Health degree evaluation system and method based on photovoltaic string

Abstract

The invention discloses a health degree evaluation system and method based on a photovoltaic string, and belongs to the technical field of photovoltaic operation and maintenance intellectualization, and the method comprises the steps: firstly collecting SCADA operation data of a plurality of strings of a photovoltaic power station, including voltage, current, assembly temperature, environment irradiance and power factors; secondly, constructing a whole-station feature reference model, and generating a state vector fusing environment and load characteristics through a nonlinear projection algorithm; aiming at the target group string, extracting a health reference track, calculating a disturbance propagation factor, and identifying a health abnormal state; when the deviation degree exceeds a threshold value, dynamically screening heterogeneous reference group strings, constructing a health score distribution model, and further calculating a confidence score and a prediction residual error; if the score is low and the residual error is unstable, determining that the string is in a sub-health or hidden fault state, and generating an intervention instruction; the method has high accuracy, strong adaptability and good interpretability, and can be widely applied to intelligent diagnosis and refined operation and maintenance management of the photovoltaic power station.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of photovoltaic operation and maintenance intelligentization, and particularly relates to a photovoltaic string health degree evaluation system and method. BACKGROUND

[0002] At present, photovoltaic power stations have been widely used in global energy systems, especially in centralized mountainous station scenes, and limited by complex terrain and difficult maintenance, the long-term stable operation of photovoltaic string components becomes a core factor to ensure power generation efficiency. The existing technology mainly relies on the SCADA system to regularly obtain voltage, current, power and other data, and assists with artificial inspection means to evaluate the health status of photovoltaic strings. However, in actual application, the following problems have not been effectively solved:

[0003] Due to the diversity of equipment and complexity of data, the existing technology is difficult to realize dynamic and fine analysis of the health status of each string, which often leads to "abnormal detection lag" or "misjudgment". Most evaluation methods do not establish a unified health degree index system, which is difficult to quantify the actual impact of different types of low-efficiency faults (such as dust accumulation, shading, device aging) on power generation, resulting in that the subsequent maintenance scheduling cannot effectively rely on the evaluation results. In addition, the existing methods mainly focus on historical data backtracking or expert experience judgment, lack of "feature-decision" integrated design path, and it is difficult to directly convert the evaluation results into operation and maintenance optimization behavior.

[0004] Especially in the mountainous photovoltaic scene, the nonlinear and multi-coupling characteristics of string operation are more obvious, and the traditional health judgment based on a single index (such as voltage and current deviation) is easy to fail. SUMMARY

[0005] The purpose of the present application is to provide a photovoltaic string health degree evaluation system and method to solve the problems in the background art.

[0006] In order to achieve the above purpose, the present application provides the following technical scheme: a photovoltaic string health degree evaluation method, comprising:

[0007] Obtaining SCADA operation data of a plurality of photovoltaic strings in a target photovoltaic power station within a preset time interval, the data including voltage, current, component temperature, environmental irradiance and load end power factor;

[0008] Based on the operation data of the plurality of photovoltaic strings, a station-level feature reference model is constructed, and a state vector of each string is generated by a nonlinear projection algorithm, the state vector fusing environmental variables and load characteristic change trend;

[0009] For the target string to be evaluated, its corresponding health reference trajectory is extracted from the state vector, and a disturbance propagation factor is calculated within a local time window;

[0010] Based on the perturbation propagation factor and the health baseline trajectory, it is identified whether the current target string has an abnormal health state. If its deviation exceeds a set threshold, a reference string group with similar configuration parameters to the target string is dynamically selected from the heterogeneous reference strings.

[0011] Based on the historical health trajectory of the reference string and the current input value, a health score distribution model in a multi-dimensional space is constructed to calculate the health confidence score and prediction residual of the target string.

[0012] If the confidence value of the health score is lower than the set threshold, and the prediction residual shows a convergence instability trend in multiple consecutive windows, the target string is determined to be in a sub-healthy or latent fault state, and a health intervention instruction including type label, location number and response level is generated.

[0013] Preferably, the step of constructing a station-wide feature benchmark model based on the operating data of multiple photovoltaic strings, and generating the state vector of each string using a nonlinear projection algorithm, includes:

[0014] The operating data of multiple photovoltaic strings are synchronized and normalized in time, and steady-state and transient characteristic parameters of voltage, current, module temperature and ambient irradiance are extracted to form a feature matrix;

[0015] An autoencoder neural network is used to perform nonlinear dimensionality reduction on the feature matrix to obtain the potential correlation feature space between strings, and a clustering algorithm is used to determine the feature benchmark center for the entire station operation.

[0016] Map the running feature vector of each string to the feature reference center, calculate its deviation and generate the corresponding state vector;

[0017] The state vector is fused with the trend of environmental variable changes and the load-side power factor change curve to form a comprehensive vector reflecting the health status of the string operation.

[0018] Preferably, extracting the corresponding health baseline trajectory from the state vector and calculating the perturbation propagation factor within a local time window includes:

[0019] Extract the comprehensive state vector sequence of the target string within T consecutive time windows, and construct the health baseline trajectory curve of the string using spline interpolation method;

[0020] Calculate the dynamic offset between the trajectory and the center of the whole station feature reference, and extract its gradient in each dimension as the perturbation initiation vector.

