Distributed photovoltaic power station monitoring management method and system based on cloud platform

Abstract

The invention relates to the technical field of photovoltaic power station intelligent monitoring, and discloses a distributed photovoltaic power station monitoring management method and system based on a cloud platform, and the method comprises the steps: collecting the operation parameters of a photovoltaic string; applying bounded perturbation in a stable interval of maximum power point tracking to obtain a feature vector; and constructing a topological graph, and reconstructing a virtual current and voltage characteristic curve and key characteristic points. And obtaining a node-level health index on the topological graph, and generating an abnormal score in combination with the reference distribution. And according to the shape deviation of the virtual current and voltage characteristic curve and the distribution of the node-level health indexes, positioning an abnormal string and an affiliation node thereof, solving the optimization problem of power generation income, degradation risk and power fluctuation under the condition of satisfying parameter constraints, and obtaining an active power limit and a power factor control track to be issued to an edge node for execution. According to the invention, the diagnosis precision and the grid-connected stability of the distributed photovoltaic power station in a large-scale scene are improved, and the expandability and the intelligent level are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent monitoring of photovoltaic power stations, in particular to a distributed photovoltaic power station monitoring management method and system based on a cloud platform. BACKGROUND

[0002] Photovoltaic power stations are an important part of new energy power generation systems. With the rapid expansion of installed capacity, the operation monitoring and intelligent management capabilities directly affect the power generation efficiency of the power station and the safety and stability of the power grid. At the present stage, the industry is gradually migrating from traditional localized monitoring methods to centralized architectures based on cloud platforms. Through edge nodes, operation parameters are collected and uploaded to the cloud to achieve comprehensive monitoring of power station equipment status, power output, and environmental conditions. This method can improve the real-time monitoring and centralized management capabilities to a certain extent, providing necessary data support for power grid dispatching 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 station based on cloud platform, still have deficiencies. The existing system architecture relies on fixed hardware configurations, has limited scalability, and requires additional modifications when the power station scale expands, increasing operation and maintenance costs. Data processing methods are mostly limited to threshold comparisons, making it difficult to accurately identify complex problems such as mismatched strings and component degradation, resulting in potential risks that cannot be discovered in a timely manner. The monitoring method lacks dynamic modeling capabilities based on physical mechanisms, and the use of current-voltage characteristic curves is limited to static comparisons, which cannot reflect real-time degradation status. In addition, the data processing process lacks flexibility in high-concurrency scenarios, which can cause delays and losses, restricting the reliability of cloud platforms in large-scale monitoring of distributed photovoltaic power stations.

[0004] Therefore, it is necessary to design a distributed photovoltaic power station monitoring management method and system based on a cloud platform to solve the problems existing in the current technology. SUMMARY

[0005] In view of this, the present application proposes a distributed photovoltaic power station monitoring management method and system based on a cloud platform, aiming to solve the problems of lack of dynamic modeling capability, insufficient abnormality recognition accuracy, and limited grid-connected control optimization of existing photovoltaic power station monitoring methods in large-scale distributed scenarios.

[0006] In one aspect, the present application proposes a distributed photovoltaic power station monitoring management method based on a cloud platform, comprising: collecting operation parameters of photovoltaic strings at the edge nodes of the power station; applying a bounded perturbation to the direct current operating point within the stable interval of maximum power point tracking, obtaining the voltage variation and current variation, and calculating the conductance and estimated equivalent electrical parameters to obtain the feature vector; A topology graph is constructed based on the electrical connection relationship of photovoltaic strings, combiner boxes and inverters, the feature vector is mounted as a node attribute, and a model with physical constraints is called to reconstruct a virtual current-voltage characteristic curve and key feature points; On the topology graph, based on the preset neighborhood selection condition and weight calculation rule, the feature vectors of neighbor nodes that meet the comparable condition in the same combiner box, the same inverter or the same azimuth cluster are aggregated with weights to obtain a node-level health index, and an abnormal score is generated in combination with a reference distribution constructed based on historical normal operation samples; When the abnormal score meets the preset condition, active sparse sampling is performed, and the virtual current-voltage characteristic curve and the node-level health index are updated; According to the shape deviation of the virtual current-voltage characteristic curve and the distribution of the node-level health index on the topology graph, an abnormal photovoltaic string and its belonging node are located, and under the parameter constraints meeting the grid-connected specification, an optimization problem of power generation benefit, degradation risk and power fluctuation is solved to obtain an active power limit and a power factor control trajectory, which is issued to the edge node for execution.

[0007] Further, when collecting the operating parameters of the photovoltaic string at the edge node of the power station, including: The operating parameters include voltage, current, power, active power of the inverter, power factor, environmental irradiance and component temperature; In the stable interval of maximum power point tracking, the voltage of the photovoltaic string, the current of the photovoltaic string, the power of the photovoltaic string, the active power of the inverter, the power factor of the inverter, the environmental irradiance and the component temperature are sampled synchronously with a sampling period not higher than 1S and are given a uniform timestamp; when the stable interval criterion is not met, sampling is suspended until the stable interval is restored; the synchronous sampling data are sequentially subjected to amplitude limiting and denoising, median filtering and sliding window de-drifting processing and complete unit uniformity and range calibration, wherein the environmental irradiance is effectively self-calibrated through alternating current side metering power combined with inverter efficiency characteristics; when parameter loss or sampling anomaly is detected, only short window interpolation is performed in the stable interval and the stable interval boundary is not crossed.

[0008] Further, in the stable interval of maximum power point tracking, a bounded perturbation is applied to the DC operating point, and when the feature vector is obtained, including: The bounded perturbation is a small bidirectional disturbance on the DC operating point voltage or the 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 between adjacent two disturbances not less than 1s; The voltage and current are synchronously collected with unified time stamp before and after the disturbance, the voltage variation and current variation are obtained by using adjacent time window difference regression, and the equivalent series resistance, equivalent parallel resistance, photo-generated current, reverse saturation current and ideal factor are obtained by implementing online estimation under the physical constraint of single diode equivalent model; the operating parameters, variation rates and equivalent electrical parameters are combined into a feature vector; wherein, the bounded disturbance is only enabled within the stable interval, and is automatically terminated beyond the stable interval.

[0009] Further, when constructing a topology graph and reconstructing a virtual current-voltage characteristic curve, comprising: Based on the primary wiring data, a topology graph containing string nodes, combiner box nodes and inverter nodes is generated, edges are established according to electrical connection, and edge weights are determined according to cable length, parallel loop ownership and orientation consistency; the feature vector is mounted to the corresponding node after normalization and time alignment; The weighted aggregated features of node characteristics, environmental irradiation, component temperature and topology neighbors are taken as input, and a representation learning model with physical constraints is called for inference, the physical constraints include that in the short-circuit to low-voltage range, the voltage is not higher than 30% of the full range, and the current relative to the short-circuit current is not more than 10%; in the high-voltage range, the voltage is not lower than 70% of the full range, the current decreases obviously with the increase of voltage and the decrease amplitude is higher than that in the low-voltage range; in the maximum power point neighborhood, the power increases with the increase of voltage below the point and decreases with the increase of voltage above the point; when the environmental irradiation increases, the short-circuit current increases and the open-circuit voltage does not decrease, when the component temperature increases, the open-circuit voltage decreases; The topology consistency constraint is added, and the virtual current-voltage characteristic curve and the key feature points are output, the key feature points include the open-circuit voltage, the short-circuit current and the voltage and current of the maximum power point.

[0010] Further, when generating a node-level health index on the topology graph, comprising: Based on the preset neighborhood selection condition, the neighbor nodes satisfying the comparable condition are selected within the same combiner box, the same inverter or the same orientation cluster; the comparable condition includes: the environmental irradiation difference is not more than 15%, the component temperature difference is not more than 3℃, the azimuth angle difference is not more than 15°, and the sampling time alignment error is not more than 1ms; the nodes not satisfying the comparable condition or in offline state are assigned zero weight and removed; the feature vectors of each neighbor node are determined according to the orientation consistency, irradiation similarity, historical correlation and electrical connection strength, and the aggregation weight is normalized, at least two layers of weighted aggregation are performed on the neighbor node feature vectors, and the node-level health index of the corresponding photovoltaic string node is obtained.

[0011] Further, when generating an abnormal score, comprising: Collecting node-level health indicators of each photovoltaic string within a preset time window, using the node-level health indicators meeting preset normal criteria to calculate normal reference centers and corresponding variance or covariance indicators; taking a weighted distance between a current node-level health indicator and the normal reference center as a first component, taking a dispersion indicator of the node-level health indicators within a current time window as a second component, and performing weighted synthesis on the first component and the second component to obtain an abnormal score of the photovoltaic string node.

[0012] Further, when performing active sparse sampling and updating the virtual current-voltage characteristic curve and the node-level health indicator, the method comprises: The active sparse sampling comprises: constructing a comprehensive priority based on the abnormal score and the dispersion of the node-level health indicator, selecting a target photovoltaic string according to the comprehensive priority within the power station, and the daily coverage ratio is not more than 5%, and the same photovoltaic string is not more than once per hour; Performing enhanced measurement on the target photovoltaic string within a stable interval of maximum power point tracking, the enhanced measurement is composed of a short window bounded perturbation not more than five times, the amplitude of a single perturbation is not more than 0.5% of the rated DC operating point, the single duration is not more than 200ms, and the total duration is not more than 2s; After the enhanced measurement is completed, recalculating the derivative quantity and the equivalent electrical parameter according to the data collected before and after the perturbation, updating the feature vector and reconstructing the virtual current-voltage characteristic curve, and updating the node-level health indicator in a sliding fusion manner, so that the new node-level health indicator is a weighted sum of the old node-level health indicator and the node-level health indicator obtained based on the latest feature reasoning, and the weight coefficient is in the range of (0.6-0.9).

[0013] Further, when locating the abnormal string and its belonging node, the method comprises: Calculating a shape deviation indicator based on the virtual current-voltage characteristic curve, the shape deviation indicator is composed of an open-circuit voltage offset ratio, a short-circuit current offset ratio, a maximum power point position offset ratio, and a segment slope change degree; Synthesizing the shape deviation indicator and the weighted distance of the node-level health indicator of the corresponding photovoltaic string node relative to the normal reference center to obtain an abnormal intensity; Applying connectivity constraints and orientation consistency constraints in the topology graph, performing neighborhood aggregation on the abnormal intensity, and extracting an abnormal cluster; Determining the string with the maximum abnormal intensity and the duration satisfying the shortest duration threshold in the abnormal cluster as an abnormal string, and outputting the junction box node and the inverter node to which the abnormal string belongs.

