An Early Warning Method for Implicit Degradation of Photovoltaic Strings Based on Temporal Feature Decoupling
By cleaning and extracting features from photovoltaic string operation and environmental time-series data, a healthy power benchmark model is constructed, which solves the problem of early warning of implicit degradation of photovoltaic strings and realizes accurate string-level early warning and operation and maintenance optimization.
Patent Information
- Application Number
- CN202611045803.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-14
- Publication Date
- 2026-11-17
AI Technical Summary
Existing photovoltaic string operation monitoring methods are unable to effectively distinguish between the impact of environmental factors and latent degradation, resulting in false alarms, missed alarms, and operation and maintenance delays, and are unable to achieve continuous monitoring and string-level early warning under minute-level data conditions.
By acquiring the operating sequence data of photovoltaic strings and environmental time-series data, performing timestamp alignment and cleaning, constructing a healthy power benchmark model for dynamic peer reference groups, calculating residual sequences and extracting trend, fluctuation, and persistence features, using three-dimensional residual joint judgment for early warning, and periodically updating the model.
It enables early identification and warning of latent degradation of photovoltaic strings in photovoltaic power plants, reduces false alarms, improves operation and maintenance efficiency, adapts to environmental changes, reduces the impact of equipment differences, and provides accurate string-level early warning information.
Smart Images

Figure CN123122444A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent operation and maintenance technology of photovoltaic power plants, and specifically relates to an early warning method for the implicit degradation of photovoltaic strings based on time-series feature decoupling. Background Technology
[0002] Photovoltaic strings are the basic power generation units in a photovoltaic power plant system, and their operating status directly affects the DC input power of the inverter and the overall power generation efficiency of the plant. During the operation of a photovoltaic power plant, factors such as module aging, hot spots, degradation of encapsulation materials, poor wiring contact, abnormal bypass diodes, localized dust accumulation, or long-term slight shading can all lead to a gradual decline in the output capacity of the photovoltaic strings. Some of these degradation processes do not manifest as obvious fault alarms in the early stages, nor do they necessarily trigger inverter shutdown or protection actions. Instead, they are reflected in a sustained output power level below normal under the same environmental conditions; this type of problem is usually considered a latent degradation issue in photovoltaic strings.
[0003] Current photovoltaic (PV) string operation monitoring typically relies on operational data collected by the inverter, such as string current, string voltage, and DC power, combined with environmental data collected by weather stations, including irradiance, module temperature, and wind speed. In actual operation and maintenance, common methods include fixed threshold judgment, horizontal comparison of string currents, statistical analysis of power generation deviations, performance ratio analysis, and manual inspection and verification. Some systems also use infrared detection, electroluminescence detection, or on-site spot checks to further confirm suspicious modules or strings.
[0004] While the above methods can detect obvious anomalies such as short circuits, open circuits, severe shading, and inverter failures, they still have the following problems regarding early, slow, and continuous latent degradation: First, the output power of photovoltaic strings is greatly affected by irradiance, module backsheet temperature, wind speed, installation tilt angle, and local weather changes. If judgment is based solely on fixed power or current thresholds, power fluctuations caused by environmental changes can easily be misjudged as string degradation, or slight degradation can be masked by environmental fluctuations. Second, even within the same power plant, different strings may have different output benchmarks due to differences in spatial location, commissioning time, module type, inverter type, and historical operating conditions. Simple horizontal comparisons are insufficient to eliminate differences between equipment and natural aging stages; third, existing methods often focus on power deviations at a single moment or within a short time window, lacking joint analysis of the long-term trend, fluctuation level, and duration of residuals, making it difficult to distinguish between temporary shading, cloud shadow disturbances, short-term dust accumulation, and permanent latent degradation; fourth, output deviations in the early stages of latent degradation are usually small, and if alarms are only triggered when the deviation reaches a significant fault level, the time for operation and maintenance intervention is delayed, making it difficult to locate string-level problems in a timely manner; fifth, the operating conditions of photovoltaic power plants change with seasons, weather conditions, and the natural aging of components, and if the model or judgment threshold remains fixed for a long time without being updated, baseline drift, false alarms, or missed alarms are likely to occur.
[0005] Furthermore, in large-scale photovoltaic power plants, the number of strings is large. If the identification of latent degradation is mainly based on manual inspection, spot checks, or offline testing, it is not only labor-intensive but also difficult to achieve continuous monitoring and string-level early warning under minute-level data conditions. Therefore, it is necessary to use photovoltaic string operation time series data and environmental time series data to establish a healthy output benchmark that can eliminate the influence of environmental factors and identify persistent power degradation from long-term residual characteristics.
