A cloud platform-based distributed photovoltaic power station monitoring management method and system

Abstract

This invention relates to the field of intelligent monitoring technology for photovoltaic power plants, and discloses a cloud-based method and system for monitoring and managing distributed photovoltaic power plants. The method includes: collecting operating parameters of photovoltaic strings; applying bounded perturbations within the stable range of maximum power point tracking to obtain feature vectors; constructing a topology graph and reconstructing virtual current-voltage characteristic curves and key feature points; obtaining node-level health indicators on the topology graph and generating anomaly scores based on a reference distribution; locating abnormal strings and their associated nodes based on the shape deviation of the virtual current-voltage characteristic curves and the distribution of node-level health indicators; solving an optimization problem of power generation revenue, degradation risk, and power fluctuation under parameter constraints; obtaining active power limits and power factor control trajectories, and distributing them to edge nodes for execution. This application improves the diagnostic accuracy and grid connection stability of distributed photovoltaic power plants in large-scale scenarios, and enhances scalability and intelligence.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for photovoltaic power plants, and more specifically, to a distributed photovoltaic power plant monitoring and management method and system based on a cloud platform. Background Technology

[0002] Photovoltaic power plants are a crucial component of the new energy power generation system. With the rapid expansion of installed capacity, their operation monitoring and intelligent management capabilities directly impact the power plant's generation efficiency and the grid's safety and stability. Currently, the industry is gradually migrating traditional localized monitoring methods to a centralized architecture based on cloud platforms. This involves collecting operating parameters through edge nodes and uploading them to the cloud, enabling comprehensive monitoring of power plant equipment status, power output, and environmental conditions. This approach can improve the real-time performance and centralized management capabilities of monitoring to a certain extent, providing necessary data support for grid dispatch and new energy consumption.

[0003] However, existing technologies, such as the invention patent CN120414873A – Real-time Monitoring and Management System and Method for Photovoltaic Power Plants Based on Cloud Platform, still have shortcomings. Existing system architectures rely on fixed hardware configurations, have limited scalability, and require additional modifications when the power plant scales up, increasing operation and maintenance costs. Data processing methods mostly rely on threshold comparisons, making it difficult to accurately identify complex issues such as string mismatch and component degradation, resulting in potential risks not being detected in a timely manner. Monitoring methods lack dynamic modeling capabilities that incorporate physical mechanisms; the use of current and voltage characteristic curves is limited to static comparisons and cannot reflect real-time degradation states. Furthermore, the data processing flow lacks flexibility in high-concurrency scenarios, easily leading to delays and data loss, which restricts the reliability of cloud platforms in large-scale monitoring of distributed photovoltaic power plants.

[0004] Therefore, it is necessary to design a cloud-based distributed photovoltaic power plant monitoring and management method and system to solve the problems existing in the current technology. Summary of the Invention

[0005] In view of this, the present invention proposes a distributed photovoltaic power plant monitoring and management method and system based on a cloud platform, aiming to solve the problems of existing photovoltaic power plant monitoring methods lacking dynamic modeling capabilities, insufficient anomaly identification accuracy, and limited grid-connected control optimization in large-scale distributed scenarios.

[0006] In one aspect, this invention proposes a cloud-based method for monitoring and managing distributed photovoltaic power plants, comprising:

[0007] The operating parameters of the photovoltaic strings are collected at the edge nodes of the power station;

[0008] Within the stable range of maximum power point tracking, a bounded perturbation is applied to the DC operating point to obtain the voltage and current changes, and the derivatives are calculated and the equivalent electrical parameters are estimated, thereby obtaining the eigenvector.

[0009] A topology graph is constructed based on the electrical connection relationship between photovoltaic strings, combiner boxes and inverters. The feature vectors are attached as node attributes, and a model with physical constraints is called to reconstruct the virtual current and voltage characteristic curves and key feature points.

[0010] Based on preset neighborhood selection conditions and weight calculation rules, the feature vectors of neighboring nodes that meet the comparable conditions within the same combiner box, the same inverter, or the same directional cluster are weighted and aggregated on the topology map to obtain node-level health indicators. Anomaly scores are generated by combining the reference distribution constructed from historical normal operation samples.

[0011] When the anomaly score meets the preset conditions, active sparse sampling is performed, and the virtual current-voltage characteristic curve and the node-level health index are updated.

[0012] Based on the shape deviation of the virtual current-voltage characteristic curve and the distribution of the node-level health indicators on the topology graph, the abnormal photovoltaic strings and their respective nodes are located. Under the parameter constraints of the grid connection specifications, the optimization problem of power generation revenue, degradation risk and power fluctuation is solved to obtain the active power limit and power factor control trajectory, which is then sent to the edge nodes for execution.

[0013] Furthermore, when collecting operating parameters of the photovoltaic strings at the edge nodes of the power station, this includes:

[0014] The operating parameters include voltage, current, power, inverter active power, power factor, ambient irradiance, and component temperature.

[0015] Within the stable range of maximum power point tracking, the voltage, current, power, active power of the photovoltaic string, power factor of the photovoltaic string, ambient irradiance, and module temperature are synchronously sampled and assigned a unified timestamp at a sampling period of no more than 1 second. When the stable range criterion is not met, sampling is paused until the stable range is restored. The synchronously sampled data is sequentially subjected to amplitude limiting and noise reduction, median filtering, and sliding window drift removal processing, and unit unification and range calibration are completed. Ambient irradiance is effectively self-calibrated by combining AC side metering power with inverter efficiency characteristics. When parameter missing or sampling abnormality is detected, short-window interpolation is performed only within the stable range and does not cross the stable range boundary.

[0016] Furthermore, the application of a bounded perturbation to the DC operating point within the stable range of maximum power point tracking, and the acquisition of the eigenvector, include:

[0017] The bounded perturbation is a small bidirectional perturbation on the DC operating point voltage or DC operating point current, with an amplitude not exceeding 0.5% of the rated DC operating point, a single duration not exceeding 200ms, and a time interval of not less than 1s between two adjacent perturbations;

[0018] Voltage and current are synchronously collected with a unified timestamp before and after the disturbance. The voltage and current changes are obtained by differential regression with adjacent time windows. Online estimation is performed under the physical constraints of the single diode equivalent model to obtain the equivalent series resistance, equivalent parallel resistance, photocurrent, reverse saturation current and ideality factor. The operating parameters, rate of change and equivalent electrical parameters are combined into a feature vector. The bounded perturbation is only activated within the stable range and automatically terminated outside the stable range.

[0019] Furthermore, when constructing the topology graph and reconstructing the virtual current-voltage characteristic curves, the following steps are included:

[0020] A topology graph containing string nodes, combiner box nodes and inverter nodes is generated based on the primary wiring data. Edges are established according to electrical connections and edge weights are determined by cable length, parallel circuit affiliation and orientation consistency. The feature vectors are normalized and time-aligned and then attached to the corresponding nodes.

[0021] Using node features, ambient irradiance, module temperature, and weighted aggregation features of topological neighbors as input, a representation learning model with physical constraints is invoked for inference. These physical constraints include: in the short-circuit to low-voltage range, the voltage is no higher than 30% of the full range, and the current decrease relative to the short-circuit current does not exceed 10%; in the high-voltage range, the voltage is no lower than 70% of the full range, and the current decreases significantly with increasing voltage, with a greater decrease than in the low-voltage range; in the neighborhood of the maximum power point, power increases with increasing voltage below this point and decreases with increasing voltage above it; short-circuit current increases with increasing ambient irradiance without decreasing open-circuit voltage, and open-circuit voltage decreases with increasing module temperature.

[0022] By incorporating topology consistency constraints, the virtual current-voltage characteristic curves and key feature points are output. These key feature points include open-circuit voltage, short-circuit current, and the voltage and current at the maximum power point.

[0023] Furthermore, when generating node-level health indicators on the topology graph, the following steps are included:

[0024] Based on preset neighborhood selection conditions, neighboring nodes that meet comparable conditions are selected within the same combiner box, the same inverter, or the same azimuth cluster. The comparable conditions include: environmental irradiance difference not exceeding 15%, component temperature difference not exceeding 3℃, azimuth angle difference not exceeding 15°, and sampling time alignment error not exceeding 1ms. Nodes that do not meet the comparable conditions or are offline are assigned zero weight and removed. The feature vectors of each neighboring node are aggregated and normalized according to azimuth consistency, irradiance similarity, historical correlation, and electrical connection strength. The feature vectors of the neighboring nodes are weighted and aggregated at least twice to obtain the node-level health indicators of the corresponding photovoltaic string nodes.

[0025] Furthermore, when generating anomaly scores, the following are included:

[0026] Within a preset time window, node-level health indicators of each photovoltaic string are collected. Node-level health indicators whose operating status meets preset normality criteria are used to calculate the normal reference center and the corresponding variance or covariance index. The weighted distance between the current node-level health indicator and the normal reference center is used as the first component, and the dispersion index of the node-level health indicator within the current time window is used as the second component. The first component and the second component are weighted and synthesized to obtain the anomaly score of the photovoltaic string node.

[0027] Furthermore, when performing active sparse sampling and updating the virtual current-voltage characteristic curves and node-level health indicators, the following steps are included:

[0028] The active sparse sampling includes: based on the dispersion of the anomaly score and the node-level health index to form a comprehensive priority, selecting target photovoltaic strings within the power station according to the comprehensive priority, with a daily coverage ratio not exceeding 5% and the same photovoltaic string not exceeding once per hour;

[0029] Enhanced measurements are performed on the target photovoltaic string within the stable range of maximum power point tracking. The enhanced measurements consist of no more than five short-window bounded perturbations, with the amplitude of a single perturbation not exceeding 0.5% of the rated DC operating point, the duration of a single perturbation not exceeding 200 ms, and the total duration not exceeding 2 s.

[0030] After the enhanced measurement is completed, the derivative quantity and equivalent electrical parameters are recalculated based on the data collected before and after the disturbance. The feature vector is updated and the virtual current-voltage characteristic curve is reconstructed. At the same time, the node-level health index is updated in a sliding fusion manner, so that the new node-level health index is a weighted sum of the old node-level health index and the node-level health index obtained based on the latest feature inference. The weight coefficient is in the range of (0.6-0.9).

[0031] Furthermore, when locating abnormal strings and their parent nodes, the following steps are included:

[0032] The shape deviation index is calculated based on the virtual current-voltage characteristic curve. The shape deviation index consists of the open-circuit voltage offset ratio, the short-circuit current offset ratio, the maximum power point position offset ratio, and the degree of change of the segmented slope.

[0033] The anomaly intensity is obtained by combining the shape deviation index with the weighted distance of the node-level health index of the corresponding photovoltaic string node relative to the normal reference center.

[0034] Apply connectivity constraints and orientation consistency constraints within the topology graph, perform neighborhood aggregation on anomaly intensity, and extract anomaly clusters.

[0035] The string with the highest abnormal intensity and the longest duration within the abnormal cluster is identified as the abnormal string, and the combiner box node and inverter node to which the abnormal string belongs are output.