[0014] Further, when solving the optimization problem and generating the control trajectory under the parameter constraints meeting the grid connection specification, the method comprises: The parameter constraints include voltage, current, ramp rate and power factor; A rolling control cycle is established, and in each control cycle, the upper and lower voltage constraints, the upper current constraint, the power ramp rate constraint and the power factor range constraint are simultaneously satisfied; The maximum generation benefit is taken as the target, and a degradation risk penalty term determined by the abnormal intensity and the dispersion index of the node-level health index and a power fluctuation penalty term determined by the active power change measure of the adjacent cycle are added to form a comprehensive target; The key feature points of the virtual current-voltage characteristic curve, the node-level health index and the environment prediction are taken as optimization inputs, the active power limit and the power factor control trajectory are solved, and the solving result is projected to the feasible region of the grid connection specification and then issued to the edge node, which executes according to the control trajectory.

[0015] Compared with the prior art, the beneficial effects of the present application are that: by collecting photovoltaic string operation parameters in real time at the edge node and applying bounded perturbation, equivalent electrical parameters are obtained to form a physically credible feature vector; a topology graph is constructed, and the virtual current-voltage characteristic curve and key feature points are reconstructed under physical constraints to realize dynamic modeling consistent with the real electrical behavior; the node-level health index is obtained by weighted aggregation of the feature vectors of the comparable neighbor nodes on the topology graph, and an abnormal score is generated to improve the identification capability of component mismatch, performance degradation and abnormal operating conditions; when the abnormal score triggers the threshold value, enhanced measurement is implemented on the target string with higher priority through active sparse sampling, and the virtual current-voltage characteristic curve and the node-level health index are dynamically updated, so that the monitoring model continuously corrects and converges under the condition of controllable disturbance cost; the abnormal photovoltaic string and its belonging node are accurately located on the topology graph, and under the premise of meeting the grid connection specification parameter constraints, the generation benefit, the degradation risk and the power fluctuation are uniformly included in the rolling optimization solution to generate the active power limit and the power factor control trajectory closed loop issued to the edge node for execution, thereby constructing a full-link closed-loop monitoring system of edge collection-physical modeling-topology correlation diagnosis-abnormal driving enhanced measurement-optimization control under grid connection constraints, which improves the power station diagnosis accuracy, abnormal positioning reliability and grid operation stability in the large-scale distributed photovoltaic scene, and improves the scalability and intelligent level of the system.

[0016] On the other hand, the present application also provides a cloud platform-based distributed photovoltaic power station monitoring and management system for applying the cloud platform-based distributed photovoltaic power station monitoring and management method described above, which comprises: The acquisition unit is configured to collect the operation parameters of the photovoltaic string at the edge node of the power station; The first processing unit is configured to apply a bounded perturbation to the direct current operating point within a stable interval of maximum power point tracking, obtain a voltage variation and a current variation, and calculate a conductance and an estimated equivalent electrical parameter, so as to obtain a feature vector; The reconstruction unit is configured to construct a topology graph based on an electrical connection relationship of the photovoltaic strings, the combiner boxes and the inverters, mount the feature vector as a node attribute, and call a model to reconstruct a virtual current-voltage characteristic curve and key feature points with physical constraints; The evaluation unit is configured to implement weighted aggregation on the feature vectors of neighbor nodes that meet comparable conditions in the same combiner box, the same inverter or the same azimuth cluster based on a preset neighborhood selection condition and a weight calculation rule on the topology graph, obtain a node-level health index, and generate an anomaly score in combination with a reference distribution constructed based on historical normal operation samples; The judgment unit executes active sparse sampling and updates the virtual current-voltage characteristic curve and the node-level health index when the anomaly score meets a preset condition. The control unit is configured to locate an abnormal photovoltaic string and its belonging node according to a shape deviation of the virtual current-voltage characteristic curve and a distribution of the node-level health index on the topology graph, and solve an optimization problem of power generation benefit, degradation risk and power fluctuation under a parameter constraint meeting a grid-connected specification, to obtain an active power limit and a power factor control trajectory, and issue to an edge node for execution.

[0017] It can be understood that the cloud platform-based distributed photovoltaic power station monitoring and management method and system have the same beneficial effects, which will not be repeated here. BRIEF DESCRIPTION OF DRAWINGS

[0018] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The accompanying drawings are included to provide a description of the preferred embodiments and are not meant to limit the present application. Moreover, the same reference numerals in the attached drawings indicate the same or similar components. In the drawings: Figure 1 A flowchart of the cloud platform-based distributed photovoltaic power station monitoring and management method provided by the embodiments of the present application; Figure 2 A functional block diagram of the cloud platform-based distributed photovoltaic power station monitoring and management system provided by the embodiments of the present application. DETAILED DESCRIPTION

[0019] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it is understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that the present disclosure can be more thoroughly understood, and so that the scope of the present disclosure can be conveyed completely to those skilled in the art. It should be noted that the embodiments in the present disclosure and the features in the embodiments can be combined with each other without conflict. The present disclosure will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

[0020] In the conventional existing distributed photovoltaic power station monitoring system, the scalability and data processing capability of the cloud platform architecture have bottlenecks. With the expansion of the power station scale, the number of edge nodes grows exponentially, and it is difficult for fixed hardware configuration to dynamically adapt to the access requirements of heterogeneous devices, resulting in synchronous rise of data acquisition cycle and processing delay. The existing method relies on static threshold and offline model for anomaly detection, which cannot capture the dynamic characteristic changes of photovoltaic strings in real time, such as the distortion of current-voltage characteristic curve caused by local shadow shielding or the drift of equivalent series resistance caused by component aging. At the same time, in high-concurrency scenarios, massive data streams in the transmission and storage links are prone to packet loss and timestamp misalignment, making the topology correlation analysis invalid, and thus affecting the accuracy of abnormal positioning.

[0021] For example, in a distributed photovoltaic power station containing thousands of strings, multiple edge nodes collect operating parameters at a frequency of seconds and upload them to the cloud. Due to the lack of dynamic modeling mechanism, only the fixed threshold can be used to judge the string state, and the slow decay of photogenerated current caused by microcracks cannot be identified. When multiple strings under a certain inverter have different orientations, resulting in uneven irradiance distribution, the existing method cannot distinguish between normal power fluctuations and string mismatch, and mistakenly determines the power drop caused by irradiance fluctuations as an anomaly, triggering redundant alarms. In addition, during the midday irradiance peak period, the data concurrency increases, causing the sampling timestamp of some nodes to deviate by more than 100 milliseconds, the node attributes of the topology graph cannot be accurately aligned, and the key inflection point in the virtual current-voltage characteristic curve reconstruction process is shifted, causing the maximum power point tracking control instruction to mismatch with the real working condition.

[0022] If the above problems are not solved, the operation and maintenance cost of the power station will increase with the frequency of invalid alarms, and the missed detection rate of abnormal strings will continue to accumulate, which may cause hot spot effect, accelerate component aging, and even cause fire hazards. The deviation of power control instructions from real working conditions will cause the inverter output to frequently exceed the limit, triggering protective disconnection, and reducing the power generation income of the power station. In the long run, the lack of accurate degradation state perception will affect life prediction and preventive maintenance planning, further increasing the total life cycle operation and maintenance cost. In addition, the failure of topology correlation analysis will weaken the abnormal positioning capability, delay the fault handling opportunity, increase the power fluctuation risk of the power grid, and threaten the transient stability of the regional power system.

[0023] For this, see Figure 1 As shown in the application, a cloud platform-based distributed photovoltaic power station monitoring and management method is proposed, comprising: S100: Collecting operation parameters of photovoltaic strings at the edge node of the power station.

[0024] S200: Applying a bounded perturbation to the DC operating point within the stable interval of the maximum power point tracking, obtaining the voltage and current variation, and calculating the conductance and estimated equivalent electrical parameters to obtain the feature vector.

[0025] S300: Based on the electrical connection relationship of photovoltaic strings, combiner boxes and inverters, a topology graph is constructed, the feature vector is mounted as a node attribute, and a model with physical constraints is called to reconstruct the virtual current-voltage characteristic curve and key feature points.

[0026] S400: On the topology graph, based on the preset neighborhood selection condition and weight calculation rule, the feature vectors of neighbor nodes that meet the comparable conditions in the same combiner box, the same inverter or the same azimuth cluster are weighted and aggregated to obtain node-level health indicators, and combined with the reference distribution constructed based on historical normal operation samples to generate abnormal scores.

[0027] S500: When the abnormal score meets the preset condition, active sparse sampling is performed, and the virtual current-voltage characteristic curve and node-level health indicators are updated.

[0028] S600: According to 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 belonging nodes are located, and under the parameter constraints that meet the grid-connected specification, the optimization problem of power generation benefit, degradation risk and power fluctuation is solved to obtain the active power limit and power factor control trajectory, which is issued to the edge node for execution.

[0029] Specifically, in step S100, the edge node collects the operation parameters of the photovoltaic strings, which means deploying a data collection module at the local device end of the photovoltaic power station. Specifically, it can be realized by using embedded sensors, smart meters or inverter built-in measurement units to realize real-time acquisition of voltage, current, power of photovoltaic strings, active power of inverters, power factor of inverters, and environmental irradiance and component temperature, etc. Raw operation data, and complete time stamp alignment and preliminary quality control locally to reduce cloud data transmission pressure and improve real-time monitoring and data reliability.

[0030] In step S200, the bounded perturbation is applied to the DC operating point in the stable interval of the maximum power point tracking, which means that the DC operating point is changed by a controllable perturbation signal with limited amplitude and duration in a safe range near the maximum power point. Specifically, the DC side voltage can be finely adjusted by pulse width modulation, or a small current perturbation can be given by adjusting the DC current of the inverter to realize short-time deviation of the operating point. The voltage and current are synchronously collected before and after the perturbation, the conductance is calculated according to the voltage and current changes, and the equivalent series resistance, equivalent parallel resistance, photo-generated current, reverse saturation current and ideal factor and other equivalent electrical parameters are estimated online according to the physical laws such as the single-diode equivalent circuit model, so as to form a feature vector that can represent the internal state of the photovoltaic string, and more detailed dynamic electrical information is obtained without performing integral curve scanning.

[0031] In step S300, the topology graph is constructed and the feature vector is mounted as a node attribute, which means that the graph structure is established according to the physical wiring relationship among the photovoltaic strings, the combiner box and the inverter. Specifically, the connection relationship between each string and the combiner box, and the combiner box and the inverter can be obtained by analyzing the one-time wiring diagram, and a topology graph containing photovoltaic string nodes, combiner box nodes and inverter nodes is constructed. The edge weight can be determined according to factors such as cable length, parallel loop ownership and orientation consistency, which is used to reflect the electrical coupling strength and spatial correlation. The normalized and time-aligned feature vector is mounted to the corresponding photovoltaic string node, which is used to associate the electrical parameters with the device topology, and support subsequent statistical analysis and diagnostic calculation based on the topology structure. The model reconstruction of the virtual current-voltage characteristic curve and the key feature points with physical constraints means that a parameterized model is constructed based on physical laws such as the single-diode equivalent circuit model, and the virtual current-voltage characteristic curve of each photovoltaic string is numerically solved or fitted by combining voltage, current, equivalent electrical parameters and environmental irradiation, component temperature and other information. The open-circuit voltage, short-circuit current and maximum power point voltage and current and other key feature points are outputted, so as to ensure that the reconstructed curve meets the basic physical characteristics of the photovoltaic module, and avoid obvious distortion of the pure data-driven model under extrapolation or noise interference.