[0021] For reference strings adjacent to the target string, analyze the degree of synchronization response of their state vectors within the same time window and calculate the disturbance response distance;

[0022] Based on the disturbance initiation vector and the disturbance response distance, the disturbance propagation factor value is calculated. The disturbance propagation factor value is used to assess the degree of linkage impact of the target string state change on the surrounding strings.

[0023] Preferably, based on the perturbation propagation factor and the health baseline trajectory, identifying whether the current target string has an abnormal health state includes:

[0024] Compare the deviation between the current point of the target string's health baseline trajectory and the real-time state vector. When the deviation exceeds the preset health anomaly threshold, it is determined that there is a health anomaly.

[0025] Simultaneously, by combining the perturbation propagation factor value, abnormal candidate strings with propagation intensity exceeding the set threshold are screened;

[0026] A self-verification mechanism is initiated for abnormal candidate strings, and structural parameters, electrical parameters and historical operating performance indicators are extracted from the power plant database;

[0027] Based on a multi-attribute similarity matching algorithm, heterogeneous reference string groups with similarity to the target string configuration parameters higher than a set threshold are dynamically filtered from all the strings on the site.

[0028] Preferably, by combining the perturbation propagation factor value, abnormal candidate strings with propagation intensity exceeding a set threshold are screened, including:

[0029] Compare the perturbation propagation factor of the current target string with its average propagation factor during its historical stable operating cycle, and calculate its relative amplification factor.

[0030] If the relative magnification factor is greater than the set threshold, the string is marked as a potential abnormal string;

[0031] Traverse the neighboring strings that are physically connected or logically associated with the target string, and calculate their disturbance response magnitude and delay characteristics within the same time window;

[0032] By using a propagation heat matrix model, combined with propagation intensity and response synchronicity, strings with heat values ​​higher than a set threshold in the propagation path are selected as abnormal candidate strings.

[0033] Preferably, the step of constructing a health score distribution model in a multidimensional space and calculating the health confidence score and prediction residual of the target string includes:

[0034] The historical health trajectories of the selected heterogeneous reference groups are standardized and uniformly mapped to a low-dimensional state space to form a multi-dimensional scoring training set.

[0035] The training set was modeled using the kernel density estimation method to generate a probability distribution model of health scores under each dimension;

[0036] Input the current comprehensive state vector of the target string into the health score probability distribution model, calculate its probability density value in each distribution dimension, and synthesize a confidence score;

[0037] The predicted state vector is generated by fitting historical trajectories and compared with the current actual state vector to obtain the predicted residual.

[0038] Preferably, the step of modeling the training set using the kernel density estimation method to generate a health score probability distribution model for each dimension includes:

[0039] The historical health trajectory of each reference group is divided into multiple fixed windows according to the time series, and the statistical features of the state vector within the window are extracted as sample points;

[0040] A kernel density function is established for each feature dimension for all sample points. The kernel density function is constructed using a Gaussian kernel function, and the bandwidth parameter is determined by combining the Silverman rule.

[0041] A continuous probability density curve is generated on each feature dimension to estimate the probability of a health score for any input state.

[0042] By combining the probability density functions of all feature dimensions to form a joint distribution model, a multidimensional probability estimation of the overall state of the target string is achieved.

[0043] Preferably, if the confidence value of the health score is lower than a set threshold, and the prediction residual shows a convergence instability trend over multiple consecutive windows, then the target string is determined to be in a sub-healthy or latent fault state, including:

[0044] Within a sliding time window, the health score confidence value and the corresponding prediction residual of the target string are continuously recorded to form a time series sample.

[0045] Determine whether the health score is below a set confidence threshold within M consecutive windows, where the confidence threshold is the P quantile of the historical score distribution;

[0046] Curve fitting is performed on the predicted residual sequence to extract trend indicators. When the residuals show non-monotonic convergence or increased volatility, they are marked as unstable.

[0047] When both the continuous low score and residual instability conditions are met, the target string is automatically identified as being in a sub-healthy state or having a hidden fault, and a diagnostic label is output.

[0048] The present invention also provides a photovoltaic string health assessment system, comprising:

[0049] Data acquisition module: acquires SCADA operation data of multiple photovoltaic strings in the target photovoltaic power station within a preset time interval, including voltage, current, module temperature, ambient irradiance and load power factor;

[0050] Feature extraction module: Based on the operating data of multiple photovoltaic strings, a full-site feature benchmark model is constructed, and a state vector of each string is generated using a nonlinear projection algorithm. The state vector integrates environmental variables and load characteristic change trends.

[0051] Perturbation calculation module: For the target string to be evaluated, extract its corresponding health baseline trajectory from the state vector and calculate the perturbation propagation factor within a local time window;

[0052] Reference string filtering module: Based on the perturbation propagation factor and health baseline trajectory, identify whether the current target string has an abnormal health state. If its deviation exceeds the set threshold, dynamically filter reference string groups with similar configuration parameters to the target string from the heterogeneous reference strings.