[0036] Furthermore, when solving the optimization problem and generating the control trajectory while satisfying the parameter constraints of the grid connection specifications, the following steps are included:

[0037] Parameter constraints include voltage, current, ramp rate, and power factor;

[0038] Establish a rolling control cycle, which simultaneously satisfies the upper and lower limits of voltage, the upper limit of current, the power ramp rate constraint, and the power factor range constraint within each control cycle.

[0039] The goal is to maximize power generation revenue, and a comprehensive objective is formed by adding a degradation risk penalty term determined by the dispersion index of abnormal intensity and node-level health indicators, and a power fluctuation penalty term determined by the active power change measurement of adjacent cycles.

[0040] Using the key feature points of the virtual current-voltage characteristic curve, node-level health indicators, and environmental predictions as optimization inputs, the active power limit and power factor control trajectory are obtained by solving. The solution results are then projected onto the feasible domain of the grid connection specification and sent to the edge nodes, which then execute the control trajectory.

[0041] Compared with existing technologies, the advantages of this invention are as follows: By acquiring photovoltaic string operating parameters in real time at edge nodes and applying bounded perturbations, equivalent electrical parameters are obtained, forming physically reliable feature vectors; a topology graph is constructed, and virtual current-voltage characteristic curves and key feature points are reconstructed under physical constraints, achieving dynamic modeling consistent with real electrical behavior; weighted aggregation of feature vectors of comparable neighbor nodes on the topology graph yields node-level health indicators and generates anomaly scores, improving the ability to identify component mismatch, performance degradation, and abnormal operating conditions; when the anomaly score triggers a threshold, enhanced measurements are performed on high-priority target strings through active sparse sampling, dynamically updating the virtual current-voltage characteristic curves. Line-level and node-level health indicators enable the monitoring model to continuously correct and converge under controllable disturbance costs. Abnormal photovoltaic strings and their associated nodes are accurately located on the topology map. Under the premise of meeting the grid connection specification parameter constraints, power generation revenue, degradation risk and power fluctuation are uniformly incorporated into the rolling optimization solution. The active power limit and power factor control trajectory are generated and sent to the edge nodes for execution. A full-link closed-loop monitoring system is constructed, which includes edge acquisition, physical modeling, topology correlation diagnosis, anomaly-driven enhanced measurement and grid-connection constraint-optimized control. In large-scale distributed photovoltaic scenarios, the system improves the accuracy of power plant diagnosis, the reliability of anomaly location and grid-connected operation stability, and enhances the scalability and intelligence level of the system.

[0042] On the other hand, this application also provides a cloud-based distributed photovoltaic power plant monitoring and management system for applying the above-mentioned cloud-based distributed photovoltaic power plant monitoring and management method, including:

[0043] The acquisition unit is configured to acquire the operating parameters of the photovoltaic strings at the edge nodes of the power plant;

[0044] The first processing unit is configured to apply a bounded perturbation to the DC operating point within the stable range of maximum power point tracking, obtain the voltage and current changes, calculate the derivatives and estimate the equivalent electrical parameters, thereby obtaining the eigenvector.

[0045] The reconstructing unit is configured to build a topology graph based on the electrical connection relationship between the photovoltaic string, combiner box and inverter, attach the feature vector as node attributes, and call the model with physical constraints to reconstruct the virtual current and voltage characteristic curves and key feature points;

[0046] The evaluation unit is configured to perform weighted aggregation on the feature vectors of neighboring nodes that meet the comparable conditions within the same combiner box, the same inverter, or the same directional cluster on the topology map based on preset neighborhood selection conditions and weight calculation rules, to obtain node-level health indicators, and generate anomaly scores by combining them with a reference distribution constructed from historical normal operation samples.

[0047] The judgment unit performs active sparse sampling and updates the virtual current-voltage characteristic curve and the node-level health index when the abnormal score meets the preset conditions.

[0048] The control unit is configured to locate abnormal photovoltaic strings and their respective nodes on the topology graph based on the shape deviation of the virtual current-voltage characteristic curve and the distribution of the node-level health indicators, and to solve the optimization problem of power generation revenue, degradation risk and power fluctuation under the parameter constraints of grid connection specifications, to obtain the active power limit and power factor control trajectory, and to send it to the edge nodes for execution.

[0049] It is understandable that the aforementioned cloud-based distributed photovoltaic power station monitoring and management methods and systems have the same beneficial effects, and will not be elaborated further here. Attached Figure Description

[0050] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0051] Figure 1 A flowchart illustrating a cloud-based distributed photovoltaic power plant monitoring and management method provided in an embodiment of the present invention;

[0052] Figure 2 This is a functional block diagram of a cloud-based distributed photovoltaic power plant monitoring and management system provided in an embodiment of the present invention. Detailed Implementation

[0053] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0054] In traditional distributed photovoltaic (PV) power plant monitoring systems, the scalability and data processing capabilities of cloud platform architectures face bottlenecks. As the scale of power plants expands, the number of edge nodes grows exponentially, and fixed hardware configurations struggle to dynamically adapt to the access needs of heterogeneous devices, leading to a simultaneous increase in data acquisition cycles and processing latency. Existing methods rely on static thresholds and offline models for anomaly detection, failing to capture real-time changes in the dynamic characteristics of PV strings, such as distortions in current and voltage characteristic curves caused by localized shading or drift in equivalent series resistance due to module aging. Furthermore, in high-concurrency scenarios, massive data streams are prone to packet loss and timestamp misalignment during transmission and storage, rendering topology correlation analysis ineffective and consequently affecting the accuracy of anomaly location.

[0055] For example, in a distributed photovoltaic power station containing thousands of strings, multiple edge nodes collect operating parameters at a rate of seconds and upload them to the cloud. Due to the lack of a dynamic modeling mechanism, the string status can only be judged by a fixed threshold, making it impossible to identify the slow decay of photocurrent caused by microcracks. When multiple strings under a certain inverter have uneven irradiance distribution due to azimuth differences, existing methods struggle to distinguish between normal power fluctuations and string mismatch, mistakenly classifying the power drop caused by irradiance fluctuations as abnormal and triggering redundant alarms. Furthermore, during the midday peak irradiance period, the surge in data concurrency causes the sampling timestamp deviation of some nodes to exceed 100 milliseconds, making it impossible to accurately align the node attributes in the topology graph. This results in a critical inflection point shift during the reconstruction of the virtual current and voltage characteristic curves, causing a mismatch between the maximum power point tracking control command and the actual operating conditions.

[0056] If the above issues are not addressed, power plant operation and maintenance costs will increase with the rising frequency of invalid alarms. The continuous accumulation of missed detections of abnormal strings may trigger hot spot effects, accelerate component aging, and even create fire hazards. Deviations between power control commands and actual operating conditions will cause frequent inverter output overruns, triggering protective grid disconnection and reducing power plant revenue. In the long term, the lack of accurate degradation state awareness will affect lifespan prediction and preventative maintenance planning, further increasing total lifecycle operation and maintenance costs. Furthermore, failures in topology correlation analysis will weaken anomaly localization capabilities, delay fault handling, increase grid power fluctuation risks, and threaten the transient stability of the regional power system.

[0057] For this, please refer to Figure 1 As shown, this application proposes a cloud-based method for monitoring and managing distributed photovoltaic power plants, including:

[0058] S100: Collects operating parameters of photovoltaic strings at the edge nodes of the power station.

[0059] S200: Within the stable range of maximum power point tracking, a bounded perturbation is applied to the DC operating point to obtain the voltage and current changes, and the derivatives are calculated and the equivalent electrical parameters are estimated, thereby obtaining the eigenvector.

[0060] S300: Based on the electrical connection relationship between photovoltaic strings, combiner boxes and inverters, a topology graph is constructed, feature vectors are attached as node attributes, and a model with physical constraints is called to reconstruct virtual current and voltage characteristic curves and key feature points.

[0061] S400: Based on preset neighborhood selection conditions and weight calculation rules, the feature vectors of neighboring nodes that meet the comparable conditions within the same combiner box, the same inverter, or the same directional cluster are weighted and aggregated on the topology graph to obtain node-level health indicators, and anomaly scores are generated by combining the reference distribution constructed from historical normal operation samples.

[0062] S500: When the anomaly score meets the preset conditions, perform active sparse sampling and update the virtual current-voltage characteristic curve and node-level health indicators.

[0063] S600: Based on the shape deviation of the virtual current and voltage characteristic curves and the distribution of node-level health indicators on the topology graph, it locates abnormal photovoltaic strings and their respective nodes, and solves the optimization problem of power generation revenue, degradation risk and power fluctuation under the parameter constraints of grid connection specifications, obtains the active power limit and power factor control trajectory, and sends it to the edge nodes for execution.

[0064] Specifically, in step S100, the edge node collects the operating parameters of the photovoltaic string. This refers to deploying a data acquisition module on the local equipment side of the photovoltaic power station. This can be achieved by using embedded sensors, smart meters, or the inverter's built-in measurement unit. The module is used to acquire raw operating data such as the voltage, current, and power of the photovoltaic string, the active power of the inverter, the power factor of the inverter, as well as environmental irradiance and component temperature in real time. The module also performs timestamp alignment and preliminary quality control locally to reduce the data transmission pressure on the cloud and improve the real-time performance and data reliability of the monitoring.

[0065] In step S200, applying a bounded perturbation to the DC operating point within the stable range of maximum power point tracking means changing the DC operating point within a safe range near the maximum power point by using a controllable perturbation signal with limited amplitude and duration. Specifically, pulse width modulation can be used to fine-tune the DC side voltage, or a small current perturbation can be given by adjusting the inverter DC current to achieve a short-term shift of the operating point. Voltage and current are synchronously collected before and after the perturbation. The derivative is calculated based on the voltage and current changes, and the equivalent series resistance, equivalent parallel resistance, photocurrent, reverse saturation current, and ideal factor, etc., are estimated online in combination with physical laws such as the single diode equivalent circuit model. This forms a feature vector that can characterize the internal state of the photovoltaic string, obtaining more refined dynamic electrical information without performing full curve scanning.

[0066] In step S300, constructing the topology graph and attaching feature vectors as node attributes refers to establishing a graph structure based on the physical wiring relationships between photovoltaic strings, combiner boxes, and inverters. Specifically, the connection relationships between each string and combiner box, and between combiner box and inverter, can be obtained through a primary wiring diagram analysis. Based on this, a topology graph containing photovoltaic string nodes, combiner box nodes, and inverter nodes is constructed. Edge weights can be determined based on factors such as cable length, parallel circuit affiliation, and orientation consistency to reflect electrical coupling strength and spatial correlation. The normalized and time-aligned feature vectors are then attached to the corresponding photovoltaic string nodes to associate electrical parameters with the equipment topology, supporting subsequent statistical analysis and diagnostic calculations based on the topology structure. Among them, the model reconstruction of virtual current and voltage characteristic curves and key feature points with physical constraints refers to constructing a parameterized model based on physical laws such as the single diode equivalent circuit model, combining information such as voltage, current, equivalent electrical parameters, environmental irradiance, and module temperature, to numerically solve or fit the virtual current and voltage characteristic curves of each photovoltaic string, and output key feature points such as open-circuit voltage, short-circuit current, and maximum power point voltage and current, so as to ensure that the reconstructed curves conform to the basic physical characteristics of photovoltaic modules and avoid significant distortion of pure data-driven models under extrapolation or noise interference.