[0032] In step S400, generating the node-level health index on the topology graph means that, based on the preset neighborhood selection condition and weight calculation rule, the characteristics of the neighbor nodes comparable to the target photovoltaic string are aggregated on the constructed topology graph. Specifically, in the same junction box, the same inverter, or the same azimuth cluster, the neighbor nodes with an environmental irradiation difference of not more than a first threshold value, a component temperature difference of not more than a second threshold value, an azimuth angle difference of not more than a third threshold value, and a sampling time alignment error of not more than a preset time threshold value are selected as comparable neighbors. The nodes that do not meet the comparable condition or are in an offline state are assigned a zero weight and removed. Then, according to the azimuth consistency, irradiation similarity, historical operation correlation, and electrical connection strength, the aggregation weight of each neighbor node is calculated and normalized, and one or more layers of weighted aggregation are performed on the neighbor node feature vector to obtain a node-level health index that takes into account local features and neighborhood reference information. When generating the anomaly score, the node-level health index of the node whose running state meets the normal criterion within a certain time window is used as a training sample to calculate the normal reference center and its variance or covariance index. Based on the weighted distance between the current node-level health index and the normal reference center and the dispersion index of the node-level health index within the current time window, a comprehensive function is constructed to output the anomaly score, so as to reflect the degree of deviation of the string state from the normal operation mode without relying on a complex deep network structure.

[0033] In step S500, actively sampling and updating the virtual current-voltage characteristic curve and the node-level health index means that when the anomaly score and the dispersion of a certain photovoltaic string indicate that it has a high abnormal risk or great uncertainty, the measurement of the string is preferentially enhanced. Specifically, the anomaly score and the node-level health index dispersion can be weighted to form a comprehensive priority, and a small number of target photovoltaic strings with high comprehensive priority are selected in the entire power station. The daily coverage ratio is limited to not more than a preset value, and the triggering frequency of the same photovoltaic string is limited to not more than a preset upper limit. In the stable interval of the maximum power point tracking, a short-window bounded perturbation of not more than a preset number of times is added to the target photovoltaic string, and the perturbation amplitude and duration still satisfy the safety constraint. The data before and after the enhanced measurement are used to recalculate the derivative quantity and the equivalent electrical parameter, update the feature vector, and reconstruct the virtual current-voltage characteristic curve. The node-level health index obtained based on the enhanced measurement is fused with the historical node-level health index in a sliding weighted manner, so that the node-level health index gradually converges over time and remains sensitive to the latest state.

[0034] In step S600, the shape deviation and the distribution of the node-level health index on the topology graph locate the abnormal photovoltaic string and its belonging node, which means that the open-circuit voltage offset ratio, short-circuit current offset ratio, maximum power point position offset ratio, and segment slope change degree are extracted based on the virtual current-voltage characteristic curve, and the shape deviation indexes are combined with the weighted distance of the node-level health index of the corresponding photovoltaic string node relative to the normal reference center to obtain abnormal intensity; the connectivity constraint and the orientation consistency constraint are applied in the topology graph, the abnormal intensity is aggregated in the neighborhood, and the abnormal cluster is extracted, so that the photovoltaic string with greater abnormal intensity and duration satisfying the shortest duration threshold in the abnormal cluster is determined as an abnormal photovoltaic string, and the junction box node and the inverter node where the abnormal photovoltaic string is located are output as the belonging node, realizing abnormal tracing from the string layer to the device layer. On this basis, a rolling optimization model is established under the constraint of grid-connected specification parameters such as voltage, current, ramp rate, and power factor, the maximum power generation income is taken as the target, and a degradation risk penalty term composed of abnormal intensity and node-level health index dispersion and a power fluctuation penalty term composed of active power change measure are introduced to form a comprehensive objective function, so that the active power limit and power factor control trajectory of each control period are obtained, and the optimization results are projected into the feasible region satisfying the grid-connected specification and then issued to the edge node for execution, so as to improve the power generation income under the premise of ensuring the safety of the power grid and the service life of the equipment.

[0035] The application fuses the topology graph structure, the physical constraint model, and the graph attention mechanism to construct a dynamic closed-loop monitoring optimization system, and realizes full-link cooperation from data acquisition, state evaluation to control decision. Specifically, the dynamic electrical parameters are obtained by bounded perturbation, which are mounted on the photovoltaic string node with environmental information to form a feature vector, the virtual current-voltage characteristic curve and key feature points are reconstructed in the topology graph combined with the physical constraint, and the node-level health index and abnormal score are generated by comparing the neighborhood weighted aggregation with the normal reference distribution; further, the active sparse sampling mechanism driven by abnormality is used to continuously update the key node data under the premise of controllable cost, and finally the control instruction satisfying the grid-connected specification is generated based on multi-objective optimization, so as to solve the problems of rigid data processing, inaccurate abnormal positioning, and single control strategy in the prior art.

[0036] The working process and principle of the present application are as follows: the photovoltaic string operation parameters are collected at the power station edge node, a bounded perturbation is applied to the direct current operating point in the maximum power point tracking stable interval, the voltage and current variation is obtained, the conductance and equivalent electrical parameters are calculated, and the characteristic vector is obtained; a topological graph is constructed based on the electrical connection relationship, the characteristic vector is mounted as a node attribute, a model with physical constraints is called to reconstruct the virtual current-voltage characteristic curve and key feature points; the comparable neighbor nodes are selected on the topological graph according to the preset neighborhood selection condition, and the neighbor node characteristic vectors are weighted and aggregated according to the rules of orientation consistency, irradiation similarity, historical correlation and electrical connection strength, to obtain the node-level health index, the reference distribution is constructed combined with the historical normal operation sample, and the abnormal score is output; when the abnormal score meets the preset condition, active sparse sampling is performed, and the virtual current-voltage characteristic curve and the node-level health index are updated; the abnormal string and the belonging node are located according to the shape deviation of the virtual current-voltage characteristic curve and the distribution of the node-level health index on the topological graph, and the optimization problem is solved under the constraint of grid connection specification to obtain the control trajectory for execution.

[0037] The scheme captures the dynamic characteristics of the string through bounded perturbation and equivalent parameter estimation, enhances the abnormal detection capability by using the topological graph and the neighborhood weighted statistical analysis, and actively updates the data of high-risk nodes under limited resources through the active sparse sampling mechanism. The shape deviation of the virtual current-voltage characteristic curve and the node-level health index are combined to improve the abnormal positioning accuracy. On this basis, the multi-objective optimization is used to balance the power generation income, degradation risk and power fluctuation, so as to realize the intelligent monitoring and management of the distributed photovoltaic power station.

[0038] As a preferred embodiment, the scheme of the present application is implemented as follows: The photovoltaic string voltage, current, power and other operation parameters are collected at the power station edge node. In the maximum power point tracking stable interval, a small bidirectional perturbation with an amplitude of not more than 0.5% and a duration of not more than 200ms is applied to the direct current operating point. The voltage and current variation is calculated through the data collected before and after the perturbation, the equivalent series resistance, parallel resistance, photo-generated current and other equivalent electrical parameters are estimated, and the characteristic vector is constructed. Based on the primary wiring data, a topological graph containing photovoltaic strings, combiner boxes and inverter nodes is generated, the normalized characteristic vector is mounted to the corresponding photovoltaic string node, and a model with physical constraints is called to reconstruct the virtual current-voltage characteristic curve and key feature points.

[0039] On the topology graph, the neighbor nodes meeting the comparable condition are selected according to the restriction condition of the same junction box, the same inverter or the same azimuth cluster, and the nodes not meeting the comparable condition or offline nodes are assigned zero weight and removed; the aggregation weight of each neighbor node is determined according to the azimuth consistency, irradiation similarity, historical correlation and electrical connection strength and is normalized, at least two layers of weighted aggregation are performed on the neighbor node feature vectors, and the node-level health index is obtained. In a preset time window, the node-level health index of the node in a normal operating state is selected to calculate the normal reference center and variance or covariance index, and the weighted distance between the current node-level health index and the normal reference center and the dispersion index thereof in the time window are weighted and synthesized to obtain an abnormal score.

[0040] When the abnormal score exceeds a threshold value, active sparse sampling is performed. The target group string is selected according to the priority synthesized by the abnormal score and uncertainty, and the daily coverage ratio is not more than 5%. The target group string is subjected to short-window bounded perturbation enhancement measurement not more than 5 times, and the feature vector, virtual curve and node-level health index are updated.

[0041] The shape deviation index of the virtual current-voltage characteristic curve is calculated, and the abnormal strength is synthesized with the distance of the node-level health index from the normal reference center, the abnormal strength is aggregated on the topology graph under the constraints of connectivity and azimuth consistency, and the abnormal cluster is extracted, and the abnormal photovoltaic group string and the belonging junction box node and inverter node are located.

[0042] Under the constraint of grid connection specification, the optimization problem is solved by taking the maximum power generation benefit as the target and adding a degradation risk and power fluctuation penalty term. The active power limit and power factor control trajectory are obtained and issued to the edge node for execution.

[0043] Through the above scheme, the photovoltaic group string dynamic characteristic change can be captured in real time, and the abnormal detection accuracy is improved. The topology correlation analysis is used to enhance the data consistency, reduce the processing delay and packet loss risk in a high concurrency scenario. The active sparse sampling mechanism reduces invalid alarms and optimizes the control strategy to balance the power generation efficiency and stability. Therefore, the operation reliability and economy of the distributed photovoltaic power station are improved.

[0044] The application further proposes that when the power station edge node collects the operating parameters of the photovoltaic string, the operating parameters include voltage, current, power, active power of the inverter, power factor, environmental irradiance, and component temperature. In the stable interval of maximum power point tracking, the voltage of the photovoltaic string, the current of the photovoltaic string, the power of the photovoltaic string, the active power of the inverter, the power factor of the inverter, the environmental irradiance, and the component temperature are synchronously sampled at a sampling period not higher than 1S and are given a uniform timestamp. When the stable interval criterion is not satisfied, the sampling is suspended until the stable interval is resumed. The synchronously sampled data is sequentially subjected to amplitude limiting denoising, median filtering, and sliding window drift removal processing and completes unit uniformity and range calibration, wherein the environmental irradiance is subjected to effective irradiance self-calibration through the alternating current side metering power combined with the inverter efficiency characteristic. When a parameter is missing or sampling is abnormal, only short window interpolation is performed in the stable interval and the stable interval boundary is not crossed.