[0053] Confidence modeling module: Based on the historical health trajectory of the reference string and the current input value, construct a health score distribution model in a multi-dimensional space, and calculate the health confidence score and prediction residual of the target string;

[0054] Early warning module: If the confidence value of the health score is lower than the set threshold, and the prediction residual shows a convergence instability trend in multiple consecutive windows, the target string is determined to be in a sub-healthy or latent fault state, and a health intervention instruction including type label, location number and response level is generated.

[0055] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0056] 1. This invention constructs a state vector that integrates environmental variables and load characteristic trends, and combines a nonlinear projection algorithm and a disturbance propagation modeling mechanism to achieve a high-dimensional, multi-angle dynamic representation of the target string's operating state. Compared with traditional health assessment methods based on single-parameter thresholds or empirical models, this invention can autonomously identify abnormal operating trends under conditions of no fault labels or minimal human intervention, significantly improving the response sensitivity to latent faults and sub-health states in the complex operating environment of photovoltaic systems.

[0057] 2. This invention introduces a health score distribution model and a joint judgment mechanism for prediction residuals. Utilizing a multi-dimensional probability space constructed from reference strings, combined with the actual state evolution trend of the target string, a health confidence assessment system with high interpretability and temporal stability is formed. Through state fluctuation analysis within a continuous window and adaptive reference group selection, this invention achieves broad adaptability to different topologies and heterogeneous component scenarios, effectively supporting refined operation and maintenance, tiered early warning, and optimal intervention strategy formulation for photovoltaic power plants. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0059] Figure 1 This is a flowchart of the method of the present invention.

[0060] Figure 2 This is a flowchart of the system modules of the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] Example 1, please refer to Figure 1 As shown in this embodiment, a photovoltaic string health assessment method includes:

[0063] Acquire SCADA operation data of multiple photovoltaic strings in the target photovoltaic power station within a preset time interval. The data includes voltage, current, module temperature, ambient irradiance, and load-side power factor.

[0064] Based on the operating data of multiple photovoltaic strings, a full-station characteristic benchmark model is constructed, and a state vector for each string is generated using a nonlinear projection algorithm. The state vector integrates environmental variables and load characteristic change trends.

[0065] For the target string to be evaluated, its corresponding health baseline trajectory is extracted from the state vector, and the perturbation propagation factor is calculated within a local time window;

[0066] Based on the perturbation propagation factor and the health baseline trajectory, it is identified whether the current target string has an abnormal health state. If its deviation exceeds the set threshold, a self-verification mechanism is triggered to dynamically filter reference string groups with similar configuration parameters to the target string from the heterogeneous reference strings.

[0067] Based on the historical health trajectory of the reference string and the current input value, a health score distribution model in a multi-dimensional space is constructed to calculate the health confidence score and prediction residual of the target string.

[0068] If the confidence value of the health score is lower than the set threshold, and the prediction residual shows a convergence instability trend in multiple consecutive windows, the target string is determined to be in a sub-healthy or latent fault state, and a health intervention instruction including type label, location number and response level is generated.

[0069] In one embodiment of the present invention, it is first necessary to obtain operational data of multiple strings within a preset time interval from the monitoring system of the photovoltaic power plant. The monitoring system described in this invention is a SCADA (Supervisory Control and Data Acquisition) system, which can collect and store key parameter data of the power plant operation at a frequency of minutes or seconds, and is the main data source for health modeling and assessment in this invention. Specifically, the operational data includes:

[0070] Voltage data: This includes the output voltage value of each photovoltaic string at each sampling point in time, which is used to reflect the output status of the photovoltaic module and possible abnormal problems such as open circuit, shading, and loose connection.

[0071] Current data: This represents the current output of the photovoltaic string at a specific moment. This parameter, together with the voltage, determines the instantaneous output power and can reflect the load response characteristics and the system operating point.

[0072] Module temperature: By monitoring the temperature in real time through temperature sensors installed on the back of the photovoltaic module, performance degradation caused by abnormal operating temperature rise, hot spot effect, or environmental abnormalities (such as poor ventilation) can be identified.

[0073] Environmental irradiance: This includes data recorded by irradiance monitoring devices deployed in the area where the strings are located. It is mainly used to construct a power generation model under standard operating conditions and to provide a standard reference for assessing the health of the strings.

[0074] Load-side power factor: This indicates the power quality after the string is connected to the inverter input. It reflects the load matching on the grid side and is helpful in identifying abnormal power generation caused by external interference (such as grid fluctuations).

[0075] The "preset time interval" described in this invention can be flexibly set according to actual application needs. For example, the commonly used time granularity is 5 minutes, 10 minutes, or 30 minutes, and the duration can be 1 day, 1 week, or a rolling window cycle (e.g., 7 days or 14 days). The multiple photovoltaic strings can be multiple series subarrays under the same inverter, or they can form a local string cluster across multiple inverters.