[0067] In step S400, generating node-level health indicators on the topology graph refers to performing feature aggregation on the constructed topology graph based on preset neighborhood selection conditions and weight calculation rules for neighboring nodes comparable to the target photovoltaic string. Specifically, within the same combiner box, the same inverter, or the same azimuth cluster, neighboring nodes whose environmental irradiance difference does not exceed a first threshold, whose module temperature difference does not exceed a second threshold, whose azimuth angle difference does not exceed a third threshold, and whose sampling time alignment error is not greater than a preset time threshold are selected as comparable neighbors. Nodes that do not meet the comparability conditions or are offline are assigned zero weight and removed. Then, based on indicators such as azimuth consistency, irradiance similarity, historical operation correlation, and electrical connection strength, the aggregation weight of each neighboring node is calculated and normalized. One or more layers of weighted aggregation are performed on the feature vectors of the neighboring nodes to obtain node-level health indicators that take into account both local characteristics and neighborhood reference information. When generating anomaly scores, node-level health indicators that meet the normal criteria within a certain time window are used as training samples. The normal reference center and its variance or covariance index are calculated. Based on the weighted distance between the current node-level health indicator and the normal reference center and the dispersion index of the node-level health indicator within the current time window, a comprehensive function is constructed to output the anomaly score. This reflects the degree to which the string state deviates from the normal operating mode without relying on a complex deep network structure.

[0068] In step S500, active sparse sampling and updating the virtual current-voltage characteristic curves and node-level health indicators refers to implementing priority enhanced measurements when the anomaly score and dispersion of a photovoltaic string indicate a high anomaly risk or significant uncertainty. Specifically, a comprehensive priority can be constructed based on a weighted average of the anomaly score and the dispersion of the node-level health indicators. A small number of target photovoltaic strings with high comprehensive priority are selected across the entire power plant, with the daily coverage ratio limited to a preset value and the triggering frequency of the same photovoltaic string not exceeding a preset upper limit. Within the stable range of maximum power point tracking, short-window bounded perturbations are added to the target photovoltaic strings no more than a preset number, while the amplitude and duration of the perturbations still meet safety constraints. The derivative and equivalent electrical parameters are recalculated using data before and after the enhanced measurement, the eigenvector is updated, and the virtual current-voltage characteristic curves are reconstructed. Simultaneously, a sliding weighted method is used to fuse the node-level health indicators obtained based on the enhanced measurement inference with historical node-level health indicators, so that the node-level health indicators gradually converge over time and maintain sensitivity to the latest state.

[0069] In step S600, the distribution of shape deviations and node-level health indicators on the topology map to locate abnormal photovoltaic strings and their associated nodes refers to extracting shape deviation indicators such as open-circuit voltage offset ratio, short-circuit current offset ratio, maximum power point position offset ratio, and segmented slope change degree based on the virtual current-voltage characteristic curve. These shape deviation indicators are then combined with the weighted distance of the corresponding photovoltaic string node's node-level health indicators relative to the normal reference center to obtain the abnormal intensity. Connectivity constraints and orientation consistency constraints are applied within the topology map, and the abnormal intensity is aggregated into neighborhoods and abnormal clusters are extracted. Photovoltaic strings with larger abnormal intensity and durations that meet the shortest duration threshold within the abnormal cluster are identified as abnormal photovoltaic strings. At the same time, the combiner box node and inverter node where the abnormal string is located are output as the associated nodes, realizing abnormal source tracing from the string layer to the device layer. Based on this, a rolling optimization model is established under the constraints of grid connection specification parameters such as voltage, current, ramp rate, and power factor. With the goal of maximizing power generation revenue, a comprehensive objective function is introduced, consisting of a degradation risk penalty term composed of anomaly intensity and node-level health index dispersion, and a power fluctuation penalty term composed of active power change measurement. The active power limit and power factor control trajectory for each control cycle are obtained by solving the model. The optimization results are then projected to the feasible domain that meets the grid connection specifications and distributed to the edge nodes for execution, thereby improving power generation revenue while ensuring grid safety and equipment lifespan.

[0070] This application constructs a dynamic closed-loop monitoring and optimization system by integrating topology graph structure, physical constraint model, and graph attention mechanism, achieving end-to-end collaboration from data acquisition and state assessment to control decision-making. Specifically, it obtains dynamic electrical parameters through bounded perturbation, attaches them and environmental information to photovoltaic string nodes to form feature vectors, reconstructs virtual current and voltage characteristic curves and key feature points in the topology graph in conjunction with physical constraints, and generates node-level health indicators and anomaly scores by comparing neighborhood-weighted aggregation with normal reference distribution. Furthermore, it utilizes an anomaly-driven active sparse sampling mechanism to continuously update key node data under controllable cost conditions, and finally generates control commands that meet grid connection specifications based on multi-objective optimization, thereby solving the problems of rigid data processing, inaccurate anomaly location, and single control strategy in existing technologies.

[0071] The working process and principle of this application are as follows: Operating parameters of photovoltaic strings are collected at the edge nodes of the power plant; bounded perturbations are applied to the DC operating point within the maximum power point tracking stability interval to obtain voltage and current changes; derivatives are calculated and equivalent electrical parameters are estimated to obtain feature vectors; a topology graph is constructed based on electrical connection relationships; the feature vectors are attached as node attributes; a model with physical constraints is called to reconstruct virtual current and voltage characteristic curves and key feature points; comparable neighbor nodes are selected on the topology graph according to preset neighborhood selection conditions; and the feature vectors of neighbor nodes are weighted and aggregated according to rules such as orientation consistency, irradiance similarity, historical correlation, and electrical connection strength to obtain node-level health indicators; a reference distribution is constructed by combining historical normal operation samples, and an anomaly score is output; when the anomaly score meets preset conditions, active sparse sampling is performed to update the virtual current and voltage characteristic curves and node-level health indicators; based on the shape deviation of the virtual current and voltage characteristic curves and the distribution of node-level health indicators on the topology graph, abnormal strings and their associated nodes are located; an optimization problem is solved under grid connection specification constraints to obtain the control trajectory for execution.

[0072] This scheme captures the dynamic characteristics of the photovoltaic string by using bounded perturbation and equivalent parameter estimation, and enhances the anomaly detection capability by using topology graphs and neighborhood-weighted statistical analysis. The active sparse sampling mechanism prioritizes updating the data of high-risk nodes under limited resources, and combines the shape deviation of virtual current and voltage characteristic curves with node-level health indicators to improve the accuracy of anomaly location. On this basis, the scheme achieves intelligent monitoring and management of distributed photovoltaic power stations by balancing power generation revenue, degradation risk and power fluctuation through multi-objective optimization.

[0073] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0074] Operating parameters such as voltage, current, and power of the photovoltaic (PV) strings are collected at the edge nodes of the power plant. Within the stable range of maximum power point tracking (MPPT), a small bidirectional perturbation with an amplitude not exceeding 0.5% and a duration not exceeding 200ms is applied to the DC operating point. The voltage and current changes are calculated using the data collected before and after the perturbation, and equivalent electrical parameters such as equivalent series resistance, parallel resistance, and photovoltaic current are estimated to form a feature vector. A topology diagram including PV strings, combiner boxes, and inverter nodes is generated based on the primary wiring data. The normalized feature vector is then attached to the corresponding PV string nodes, and a physically constrained model is used to reconstruct the virtual current-voltage characteristic curves and key feature points.

[0075] On the topology map, neighboring nodes that meet the comparability criteria are selected based on the constraints of the same combiner box, the same inverter, or the same azimuth cluster. Nodes that do not meet the comparability criteria or are offline are assigned zero weight and removed. The aggregation weight of each neighboring node is determined and normalized according to azimuth consistency, irradiance similarity, historical correlation, and electrical connection strength. The feature vectors of neighboring nodes are aggregated at least twice to obtain node-level health indicators. Within a preset time window, the normal reference center and variance or covariance indicators of the node-level health indicators that are operating normally are calculated. The weighted distance between the current node-level health indicator and the normal reference center and its dispersion index within the time window are weighted and synthesized into an anomaly score.

[0076] When the anomaly score exceeds the threshold, active sparse sampling is performed. Target strings are selected based on the priority of anomaly score and uncertainty synthesis, with a daily coverage rate not exceeding 5%. Short-window bounded perturbation enhancement measurements are performed on the target strings no more than 5 times to update the feature vector, virtual curve, and node-level health indicators.

[0077] The shape deviation index of the virtual current and voltage characteristic curve is calculated and combined with the distance of the node-level health index relative to the normal reference center to synthesize the anomaly intensity. Connectivity and orientation consistency constraints are applied to the topology graph. The anomaly intensity is aggregated and anomaly clusters are extracted to locate the abnormal photovoltaic strings and their respective combiner box nodes and inverter nodes.

[0078] Under grid connection regulations, with the goal of maximizing power generation revenue, degradation risk and power fluctuation penalties are incorporated to solve the optimization problem. The active power limit and power factor control trajectory are obtained and distributed to edge nodes for execution.

[0079] Through the above-described scheme, this application can capture real-time changes in the dynamic characteristics of photovoltaic strings, improving the accuracy of anomaly detection. Topology correlation analysis enhances data consistency, reducing processing latency and packet loss risks in high-concurrency scenarios. An active sparse sampling mechanism reduces invalid alarms and optimizes control strategies to balance power generation efficiency and stability. Therefore, the operational reliability and economy of distributed photovoltaic power plants are improved.

[0080] This application further proposes a method for collecting operating parameters of photovoltaic (PV) strings at power plant edge nodes, including voltage, current, power, inverter active power, power factor, ambient irradiance, and module temperature. Within the stable range of maximum power point tracking (MPPT), the voltage, current, power, inverter active power, inverter power factor, ambient irradiance, and module temperature of the PV strings are synchronously sampled and assigned a unified timestamp at a sampling period not exceeding 1 second. Sampling is paused when the stable range criterion is not met until the stable range is restored. The synchronously sampled data undergoes amplitude limiting and noise reduction, median filtering, and sliding window drift correction, and unit unification and range calibration are performed. Ambient irradiance is effectively self-calibrated by combining AC-side power metering with inverter efficiency characteristics. When missing parameters or sampling anomalies are detected, short-window interpolation is performed only within the stable range and does not cross the stable range boundary.

[0081] The operating parameters include voltage, current, power, inverter active power, power factor, ambient irradiance, and component temperature, covering multi-dimensional data on electrical, environmental, and equipment status. The sampling period is set to no more than 1 second, dynamically adapting to the stable range of maximum power point tracking to ensure real-time and stable data acquisition. Synchronous sampling achieves multi-parameter time alignment through a unified timestamp, avoiding asynchronous errors. Amplification and noise reduction filters out abnormal jumps within a reasonable parameter range, median filtering eliminates impulse noise, and sliding window drift suppression inhibits slow-varying interference. Ambient irradiance self-calibration uses the AC-side power and inverter efficiency curves to infer the effective irradiance value, eliminating sensor bias. Short-window interpolation is performed only within the stable range, avoiding the introduction of non-physical data through cross-range interpolation.