[0045] The operating parameters include voltage, current, power, active power of the inverter, power factor, environmental irradiance, and component temperature, covering multi-dimensional data of electricity, environment, and equipment state. The sampling period is set to be not higher than 1S, dynamically adapts to the stable interval of maximum power point tracking, and ensures the real-time and stability of data collection. Synchronous sampling realizes time alignment of multiple parameters through a uniform timestamp, avoiding asynchronous errors. Amplitude limiting denoising filters out abnormal jumps by setting a reasonable range of parameters, median filtering eliminates impulse noise, and sliding window drift removal suppresses slow interference. Environmental irradiance self-calibration uses the alternating current side power and the inverter efficiency curve to back-calculate the effective irradiance value, eliminating sensor bias. Short window interpolation is only performed in the stable interval, avoiding the introduction of non-physical data by cross-interval interpolation.

[0046] Specifically, in the stable interval of maximum power point tracking, voltage, current, power, inverter parameters, and environmental data are synchronously collected at a fixed time interval, ensuring that the timestamps of various parameters are consistent. When it is detected that the stable interval condition is not satisfied, the sampling is immediately suspended to prevent non-steady-state data from being mixed in. The collected data is sequentially subjected to amplitude limiting processing to remove abnormal values exceeding the physical range. Median filtering is used to eliminate transient noise interference. The mean or trend item is calculated through a sliding window to eliminate data drift. The environmental irradiance value is dynamically calibrated through the inverter output power and the efficiency characteristic curve to improve the accuracy of the irradiance data. If some parameters are missing or sampling is abnormal, only the effective data adjacent in time in the current stable interval is used for interpolation to complete the data, and the interpolation window does not exceed the stable interval boundary, ensuring data continuity while avoiding the introduction of non-steady-state errors.

[0047] As a preferred embodiment, the scheme of the application is implemented as follows: The operation parameters of a photovoltaic string are collected at an edge node of a power station, including voltage, current, power, active power of an inverter, power factor, environmental irradiance, and component temperature. In a stable interval of maximum power point tracking, these parameters are synchronously sampled at a sampling period of 0.5 S and assigned a uniform timestamp. Sampling is suspended when the stable interval criterion is not met until the stable interval is restored. The synchronously sampled data is sequentially subjected to amplitude limiting denoising, median filtering, and sliding window drift removal processing, and uniformity and range calibration is completed. The environmental irradiance is self-calibrated by effective irradiance through the alternating current side metering power combined with the inverter efficiency characteristics. When parameter loss or sampling anomaly is detected, only short window interpolation of 10 seconds is performed within the stable interval and the stable interval boundary is not crossed.

[0048] Specifically, the amplitude limiting denoising adopts a 3σ criterion to remove abnormal values. The median filtering uses a 5-point median filter. The sliding window drift removal adopts a 60-second window moving average. The range calibration is performed by a calibration coefficient calibrated periodically. The effective irradiance self-calibration is obtained by dividing the inverter alternating current side power by the inverter efficiency curve interpolation. The short window interpolation adopts a linear interpolation method.

[0049] Through the above technical solutions, the present application realizes high-quality collection of photovoltaic string operation parameters. Thereby, the accuracy and reliability of subsequent data analysis are improved. Further, through stable interval judgment, synchronous sampling, data preprocessing, and self-calibration, the influence of sampling noise and abnormal data is effectively reduced. For example, the self-calibration of environmental irradiance avoids errors caused by irradiance sensor failure. The short window interpolation avoids unreasonable interpolation across the stable interval while ensuring data continuity. These measures together ensure the time consistency, numerical accuracy, and physical reasonableness of the collected data.

[0050] The present application further proposes that a bounded perturbation is applied to the direct current operating point within the stable interval of maximum power point tracking, and when obtaining the feature vector, including: the bounded perturbation is a small bidirectional disturbance on the direct current operating point voltage or direct current operating point current, the amplitude does not exceed 0.5% of the rated direct current operating point, the single duration is not more than 200 ms, and the time interval of adjacent two disturbances is not less than 1 s. The voltage and current are synchronously collected before and after the disturbance with a uniform timestamp, the voltage variation and current variation are obtained by using adjacent time window difference regression, and online estimation is implemented under the physical constraint of the single diode equivalent model, to obtain the equivalent series resistance, the equivalent parallel resistance, the photo-generated current, the reverse saturation current, and the ideal factor. The operation parameters, the variation rates, and the equivalent electrical parameters are combined into a feature vector. The bounded perturbation is only enabled within the stable interval and automatically terminated when exceeding the stable interval.

[0051] Wherein, the bidirectional perturbation is achieved by alternatingly applying positive and negative voltage or current offsets, and the perturbation amplitude is limited to 0.5% of the rated value to avoid exceeding the stable interval boundary. The differential regression adopts the least square method to linearly fit the voltage and current sequences before and after the perturbation, and extracts the slope as the change. The single-diode equivalent model describes the physical characteristics of the photovoltaic module through a nonlinear equation set, and the online estimation process solves the equation set through the Newton iteration method, with the constraints including the non-negative photo-generated current, the positive definite series resistance and parallel resistance. The feature vector is constructed by splicing the original operating parameters, differential regression results and equivalent parameters in the standardized format to form a multi-dimensional data matrix.

[0052] Specifically, in the stable operating state, a positive voltage perturbation of 0.3V and a negative voltage perturbation of 0.3V are alternately applied, and each perturbation lasts for 150ms before returning to the original state, and the interval between the two perturbations is set to 1.2s. The voltage and current data are synchronously collected during the perturbation, and the differential calculation is performed on the data before and after the perturbation using a sliding window with a length of 50ms to obtain the change of ΔV=0.6V and ΔI=0.02A. Based on the single-diode model, an equation set containing five unknown parameters is established, and through online iteration calculation, the equivalent series resistance is 0.25Ω, the equivalent parallel resistance is 125Ω, and the photo-generated current is 5.2A. The original voltage value 28.5V, the current value 4.8A and the equivalent parameters are jointly coded as a 16-dimensional feature vector. When it is detected that the sudden change of environmental irradiation causes the stable interval to fail, the perturbation sequence is immediately terminated and the direct current operating point is kept stable.

[0053] As a preferred embodiment, the scheme of the application is implemented as follows: A bounded perturbation is applied to the direct current operating point within the stable interval of the maximum power point tracking. Specifically, the bounded perturbation is a small bidirectional perturbation on the direct current operating point voltage or direct current operating point current. The perturbation amplitude does not exceed 0.5% of the rated direct current operating point, the single duration does not exceed 200ms, and the time interval between the adjacent two perturbations is not less than 1s.

[0054] Further, the voltage and current are synchronously collected before and after the perturbation with a unified timestamp. The voltage change and current change are obtained by differential regression of adjacent time windows. The online estimation is implemented under the physical constraints of the single-diode equivalent model to obtain the equivalent series resistance, the equivalent parallel resistance, the photo-generated current, the reverse saturation current and the ideality factor.

[0055] Thus, the operating parameters, the change rates and the equivalent electrical parameters are combined into a feature vector. Among them, the bounded perturbation is only enabled within the stable interval, and is automatically terminated when it exceeds the stable interval.

[0056] For example, in practical applications, a perturbation of ±0.3% can be applied to the DC operating point voltage for a duration of 150 ms, with an interval of 1.5 s between adjacent perturbations. The voltage and current data are synchronously collected with a sampling period of 10 ms. Through differential regression with a time window of 500 ms, the voltage and current variations are obtained. Based on the single-diode model, the equivalent electrical parameters are estimated using the least squares method. Finally, a 13-dimensional feature vector containing voltage, current, power, variation rate, and 5 equivalent electrical parameters is generated.

[0057] Through the above technical solutions, the present application realizes dynamic capture and accurate quantification of the electrical characteristics of photovoltaic strings. Through bounded perturbation and synchronous sampling, high-precision measurement data of voltage and current response are obtained. Based on online parameter estimation constrained by physical models, key features reflecting the real-time state of the string are extracted. This method avoids the limitations of traditional static testing and can timely reflect the performance changes of the string, providing a reliable data foundation for subsequent anomaly detection and diagnosis. At the same time, the amplitude and duration of the perturbation are strictly controlled, minimizing the impact on normal power generation. In addition, the stable interval judgment and automatic termination mechanism further ensure the effectiveness and safety of the measurement.

[0058] The present application further proposes generating a topology graph containing string nodes, combiner box nodes, and inverter nodes based on one-time wiring data, establishing edges according to electrical connections, and determining edge weights based on cable length, parallel loop ownership, and orientation consistency. After normalizing and time-aligning the feature vectors, they are mounted to the corresponding nodes. Taking the weighted aggregated features of node characteristics, environmental irradiance, component temperature, and topology neighbors as input, a representation learning model with physical constraints is called for inference. The physical constraints include that in the short-circuit to low-voltage range, the voltage is not higher than 30% of the full-scale range, and the current drop relative to the short-circuit current is not more than 10%. In the high-voltage range, the voltage is not lower than 70% of the full-scale range, and the current decreases significantly with increasing voltage and the decrease amplitude is higher than that in the low-voltage range. In the maximum power point neighborhood, the power increases with increasing voltage below the point and decreases above the point. When environmental irradiance increases, the short-circuit current increases and the open-circuit voltage does not decrease, and when the component temperature increases, the open-circuit voltage decreases. With the addition of topology consistency constraints, the virtual current-voltage characteristic curve and key feature points, including open-circuit voltage, short-circuit current, and voltage and current of the maximum power point, are output.

[0059] The cable length is obtained by measurement or design drawing, and the loop belongs to the determination according to the primary wiring topology, and the azimuth consistency is calculated through the installation angle and the geographical position of the string. The normalization adopts the maximum and minimum value scaling method, and the time alignment is based on the unified timestamp for interpolation synchronization. The representation learning model of the physical constraint adopts a neural network structure, and the loss function of the model is embedded with the physical relationship constraint of short-circuit current and open-circuit voltage, and the voltage-current segment slope condition is forced to be met during back propagation. The topology consistency constraint is realized through the graph regularization term, and the change trend of the characteristic curves of the adjacent nodes is ensured to be consistent.

[0060] Specifically, the primary wiring data is parsed into a graph structure containing string nodes, combiner box nodes and inverter nodes, the edge weight is inversely proportional to the cable length, and the loop belongs to the same node between the edge weight increases the correction coefficient, and the edge weight between the nodes with an azimuth angle difference less than 15 degrees is increased in weight. After the feature vector is normalized to eliminate the dimensional difference, the data synchronization is ensured through timestamp alignment. During the training of the model with physical constraints, the output current in the short-circuit to low-voltage range is limited to not less than 90% of the rated value, the lower limit of the voltage in the high-voltage range is set to 70% of the full range, and the positive correlation between irradiation and short-circuit current and the negative correlation between temperature and open-circuit voltage are introduced as soft constraints. The topology consistency constraint is realized by calculating the similarity loss of the characteristic curves of adjacent nodes to suppress local abnormal fluctuations. The error of the virtual curve output by the model inference at the key points is controlled to be within 2% of the measured data, and the estimated values of the open-circuit voltage and the short-circuit current are calibrated through closed-loop feedback, and finally the characteristic curve conforming to the physical law and the topology relationship is generated.