[0076] In an embodiment of the present invention, a full-site characteristic benchmark model is constructed based on the operating data of multiple photovoltaic strings, and a state vector for each string is generated using a nonlinear projection algorithm. The state vector integrates environmental variables and load characteristic change trends, specifically including:

[0077] First, the operating data of multiple photovoltaic strings acquired from the SCADA system are uniformly timestamped to ensure that all data points form a complete time-series data set at the same sampling frequency. Then, various operating parameters (including voltage, current, module temperature, and ambient irradiance) are normalized to make them comparable on a numerical scale. A commonly used normalization method is the min-max normalization method, with the conversion formula: Normalized value = (Original value - Minimum value) / (Maximum value - Minimum value). Based on the preprocessed data, the following two types of features are extracted:

[0078] Steady-state characteristics: such as daily average voltage, daily average current, average module temperature, and total daily irradiance;

[0079] Transient characteristics: such as the rate of change of voltage during morning and evening, the daily fluctuation range of temperature, the peak current, and the maximum voltage slope.

[0080] All features are combined into a multidimensional feature matrix, which describes the running state of each string within different time windows.

[0081] In constructing the feature benchmark model, an autoencoder neural network is used to achieve nonlinear dimensionality reduction of high-dimensional features. This neural network consists of an encoder and a decoder. The encoder compresses the input features into a low-dimensional latent vector, while the decoder attempts to reconstruct the original features, thus ensuring the effectiveness of the latent space. The objective function of network training is to minimize the reconstruction error, and the optimization method is a backpropagation algorithm based on gradient descent.

[0082] After training, the encoder part maps the feature vector of each string to a low-dimensional space to obtain its latent feature representation. Next, the K-means clustering algorithm is used to cluster the low-dimensional vectors of all strings. The number of clusters K is determined by silhouette coefficient analysis. The cluster center vector is the overall feature reference center, used to measure the similarity between it and other strings in the feature space.

[0083] The degree of health deviation of each photovoltaic string's low-dimensional potential eigenvector can be measured by the distance between it and the feature reference center. Specifically, Euclidean distance is used as the deviation measurement method. If the distance between the potential eigenvector of a string and the reference center exceeds a set threshold (e.g., the mean plus 1.5 times the standard deviation), it can be preliminarily judged that the string has an abnormal trend in operation.

[0084] In this way, a multi-dimensional state vector containing "deviation metric", "operation category label" and "feature directionality" is generated for each string. This state vector can be used to represent the relative operating state of the current string in the entire photovoltaic power plant operating environment.

[0085] The present invention further integrates the generated state vector with the trend of environmental variable changes and the load-side power factor change curve.

[0086] The environmental variable trends are composed of the slope and volatility of ambient irradiance and component temperature within different time windows; the load-side power factor variation curve is obtained by calculating the standard deviation and rate of change of the power factor per unit time. This part can be dynamically updated using sliding window technology, with the window size typically set to 24 hours or 48 hours.

[0087] The fusion method employs a weighted vector superposition model, where the weight of the original state vector is set to 0.6, the weight of the environmental trend feature is set to 0.25, and the weight of the load factor change feature is set to 0.15. The weight values ​​can be adjusted and optimized according to the principle of minimizing historical regression error. The fused vector is the comprehensive state vector.

[0088] In a preferred embodiment of the present invention, for the target string to be evaluated, the corresponding health baseline trajectory is extracted from the state vector, and the perturbation propagation factor is calculated within a local time window, specifically including:

[0089] First, for the target string to be evaluated, within a set continuous time window T (e.g., 72 hours, with 864 data sets sampled every 5 minutes), a comprehensive state vector is extracted for each time point. The comprehensive state vector is a multi-dimensional vector containing information such as the target string's feature deviation, environmental trend indicators, and load disturbance factors.

[0090] To describe the continuous change trend of the state vector over time, this invention employs spline interpolation to smoothly fit the state vector sequence. Spline interpolation is performed separately in each dimension, ensuring the continuity and differentiability of the health trajectory across all feature dimensions, thereby constructing the health baseline trajectory curve of the target string. This trajectory represents the multidimensional evolution path that the target string should exhibit under normal operating conditions.

[0091] After obtaining the healthy baseline trajectory, the dynamic offset is calculated by comparing the state vector at the current time point with the vector at the corresponding time point on the healthy trajectory dimension by dimension. This dynamic offset represents the degree of deviation between the actual state and the ideal trajectory. Let the offset in a certain dimension be Δx, then the disturbance initiation vector consists of all Δx values.

[0092] Furthermore, the first derivative of the initial perturbation vector is approximated to extract its gradient over time, which is used to characterize the intensity of the state transition. If the gradient value in a certain dimension exceeds a preset threshold (e.g., the average plus twice the standard deviation), then that dimension is marked as having a significant perturbation trend, serving as the initial condition for subsequent perturbation propagation calculations.

[0093] Select a set of reference strings that are electrically connected to the target string or are on the same branch (e.g., the other N strings downstream of the inverter), and extract their state vector sequences within the same time window T.