[0082] Specifically, within the stable range of maximum power point tracking (MPPT), voltage, current, power, inverter parameters, and environmental data are synchronously collected at fixed time intervals to ensure consistent timestamps for each parameter. Sampling is immediately paused when the stable range conditions are not met to prevent the intrusion of non-steady-state data. The collected data undergoes amplitude limiting processing to remove outliers exceeding the physical range. Median filtering is used to eliminate transient noise interference. A sliding window is used to calculate the mean or trend term to eliminate data drift. Environmental irradiance values ​​are dynamically calibrated using the inverter's output power and efficiency characteristic curves to improve the accuracy of irradiance data. If some parameters are missing or sampling anomalies are detected, interpolation is performed only using valid data from the current stable range, with the interpolation window not exceeding the stable range boundary to ensure data continuity while avoiding the introduction of non-steady-state errors.

[0083] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0084] When collecting operating parameters of photovoltaic strings at the edge nodes of the power plant, these parameters include voltage, current, power, inverter active power, power factor, ambient irradiance, and module temperature. Within the stable range of maximum power point tracking, these parameters are synchronously sampled at a sampling period of 0.5 seconds and assigned a unified timestamp. Sampling is paused when the stable range criterion is not met until the stable range is restored. The synchronously sampled data is sequentially subjected to amplitude limiting and noise reduction, median filtering, and sliding window drift removal processing, and unit unification and range calibration are completed. Ambient irradiance is effectively self-calibrated by combining AC side power metering with inverter efficiency characteristics. When missing parameters or sampling anomalies are detected, short-window interpolation with a 10-second window is performed only within the stable range, without crossing the stable range boundary.

[0085] Specifically, amplitude limiting and noise reduction uses the 3σ criterion to remove outliers. Median filtering uses a 5-point median filter. Sliding window drift reduction uses a 60-second moving average. Range calibration is performed using periodically calibrated calibration coefficients. Effective irradiance self-calibration is obtained by interpolating the inverter's AC side power by the inverter efficiency curve. Short-window interpolation uses a linear interpolation method.

[0086] Through the above technical solutions, this application achieves high-quality acquisition of photovoltaic string operating parameters. This improves the accuracy and reliability of subsequent data analysis. Furthermore, by employing measures such as stability interval determination, synchronous sampling, data preprocessing, and self-calibration, the impact of sampling noise and abnormal data is effectively reduced. For example, self-calibration of environmental irradiance avoids errors caused by irradiance sensor malfunctions. Short-window interpolation ensures data continuity while avoiding unreasonable interpolation across stability intervals. These measures collectively ensure the temporal consistency, numerical accuracy, and physical rationality of the acquired data.

[0087] This application further proposes a method for applying bounded perturbations to the DC operating point within the stable range of maximum power point tracking (MPPT) and obtaining the eigenvector. The bounded perturbation is defined as a small bidirectional disturbance in the DC operating point voltage or current, with an amplitude not exceeding 0.5% of the rated DC operating point, a single duration not exceeding 200 ms, and a time interval of not less than 1 s between adjacent disturbances. Voltage and current are synchronously acquired before and after the disturbance using a unified timestamp. The voltage and current changes are obtained using differential regression with adjacent time windows, and online estimation is performed under the physical constraints of a single-diode equivalent model to obtain the equivalent series resistance, equivalent parallel resistance, photocurrent, reverse saturation current, and ideality factor. The operating parameters, rate of change, and equivalent electrical parameters are combined into the eigenvector. The bounded perturbation is only activated within the stable range and automatically terminates outside the stable range.

[0088] The bidirectional disturbance is achieved by alternately applying positive and reverse voltage or current offsets, with the disturbance amplitude limited to 0.5% of the rated value to avoid exceeding the stability interval boundary. Differential regression uses the least squares method to linearly fit the voltage and current sequences before and after the disturbance, extracting the slope as the change. The single-diode equivalent model describes the physical characteristics of the photovoltaic module through a set of nonlinear equations. The online estimation process solves the equations using Newton's iteration method, with constraints including non-negative photocurrent and positive definite series and parallel resistances. During eigenvector construction, the original operating parameters, differential regression results, and equivalent parameters are concatenated in a standardized format to form a multidimensional data matrix.

[0089] Specifically, under stable operating conditions, alternating positive and negative voltage perturbations of 0.3V are applied, each lasting 150ms before returning to the original state, with an interval of 1.2s between perturbations. Voltage and current data are simultaneously acquired during the perturbation period. A 50ms sliding window is used to perform differential calculations on the data before and after the perturbation, obtaining changes in ΔV = 0.6V and ΔI = 0.02A. Based on a single-diode model, a system of equations containing five unknown parameters is established. Through online iterative calculations, the equivalent series resistance is determined to be 0.25Ω, the equivalent parallel resistance to be 125Ω, and the photocurrent to be 5.2A. The original voltage value of 28.5V, the current value of 4.8A, and the equivalent parameters are encoded together into a 16-dimensional feature vector. When a sudden change in environmental irradiance causes the stable operating range to fail, the perturbation sequence is immediately terminated, and the DC operating point is kept stable.

[0090] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0091] A bounded perturbation is applied to the DC operating point within the stable range of maximum power point tracking. Specifically, the bounded perturbation is a small bidirectional disturbance in the DC operating point voltage or DC operating point current. The perturbation amplitude does not exceed 0.5% of the rated DC operating point, the duration of a single perturbation does not exceed 200 ms, and the time interval between two adjacent perturbations is not less than 1 s.

[0092] Furthermore, voltage and current are synchronously acquired with a unified timestamp before and after the disturbance. Differential regression using adjacent time windows is used to obtain the voltage and current changes. Online estimation is performed under the physical constraints of a single-diode equivalent model to obtain the equivalent series resistance, equivalent parallel resistance, photocurrent, reverse saturation current, and ideality factor.

[0093] Therefore, the operating parameters, rate of change, and equivalent electrical parameters are combined into an eigenvector. The bounded perturbation is activated only within the stable region and automatically terminates outside the stable region.

[0094] For example, in practical applications, a ±0.3% perturbation can be applied to the DC operating point voltage for 150 ms, with an interval of 1.5 s between adjacent perturbations. Voltage and current data are simultaneously acquired using a 10 ms sampling period. Differential regression is performed through a 500 ms time window to obtain the voltage and current changes. Based on a single diode model, the equivalent electrical parameters are estimated using the least squares method. Finally, a 13-dimensional feature vector containing voltage, current, power, rate of change, and five equivalent electrical parameters is generated.

[0095] Through the above technical solution, this application achieves dynamic capture and precise quantification of the electrical characteristics of photovoltaic strings. High-precision measurement data of voltage and current responses are obtained through bounded perturbation and synchronous sampling. Online parameter estimation based on physical model constraints extracts key features reflecting the real-time state of the strings. This method avoids the limitations of traditional static testing, can reflect string performance changes in a timely manner, and provides a reliable data foundation for subsequent anomaly detection and diagnosis. Simultaneously, the amplitude and duration of the perturbation are strictly controlled, minimizing the impact on normal power generation. Furthermore, the stability interval judgment and automatic termination mechanism further ensure the effectiveness and safety of the measurement.

[0096] This application further proposes generating a topology graph containing string nodes, combiner box nodes, and inverter nodes based on primary wiring data. Edges are established according to electrical connections, and edge weights are determined by cable length, parallel circuit affiliation, and orientation consistency. Feature vectors are normalized and time-aligned before being attached to the corresponding nodes. Using node features, ambient irradiance, module temperature, and weighted aggregated features of topological neighbors as input, a representation learning model with physical constraints is invoked for inference. These physical constraints include: in the short-circuit to low-voltage range, the voltage does not exceed 30% of the full range, and the current decrease relative to the short-circuit current does not exceed 10%. In the high-voltage range, the voltage does not fall below 70% of the full range, and the current decreases significantly with increasing voltage, with a larger decrease than in the low-voltage range. In the neighborhood of the maximum power point, power increases with increasing voltage below this point and decreases with increasing voltage above this point. When ambient irradiance increases, the short-circuit current increases without decreasing the open-circuit voltage; when module temperature increases, the open-circuit voltage decreases. By incorporating topology consistency constraints, the virtual current-voltage characteristic curves and key feature points are output. The key feature points include open-circuit voltage, short-circuit current, and voltage and current at the maximum power point.

[0097] Cable lengths are obtained through measurement or design drawings; parallel circuit assignments are determined based on the primary wiring topology; and orientation consistency is calculated using string installation angles and geographical locations. Normalization employs a maximum-minimum scaling method, and time alignment is based on interpolation synchronization using a unified timestamp. The physical constraint representation learning model uses a neural network structure, embedding the physical relationship constraints between short-circuit current and open-circuit voltage into its loss function, and forcibly satisfying the voltage-current piecewise slope condition during backpropagation. Topology consistency constraints are implemented through graph regularization terms to ensure that the characteristic curves of adjacent nodes exhibit consistent trends.

[0098] Specifically, the primary wiring data is parsed into a graph structure containing strings, combiner boxes, and inverter nodes. Edge weights are calculated inversely proportional to cable lengths. Edge weights between nodes belonging to the same parallel circuit are adjusted with a correction coefficient, and edge weights between nodes with an azimuth difference of less than 15 degrees are increased. After normalization to eliminate dimensional differences, feature vectors are aligned using timestamps to ensure data synchronization. During training, the physically constrained model limits the output current in the short-circuit to low-voltage range to no less than 90% of the rated value, and sets the lower voltage limit in the high-voltage range to 70% of the full range. Simultaneously, a positive correlation between irradiance and short-circuit current and a negative correlation between temperature and open-circuit voltage are introduced as soft constraints. Topology consistency constraints suppress local abnormal fluctuations by calculating the similarity loss of characteristic curves of adjacent nodes. The error between the virtual curve output after model inference and the measured data at key points is controlled within 2%. The estimated values ​​of open-circuit voltage and short-circuit current are calibrated through closed-loop feedback, ultimately generating characteristic curves that conform to physical laws and topological relationships.

[0099] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0100] A topology graph containing string nodes, combiner box nodes, and inverter nodes is generated based on the initial wiring data. Edges are established according to electrical connections, and edge weights are determined by cable length, parallel circuit affiliation, and orientation consistency. The feature vectors are normalized and time-aligned before being attached to the corresponding nodes.