[0061] As a preferred embodiment, the scheme of the present application is implemented as follows: A topology graph containing string nodes, combiner box nodes and inverter nodes is generated based on primary wiring data. Edges are established according to electrical connections, and edge weights are determined based on cable length, loop belonging and azimuth consistency. The feature vector is normalized and time-aligned before being mounted to the corresponding node.

[0062] The weighted aggregation features of node characteristics, environmental irradiation, component temperature and topology neighbors are taken as inputs, and a representation learning model with physical constraints is called for inference. The physical constraints include: in the short-circuit to low-voltage range, the voltage is not higher than 30% of the full range, and the decline rate of the current relative to the short-circuit current is not more than 10%. In the high-voltage range, the voltage is not less than 70% of the full range, and the current decreases obviously with the increase of the voltage and the decrease amplitude is higher than that in the low-voltage range. In the maximum power point neighborhood, the power increases with the increase of the voltage below the point and decreases with the increase of the voltage above the point. The short-circuit current increases and the open-circuit voltage does not decrease with the increase of the environmental irradiation, and the open-circuit voltage decreases with the increase of the component temperature.

[0063] The topology consistency constraint is added, and the virtual current-voltage characteristic curve and key feature points are output. The key feature points include open circuit voltage, short circuit current, and voltage and current of the maximum power point.

[0064] Specifically, first, a topology graph is constructed based on the primary wiring diagram of the power station, and the strings, combiner boxes, and inverters are taken as different types of nodes. According to the actual electrical connection relationship, edges are established between the nodes, and the edges are given weights. The edge weights are determined by factors such as cable length, parallel loop ownership, and orientation consistency. Then the collected feature vectors are normalized and aligned according to the unified timestamp before being mounted on the corresponding nodes.

[0065] Next, taking the node features, environmental irradiation, and component temperature as inputs, while considering the weighted aggregated features of the neighboring nodes in the topology, a pre-trained representation learning model with physical constraints is called to perform inference. The model needs to meet a series of physical constraint conditions during training and inference to ensure that the output results conform to the basic electrical characteristics of photovoltaic components.

[0066] Finally, based on the model output, topology consistency constraints are added 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.

[0067] Through the above technical solutions, the application realizes the reconstruction of photovoltaic string characteristic curves based on a topology graph. Thus, the complete current-voltage characteristic curve of the string can be obtained without affecting normal power generation, providing an important basis for subsequent anomaly detection and performance evaluation. At the same time, by introducing physical constraints and topology consistency constraints, the accuracy and reliability of the reconstruction results are improved. In addition, this method can adapt to photovoltaic power stations of different sizes and topologies, and has strong universality and expandability.

[0068] The application further proposes that when generating node-level health indicators on the topology graph, the following steps are included: based on a preset neighborhood selection condition, select neighbor nodes that meet the comparable condition within the same combiner box, the same inverter, or the same orientation cluster, the comparable condition including an environmental irradiation difference of no more than fifteen percentage points, a component temperature difference of no more than three degrees Celsius, an orientation angle difference of no more than fifteen degrees, and a sampling time alignment error of no more than one millisecond; give zero weight to neighbor nodes that do not meet the comparable condition or are in an offline state and exclude them from the aggregation calculation; determine the aggregation weights of each neighbor node according to factors such as orientation consistency, irradiation similarity, historical correlation, and electrical connection strength, and normalize them, and at least two layers of weighted aggregation are performed on the feature vectors of the neighbor nodes that meet the comparable condition to obtain the node-level health indicators of the corresponding photovoltaic string nodes.

[0069] The comparable condition is set by setting thresholds for the environmental irradiance difference, the component temperature difference, the azimuth angle difference, and the sampling time alignment error, the neighbor nodes with similar environmental conditions and time synchronization are screened, and the nodes with large environmental differences are avoided to be included in the comparison range. The nodes that do not meet the comparable condition or are in an offline state are assigned a zero weight and are removed, so as to prevent invalid data or missing data from interfering with the aggregation result. The aggregation weight comprehensively considers factors such as azimuth consistency (the closer the azimuth angle, 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 the sum of the weights of each neighbor node is ensured to be one through normalization. In multi-layer aggregation, the first layer aggregation is used to collect 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 in the same azimuth cluster, so that the node-level health index takes into account both local features and larger range topological relations.

[0070] As a preferred embodiment, the scheme of the application is implemented as follows: As a preferred embodiment, the scheme of the application is implemented as follows: in the topology graph, for each photovoltaic string node, first, the candidate neighbor set is determined according to the primary wiring relationship and the component installation azimuth, 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 azimuth cluster; then, based on the collected environmental irradiance, component temperature, azimuth angle, and sampling time stamp, nodes with an environmental irradiance difference of not more than 15%, a component temperature difference of not more than 3℃, an azimuth angle difference of not more than 15°, and a sampling time alignment error of not more than 1ms are selected as effective neighbors, and nodes that do not meet the above conditions or are in an offline state are set to have a weight of zero and are removed. On this basis, the aggregation weight of each effective neighbor node is calculated, the azimuth angle cosine similarity can be used to measure the azimuth consistency, the inverse of the relative difference of the effective irradiance can be used to measure the irradiance similarity, the correlation coefficient of the power generation or active power in the past preset time window can be used to measure the historical correlation, and the function of the line impedance or the number of parallel branches can be used to measure the electrical connection strength, then the above factors are weighted and summed and normalized to obtain the final weight for feature aggregation. Subsequently, the neighbor node feature vector is subjected to first layer weighted summation to obtain intermediate aggregation features, and then the intermediate aggregation features and the effective neighbor features in a larger neighborhood are subjected to second layer weighted aggregation, and finally, the second-order aggregation features as the node-level health index are obtained, which are used to represent the health status of the corresponding photovoltaic string under the current working condition.

[0071] The application further proposes, when generating the anomaly score, comprising: collecting node-level health indicators of each photovoltaic string node within a preset time window, using node-level health indicator samples meeting preset normal criteria in the running state to calculate normal reference center and corresponding variance or covariance indicators; taking the weighted distance between the current node-level health indicator and the normal reference center as the first component, taking the dispersion indicator of the node-level health indicator within the current time window as the second component, and weighting and synthesizing the first component and the second component to obtain the anomaly score of the photovoltaic string node.

[0072] The normal reference center can be obtained by taking a weighted average of the node-level health indicators of the historical normal samples, and the variance or covariance indicators are used to characterize the natural fluctuation range of the normal samples in each dimension. The weighted distance can adopt Euclidean distance, Mahalanobis distance or other distance measurement forms considering correlation, and is used to reflect the deviation degree of the current node-level health indicator from the normal reference center; the dispersion indicator can be obtained by statistics of the variance or range of the node-level health indicators on the time axis within the current time window, and is used to reflect the fluctuation degree of the current node state in the short term. By weighting and synthesizing the deviation degree and the fluctuation degree, the anomaly score not only considers the deviation of the current state from the long-term normal mode, but also considers the instability of the recent state, so that the abnormal identification result is more robust.

[0073] As a preferred embodiment, the scheme of the application is implemented as follows: in the normal operation stage, one or more photovoltaic string nodes meeting the conditions that the voltage, current, power and power factor are in the normal range and no protection action occurs within a preset time window are selected, the corresponding node-level health indicators are taken as a normal sample set, the mean vector in each dimension is calculated as the normal reference center, and the covariance matrix is calculated as the description of the normal fluctuation range. In real-time operation, for each photovoltaic string node, the Mahalanobis distance between the current node-level health indicator and the normal reference center is obtained as the first component; at the same time, within the latest time window of a preset length, the variance or comprehensive dispersion of the node-level health indicator in each dimension is calculated as the second component. The first component and the second component are linearly combined according to the preset weight to obtain the anomaly score at the current time. When the anomaly score exceeds the preset threshold, it is considered that the photovoltaic string node has an abnormal risk, and the subsequent active sparse sampling and enhanced measurement process is triggered.

[0074] By the technical solution, the application fully utilizes the topological structure information and multi-dimensional similarity between nodes without relying on complex deep learning structures such as graph neural networks and contrast learning, generates a node-level health index with clear physical meaning through comparable condition filtering and multi-layer weighted aggregation, and constructs an anomaly score mechanism based on a normal reference center and a dispersion index, thereby achieving fine evaluation of the health status of the photovoltaic string. On the one hand, the method improves the accuracy and interpretability of anomaly detection and diagnosis, and on the other hand, avoids the deployment cost and uncertainty brought by complex models, and is more suitable for integration into actual cloud platform monitoring systems.

[0075] The application further proposes active sparse sampling, which includes an integrated priority constituted by anomaly score and dispersion of node-level health index as a selection basis, selects a target photovoltaic string from top to bottom according to the integrated priority in the power station, and the daily coverage ratio is not more than 5%, and the same photovoltaic string is not more than once per hour; the enhanced measurement is performed on the target photovoltaic string in the stable interval of the maximum power point tracking, the enhanced measurement is composed of no more than five short-window bounded perturbations, the single perturbation amplitude is not more than 0.5% of the rated direct-current working point, the single duration is not more than 200 ms, and the total duration is not more than 2 s; after the enhanced measurement is completed, the conductance and equivalent electrical parameters are recalculated according to the data collected before and after the perturbation, the feature vector is updated and the virtual current-voltage characteristic curve is reconstructed, and 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 measurement reasoning, and the weight coefficient is in the range of (0.6-0.9).

[0076] The determination of the integrated priority considers the size of the anomaly score and the dispersion of the node-level health index in the preset time window, realizes the unified quantification of the "abnormal degree" and the "state instability degree", and avoids the sampling resource bias to a few extreme samples caused by relying on a single index. The dispersion of the node-level health index can be obtained by calculating the standard deviation or other dispersion measures of the index at the last several sampling times, and the greater the dispersion, the less stable the current diagnosis result and the higher the potential information value. By limiting the daily coverage ratio and the number of triggers of the same photovoltaic string within a unit time, the active sparse sampling does not have a large impact on normal power generation in a large-scale power station scenario, while ensuring that the sampling load is within the bearing range of the cloud platform and the edge node. The enhanced measurement adopts multiple low-amplitude, short-time bounded perturbations to obtain multiple groups of high-quality perturbation data within a limited time window, which is used to improve the accuracy of conductance calculation and equivalent electrical parameter estimation. The sliding fusion update realizes the smooth transition and adaptive update of the state evaluation result by weighted sum of the new and old node-level health indexes while retaining the continuity of the historical state.