[0094] Dynamic Time Warping (DTW) is used to compare the similarity of the state vector trajectories of the target string and each reference string. DTW can identify sequence changes with similar trends but temporal offsets in different time periods. The smaller the DTW distance, the stronger the synchronous response of the reference string to the perturbation of the target string.

[0095] This invention calculates the Disturbance Propagation Factor (DPF) based on the magnitude of the initial disturbance vector and the response distance of the reference string. The formula for calculating the DPF is as follows: DPF = Weighting coefficient × Initial disturbance magnitude ÷ Average response distance; where: the initial disturbance magnitude is the Euclidean magnitude of the initial disturbance vector; the average response distance is the mean of the DTW distances in the reference string set; the weighting coefficient can be set to an empirical value, such as 0.8~1.2, to adjust the model's sensitivity to disturbance intensity and response breadth.

[0096] A higher DPF indicates that abnormal changes in the target string are more likely to have a chain reaction effect on the system. Therefore, DPF can not only serve as a precursor indicator of target string health decline, but also be used to identify key risk propagation paths, effectively supporting early warning mechanisms and intelligent scheduling optimization.

[0097] Based on the perturbation propagation factor and the health baseline trajectory, it is determined whether the current target string has an abnormal health state. If its deviation exceeds a set threshold, a group of reference strings with similar configuration parameters to the target string is dynamically selected from the heterogeneous reference strings. Specifically, this includes:

[0098] In this step, the real-time integrated state vector of the current target string is first compared dimension by dimension with the vector at the same time point on its corresponding health baseline trajectory curve. The deviation of each dimension is calculated, and the overall deviation is calculated by weighted Euclidean distance. The calculation method is as follows: the difference between the current actual value and the baseline trajectory value of each feature dimension (such as voltage, current, component temperature, irradiance, power factor, etc.) is squared and multiplied by the weight coefficient corresponding to that dimension. Then, all weighted squared differences are summed, and the square root of the sum is calculated to obtain the state deviation.

[0099] The weighting coefficients can be preset according to the degree of influence of each dimension on power generation. For example, the weights of voltage and current dimensions can be set to 0.3, temperature and irradiance to 0.2, and power factor to 0.1.

[0100] If the aforementioned deviation exceeds a preset health anomaly threshold, for example, set to the average historical deviation of the string plus twice the standard deviation (μ+2σ), then the string is determined to be in a health anomaly state. This invention combines a disturbance propagation factor for linked screening. The specific implementation is as follows:

[0101] The current target string's disturbance propagation factor is compared with its average propagation factor calculated during historical stable operating cycles (e.g., no alarms in the past 7 days), and the propagation amplification factor (β) is calculated: β = current propagation factor ÷ historical average propagation factor; if the β value exceeds the preset threshold (e.g., 2.0), it indicates that the string's status change has a significant propagation trend throughout the station, and the string is marked as a potential abnormal string.

[0102] Traverse the set of neighboring strings that have a physical connection (same inverter, combiner box) or a communication logical association (same region or subsystem) with the target string, and extract their state vector sequence within the same time window.

[0103] The disturbance response amplitude (i.e., the magnitude of the deviation) and response delay characteristics (i.e., the time offset of the maximum change point) of these strings are calculated respectively to form a response attribute vector.

[0104] This invention further constructs a propagation heat matrix model based on propagation relationships. This model starts with the target string and records the propagation impact score of each adjacent string, defined as: Heat value Hij = Propagation factor × Response amplitude × (1 ÷ Response delay); where Hij represents the propagation heat value of target string i to string j; the response delay is measured in minutes, and a minimum value of 0.1 is set if it is zero to prevent division by zero errors. A propagation heat threshold is set, for example, 1.5 times the average heat value of the entire site. Strings with heat values ​​higher than the threshold are filtered out and marked as an abnormal candidate string set, serving as the analysis object for subsequent self-verification mechanisms.

[0105] For each abnormal candidate string, a self-verification process is initiated, which includes: extracting the string's structural parameters (such as component type, string capacity, and installation angle), electrical parameters (such as inverter model and wiring method), and historical operating performance indicators (such as annual average PR and power generation curve characteristic values) from the photovoltaic power station's backend database; constructing a feature vector for each string; calculating the similarity of all strings in the entire station; and using a weighted cosine similarity algorithm or K-nearest neighbor clustering method for screening. The weights can be set according to the parameter sensitivity, such as 0.5 for structural parameters, 0.3 for electrical parameters, and 0.2 for performance indicators. String groups with a similarity higher than a preset threshold (such as 0.85) with the target string are selected as a heterogeneous reference string set for subsequent health score correction and prediction model updates.

[0106] Based on the historical health trajectory of the reference string and the current input value, a multi-dimensional health score distribution model is constructed to calculate the health confidence score and prediction residual of the target string. Specifically, this includes:

[0107] First, the historical health trajectories of heterogeneous reference strings are uniformly preprocessed. The health trajectory is a comprehensive state vector sequence of the string during stable operation, with a time span that can be set from 7 to 30 days.

[0108] To eliminate the differences in dimensions and scales between different strings, the max-min normalization method is used to standardize all state vectors, that is, to linearly scale each feature value according to its maximum and minimum values ​​so that it falls into the [0,1] interval.