[0101] Using node features, ambient irradiance, module temperature, and weighted aggregate features of topological neighbors as input, a representation learning model with physical constraints is invoked for inference. The physical constraints include: within the short-circuit to low-voltage range, the voltage does not exceed 30% of the full range, and the decrease in current relative to the short-circuit current does not exceed 10%. Within the high-voltage range, the voltage does not fall below 70% of the full range, and the current decreases significantly with increasing voltage, with a larger decrease than in the low-voltage range. In the neighborhood of the maximum power point, power increases with increasing voltage below this point and decreases with increasing voltage above this point. With increasing ambient irradiance, the short-circuit current increases while the open-circuit voltage does not decrease; with increasing module temperature, the open-circuit voltage decreases.

[0102] By incorporating topology consistency constraints, the system outputs virtual current-voltage characteristic curves and key feature points. Key feature points include open-circuit voltage, short-circuit current, and the voltage and current at the maximum power point.

[0103] Specifically, a topology diagram is first constructed based on the primary wiring diagram of the power plant, with strings, combiner boxes, and inverters treated as different types of nodes. Edges are established between nodes according to the actual electrical connections, and edge weights are assigned. The edge weights are determined by a combination of factors such as cable length, parallel circuit affiliation, and orientation consistency. Then, the collected feature vectors are normalized, aligned according to a unified timestamp, and attached to the corresponding nodes.

[0104] Next, using node features, ambient irradiance, and module temperature as inputs, and considering the weighted aggregation features of topological neighbor nodes, a pre-trained representation learning model with physical constraints is invoked for inference. This model needs to satisfy a series of physical constraints during both training and inference to ensure that the output conforms to the basic electrical characteristics of photovoltaic modules.

[0105] Finally, topological consistency constraints are added to the model output to generate virtual current-voltage characteristic curves, and key feature points such as open-circuit voltage, short-circuit current, maximum power point voltage, and current are extracted. These feature points and curves can be used for subsequent anomaly detection and performance evaluation.

[0106] Through the above technical solution, this application achieves the reconstruction of photovoltaic string characteristic curves based on topology graphs. This allows for the acquisition of complete current-voltage characteristic curves of the strings without affecting normal power generation, providing crucial information for subsequent anomaly detection and performance evaluation. Furthermore, the introduction of physical constraints and topology consistency constraints improves the accuracy and reliability of the reconstruction results. In addition, this method is adaptable to photovoltaic power plants of different scales and topologies, exhibiting strong versatility and scalability.

[0107] This application further proposes that when generating node-level health indicators on the topology graph, the following steps are included: based on preset neighborhood selection conditions, selecting neighboring nodes that meet comparable conditions within the same combiner box, the same inverter, or the same azimuth cluster. The comparable conditions include an environmental irradiance difference of no more than 15 percentage points, a component temperature difference of no more than 3 degrees Celsius, an azimuth angle difference of no more than 15 degrees, and a sampling time alignment error of no more than one millisecond; assigning zero weight to neighboring nodes that do not meet the comparable conditions or are offline and removing them from the aggregation calculation; determining and normalizing the aggregation weight of each neighboring node based on factors such as azimuth consistency, irradiance similarity, historical correlation, and electrical connection strength; and performing at least two layers of weighted aggregation on the feature vectors of neighboring nodes that meet the comparable conditions to obtain the node-level health indicators of the corresponding photovoltaic string nodes.

[0108] The comparability criteria are defined by setting thresholds for environmental irradiance difference, component temperature difference, azimuth difference, and sampling time alignment error to filter neighboring nodes with similar environmental conditions and synchronized time, avoiding the inclusion of nodes with significant environmental differences in the comparison range. Nodes that do not meet the comparability criteria or are offline are assigned zero weight and removed to prevent invalid or missing data from interfering with the aggregation results. The aggregation weight comprehensively considers factors such as azimuth consistency (the closer the azimuth, the greater the weight), irradiance similarity (the closer the effective irradiance, the greater the weight), historical correlation (the more similar the historical power generation behavior, the greater the weight), and electrical connection strength (the closer the electrical coupling relationship, the greater the weight), and normalization ensures that the sum of the weights of each neighboring node is one. In multi-layer aggregation, the first layer aggregation is used to gather local neighbor features under the same combiner box or the same inverter, and the second layer aggregation is used to further fuse indirect neighbor features within the same azimuth cluster, so that the node-level health indicators take into account both local features and larger-scale topological relationships.

[0109] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0110] As a preferred embodiment, the specific implementation of the scheme in this application is as follows: In the topology diagram, for each photovoltaic string node, firstly, a candidate neighbor set is determined based on the primary wiring relationship and the component installation orientation, including other photovoltaic string nodes in the same combiner box, other photovoltaic string nodes under the same inverter, and photovoltaic string nodes in the same orientation cluster; then, based on the collected environmental irradiance, component temperature, azimuth angle, and sampling timestamp, nodes with an environmental irradiance difference of no more than 15%, a component temperature difference of no more than 3℃, an azimuth angle difference of no more than 15°, and a sampling time alignment error of no more than 1ms are selected as valid neighbors, and nodes that do not meet the above conditions or are offline are set to zero weight and removed. Based on this, the aggregation weights of each effective neighbor node are calculated. Azimuth cosine similarity is used to measure azimuth consistency; the reciprocal of the relative difference in effective irradiance is used to measure irradiance similarity; the correlation coefficient of power generation or active power within a preset time window is used to measure historical correlation; and a function of line impedance or the number of parallel branches is used to measure electrical connection strength. These factors are then weighted, summed, and normalized to obtain the final weights used for feature aggregation. Subsequently, a first-level weighted summation is performed on the neighbor node feature vectors to obtain intermediate aggregated features. These intermediate aggregated features are then combined with the effective neighbor features within a larger neighborhood for a second-level weighted aggregation, ultimately yielding second-order aggregated features that serve as node-level health indicators, characterizing the health status of the corresponding photovoltaic string under current operating conditions.

[0111] This application further proposes that, when generating anomaly scores, the following steps are included: collecting node-level health indicators for each photovoltaic string node within a preset time window; using node-level health indicator samples whose operating status meets preset normal criteria to calculate a normal reference center and its corresponding variance or covariance index; using the weighted distance between the current node-level health indicator and the normal reference center as a first component, and the dispersion index of the node-level health indicator within the current time window as a second component, and weighting and synthesizing the first component and the second component to obtain the anomaly score of the photovoltaic string node.

[0112] The normal reference center can be obtained by taking a weighted average of the node-level health indicators of historical normal samples. Variance or covariance indicators are used to characterize the natural fluctuation range of normal samples in each dimension. The weighted distance can be Euclidean distance, Mahalanobis distance, or other distance metrics that consider correlation, to reflect the degree of deviation of the current node-level health indicator from the normal reference center. The dispersion indicator can be obtained by statistically analyzing the variance or range of the node-level health indicator on the time axis within the current time window, to reflect the degree of fluctuation of the current node state in the short term. By weighting and synthesizing the degree of deviation and the degree of fluctuation, the anomaly score considers both the deviation of the current state from the long-term normal pattern and the instability of the recent state, making the anomaly identification results more robust.

[0113] As a preferred embodiment, the specific implementation of this application is as follows: During normal operation, photovoltaic string nodes that meet the requirements of voltage, current, power, and power factor all being within the normal range and without any protection action within one or more preset time windows are selected. The corresponding node-level health indicators are used as a normal sample set, and the mean vector in each dimension is calculated as the normal reference center. The covariance matrix is ​​also calculated as a description of the normal fluctuation range. During real-time operation, for each photovoltaic string node, a first component is obtained based on the Mahalanobis distance between the current node-level health indicator and the normal reference center. Simultaneously, within the most recent preset time window, the variance or comprehensive dispersion of the node-level health indicator in each dimension is calculated as a second component. The first component and the second component are linearly combined according to preset weights to obtain the anomaly score at the current moment. When the anomaly score exceeds a preset threshold, the photovoltaic string node is considered to have an anomaly risk, and the subsequent active sparse sampling and enhanced measurement process is triggered.

[0114] Through the above technical solution, this application, without relying on complex deep learning structures such as graph neural networks and contrastive learning, fully utilizes topological information and multidimensional similarity between nodes. It generates node-level health indicators with clear physical meaning through comparable condition screening and multi-layer weighted aggregation, and constructs an anomaly scoring mechanism based on normal reference centers and dispersion indicators, achieving a precise assessment of the health status of photovoltaic strings. This method improves the accuracy and interpretability of anomaly detection and diagnosis while avoiding the deployment costs and uncertainties associated with complex models, making it more suitable for integration into practical cloud platform monitoring systems.

[0115] This application further proposes active sparse sampling, which includes using a comprehensive priority system weighted by the dispersion of anomaly scores and node-level health indicators as the selection criterion. Target photovoltaic strings are selected from top to bottom according to the comprehensive priority within the power plant, with a daily coverage ratio not exceeding 5% and the same photovoltaic string not being selected more than once per hour. Enhanced measurements are performed on the target photovoltaic strings within the stable range of maximum power point tracking. The enhanced measurements consist of no more than five short-window bounded perturbations, with the amplitude of a single perturbation not exceeding 0.5% of the rated DC operating point, the duration of a single perturbation not exceeding 200 ms, and the total duration not exceeding 2 s. After the enhanced measurements are completed, the derivative and equivalent electrical parameters are recalculated based on the data collected before and after the perturbation, the eigenvector is updated, and the virtual current-voltage characteristic curve is reconstructed. Simultaneously, the node-level health indicators are updated using a sliding fusion method, so that the new node-level health indicators are the weighted sum of the old node-level health indicators and the node-level health indicators obtained based on the inference of this measurement, with the weight coefficient ranging from 0.6 to 0.9.

[0116] The comprehensive priority determination considers both the magnitude of the anomaly score and the dispersion of node-level health indicators within a preset time window, achieving a unified quantification of "anomaly degree" and "state instability degree," avoiding reliance on a single indicator that could lead to sampling resources biased towards a few extreme samples. The dispersion of node-level health indicators can be obtained by calculating the standard deviation of the indicator over the most recent sampling times or other dispersion measures; a larger dispersion indicates less stable current diagnostic results and higher potential information value. By limiting the daily coverage ratio and the number of triggers of the same photovoltaic string per unit time, active sparse sampling is ensured not to have an excessive impact on normal power generation in large-scale power plant scenarios, while ensuring that the sampling load is within the tolerance range of the cloud platform and edge nodes. Enhanced measurement employs multiple low-amplitude, short-duration bounded perturbations to acquire multiple sets of high-quality perturbation data within a limited time window, improving the accuracy of derivative calculation and equivalent electrical parameter estimation. Sliding fusion update, through weighted summation of old and new node-level health indicators, quickly reflects the latest measurement information while preserving the continuity of historical states, achieving a smooth transition and adaptive update of state assessment results.

[0117] As a preferred embodiment, the solution of this application is implemented as follows: During operation, the cloud platform periodically calculates the anomaly score of each photovoltaic string node and statistically analyzes the variance or standard deviation of the node-level health index within the most recent preset time window. The anomaly score and the dispersion are weighted and added together according to a preset ratio to obtain the comprehensive priority of each photovoltaic string. For example, the comprehensive priority can be set to the form of "anomaly score plus dispersion multiplied by a weighting coefficient". All photovoltaic strings in the power station are sorted from high to low according to the comprehensive priority. Each day, the photovoltaic strings with the highest comprehensive priority (approximately 5%) are selected as the target set for active sparse sampling, and the same photovoltaic string is constrained to be selected at most once in any given hour.