[0077] As a preferred embodiment, the scheme of the application is implemented as follows: during operation, the cloud platform periodically calculates the anomaly score of each photovoltaic string node, and calculates the variance or standard deviation of the node-level health index within a preset length of time window, adds the anomaly score and the dispersion by a preset proportion, and obtains the comprehensive priority of each photovoltaic string. For example, the comprehensive priority can be set in the form of “anomaly score plus dispersion multiplied by weight coefficient”. All photovoltaic strings in the power station are sorted according to the comprehensive priority from high to low, and about 5% of the photovoltaic strings with the highest comprehensive priority are selected as the target set of active sparse sampling every day, and the same photovoltaic string is limited to be selected at most once in any hour.

[0078] For the selected target photovoltaic string, when the maximum power point tracking is in a stable interval, an enhanced measurement process is triggered, and three to five short-window bounded perturbations are applied to the direct-current operating point within an enhanced measurement period. The amplitude of each perturbation is controlled between 0.3% and 0.5% of the rated direct-current operating point, the duration of a single perturbation is 150 to 200 ms, and the total duration is controlled within 1.5 s to 2 s, so as to ensure that the perturbation does not cause significant power fluctuation or deviation beyond the stable interval. Voltage and current data are synchronously collected at a fixed sampling period before and after the enhanced measurement, the voltage variation and current variation are recalculated through difference regression of adjacent time windows, and equivalent electrical parameters such as equivalent series resistance, equivalent parallel resistance, photogenerated current, reverse saturation current and ideal factor are estimated online under the physical constraint of single-diode equivalent model. The updated operating parameters, variation rates and equivalent electrical parameters are combined into a new feature vector, and the model reconstruction with physical constraints is called to reconstruct the virtual current-voltage characteristic curve and key feature points of the corresponding photovoltaic string.

[0079] When updating the node-level health index, the node-level health index recalculated based on the enhanced measurement data is slidingly fused with the node-level health index before the enhanced measurement according to a preset weight. For example, the new node-level health index can be set as the weighted sum of the old node-level health index and the updated node-level health index according to a weight coefficient of 0.6 to 0.9, so that the updated node-level health index not only inherits the continuity of the historical state, but also has high sensitivity to the latest enhanced measurement result. Subsequently, the updated feature vector, virtual current-voltage characteristic curve and node-level health index are stored in the monitoring database in the cloud platform, which are used for subsequent anomaly positioning, trend analysis and linkage with the optimization control module.

[0080] By the technical solution, the application realizes the key enhanced measurement of high-risk or high-uncertainty photovoltaic string under the premise of not increasing the overall sampling burden. Through the active sparse sampling and sliding fusion updating mechanism, on the one hand, the accuracy of the estimated equivalent electrical parameters and the reliability of the virtual current-voltage characteristic curve are improved, and on the other hand, the smooth evolution and rapid response capability of the node-level health index over time are ensured, thereby improving the accuracy of abnormality detection and state assessment, providing more reliable input data for subsequent optimization control, and taking into account the scalability and operation efficiency in the large-scale distributed photovoltaic power station monitoring scenario.

[0081] The application further proposes a scheme for positioning abnormal photovoltaic strings and their belonging nodes, including: calculating a shape deviation index based on the virtual current-voltage characteristic curve, combining the weighted distance of the shape deviation index and the node-level health index relative to the normal reference center into an abnormal intensity, applying connectivity constraints and orientation consistency constraints in the topology graph to aggregate the neighborhood of the abnormal intensity and extract an abnormal cluster, determining the photovoltaic string with the maximum abnormal intensity and the duration satisfying the preset threshold in the abnormal cluster as an abnormal string, and outputting the junction box node and the inverter node to which the abnormal string belongs.

[0082] The shape deviation index can be composed of an open-circuit voltage offset ratio, a short-circuit current offset ratio, a maximum power point position offset ratio, and a segmented slope change degree of a low-voltage region and a high-voltage region, for quantifying the shape deviation of the virtual current-voltage characteristic curve relative to the normal working condition from multiple dimensions. The weighted distance of the node-level health index relative to the normal reference center reflects the degree of deviation of the photovoltaic string from the normal operation mode in the multi-dimensional health feature space. The abnormal intensity linearly or nonlinearly combines the shape deviation index and the weighted distance through a preset weight, avoiding the misjudgment risk caused by relying on a single index. The neighborhood aggregation process introduces connectivity constraints to ensure that the photovoltaic strings in the abnormal cluster are topologically connected to each other, and introduces orientation consistency constraints to ensure that the photovoltaic strings in the abnormal cluster have similar orientations and environmental conditions, thereby reducing the interference of isolated noise points and environmental fluctuations on the positioning results. Finally, through the duration criterion, the string with the continuous abnormal intensity exceeding the threshold in the preset time window is screened, the one with the maximum abnormal intensity is determined as the abnormal string, and the corresponding junction box node and inverter node are output, facilitating targeted troubleshooting and maintenance by operation and maintenance personnel.

[0083] Specifically, a reference virtual current-voltage characteristic curve under normal working conditions can be maintained for each photovoltaic string in the cloud platform, or representative reference feature points can be calculated according to historical normal samples. The open-circuit voltage, short-circuit current, maximum power point voltage and current and other features in the virtual current-voltage characteristic curve at the current time are extracted and compared with the reference values to obtain the open-circuit voltage offset ratio, short-circuit current offset ratio and maximum power point position offset ratio. Further, the virtual current-voltage characteristic curve is segmented into low-voltage and high-voltage ranges, the slope difference of the current curve and the reference curve in each segment is compared, and the weighted sum of the absolute values of the slope difference is calculated as the segment slope change degree, thereby forming the shape deviation index. The abnormal distance of the node-level health index can be measured by Mahalanobis distance, weighted Euclidean distance or other distance measurement methods considering the correlation of each dimension, to depict the deviation degree of the current node-level health index relative to the normal reference center in the multi-dimensional space, and then linearly combined with the shape deviation index according to the preset weight to generate the abnormal strength.

[0084] In the topology graph, a connected subgraph can be selected according to the electrical connection relationship, and then the area where the azimuth angle deviation and the environmental difference do not exceed the preset threshold is selected by combining the component orientation, environmental irradiation and component temperature, etc. The abnormal strength of each photovoltaic string in the area is averaged or weighted summed, and the density clustering, threshold connected domain extraction and other methods are used to identify the node set whose overall level of abnormal strength exceeds the threshold value, which is taken as an abnormal cluster. The photovoltaic strings in each abnormal cluster are sorted in descending order of abnormal strength, and the duration of abnormal strength continuously exceeding the preset threshold is checked on the time axis. The photovoltaic string with the maximum abnormal strength and the duration reaching or exceeding the minimum duration threshold is selected as an abnormal string, and the junction box node and the inverter node to which the abnormal string belongs can be determined by traversing the topology edge.

[0085] Through the above technical solutions, the present application establishes a unified abnormal strength measurement between the physical form information of the virtual current-voltage characteristic curve and the multi-dimensional statistical features of the node-level health index, and further combines the topological connectivity and orientation consistency constraints to aggregate the abnormal signals in space and time dimensions, thereby improving the accuracy of abnormal string positioning. The scheme can distinguish between power fluctuations caused by environmental changes and real abnormalities caused by component aging and electrical faults under complex working conditions and environmental fluctuations, reduce false positives and false negatives, and ensure that the abnormal positioning result is consistent with the actual electrical connection relationship and physical layout, thereby providing reliable support for fine operation and maintenance of distributed photovoltaic power stations.

[0086] Even after the aforementioned abnormal scoring and abnormal string positioning scheme based on the virtual current-voltage characteristic curve and the node-level health index, how to further use these state information for control strategy generation under the premise of meeting the grid connection specification parameter constraints, so as to balance the power generation benefit, equipment degradation risk and grid stability, is still a problem to be solved.

[0087] To this end, the application further proposes a scheme for solving an optimization problem and generating a control trajectory under the premise of meeting the parameter constraints of the grid connection specification, including: the parameter constraints at least include voltage constraints, current constraints, power ramp rate constraints and power factor range constraints; a rolling control period is established, and in each control period, the voltage upper limit and the voltage lower limit constraints, the current upper limit constraints, the power ramp rate constraints and the power factor range constraints are satisfied at the same time; the maximum power generation benefit is taken as the target, and a degradation risk penalty term determined by the abnormal intensity and the uncertainty and a power fluctuation penalty term determined by the active power change measure of adjacent control periods are added to form a comprehensive objective function; the key feature points of the virtual current-voltage characteristic curve, the node-level health index and the environmental prediction data are taken as the optimization input, the active power limit and the power factor control trajectory are solved, and the solving result is projected to the feasible region defined by the grid connection specification and then issued to the edge node for execution.

[0088] The parameter constraints guarantee that the grid-connected operation does not exceed the rated capacity of the electrical equipment by setting the upper and lower limits of the voltage on the photovoltaic string side or the inverter side, and the upper limit of the inverter output current; the power ramp rate constraint is used to limit the change rate of active power between adjacent control periods, so as to avoid the impact of rapid fluctuations of the power station output on the power grid; and the power factor range constraint is used to ensure that the operating power factor of the inverter remains within the range allowed by the grid specification. The rolling control period can be set to several minutes according to the scale of the power station and the scheduling requirements, and the control strategy can be dynamically adjusted according to the changes in environmental conditions and equipment states through periodic rolling updates. The weight coefficient of the degradation risk penalty term can be determined by the abnormal intensity and the uncertainty, wherein the abnormal intensity can be derived from the deviation of the shape deviation index of the virtual current-voltage characteristic curve and the node-level health index from the normal reference center, and the 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, and is used to reflect the uncertainty level of the current state evaluation result; the power fluctuation penalty term can be constructed based on the difference or second-order difference of the active power set value of adjacent control periods, and is used to suppress frequent and large amplitude output adjustments. In the optimization input, the key feature points of the virtual current-voltage characteristic curve can include the open-circuit voltage, the short-circuit current, and the voltage and current of the maximum power point, the node-level health index provides quantitative information of the health state of each photovoltaic string or each topological node, and the environmental prediction data can include the irradiance prediction and component temperature prediction for a period of time in the future. The de-projection process can check whether each control quantity in the optimization result satisfies the constraints of voltage, current, power factor, etc., and when it is found that the active power limit or the power factor exceeds the specification boundary, it is projected to the inside of the feasible region along the constraint boundary direction, to ensure that the final control trajectory completely meets the requirements of the grid connection specification.