[0109] Subsequently, principal component analysis (PCA) or an autoencoder network is used to map the high-dimensional state vector to a low-dimensional state space (e.g., reduce the dimensionality to 3-5 dimensions) to reduce redundant feature interference and obtain a multi-dimensional score training set in a unified format.

[0110] To perform nonparametric modeling on the scoring training set, a continuous probability distribution is established across each feature dimension using kernel density estimation (KDE). The specific steps are as follows:

[0111] The standardized state sequence of each reference group is divided into multiple fixed-width sliding windows (e.g., one window every 24 hours) according to the time series, and statistical features such as mean, maximum value, and rate of change are extracted in each window as a single sample point.

[0112] For the distribution of all sample points along a certain feature dimension, a Gaussian kernel function is used to estimate the kernel density. The form of the Gaussian kernel function is: Where x represents the deviation of the state variable in that dimension. The choice of the kernel function bandwidth parameter has a significant impact on the estimation effect. The Silverman rule is used to automatically calculate the bandwidth, and the calculation formula is: bandwidth equals 1.06 multiplied by the standard deviation multiplied by the square of the sample size (-1 / 5);

[0113] After establishing the kernel density functions for all feature dimensions, a multidimensional joint probability distribution model of the health status of this set of features is constructed. If we assume that each dimension is independent, the joint density value can be obtained by directly multiplying the probability density functions of each dimension, forming a multivariate distribution model for health scoring.

[0114] The comprehensive state vector of the target string at the current time is input into the joint probability distribution model mentioned above. The corresponding probability density value is obtained in each dimension, and the values ​​are weighted and fused according to the set weights to generate the final health confidence score (HCS).

[0115] The Health Confidence Score is calculated as follows: HCS = Σ (weight i × kernel density value i); where weight i reflects the contribution of each dimension to the health status and can be set through cross-validation or expert experience. For example, electrical parameters have a higher weight, followed by environmental parameters.

[0116] A higher score indicates that the current state is more consistent with the distribution characteristics of the health model; if the score is lower than a set threshold (e.g., less than the 10th percentile of the historical reference score distribution), it can be considered that there is a potential risk of health abnormality.

[0117] Building upon this foundation, to enhance model stability and misjudgment suppression capabilities, this invention further introduces a prediction residual mechanism. Specifically, based on the trajectory patterns of historical reference sequences, a regression model (such as random forest regression or support vector regression) is used to predict the state vector that the target sequence should present at the current moment, denoted as the prediction vector. This prediction vector is then compared dimension-by-dimensionally with the actual current state vector, the difference is calculated, and the Euclidean distance is obtained, serving as the prediction residual value. A larger prediction residual value indicates a more severe deviation of the current state from the historical trajectory pattern, which can be used to adjust the credibility of the health score or directly trigger an anomaly warning.

[0118] If the confidence value of the health score is lower than a set threshold, and the prediction residual shows a convergence instability trend over multiple consecutive windows, the target string is determined to be in a sub-healthy or latent fault state, and a health intervention instruction including a type label, location number, and response level is generated, specifically including:

[0119] During the continuous operation of the photovoltaic string, the current health score confidence value and prediction residual are calculated at preset time intervals (e.g., every 5 minutes), and the results are recorded in a time-ordered sliding window. The sliding window width can be set to 3 to 6 hours, corresponding to 36 to 72 sampling points. The health score confidence value represents the probability density of the current state in the health score distribution model; the lower the value, the less typical the state.

[0120] The prediction residual is the Euclidean distance between the current state vector and the historical trajectory fitting value, which is used to measure the deviation trend of the running state.

[0121] Within a sliding window, the confidence value of the health score is dynamically monitored to determine whether it falls below a set lower threshold for a score within M consecutive sampling periods. This threshold can be determined by referring to the P-th percentile of the historical reference string score distribution. P is recommended to be between 10 and 20 to represent the bottom critical value of the normal score range. For example, if P is 10, the score confidence value is lower than the top 10% of the scores in the distribution, indicating a significant deviation from the healthy model. If the health score remains below this threshold for M consecutive window periods (e.g., M=5), the target string is considered to be in a long-term abnormal borderline state, requiring further analysis based on residual trends.

[0122] The predicted residual value sequence within the sliding window is used as the time series input, and its changing trend is evaluated using methods such as linear regression, moving variance analysis, or spline fitting.

[0123] This invention introduces two typical types of instability behaviors:

[0124] Non-monotonic convergence: This means that the residual value exhibits a fluctuating pattern of rising-falling-rising again within the window, indicating that the state repeatedly deviates and cannot stabilize.

[0125] Increased volatility trend: This means that the standard deviation of the residuals within the window continues to widen, or the moving variance increases by more than twice the historical mean standard deviation across multiple windows.

[0126] If any of the behaviors is satisfied, it can be determined that the state fluctuation of the string within the current window shows an unstable trend.