[0118] For the selected target photovoltaic string, when its maximum power point tracking is in a stable range, an enhanced measurement process is triggered. Within one enhanced measurement cycle, three to five short-window bounded perturbations are applied to the DC operating point. The amplitude of each perturbation is controlled between 0.3% and 0.5% of the rated DC operating point, the duration of each perturbation is controlled between 150 and 200 ms, and the total duration is controlled within 1.5 s to 2 s, ensuring that the perturbation does not cause significant power fluctuations or deviations beyond the stable range. Voltage and current data are synchronously collected at a fixed sampling period before and after the enhanced measurement. The voltage and current changes are recalculated through differential regression of adjacent time windows. Under the physical constraints of the single-diode equivalent model, the equivalent series resistance, equivalent parallel resistance, photocurrent, reverse saturation current, and ideality factor, etc., are estimated online. The updated operating parameters, rate of change, and equivalent electrical parameters are combined into a new feature vector. The virtual current-voltage characteristic curves and key feature points of the corresponding photovoltaic string are reconstructed by calling the physically constrained model.

[0119] When updating node-level health indicators, the recalculated node-level health indicators based on enhanced measurement data are fused with the node-level health indicators before enhanced measurement according to preset weights. For example, the new node-level health indicators can be set as the weighted sum of the old node-level health indicators and the updated node-level health indicators using weight coefficients ranging from 0.6 to 0.9. This ensures that the updated node-level health indicators both inherit the continuity of historical states and have high sensitivity to the latest enhanced measurement results. Subsequently, the updated feature vectors, virtual current and voltage characteristic curves, and node-level health indicators are stored in the monitoring database of the cloud platform for subsequent anomaly location, trend analysis, and linkage with the optimization control module.

[0120] Through the above technical solution, this application achieves enhanced measurement of high-risk or high-uncertainty photovoltaic strings without increasing the overall sampling burden. By using active sparse sampling and sliding fusion update mechanism, it improves the accuracy of conductor quantity and equivalent electrical parameter estimation and the reliability of virtual current and voltage characteristic curves. On the other hand, it ensures the smooth evolution of node-level health indicators over time and the ability to respond quickly, thereby improving the accuracy of anomaly detection and state assessment. This provides more reliable input data for subsequent optimization control and takes into account the scalability and operational efficiency in large-scale distributed photovoltaic power plant monitoring scenarios.

[0121] This application further proposes a scheme for locating abnormal photovoltaic strings and their associated nodes, including: calculating a shape deviation index based on a virtual current-voltage characteristic curve; combining the shape deviation index and the weighted distance of the node-level health index relative to the normal reference center to form an abnormal intensity; applying connectivity constraints and orientation consistency constraints in the topology graph to aggregate the neighborhood of the abnormal intensity and extract abnormal clusters; identifying the photovoltaic string with the largest abnormal intensity and the duration meeting a preset threshold within the abnormal cluster as the abnormal string; and outputting the combiner box node and inverter node to which the abnormal string belongs.

[0122] The shape deviation index can be composed of the open-circuit voltage offset ratio, short-circuit current offset ratio, maximum power point position offset ratio, and the degree of change in the segmented slope between the low-voltage and high-voltage regions. This is used to quantify the shape deviation of the virtual current-voltage characteristic curve relative to normal operating conditions from multiple dimensions. The weighted distance of the node-level health index relative to the normal reference center reflects the degree to which the photovoltaic string deviates from the normal operating mode in the multi-dimensional health feature space. The anomaly intensity is determined by linearly or non-linearly combining the shape deviation index with the aforementioned weighted distance using preset weights, avoiding the risk of misjudgment caused by relying on a single index. The neighborhood aggregation process introduces connectivity constraints to ensure that photovoltaic strings within anomaly clusters are topologically interconnected; simultaneously, it introduces orientation consistency constraints to ensure that photovoltaic strings within anomaly clusters have similar orientations and environmental conditions, thereby reducing the interference of isolated noise points and environmental fluctuations on the positioning results. Finally, the duration criterion is used to filter strings whose anomaly intensity continuously exceeds a threshold within a preset time window. The string with the largest anomaly intensity is identified as the anomaly string, and its corresponding combiner box node and inverter node are output, facilitating targeted troubleshooting and maintenance by operation and maintenance personnel.

[0123] Specifically, a baseline virtual current-voltage characteristic curve under historical normal operating conditions can be maintained for each photovoltaic string in the cloud platform, or representative baseline feature points can be calculated based on historical normal samples. Features such as open-circuit voltage, short-circuit current, and maximum power point voltage and current are extracted from the current virtual current-voltage characteristic curve and compared with the baseline values ​​to obtain the open-circuit voltage offset ratio, short-circuit current offset ratio, and maximum power point position offset ratio. Furthermore, the virtual current-voltage characteristic curve is segmented into low-voltage and high-voltage ranges, and the slope difference between the current curve and the baseline curve in each segment is compared. The weighted sum of the absolute values ​​of the slope differences is calculated as the segment slope variation degree, thus forming a shape deviation index. The abnormal distance of the node-level health index can be measured using Mahalanobis distance, weighted Euclidean distance, or other distance metrics that consider the correlation of various dimensions, characterizing the degree of deviation of the current node-level health index from the normal reference center in multi-dimensional space. This deviation is then linearly combined with the shape deviation index according to preset weights to generate the abnormal intensity.

[0124] In the topology graph, connected subgraphs are first selected based on electrical connectivity. Then, by combining information such as component orientation, ambient irradiance, and component temperature, regions where azimuth deviation and environmental differences do not exceed preset thresholds are selected. Within these regions, the abnormal intensity of each photovoltaic string is averaged or weighted and summed based on its neighborhood. Density clustering and threshold connectivity extraction methods are used to identify the set of nodes whose overall abnormal intensity exceeds the threshold, which are then designated as abnormal clusters. The photovoltaic strings within each abnormal cluster are sorted from highest to lowest abnormal intensity, and the duration for which the abnormal intensity continuously exceeds the preset threshold is checked on the time axis. The photovoltaic strings with the longest duration exceeding or exceeding the shortest duration threshold and the highest abnormal intensity are marked as abnormal strings. By traversing the topology edges, the combiner box node and inverter node to which the abnormal string belongs can be determined.

[0125] Through the above technical solution, this application establishes a unified anomaly intensity metric between the physical morphological information of virtual current-voltage characteristic curves and the multidimensional statistical characteristics of node-level health indicators. Combined with topological connectivity and orientation consistency constraints, it aggregates anomaly signals in both spatial and temporal dimensions, improving the accuracy of anomaly string location. This solution can distinguish between power fluctuations caused by environmental changes and genuine anomalies caused by component aging and electrical faults under complex operating conditions and environmental fluctuations, reducing false alarms and missed alarms. It ensures that the anomaly location results are consistent with the actual electrical connection relationships and physical layout, thus providing reliable support for the refined operation and maintenance of distributed photovoltaic power stations.

[0126] Even after adopting the aforementioned anomaly scoring and anomaly string location scheme based on virtual current and voltage characteristic curves and node-level health indicators, how to further use this state information for control strategy generation while meeting the grid connection specification parameter constraints, so as to achieve a balance between power generation revenue, equipment degradation risk and grid stability, remains a problem that needs to be solved.

[0127] To this end, this application further proposes a scheme for solving the optimization problem and generating the control trajectory under the parameter constraints of the grid connection specification, including: parameter constraints including at least voltage constraints, current constraints, power ramp rate constraints, and power factor range constraints; establishing a rolling control cycle, which simultaneously satisfies the upper and lower voltage limits, upper current limits, power ramp rate constraints, and power factor range constraints in each control cycle; taking the maximization of power generation revenue as the objective, and adding a degradation risk penalty term determined by the anomaly intensity and uncertainty, and a power fluctuation penalty term determined by the active power change measure of adjacent control cycles to form a comprehensive objective function; using the key feature points of the virtual current and voltage characteristic curves, node-level health indicators, and environmental prediction data as optimization inputs, solving for the active power limit and power factor control trajectory, and projecting the solution onto the feasible region defined by the grid connection specification before distributing it to the edge nodes for execution.

[0128] The parameter constraints ensure that grid-connected operation does not exceed the rated capacity of electrical equipment by setting upper and lower voltage limits on the photovoltaic string side or inverter side, as well as the upper limit of inverter output current. Power ramp-up constraints limit the rate of change of active power between adjacent control cycles, preventing rapid fluctuations in power plant output from impacting the grid. Power factor range constraints ensure that the inverter's operating power factor remains within the range allowed by grid specifications. The rolling control cycle can be set to several minutes based on the power plant scale and dispatch requirements. Through periodic rolling updates, the control strategy can be dynamically adjusted according to changes in environmental conditions and equipment status. The weighting coefficient of the degradation risk penalty term can be determined jointly by the anomaly intensity and uncertainty. The anomaly intensity can be derived from the deviation of the virtual current-voltage characteristic curve shape deviation index from the node-level health index relative to the normal reference center. Uncertainty can be quantified by the dispersion of the node-level health index within a preset time window, the fitting residual of the virtual current-voltage characteristic curve, or the prediction error, reflecting the uncertainty level of the current state assessment result. The power fluctuation penalty term can be constructed based on the difference or second-order difference of the active power setpoints between adjacent control cycles to suppress frequent and large-amplitude output adjustments. In the optimization input, key feature points of the virtual current-voltage characteristic curves can include open-circuit voltage, short-circuit current, and the voltage and current at the maximum power point. Node-level health indicators provide quantitative information on the health status of each photovoltaic string or topology node. Environmental prediction data can include irradiance prediction and module temperature prediction for a future period. The deprojection process checks whether each control variable in the optimization results meets constraints such as voltage, current, and power factor. When it is found that the active power limit or power factor exceeds the specification boundary, it is projected into the feasible region along the constraint boundary direction to ensure that the final issued control trajectory fully complies with the grid connection specification requirements.

