089 was a DC / AC module fault. According to the system's fault severity assessment criteria, this type of fault is classified as medium severity, resulting in a corresponding fault severity score of 0.65.

[0149] The first set value, 0.45, is used to determine whether a parameter update is necessary. When the product of the anomaly score and the fault severity score is greater than 0.45, the system deems the current fault characteristics sufficiently significant, the model parameters are generally reasonable, and no adjustment is required. In this example, the product of an anomaly score of 0.73 and a fault severity score of 0.65 is 0.4745, which is greater than the first set value of 0.45.

[0150] The second-level judgment step includes: judging whether the abnormality score is greater than a second set value;

[0151] When the abnormality score is greater than the second set value, according to the formula Calculate the increment of the anomaly score and use the formula Calculate anomaly scores after enhanced processing capabilities;

[0152] When the abnormality score is not greater than the second set value, the abnormality score increment is set to 0, and the abnormality score is calculated according to the formula Calculate the anomaly score after updating the processing capacity;

[0153] Where f is the increment of anomaly score, is the preset weight, b is the abnormality score of the fault source, c is the fault level score at the fault source, d is the abnormality score after enhancing the processing capability, e is the first preset weight, g is the original abnormality score of the lightweight twin model, and h is the abnormality score after updating the processing capability.

[0154] In another embodiment, the abnormality score 0.73 needs to be compared with the second set value. The second set value is set to 0.70 to distinguish different parameter update strategies. Since the abnormality score 0.73 is greater than the second set value 0.70, the update strategy with enhanced processing power will be adopted, and the first set of calculation formulas will be used to adjust the parameters. In the default weight Set to 0.8, the anomaly score b of the fault source is 0.73, the fault level score c is 0.65, and the increment of the anomaly score is calculated. .

[0155] Continue using the formula Calculate the anomaly score after enhanced processing capabilities. The first preset weight, e, is set to 1.2, and the lightweight twin model's original anomaly score, g, is 0.68 (this is the model's anomaly score for INV-089 in its normal state before the fault). Substituting the parameters into the formula, we obtain d = 0.7056. The calculation results show that after the parameter update, the model's anomaly score for failure modes like INV-089 has increased from 0.68 to 0.7056, increasing the model's sensitivity to similar faults.

[0156] Assume that in another case the anomaly score is 0.55, which is less than the second set value of 0.70. In this case, the system will set the increment f of the anomaly score to 0 and use the formula Calculation is performed. Since f = 0, h = 1.2 × 0 + 0.68 = 0.68, which means the model's anomaly score remains unchanged. This design logic takes into account that when the anomaly score is low, the fault characteristics may not be obvious enough, and drastically adjusting the model parameters may introduce unnecessary noise. Therefore, a conservative update strategy is adopted.

[0157] Example 2 is an embodiment of the present invention, which provides a holographic twin distributed photovoltaic collection and control system, including a lightweight twin module, a fault detection module, a fault diffusion prediction module, and a control strategy formulation module;

[0158] The lightweight twin module is used to simulate the operating status of photovoltaic equipment;

[0159] The fault detection module uses a lightweight twin model for fault detection;

[0160] The fault spread prediction module analyzes the initial state of the fault based on the lightweight twin model and predicts the fault spread range;

[0161] The control strategy formulation module formulates a blocking control strategy based on the prediction results of the fault propagation range.

[0162] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A holographic twin distributed photovoltaic collection and control method, characterized in that: The following steps are involved: Deploy lightweight twin models within edge nodes of photovoltaic equipment; Use lightweight twin models to detect faults, locate the source of faults, and predict the scope of fault spread; Formulate a blocking control strategy based on the prediction results of the fault propagation range and update the lightweight twin model; The steps to locate the fault source and predict the fault spread range include: Deploy a propagation tracing network on the edge nodes of photovoltaic equipment and define the fault propagation direction and data flow topology in the propagation tracing network; Establish a fault propagation association mechanism and fault propagation chain based on the fault propagation direction and data flow topology in the propagation tracing network; Performing source tracing analysis on the faulty device data based on the fault propagation chain to determine the location of the fault source in the fault propagation chain; Predicting the fault propagation range based on the propagation direction of the fault propagation chain and the data flow topology relationship; The prediction of the scope of fault spread includes: Calculating the fault diffusion speed and diffusion path according to the fault propagation chain; The degree of output impact based on the fault level of the fault source and the electrical connection relationship between the photovoltaic devices; Based on the diffusion speed and impact level, the time range of the fault diffusion and the final impact boundary are predicted.