[0089] Specifically, in a preferred embodiment, fifteen minutes can be taken as a rolling control period, at the beginning of each rolling control period, the virtual current-voltage characteristic curve key feature points, node-level health indicators and environmental prediction data in the recent period are obtained from the cloud platform, and the abnormal strength information output by the abnormal string positioning module is received. The parameter constraint module sets the upper limit of the voltage to 110% of the rated voltage, the lower limit of the voltage to 90% of the rated voltage, the upper limit of the current to 105% of the rated current, the power ramp rate constraint to no more than 2% of the rated power per minute, and the power factor range constraint to 0.95 lead to 0.95 lag. In the comprehensive objective function, the power generation income term is calculated based on the predicted power generation and the real-time electricity price; the degradation risk penalty term is constructed by multiplying the preset risk coefficient by the degradation risk factor obtained by weighting the abnormal strength and the uncertainty, and the uncertainty can be calculated by the variance, covariance or other dispersion index of the node-level health indicators in the recent several sampling periods; the power fluctuation penalty term can adopt the square of the active power limit difference value or the square of the second-order difference of the adjacent rolling control period, multiplied by the preset penalty coefficient. The open-circuit voltage, short-circuit current, maximum power point voltage and current of the virtual current-voltage characteristic curve, and the corresponding component temperature, environmental irradiation 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, and the active power limit and power factor setting value of each control period are obtained by using solvers such as interior point method and sequential quadratic programming method. After the solution is solved, the feasible region projection is performed on the obtained solution, and when it is detected that the power factor of a control period exceeds the range of 0.95 lead to 0.95 lag, the power factor is projected to the nearest feasible point along the constraint boundary; when it is detected that the voltage, current or power ramp rate has a slight over-limit trend, the corresponding penalty coefficient in the objective function is appropriately increased or the active power limit is reduced for correction, and finally the control trajectory satisfying all constraints is formed. The control trajectory includes the active power limit and the corresponding power factor setting value in each rolling control period, which is transmitted to the edge node through an encrypted communication channel, and is converted into specific control instructions of the inverter active power and reactive power by the local controller of the edge node for execution.

[0090] By the technical solution, on the basis of finely depicting the device state by using the virtual current-voltage characteristic curve and the node-level health index, a rolling optimization control mechanism under multi-parameter constraints is introduced, which can not only ensure that key operating parameters such as voltage, current, power ramp rate and power factor always meet the grid connection specification requirements, but also balance the operation strategy through considering power generation income, degradation risk and power fluctuation in the objective function. The degradation risk penalty term comprehensively depicts the potential aging and failure risk of the component or loop through abnormal intensity and uncertainty, prompting the optimization result to moderately reduce the output of high-risk strings, delaying the deterioration of the device; the power fluctuation penalty term suppresses the large change of the output in adjacent control periods, improving the stability of the power access on the grid side. Thus, the dynamic optimization control of the power station operation state under complex working conditions and large-scale distributed photovoltaic scenes is realized, the power generation income and device life utilization level are improved, and the safety and stability of the grid operation are considered.

[0091] In the above embodiment, the equivalent electrical parameters are obtained by collecting the operating parameters of the photovoltaic string in real time at the edge node and applying a bounded perturbation, and a physically reliable feature vector is formed; the topology graph is constructed, and the virtual current-voltage characteristic curve and the key feature points are reconstructed under physical constraints, realizing dynamic modeling consistent with the real electrical behavior; the node-level health index is obtained by weighted aggregation of the feature vectors of the comparable neighbor nodes on the topology graph, and an abnormal score is generated, improving the identification ability of component mismatch, performance degradation and abnormal operating conditions; when the abnormal score triggers the threshold, enhanced measurement is implemented on the target string with high priority through active sparse sampling, and the virtual current-voltage characteristic curve and the node-level health index are dynamically updated, so that the monitoring model continuously corrects and converges under the condition of controllable disturbance cost; the abnormal photovoltaic string and its belonging node are accurately located on the topology graph, and the power generation income, degradation risk and power fluctuation are uniformly included in the rolling optimization solution under the premise of meeting the grid connection specification parameter constraints, and the active power limit and power factor control trajectory are closed-loop fed back to the edge node for execution, constructing a full-link closed-loop monitoring system of edge collection-physical modeling-topology correlation diagnosis-abnormal driving enhanced measurement-optimization control under grid connection constraints, which improves the diagnosis accuracy, abnormal positioning reliability and grid operation stability of the power station in the large-scale distributed photovoltaic scene, and improves the scalability and intelligent level of the system.

[0092] In another preferred mode based on the above embodiment, referring to Figure 2 The present embodiment provides a cloud platform-based distributed photovoltaic power station monitoring and management system for applying the above cloud platform-based distributed photovoltaic power station monitoring and management method, which comprises: The acquisition unit is configured to collect operating parameters of photovoltaic strings at edge nodes of the power station; The first processing unit is configured to apply a bounded perturbation to the direct current operating point within a stable interval of maximum power point tracking, obtain a voltage variation and a current variation, and calculate a conductance and an estimated equivalent electrical parameter, thereby obtaining a feature vector; The reconstruction unit is configured to construct a topology graph based on an electrical connection relationship of the photovoltaic string, the combiner box, and the inverter, mount the feature vector as a node attribute, and call a model reconstruction virtual current-voltage characteristic curve and key feature points with physical constraints; The evaluation unit is configured to implement weighted aggregation on the feature vectors of neighbor nodes that meet comparable conditions within the same combiner box, the same inverter, or the same azimuth cluster based on a preset neighborhood selection condition and a weight calculation rule on the topology graph, obtain a node-level health index, and generate an abnormal score in combination with a reference distribution constructed based on historical normal operation samples; The judgment unit executes active sparse sampling and updates the virtual current-voltage characteristic curve and the node-level health index when the abnormal score meets a preset condition. The control unit is configured to locate an abnormal photovoltaic string and its belonging node according to a shape deviation of the virtual current-voltage characteristic curve and a distribution of the node-level health index on the topology graph, and solve an optimization problem of power generation benefit, degradation risk, and power fluctuation under a parameter constraint meeting a grid-connected specification to obtain an active power limit and a power factor control trajectory, which are issued to an edge node for execution.

[0093] Specifically, the collection unit synchronously collects multi-dimensional operating parameters at an edge node at a sampling period not higher than one second, and pauses sampling when a stable interval criterion of maximum power point tracking is not met, so as to avoid introducing non-steady-state data into subsequent analysis and ensure effectiveness and comparability of the collected data. The first processing unit obtains dynamic response data of voltage and current before and after a short-time bounded perturbation with a duration not exceeding 200 milliseconds within the stable interval, estimates equivalent series resistance, equivalent parallel resistance, photo-generated current, reverse saturation current, and ideal factor and other equivalent electrical parameters under the physical constraint of a single-diode equivalent model, and combines operating parameters, variation rates, and equivalent electrical parameters to form a feature vector with physical meaning, thereby providing a basis for subsequent virtual current-voltage characteristic curve reconstruction and health evaluation. The reconstruction unit constructs a topology graph containing photovoltaic string nodes, combiner box nodes, and inverter nodes according to a power station primary system diagram, establishes edges according to an electrical connection relationship, determines edge weights by referring to cable length, parallel loop belonging, and azimuth consistency, mounts the normalized and time-aligned feature vector to the corresponding node, and calls a representation learning model with physical constraints to reconstruct a virtual current-voltage characteristic curve and key feature points such as open-circuit voltage, short-circuit current, and maximum power point in compliance with physical laws under constraint conditions of voltage, current, and power relationships in short-circuit regions, low-voltage regions, high-voltage regions, and maximum power point neighborhoods.

[0094] The evaluation unit first selects neighbor nodes within the same junction box, the same inverter or the same azimuth cluster range as comparable neighbor nodes based on preset neighborhood selection conditions, where the environmental irradiance difference, component temperature difference, azimuth angle difference and sampling time alignment error are all within the preset threshold, and assigns zero weight to neighbor nodes that do not meet the comparable conditions or are in an offline state and eliminates them from the aggregation calculation. Then, the aggregation weights of each comparable neighbor node are determined according to factors such as azimuth consistency, irradiance similarity, historical correlation and electrical connection strength, and normalized, and the neighbor node feature vectors are at least two layers of weighted aggregation to obtain a node-level health index that can comprehensively reflect the local state of the device and the topological correlation. The evaluation unit further selects historical node-level health index samples that meet the preset normal criteria to construct a normal reference distribution, calculates the normal reference center and the corresponding variance or covariance index, and synthesizes the weighted distance between the current node-level health index and the normal reference center and the dispersion of the node-level health index within the preset time window to obtain the abnormal score of each photovoltaic string.

[0095] When the evaluation unit detects that the abnormal score of a certain photovoltaic string node reaches or exceeds the preset threshold, it starts the active sparse sampling process: based on the weighted combination of the abnormal score and the node-level health index dispersion, a comprehensive priority is formed, and the target photovoltaic string is selected in the power station according to the comprehensive priority from high to low, with the constraint that the daily coverage ratio does not exceed 5% of the total number of string in the power station and the same photovoltaic string is not 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 the stable interval, the first processing unit implements multiple short-window bounded perturbations within a limited time according to the enhanced measurement strategy, collects voltage and current data before and after the perturbation, recalculates the conductance and equivalent electrical parameters, updates the corresponding feature vector and reconstructs the new virtual current-voltage characteristic curve by the reconstruction unit; at the same time, the judgment unit updates the node-level health index by sliding fusion, and the latest node-level health index obtained based on the enhanced measurement is weighted and summed with the node-level health index before updating according to the preset weight coefficient, so that the node-level health index not only maintains the continuity of the historical state, but also can quickly reflect the latest measurement results.

[0096] The control unit receives the virtual current-voltage characteristic curve and key feature points output by the reconstruction unit, combines the node-level health indicators and abnormal score results output by the evaluation unit, calculates the shape deviation indicators and abnormal intensity of each photovoltaic string on the topology graph, and locuses the abnormal photovoltaic string and the junction box node and inverter node to which it belongs according to the spatial distribution of the shape deviation and the node-level health indicators on the topology graph. On this basis, the control unit establishes an optimization model within a rolling control period according to the upper and lower voltage limits, current upper limit, power ramp rate, and power factor range, etc. grid connection specification parameters, takes the maximum power generation benefit as the target, introduces a degradation risk penalty term constructed by the abnormal intensity and state uncertainty, and a power fluctuation penalty term based on the active power set value change in the adjacent control period to construct a comprehensive objective function, takes the key feature points of the virtual current-voltage characteristic curve, the node-level health indicators, and the environmental prediction data as the optimization input, and solves the active power limit and power factor control trajectory of each rolling control period. After ensuring that the optimization results meet all the grid connection specification constraints through feasible region projection, the active power limit and power factor control trajectory are issued to the local controller of the edge node for power regulation.

[0097] Through the above multi-unit cooperation, the system of the present application completes the complete closed-loop process from high-quality data acquisition, dynamic feature extraction under bounded perturbation, to virtual current-voltage characteristic curve reconstruction with physical constraints, to node-level health indicators based on topology structure and multi-dimensional similarity, active sparse sampling to enhance high-risk string measurement, and rolling optimization control under multi-parameter constraints, improves the abnormal identification accuracy and control strategy flexibility of distributed photovoltaic power stations in large-scale scenarios, reduces the false positive rate and false negative rate, and at the same time takes into account the power generation benefit, equipment degradation risk and power grid operation stability, solves the problems of monitoring lag, rough positioning and control rigidity in the existing cloud monitoring system.