[0129] Specifically, in a preferred embodiment, a 15-minute rolling control cycle can be used. At the beginning of each rolling control cycle, key feature points of the virtual current-voltage characteristic curves, node-level health indicators, and environmental prediction data for the most recent cycle are obtained from the cloud platform, and the abnormality intensity information output by the abnormal string location module is received. The parameter constraint module sets the upper voltage limit to 110% of the rated voltage, the lower voltage limit to 90% of the rated voltage, the upper current limit to 105% of the rated current, the power ramp-up rate constraint to no more than 2% of the rated power per minute, and the power factor range constraint to 0.95 leading to 0.95 lagging. In the comprehensive objective function, the power generation revenue term is calculated based on the predicted power generation and real-time electricity price; the degradation risk penalty term is constructed by multiplying the degradation risk factor obtained by weighting the abnormality intensity and uncertainty by a preset risk coefficient. The uncertainty can be calculated by the variance, covariance, or other dispersion indicators of the node-level health indicators in the most recent sampling cycles; the power fluctuation penalty term can be calculated by multiplying the square of the difference in active power limits between adjacent rolling control cycles or the square of the second-order difference by a preset penalty coefficient. The open-circuit voltage, short-circuit current, maximum power point voltage and current from the virtual current-voltage characteristic curves, along with corresponding component temperature, environmental irradiance prediction data, and node-level health indicators, are input into the optimization model to construct a quadratic programming or convex optimization problem with linear or nonlinear constraints. The active power limit and power factor setpoint for each control cycle are obtained using solvers such as the interior-point method or sequential quadratic programming. After solving, the feasible region is projected onto the solution. When the power factor of a certain control cycle exceeds the range of 0.95 lead to 0.95 lag, the power factor is projected along the constraint boundary to the nearest feasible point. When a slight overshoot trend is detected in voltage, current, or power ramp rate, corrections are made by appropriately increasing the corresponding penalty coefficient or reducing the active power limit in the objective function, ultimately forming a control trajectory that satisfies all constraints. The control trajectory includes the active power limit and corresponding power factor setpoint for each rolling control cycle, and is sent to the edge nodes via an encrypted communication channel. The local controller of the edge nodes converts this into specific control commands for the inverter's active and reactive power and executes them.

[0130] Through the aforementioned technical solution, this application, based on the detailed characterization of equipment status using virtual current-voltage characteristic curves and node-level health indicators, introduces a rolling optimization control mechanism under multi-parameter constraints. This not only ensures that key operating parameters such as voltage, current, power ramp-up rate, and power factor consistently meet grid connection specifications, but also balances the operating strategy across multiple objectives by simultaneously considering power generation revenue, degradation risk, and power fluctuations in the objective function. The degradation risk penalty term comprehensively characterizes the potential aging and failure risks of components or circuits through anomaly intensity and uncertainty, prompting the optimization result to appropriately reduce the output of high-risk strings and delay equipment degradation. The power fluctuation penalty term suppresses large changes in output between adjacent control cycles, improving the stability of power access to the grid. Thus, dynamic optimization control of the power plant's operating status is achieved under complex operating conditions and large-scale distributed photovoltaic scenarios, improving power generation revenue and equipment lifespan utilization while ensuring the safety and stability of grid operation.

[0131] In the above embodiments, by collecting photovoltaic string operating parameters in real time at edge nodes and applying bounded perturbations, equivalent electrical parameters are obtained to form physically reliable feature vectors. A topology graph is constructed, and virtual current-voltage characteristic curves and key feature points are reconstructed under physical constraints, achieving dynamic modeling consistent with real electrical behavior. Weighted aggregation of feature vectors from comparable neighboring nodes on the topology graph yields node-level health indicators and generates anomaly scores, improving the ability to identify component mismatch, performance degradation, and abnormal operating conditions. When anomaly scores trigger a threshold, enhanced measurements are performed on high-priority target strings through active sparse sampling, dynamically updating the virtual current-voltage characteristic curves and node-level health indicators. The indicators enable the monitoring model to continuously correct and converge under controllable disturbance costs; accurately locate abnormal photovoltaic strings and their associated nodes on the topology map; and, under the premise of meeting the grid connection specification parameter constraints, integrate power generation revenue, degradation risk and power fluctuation into the rolling optimization solution, generate active power limit and power factor control trajectory closed loop and send it to edge nodes for execution. This constructs a full-link closed-loop monitoring system of edge acquisition, physical modeling, topology correlation diagnosis, anomaly-driven enhanced measurement and grid connection constraint optimization control. In large-scale distributed photovoltaic scenarios, this improves the accuracy of power plant diagnosis, the reliability of anomaly location and grid connection operation stability, and enhances the scalability and intelligence level of the system.

[0132] In another preferred embodiment based on the above embodiments, see [reference] Figure 2 As shown, this embodiment provides a cloud-based distributed photovoltaic power plant monitoring and management system for applying the above-described cloud-based distributed photovoltaic power plant monitoring and management method, including:

[0133] The acquisition unit is configured to acquire the operating parameters of the photovoltaic strings at the edge nodes of the power plant;

[0134] The first processing unit is configured to apply a bounded perturbation to the DC operating point within the stable range of maximum power point tracking, obtain the voltage and current changes, calculate the derivatives and estimate the equivalent electrical parameters, thereby obtaining the eigenvector.

[0135] The reconstructing unit is configured to build a topology graph based on the electrical connection relationship between the photovoltaic string, combiner box and inverter, attach the feature vector as node attributes, and call the model with physical constraints to reconstruct the virtual current and voltage characteristic curves and key feature points;

[0136] The evaluation unit is configured to perform weighted aggregation on the feature vectors of neighboring nodes that meet the comparable conditions within the same combiner box, the same inverter, or the same directional cluster on the topology map based on preset neighborhood selection conditions and weight calculation rules, to obtain node-level health indicators, and generate anomaly scores by combining them with a reference distribution constructed from historical normal operation samples.

[0137] The judgment unit performs active sparse sampling and updates the virtual current-voltage characteristic curve and the node-level health index when the abnormal score meets the preset conditions.

[0138] The control unit is configured to locate abnormal photovoltaic strings and their respective nodes on the topology graph based on the shape deviation of the virtual current-voltage characteristic curve and the distribution of the node-level health indicators, and to solve the optimization problem of power generation revenue, degradation risk and power fluctuation under the parameter constraints of grid connection specifications, to obtain the active power limit and power factor control trajectory, and to send it to the edge nodes for execution.

[0139] Specifically, the acquisition unit synchronously collects multi-dimensional operating parameters at edge nodes with a sampling period of no more than one second. Sampling is paused when the stability interval criterion for maximum power point tracking is not met, thus avoiding the introduction of unsteady-state data into subsequent analysis and ensuring the validity and comparability of the collected data. The first processing unit obtains dynamic response data of voltage and current before and after the disturbance by applying short-term bounded perturbations with a duration of no more than two hundred milliseconds within the stable interval. Under the physical constraints of the single-diode equivalent model, it estimates equivalent series resistance, equivalent parallel resistance, photocurrent, reverse saturation current, and ideal factor, etc., equivalent electrical parameters online. The operating parameters, rate of change, and equivalent electrical parameters are combined to form a feature vector with physical meaning, providing a basis for subsequent reconstruction of virtual current-voltage characteristic curves and health assessment. The reconstructing unit builds a topology graph containing photovoltaic string nodes, combiner box nodes, and inverter nodes based on the primary system diagram of the power plant. Edges are established according to electrical connection relationships, and edge weights are determined with reference to cable length, parallel circuit affiliation, and orientation consistency. Normalized and time-aligned feature vectors are attached to the corresponding nodes. A representation learning model with physical constraints is invoked to apply constraints on voltage, current, and power relationships in the short-circuit region, low-voltage region, high-voltage region, and the neighborhood of the maximum power point. The reconstructed virtual current-voltage characteristic curves conforming to physical laws are obtained, along with key feature points such as open-circuit voltage, short-circuit current, and maximum power point.

[0140] On the aforementioned topology map, the evaluation unit first selects comparable neighbor nodes within the same combiner box, inverter, or azimuth cluster based on preset neighborhood selection criteria. These nodes are considered comparable because their environmental irradiance difference, component temperature difference, azimuth angle difference, and sampling time alignment error are all within preset thresholds. Neighbor nodes that do not meet the comparability criteria or are offline are assigned zero weight and removed from the aggregation calculation. Then, the unit determines and normalizes the aggregation weights of each comparable neighbor node based on factors such as azimuth consistency, irradiance similarity, historical correlation, and electrical connection strength. At least two layers of weighted aggregation are performed on the neighbor node feature vectors to obtain node-level health indicators that comprehensively reflect the local state of the equipment and its topology correlation. The evaluation unit further selects historical node-level health indicator samples that meet preset normality criteria to construct a normal reference distribution. It calculates the normal reference center and its corresponding variance or covariance index. The weighted distance between the current node-level health indicator and the normal reference center, along with the dispersion of the node-level health indicator within a preset time window, are weighted and synthesized to obtain the anomaly score for each photovoltaic string.

[0141] When the judgment unit detects that the abnormal score of a photovoltaic string node reaches or exceeds a preset threshold, it initiates an active sparse sampling process: based on the abnormal score and the dispersion of node-level health indicators, a comprehensive priority is formed. Target photovoltaic strings are selected from high to low priority within the power plant area, with the daily coverage ratio constrained to not exceed 5% of the total number of strings in the power plant and the same photovoltaic string not being sampled more than once per hour, to prevent excessive impact on normal power generation and communication bandwidth. When the maximum power point tracking of the target photovoltaic string is in a stable range, the first processing unit implements multiple short-window bounded micro-perturbations within a limited time according to the enhanced measurement strategy, collects voltage and current data before and after the perturbation, recalculates the derivative and equivalent electrical parameters, updates the corresponding feature vector, and the reconstruction unit reconstructs a new virtual current-voltage characteristic curve. Simultaneously, the judgment unit updates the node-level health indicators through a sliding fusion method, weighting and summing the latest node-level health indicators obtained based on enhanced measurement with the node-level health indicators before the update according to a preset weight coefficient, so that the node-level health indicators maintain the continuity of historical states while quickly reflecting the latest measurement results.

[0142] The control unit receives the virtual current-voltage characteristic curves and key feature points output by the reconfiguration unit. Combining this with the node-level health indicators and anomaly scores output by the evaluation unit, it calculates the shape deviation index and anomaly intensity of each photovoltaic string on the topology map. Based on the spatial distribution of shape deviations and node-level health indicators on the topology map, it locates the abnormal photovoltaic strings and their associated combiner box nodes and inverter nodes. On this basis, the control unit establishes an optimization model within the rolling control cycle according to grid connection specification parameters such as upper and lower voltage limits, upper current limits, power ramp-up rate, and power factor range. With the goal of maximizing power generation revenue, it introduces a degradation risk penalty term constructed from anomaly intensity and state uncertainty, and a power fluctuation penalty term based on changes in active power setpoints in adjacent control cycles to construct a comprehensive objective function. Using the key feature points of the virtual current-voltage characteristic curves, node-level health indicators, and environmental prediction data as optimization inputs, it solves for the active power limit and power factor control trajectory for each rolling control cycle. After ensuring that the optimization results meet all grid connection specification constraints through feasible region projection, the results are sent to the local controllers of the edge nodes for power regulation.

[0143] Through the aforementioned multi-unit collaboration, the system of this application, without introducing complex models such as graph neural networks and contrastive learning, completes a full closed-loop process, from high-quality data acquisition and dynamic feature extraction under bounded perturbations, to reconstruction of virtual current and voltage characteristic curves with physical constraints, to generating node-level health indicators based on topology and multi-dimensional similarity, to actively sparse sampling to enhance high-risk string measurement, and to rolling optimization control under multi-parameter constraints. This improves the anomaly identification accuracy and control strategy flexibility of distributed photovoltaic power plants in large-scale scenarios, reduces false alarm and false alarm rates, and simultaneously takes into account power generation revenue, equipment degradation risk, and grid operation stability. It also solves the problems of monitoring lag, coarse positioning, and rigid control in existing cloud monitoring systems.