2. A holographic twin distributed photovoltaic collection and control method according to claim 1, characterized in that: The steps of constructing the lightweight twin model include: Collect data information of photovoltaic equipment and build data sets; Use anomaly detection algorithms to filter dirty data in the data set and extract feature data; Constructing training data through feature data, and building a lightweight twin model based on the training data; The data information includes operation data and fault equipment data.

3. A holographic twin distributed photovoltaic collection and control method according to claim 2, characterized in that: The steps for fault detection using lightweight twin models include: Input the operating data and faulty equipment data into the lightweight twin model to obtain anomaly scores; Compare anomaly scores and set anomaly thresholds; When the anomaly score is greater than the set anomaly threshold, it is judged as abnormal data and output to the decision center; When the abnormality score is not greater than the set abnormality threshold, the abnormal data is spliced, and the abnormal data after data splicing includes data labels and fault information; The data labels and fault information are input into the lightweight twin model as new data to re-perform fault detection.

4. A holographic twin distributed photovoltaic collection and control method as claimed in claim 3, characterized in that: The steps of formulating the blocking control strategy include: Determine the scope of photovoltaic equipment that needs to be isolated based on the fault propagation range; When the spread rate exceeds the preset threshold, an emergency blocking instruction is generated; when the spread rate does not exceed the preset threshold, a progressive isolation strategy is formulated; Send the control instructions to the corresponding edge node for execution.

5. The holographic twin distributed photovoltaic collection and control method according to claim 4, characterized in that: The steps of executing the control instruction and evaluating the control effect include: The edge node receives and executes the control instruction to adjust the specified photovoltaic equipment; Real-time monitoring of the operating status and fault propagation of photovoltaic equipment after executing control instructions; Compare and analyze the actual fault location information monitored with the predicted fault source; The prediction accuracy of the fault location information is judged based on the comparative analysis results, and the judgment results are fed back to the lightweight twin model.

6. A holographic twin distributed photovoltaic collection and control method according to claim 5, characterized in that: The steps to update the lightweight twin model include: When the fault location information is correctly determined, the lightweight twin model is continuously monitored; When the fault location information is incorrectly determined, updating the parameters of the lightweight twin model; The parameter update of the lightweight twin model includes: Calculate the anomaly score and fault level score of the fault source; When the product of the abnormality score and the fault level score is greater than a first set value, no parameter update is performed on the lightweight twin model; When the product of the abnormality score and the fault level score is not greater than the first set value, a secondary judgment is performed.

7. A holographic twin distributed photovoltaic collection and control method according to claim 6, characterized in that: The steps of the secondary judgment include: Determining whether the abnormality score is greater than a second set value; When the abnormality score is greater than the second set value, calculating an increment of the abnormality score, and calculating an abnormality score after enhancing the processing capability; When the abnormality score is not greater than the second set value, the increment of the abnormality score is set to 0, and the abnormality score after the processing capability is updated is calculated.

8. A holographic twin distributed photovoltaic collection and control system, applying a holographic twin distributed photovoltaic collection and control method according to any one of claims 1 to 7, characterized in that: It includes lightweight twin module, fault detection module, fault diffusion prediction module, and control strategy formulation module; The lightweight twin module is used to simulate the operating status of the photovoltaic equipment; The fault detection module uses a lightweight twin model to perform fault detection; The fault spread prediction module analyzes the initial state of the fault based on the lightweight twin model and predicts the fault spread range; The control strategy formulation module formulates a blocking control strategy according to the prediction result of the fault diffusion range.

Images