[0127] The target string meets the following two conditions:

[0128] The health score confidence value is below the set confidence threshold for M consecutive periods;

[0129] The predicted residual sequence shows a clear trend of convergence instability;

[0130] The system will automatically classify the string as being in a sub-healthy state or having a hidden risk of failure. Such states are usually difficult to identify by traditional alarm mechanisms, but in actual operation and maintenance, they may lead to a continuous decline in power generation efficiency or the accumulation of local failures.

[0131] Once the diagnosis is confirmed, a diagnostic label containing the following information will be automatically output: status level (e.g., "sub-healthy", "latent fault"); timestamp; residual trend type; recommended action (e.g., "recommend manual review" or "included in key monitoring"); and the label will be pushed to the scheduling control system or the terminal of the operation and maintenance personnel as the basis for triggering the intelligent early warning strategy.

[0132] Example 2, please refer to Figure 2 As shown in this embodiment, a photovoltaic string health assessment system includes:

[0133] Data acquisition module: acquires SCADA operation data of multiple photovoltaic strings in the target photovoltaic power station within a preset time interval, including voltage, current, module temperature, ambient irradiance and load power factor;

[0134] Feature extraction module: Based on the operating data of multiple photovoltaic strings, a full-site feature benchmark model is constructed, and a state vector of each string is generated using a nonlinear projection algorithm. The state vector integrates environmental variables and load characteristic change trends.

[0135] Perturbation calculation module: For the target string to be evaluated, extract its corresponding health baseline trajectory from the state vector and calculate the perturbation propagation factor within a local time window;

[0136] Reference string filtering module: Based on the perturbation propagation factor and health baseline trajectory, identify whether the current target string has an abnormal health state. If its deviation exceeds the set threshold, dynamically filter reference string groups with similar configuration parameters to the target string from the heterogeneous reference strings.

[0137] Confidence modeling module: Based on the historical health trajectory of the reference string and the current input value, construct a health score distribution model in a multi-dimensional space, and calculate the health confidence score and prediction residual of the target string;

[0138] Early warning module: If the confidence value of the health score is lower than the set threshold, and the prediction residual shows a convergence instability trend in multiple consecutive windows, the target string is determined to be in a sub-healthy or latent fault state, and a health intervention instruction including type label, location number and response level is generated.

[0139] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for assessing the health of photovoltaic strings, characterized in that: include: Acquire SCADA operation data of multiple photovoltaic strings in the target photovoltaic power station within a preset time interval. The data includes voltage, current, module temperature, ambient irradiance, and load-side power factor. Based on the operating data of multiple photovoltaic strings, a full-station characteristic benchmark model is constructed, and a state vector for each string is generated using a nonlinear projection algorithm. The state vector integrates environmental variables and load characteristic change trends. For the target string to be evaluated, its corresponding health baseline trajectory is extracted from the state vector, and the perturbation propagation factor is calculated within a local time window; Based on the perturbation propagation factor and the health baseline trajectory, it is identified whether the current target string has an abnormal health state. If its deviation exceeds a set threshold, a reference string group with similar configuration parameters to the target string is dynamically selected from the heterogeneous reference strings. Based on the historical health trajectory of the reference string and the current input value, a health score distribution model in a multi-dimensional space is constructed to calculate the health confidence score and prediction residual of the target string. If the confidence value of the health score is lower than the set threshold, and the prediction residual shows a convergence instability trend in multiple consecutive windows, the target string is determined to be in a sub-healthy or latent fault state, and a health intervention instruction including type label, location number and response level is generated.

2. The method for assessing the health of photovoltaic strings according to claim 1, characterized in that: The system constructs a full-site feature benchmark model based on the operating data of multiple photovoltaic strings, and generates the state vector of each string using a nonlinear projection algorithm, including: The operating data of multiple photovoltaic strings are synchronized and normalized in time, and steady-state and transient characteristic parameters of voltage, current, module temperature and ambient irradiance are extracted to form a feature matrix; An autoencoder neural network is used to perform nonlinear dimensionality reduction on the feature matrix to obtain the potential correlation feature space between strings, and a clustering algorithm is used to determine the feature benchmark center for the entire station operation. Map the running feature vector of each string to the feature reference center, calculate its deviation and generate the corresponding state vector; The state vector is fused with the trend of environmental variable changes and the load-side power factor change curve to form a comprehensive vector reflecting the health status of the string operation.

3. The method for assessing the health of photovoltaic strings according to claim 1, characterized in that: Extract the corresponding health baseline trajectory from the state vector and calculate the perturbation propagation factor within a local time window, including: Extract the comprehensive state vector sequence of the target string within T consecutive time windows, and construct the health baseline trajectory curve of the string using spline interpolation method; Calculate the dynamic offset between the trajectory and the center of the whole station feature reference, and extract its gradient in each dimension as the perturbation initiation vector. For reference strings adjacent to the target string, analyze the degree of synchronization response of their state vectors within the same time window and calculate the disturbance response distance; Based on the disturbance initiation vector and the disturbance response distance, the disturbance propagation factor value is calculated. The disturbance propagation factor value is used to assess the degree of linkage impact of the target string state change on the surrounding strings.