[0144] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for monitoring and managing distributed photovoltaic power plants based on a cloud platform, characterized in that, include: The operating parameters of the photovoltaic strings are collected at the edge nodes of the power station; Within the stable range of maximum power point tracking, a bounded perturbation is applied to the DC operating point to obtain the voltage and current changes, and the derivatives are calculated and the equivalent electrical parameters are estimated, thereby obtaining the eigenvector. A topology graph is constructed based on the electrical connection relationship between photovoltaic strings, combiner boxes and inverters. The feature vectors are attached as node attributes, and a model with physical constraints is called to reconstruct the virtual current and voltage characteristic curves and key feature points. Based on preset neighborhood selection conditions and weight calculation rules, the feature vectors of neighboring nodes that meet the comparable conditions within the same combiner box, the same inverter, or the same directional cluster are weighted and aggregated on the topology map to obtain node-level health indicators. Anomaly scores are generated by combining the reference distribution constructed from historical normal operation samples. When the anomaly score meets the preset conditions, active sparse sampling is performed, and the virtual current-voltage characteristic curve and the node-level health index are updated. Based on the shape deviation of the virtual current-voltage characteristic curve and the distribution of the node-level health indicators on the topology graph, the abnormal photovoltaic strings and their respective nodes are located. Under the parameter constraints of the grid connection specifications, the optimization problem of power generation revenue, degradation risk and power fluctuation is solved to obtain the active power limit and power factor control trajectory, which is then sent to the edge nodes for execution.

2. The distributed photovoltaic power station monitoring and management method based on a cloud platform according to claim 1, characterized in that, When collecting operating parameters of photovoltaic strings at the edge nodes of a power station, the following are included: The operating parameters include voltage, current, power, inverter active power, power factor, ambient irradiance, and component temperature. Within the stable range of maximum power point tracking, the voltage, current, power, active power of the photovoltaic string, power factor of the photovoltaic string, ambient irradiance, and module temperature are synchronously sampled and assigned a unified timestamp at a sampling period of no more than 1 second. When the stable range criterion is not met, sampling is paused until the stable range is restored. The synchronously sampled data is sequentially subjected to amplitude limiting and noise reduction, median filtering, and sliding window drift removal processing, and unit unification and range calibration are completed. Ambient irradiance is effectively self-calibrated by combining AC side metering power with inverter efficiency characteristics. When parameter missing or sampling abnormality is detected, short-window interpolation is performed only within the stable range and does not cross the stable range boundary.

3. The method for monitoring and managing distributed photovoltaic power plants based on a cloud platform according to claim 2, characterized in that, Applying a bounded perturbation to the DC operating point within the stable range of maximum power point tracking, and obtaining the eigenvector, includes: The bounded perturbation is a small bidirectional perturbation on the DC operating point voltage or DC operating point current, with an amplitude not exceeding 0.5% of the rated DC operating point, a single duration not exceeding 200ms, and a time interval of not less than 1s between two adjacent perturbations; Voltage and current are synchronously collected with a unified timestamp before and after the disturbance. The voltage and current changes are obtained by differential regression with adjacent time windows. Online estimation is performed under the physical constraints of the single diode equivalent model to obtain the equivalent series resistance, equivalent parallel resistance, photocurrent, reverse saturation current and ideality factor. The operating parameters, changes and equivalent electrical parameters are combined into a feature vector. The bounded perturbation is only activated within the stable range and automatically terminated outside the stable range.

4. The method for monitoring and managing distributed photovoltaic power plants based on a cloud platform according to claim 3, characterized in that, When constructing the topology graph and reconstructing the virtual current-voltage characteristic curves, the following steps are included: A topology graph containing string nodes, combiner box nodes and inverter nodes is generated based on the primary wiring data. Edges are established according to electrical connections and edge weights are determined by cable length, parallel circuit affiliation and orientation consistency. The feature vectors are normalized and time-aligned and then attached to the corresponding nodes. Using node features, ambient irradiance, module temperature, and weighted aggregation features of topological neighbors as input, a representation learning model with physical constraints is invoked for inference. These physical constraints include: in the short-circuit to low-voltage range, the voltage is no higher than 30% of the full range, and the current decrease relative to the short-circuit current does not exceed 10%; in the high-voltage range, the voltage is no lower than 70% of the full range, and the current decreases significantly with increasing voltage, with a greater decrease than in the low-voltage range; in the neighborhood of the maximum power point, power increases with increasing voltage below this point and decreases with increasing voltage above it; short-circuit current increases with increasing ambient irradiance without decreasing open-circuit voltage, and open-circuit voltage decreases with increasing module temperature. By incorporating topology consistency constraints, the virtual current-voltage characteristic curves and key feature points are output. These key feature points include open-circuit voltage, short-circuit current, and the voltage and current at the maximum power point.

5. The distributed photovoltaic power station monitoring and management method based on a cloud platform according to claim 4, characterized in that, When generating node-level health metrics on the topology graph, the following are included: Based on preset neighborhood selection conditions, neighboring nodes that meet comparable conditions are selected within the same combiner box, the same inverter, or the same azimuth cluster. The comparable conditions include: environmental irradiance difference not exceeding 15%, component temperature difference not exceeding 3℃, azimuth angle difference not exceeding 15°, and sampling time alignment error not exceeding 1ms. Nodes that do not meet the comparable conditions or are offline are assigned zero weight and removed. The feature vectors of each neighboring node are aggregated and normalized according to azimuth consistency, irradiance similarity, historical correlation, and electrical connection strength. The feature vectors of the neighboring nodes are weighted and aggregated at least twice to obtain the node-level health indicators of the corresponding photovoltaic string nodes.

6. The method for monitoring and managing a distributed photovoltaic power station based on a cloud platform according to claim 5, characterized in that, When generating anomaly scores, the following are included: Within a preset time window, node-level health indicators of each photovoltaic string are collected. Node-level health indicators whose operating status meets preset normality criteria are used to calculate the normal reference center and the corresponding variance or covariance index. The weighted distance between the current node-level health indicator and the normal reference center is used as the first component, and the dispersion index of the node-level health indicator within the current time window is used as the second component. The first component and the second component are weighted and synthesized to obtain the anomaly score of the photovoltaic string node.

7. The method for monitoring and managing a distributed photovoltaic power station based on a cloud platform according to claim 6, characterized in that, When performing active sparse sampling and updating the virtual current-voltage characteristic curves and node-level health indicators, the following steps are included: The active sparse sampling includes: based on the dispersion of the anomaly score and the node-level health index to form a comprehensive priority, selecting target photovoltaic strings within the power station according to the comprehensive priority, with a daily coverage ratio not exceeding 5% and the same photovoltaic string not exceeding once per hour; Enhanced measurements are performed on the target photovoltaic string within the stable range of maximum power point tracking. The enhanced measurements consist of no more than five short-window bounded perturbations, with the amplitude of a single perturbation not exceeding 0.5% of the rated DC operating point, the duration of a single perturbation not exceeding 200 ms, and the total duration not exceeding 2 s. After the enhanced measurement is completed, the derivative quantity and equivalent electrical parameters are recalculated based on the data collected before and after the disturbance. The feature vector is updated and the virtual current-voltage characteristic curve is reconstructed. At the same time, the node-level health index is updated in a sliding fusion manner, so that the new node-level health index is a weighted sum of the old node-level health index and the node-level health index obtained based on the latest feature inference, with the weight coefficient ranging from 0.6 to 0.

9.

8. The method for monitoring and managing distributed photovoltaic power plants based on a cloud platform according to claim 1, characterized in that, When locating abnormal strings and their parent nodes, the following steps are included: The shape deviation index is calculated based on the virtual current-voltage characteristic curve. The shape deviation index consists of the open-circuit voltage offset ratio, the short-circuit current offset ratio, the maximum power point position offset ratio, and the degree of change of the segmented slope. The anomaly intensity is obtained by combining the shape deviation index with the weighted distance of the node-level health index of the corresponding photovoltaic string node relative to the normal reference center. Apply connectivity constraints and orientation consistency constraints within the topology graph, perform neighborhood aggregation on anomaly intensity, and extract anomaly clusters. The string with the highest abnormal intensity and the longest duration within the abnormal cluster is identified as the abnormal string, and the combiner box node and inverter node to which the abnormal string belongs are output.

9. The method for monitoring and managing a distributed photovoltaic power station based on a cloud platform according to claim 8, characterized in that, When solving optimization problems and generating control trajectories while satisfying the parameter constraints of grid connection specifications, the following are included: Parameter constraints include voltage, current, ramp rate, and power factor; Establish a rolling control cycle, which simultaneously satisfies the upper and lower limits of voltage, the upper limit of current, the power ramp rate constraint, and the power factor range constraint within each control cycle. The goal is to maximize power generation revenue, and a comprehensive objective is formed by adding a degradation risk penalty term determined by the dispersion index of abnormal intensity and node-level health indicators, and a power fluctuation penalty term determined by the active power change measurement of adjacent cycles. Using the key feature points of the virtual current-voltage characteristic curve, node-level health indicators, and environmental predictions as optimization inputs, the active power limit and power factor control trajectory are obtained by solving. The solution results are then projected onto the feasible domain of the grid connection specification and sent to the edge nodes, which then execute the control trajectory.

10. A cloud-based distributed photovoltaic power station monitoring and management system, used to apply the cloud-based distributed photovoltaic power station monitoring and management method as described in any one of claims 1-9, characterized in that, include: The acquisition unit is configured to acquire the operating parameters of the photovoltaic strings at the edge nodes of the power plant; The first processing unit is configured to apply a bounded perturbation to the DC operating point within the stable range of maximum power point tracking, obtain the voltage and current changes, calculate the derivatives and estimate the equivalent electrical parameters, thereby obtaining the eigenvector. The reconstructing unit is configured to build a topology graph based on the electrical connection relationship between the photovoltaic string, combiner box and inverter, attach the feature vector as node attributes, and call the model with physical constraints to reconstruct the virtual current and voltage characteristic curves and key feature points; The evaluation unit is configured to perform weighted aggregation on the feature vectors of neighboring nodes that meet the comparable conditions within the same combiner box, the same inverter, or the same directional cluster on the topology map based on preset neighborhood selection conditions and weight calculation rules, to obtain node-level health indicators, and generate anomaly scores by combining them with a reference distribution constructed from historical normal operation samples. The judgment unit performs active sparse sampling and updates the virtual current-voltage characteristic curve and the node-level health index when the abnormal score meets the preset conditions. The control unit is configured to locate abnormal photovoltaic strings and their respective nodes on the topology graph based on the shape deviation of the virtual current-voltage characteristic curve and the distribution of the node-level health indicators, and to solve the optimization problem of power generation revenue, degradation risk and power fluctuation under the parameter constraints of grid connection specifications, to obtain the active power limit and power factor control trajectory, and to send it to the edge nodes for execution.

Images