[0098] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application and not to limit them, although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that the specific embodiments of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the present application, any modification or equivalent replacement thereof should be covered within the protection scope of the claims of the present application.

Claims

1. A cloud platform-based distributed photovoltaic power station monitoring management method, characterized in that, The application relates to a photovoltaic power station health monitoring method and system. Collecting operation parameters of photovoltaic strings at edge nodes of a power station; Applying a bounded perturbation to a direct-current working point in a stable interval of maximum power point tracking to obtain a voltage variation and a current variation, calculating a conductance and an estimated equivalent electrical parameter, and thus obtaining a characteristic vector; Based on an electrical connection relationship among photovoltaic strings, a combiner box and an inverter, a topological graph is constructed, the characteristic vector is mounted as a node attribute, and a model with physical constraints is called to reconstruct a virtual current-voltage characteristic curve and key feature points; On the topological graph, based on a preset neighborhood selection condition and a weight calculation rule, the characteristic vectors of neighbor nodes meeting a comparable condition in a same combiner box, a same inverter or a same azimuth cluster are weighted and aggregated to obtain a node-level health index, and an abnormal score is generated in combination with a reference distribution constructed based on historical normal operation samples; When the abnormal score meets a preset condition, active sparse sampling is performed, and the virtual current-voltage characteristic curve and the node-level health index are updated; According to shape deviation of the virtual current-voltage characteristic curve and distribution of the node-level health index on the topological graph, an abnormal photovoltaic string and a node belonging to the abnormal photovoltaic string are located, and an optimization problem of power generation benefit, degradation risk and power fluctuation is solved under the parameter constraint meeting a grid-connected specification to obtain an active power limit and a power factor control trajectory, which are sent to the edge nodes for execution. 2.The cloud platform-based distributed photovoltaic power station monitoring management method according to claim 1, characterized in that, When collecting operation parameters of photovoltaic strings at edge nodes of a power station, the operation parameters include voltage, current, power, active power of an inverter, power factor, environmental irradiance and component temperature. In a stable interval of maximum power point tracking, voltage of a photovoltaic string, current of the photovoltaic string, power of the photovoltaic string, active power of an inverter, power factor of the inverter, environmental irradiance and component temperature are synchronously sampled at a sampling period not higher than 1S and are given a uniform timestamp; sampling is suspended when a stable interval criterion is not met until the stable interval is restored; synchronous sampling data are sequentially subjected to amplitude limiting and denoising, median filtering and sliding window de-drifting processing and are subjected to unit uniformity and range calibration, wherein environmental irradiance is subjected to effective irradiance self-calibration through alternating current side metering power and inverter efficiency characteristics; when parameter loss or sampling abnormality is detected, only short window interpolation is performed in the stable interval and the stable interval boundary is not crossed. When a bounded perturbation is applied to a direct-current working point in a stable interval of maximum power point tracking, the bounded perturbation is a small bidirectional disturbance on the direct-current working point voltage or the direct-current working point current, the amplitude is not more than 0.5% of a rated direct-current working point, the single duration is not more than 200 ms, and the time interval of adjacent two disturbances is not less than 1s. 3.The cloud platform-based distributed photovoltaic power station monitoring management method according to claim 2, characterized in that, ​ ​ The voltage and current are synchronously collected before and after the disturbance with unified time stamp, the voltage variation and current variation are obtained by using adjacent time window difference regression, and online estimation is performed under the physical constraint of single diode equivalent model to obtain equivalent series resistance, equivalent parallel resistance, photo-generated current, reverse saturation current and ideal factor; the operating parameters, variation rates and equivalent electrical parameters are combined into a feature vector; wherein, the bounded disturbance is only enabled within the stable interval and automatically terminated beyond the stable interval.

4. The cloud platform-based distributed photovoltaic power station monitoring management method according to claim 3, characterized in that, When constructing a topology graph and reconstructing a virtual current-voltage characteristic curve, 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 based on cable length, parallel loop ownership and orientation consistency; the feature vector is normalized and time-aligned and then mounted to the corresponding node; The weighted aggregated features of node characteristics, environmental irradiance, component temperature and topology neighbors are taken as inputs, and a representation learning model with physical constraints is called for inference, the physical constraints including that in the short-circuit to low-voltage range, the voltage is not higher than 30% of the full range, and the current relative to the short-circuit current decreases by no more than 10%; in the high-voltage range, the voltage is not lower than 70% of the full range, and the current decreases significantly with the increase of voltage and the decrease amplitude is higher than that in the low-voltage range; in the maximum power point neighborhood, the power increases with the increase of voltage below the point and decreases with the increase of voltage above the point; when the environmental irradiance increases, the short-circuit current increases and the open-circuit voltage does not decrease, and when the component temperature increases, the open-circuit voltage decreases; Topology consistency constraints are added, and the virtual current-voltage characteristic curve and key feature points including open-circuit voltage, short-circuit current and voltage and current of the maximum power point are output.

5. The cloud platform-based distributed photovoltaic power station monitoring management method according to claim 4, characterized in that, When generating node-level health indicators on the topology graph, the following steps are included: Based on the preset neighborhood selection conditions, neighbor nodes that meet the comparable conditions are selected within the same combiner box, the same inverter or the same orientation cluster; the comparable conditions include: environmental irradiance difference not more than 15%, component temperature difference not more than 3℃, orientation angle difference not more than 15° and sampling time alignment error not more than 1ms; nodes that do not meet the comparable conditions or are in offline state are assigned zero weight and removed; the feature vectors of each neighbor node are determined according to the orientation consistency, irradiance similarity, historical correlation and electrical connection strength to determine the aggregation weight and normalize, and at least two layers of weighted aggregation are performed on the neighbor node feature vectors to obtain the node-level health indicators of the corresponding photovoltaic string node. 6.The cloud platform-based distributed photovoltaic power station monitoring management method according to claim 5, characterized in that, When generating an abnormal score, the following steps are included: The node-level health indicators of each photovoltaic string are collected within a preset time window, and the node-level health indicators whose operating state meets the preset normal criterion are used to calculate the normal reference center and the corresponding variance or covariance indicators; the weighted distance between the current node-level health indicator and the normal reference center is taken as the first component, and the dispersion index of the node-level health indicators in the current time window is taken as the second component; the first component and the second component are weighted and combined to obtain the abnormal score of the photovoltaic string node.

7. The cloud platform-based distributed photovoltaic power station monitoring management method according to claim 6, characterized in that, The method comprises the following steps: The active sparse sampling comprises: constructing a comprehensive priority based on the anomaly score and the dispersion of the node-level health index, selecting a target photovoltaic string in the power station according to the comprehensive priority, and the daily coverage ratio is not more than 5%, and the same photovoltaic string is not more than once per hour; In the stable interval of the maximum power point tracking, the target photovoltaic string is subjected to enhanced measurement, and the enhanced measurement is composed of a short window bounded perturbation not more than five times, the amplitude of a single perturbation is not more than 0.5% of the rated DC operating point, the single duration is not more than 200ms, and the total duration is not more than 2s; After the completion of the enhanced measurement, the conductance and the equivalent electrical parameters are recalculated according to the data collected before and after the perturbation, the feature vector is updated, and the node-level health index is reconstructed in a sliding fusion manner, so that the new node-level health index is the weighted sum of the old node-level health index and the node-level health index obtained based on the latest feature reasoning, and the weight coefficient is in the range of (0.6-0.9). 8.The cloud platform based distributed photovoltaic power station monitoring management method according to claim 1, characterized in that, When the abnormal string and its belonging node are located, the method comprises the following steps: A shape deviation index is calculated based on the virtual current-voltage characteristic curve, and the shape deviation index is composed of an open-circuit voltage offset ratio, a short-circuit current offset ratio, a maximum power point position offset ratio, and a segment slope change degree; The shape deviation index and the weighted distance of the node-level health index of the corresponding photovoltaic string node relative to the normal reference center are integrated to obtain an abnormal intensity; The connectivity constraint and the orientation consistency constraint are applied in the topology graph, the abnormal intensity is neighborhood aggregated, and an abnormal cluster is extracted; The string with the maximum abnormal intensity in the abnormal cluster and the duration satisfying the shortest duration threshold is determined as the abnormal string, and the junction box node and the inverter node to which the abnormal string belongs are output. 9.The cloud platform-based distributed photovoltaic power station monitoring management method according to claim 8, characterized in that, When the optimization problem is solved and the control trajectory is generated under the parameter constraint of the grid connection specification, the method comprises the following steps: The parameter constraint includes voltage, current, ramp rate and power factor; A rolling control cycle is established, and in each control cycle, the upper and lower voltage constraints, the current upper limit constraint, the power ramp rate constraint and the power factor range constraint are satisfied at the same time; The maximum power generation benefit is taken as the target, and a degradation risk penalty term determined by the abnormal intensity and the dispersion index of the node-level health index and a power fluctuation penalty term determined by the active power change measure of the adjacent cycle are added to form a comprehensive target; The key feature points of the virtual current-voltage characteristic curve, the node-level health index and the environment prediction are taken as the optimization input, the active power limit and the power factor control trajectory are solved, and the solving result is projected to the feasible region of the grid connection specification and then sent to the edge node, which executes according to the control trajectory.

10. A cloud platform-based distributed photovoltaic power station monitoring management system for applying the cloud platform-based distributed photovoltaic power station monitoring management method according to any one of claims 1-9, characterized in that, The method comprises the following steps: The acquisition unit is configured to acquire the operating parameters of the photovoltaic string at the edge node of the power station; The first processing unit is configured to apply a bounded perturbation to the DC operating point in the stable interval of the maximum power point tracking, obtain the voltage change and the current change, calculate the conductance and the estimated equivalent electrical parameters, and obtain the feature vector; The reconstruction unit is configured to construct a topology graph based on an electrical connection relationship of photovoltaic strings, combiner boxes and inverters, mount the feature vectors as node attributes, and call a model reconstruction virtual current-voltage characteristic curve and key feature points with physical constraints; The evaluation unit is configured to implement weighted aggregation on feature vectors of neighbor nodes meeting comparable conditions in the same combiner box, the same inverter or the same azimuth cluster based on preset neighborhood selection conditions and weight calculation rules on the topology graph, obtain a node-level health index, and generate an abnormal score in combination with a reference distribution constructed based on historical normal operation samples; The judgment unit executes active sparse sampling when the abnormal score meets a preset condition, and updates the virtual current-voltage characteristic curve and the node-level health index; The control unit is configured to locate abnormal photovoltaic strings and their belonging nodes according to a shape deviation of the virtual current-voltage characteristic curve and a distribution of the node-level health index on the topology graph, and solve an optimization problem of power generation income, degradation risk and power fluctuation under the parameter constraints meeting the grid-connected specification to obtain an active power limit and a power factor control trajectory, which are issued to edge nodes for execution.

Images