4. The method for assessing the health of photovoltaic strings according to claim 1, characterized in that: Based on the perturbation propagation factor and the health baseline trajectory, identify whether the current target string has an abnormal health state, including: Compare the deviation between the current point of the target string's health baseline trajectory and the real-time state vector. When the deviation exceeds the preset health anomaly threshold, it is determined that there is a health anomaly. Simultaneously, by combining the perturbation propagation factor value, abnormal candidate strings with propagation intensity exceeding the set threshold are screened; A self-verification mechanism is initiated for abnormal candidate strings, and structural parameters, electrical parameters and historical operating performance indicators are extracted from the power plant database; Based on a multi-attribute similarity matching algorithm, heterogeneous reference string groups with similarity to the target string configuration parameters higher than a set threshold are dynamically filtered from all the strings on the site.

5. The method for assessing the health of photovoltaic strings according to claim 4, characterized in that: Simultaneously, by combining the perturbation propagation factor value, abnormal candidate strings with propagation intensity exceeding a set threshold are screened, including: Compare the perturbation propagation factor of the current target string with its average propagation factor during its historical stable operating cycle, and calculate its relative amplification factor. If the relative magnification factor is greater than the set threshold, the string is marked as a potential abnormal string; Traverse the neighboring strings that are physically connected or logically associated with the target string, and calculate their disturbance response magnitude and delay characteristics within the same time window; By using a propagation heat matrix model, combined with propagation intensity and response synchronicity, strings with heat values ​​higher than a set threshold in the propagation path are selected as abnormal candidate strings.

6. The method for assessing the health of photovoltaic strings according to claim 1, characterized in that: The construction of a multidimensional health score distribution model, and the calculation of the health confidence score and prediction residuals of the target string, include: The historical health trajectories of the selected heterogeneous reference groups are standardized and uniformly mapped to a low-dimensional state space to form a multi-dimensional scoring training set. The training set was modeled using the kernel density estimation method to generate a probability distribution model of health scores under each dimension; Input the current comprehensive state vector of the target string into the health score probability distribution model, calculate its probability density value in each distribution dimension, and synthesize a confidence score; The predicted state vector is generated by fitting historical trajectories and compared with the current actual state vector to obtain the predicted residual.

7. The method for assessing the health of photovoltaic strings according to claim 6, characterized in that: The step of modeling the training set using the kernel density estimation method to generate a probability distribution model of health scores under each dimension includes: The historical health trajectory of each reference group is divided into multiple fixed windows according to the time series, and the statistical features of the state vector within the window are extracted as sample points; A kernel density function is established for each feature dimension for all sample points. The kernel density function is constructed using a Gaussian kernel function, and the bandwidth parameter is determined by combining the Silverman rule. A continuous probability density curve is generated on each feature dimension to estimate the probability of a health score for any input state. By combining the probability density functions of all feature dimensions to form a joint distribution model, a multidimensional probability estimation of the overall state of the target string is achieved.

8. The method for assessing the health of photovoltaic strings according to claim 1, characterized in that: If the confidence value of the health score is lower than a set threshold, and the prediction residual shows a convergence instability trend over multiple consecutive windows, then the target string is determined to be in a sub-healthy or latent fault state, including: Within a sliding time window, the health score confidence value and the corresponding prediction residual of the target string are continuously recorded to form a time series sample. Determine whether the health score is below a set confidence threshold within M consecutive windows, where the confidence threshold is the P quantile of the historical score distribution; Curve fitting is performed on the predicted residual sequence to extract trend indicators. When the residuals show non-monotonic convergence or increased volatility, they are marked as unstable. When both the continuous low score and residual instability conditions are met, the target string is automatically identified as being in a sub-healthy state or having a hidden fault, and a diagnostic label is output.

9. A photovoltaic string health assessment system, used to implement the photovoltaic string health assessment method according to any one of claims 1-8, characterized in that: include: Data acquisition module: acquires SCADA operation data of multiple photovoltaic strings in the target photovoltaic power station within a preset time interval, including voltage, current, module temperature, ambient irradiance and load power factor; Feature extraction module: Based on the operating data of multiple photovoltaic strings, a full-site feature benchmark model is constructed, and a state vector of each string is generated using a nonlinear projection algorithm. The state vector integrates environmental variables and load characteristic change trends. Perturbation calculation module: For the target string to be evaluated, extract its corresponding health baseline trajectory from the state vector and calculate the perturbation propagation factor within a local time window; Reference string filtering module: Based on the perturbation propagation factor and health baseline trajectory, identify whether the current target string has an abnormal health state. If its deviation exceeds the set threshold, dynamically filter reference string groups with similar configuration parameters to the target string from the heterogeneous reference strings. Confidence modeling module: Based on the historical health trajectory of the reference string and the current input value, construct a health score distribution model in a multi-dimensional space, and calculate the health confidence score and prediction residual of the target string; Early warning module: If the confidence value of the health score is lower than the set threshold, and the prediction residual shows a convergence instability trend in multiple consecutive windows, the target string is determined to be in a sub-healthy or latent fault state, and a health intervention instruction including type label, location number and response level is generated.

Images