[0006] Therefore, the main technical problems to be solved in this application are: how to establish a unified time-series analysis basis between photovoltaic string operation data and environmental data; how to select a suitable healthy string as a reference based on the spatial location, commissioning time, equipment model and historical operating status of the target photovoltaic string to form a dynamically updatable reference benchmark under the same operating conditions; how to construct a healthy power benchmark model that can decouple the effects of irradiance, module backsheet temperature, ambient wind speed and installation tilt angle; how to extract trend features, fluctuation features and persistence features based on the residual sequence between the actual output power and the benchmark power prediction value; how to use multidimensional residual features to distinguish between temporary shading and permanent latent attenuation, and output string-level attenuation level and early warning information; and how to periodically update the relevant models during the long-term operation of the power plant to adapt to seasonal changes, natural aging and changes in operating conditions. Summary of the Invention
[0007] The technical problem to be solved by this invention is to provide an early warning method for the implicit degradation of photovoltaic strings based on time-series feature decoupling. This invention acquires the running time-series data of the target photovoltaic string, such as string current, string voltage, and DC power, as well as environmental time-series data such as irradiance, module backsheet temperature, ambient wind speed, and installation tilt angle. Then, it performs timestamp alignment, outlier removal, invalid time period removal, and missing data processing to form cleaned minute-level time-series data, so that subsequent power benchmark modeling and residual analysis have a unified and continuous data foundation.
[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: An early warning method for implicit degradation of photovoltaic strings based on time-series feature decoupling, comprising the following steps: S1. Obtain the runtime timing data and corresponding environmental timing data of the target photovoltaic string, and perform timestamp alignment, outlier removal, invalid time period removal and missing data processing on the runtime timing data and environmental timing data to form cleaned minute-level timing data; S2. Based on the spatial location, commissioning time, module model, inverter model and historical operating status of the target photovoltaic string, construct a dynamic peer reference group from adjacent healthy photovoltaic strings, and train a healthy power benchmark model decoupled from environmental factors based on the dynamic peer reference group, hereinafter referred to as the healthy power benchmark model. S3. Using the healthy power benchmark model, output the benchmark power prediction value of the target photovoltaic string under the current environmental conditions, calculate the residual sequence between the actual output power of the target photovoltaic string and the benchmark power prediction value, and extract trend features, fluctuation features and persistence features from the residual sequence. S4. Perform a three-dimensional residual joint determination based on the trend characteristics, fluctuation characteristics and persistence characteristics to distinguish between temporary shading and permanent hidden attenuation, and input the three-dimensional residual characteristics into the health grading model to output the attenuation level of the target photovoltaic string; S5. When the target photovoltaic string continuously meets the early warning triggering conditions, output early warning information including string number, string location, attenuation level and operation and maintenance suggestions, and perform periodic online self-updates on the healthy power benchmark model and health classification model.
[0009] Preferably, in step S1, the runtime timing data includes string current, string voltage, and DC power; the environmental timing data includes irradiance, module backplane temperature, ambient wind speed, and installation tilt angle; the runtime timing data is collected via Modbus TCP protocol at a sampling frequency of 1 minute / time, and the environmental timing data is collected from the power plant meteorological station via HTTP protocol at a sampling frequency of 1 minute / time.
[0010] Preferably, in step S1, the data cleaning includes: using the 3σ criterion to remove outliers with amplitudes exceeding the mean ± 3 times the standard deviation; removing invalid data corresponding to downtime, inverter failure periods, and extreme weather periods; the extreme weather periods include periods with irradiance below 50W / m² or wind speeds exceeding 25m / s; using forward filling to complete data segments with consecutive missing durations not exceeding 5 minutes, and marking data segments with consecutive missing durations exceeding 5 minutes as invalid and removing them.
[0011] Preferably, in step S2, the healthy photovoltaic strings in the dynamic peer reference group simultaneously meet the following conditions: they are located in the same power station and the same area as the target photovoltaic string; the spatial distance between them and the target photovoltaic string is no more than 50m; the difference in commissioning time between them and the target photovoltaic string is no more than 30 days; the component model and inverter model are the same as those of the target photovoltaic string; no attenuation or fault records are detected in the historical operating data; and the number of healthy photovoltaic strings in the dynamic peer reference group is no less than 3.
[0012] Preferably, in step S2, the healthy power benchmark model is trained using the LightGBM algorithm, with irradiance, module backsheet temperature, ambient wind speed, and installation tilt angle as input features, and outputs a benchmark power prediction value for the target photovoltaic string under the same operating conditions. The benchmark power prediction value is expressed according to the following formula: ; In the formula, Represents a series of baseline power prediction values. This indicates the reference power prediction sequence at the sampling time. The value of , The mapping function for the healthy power benchmark model. Sampling time The corresponding irradiance, Sampling time The corresponding component backplane temperature, Sampling time Corresponding ambient wind speed, For installation tilt angle.
[0013] Preferably, the LightGBM algorithm has a learning rate of 0.05, a number of trees of 500, a maximum depth of 6, a minimum number of samples per leaf node of 20, an L2 regularization parameter of 1.0, and a feature sampling ratio of 0.8; the loss function is mean squared error, and the training termination condition is that the validation set loss no longer decreases after 10 consecutive rounds; the training samples are no less than 30 consecutive days of fault-free historical operating data.
[0014] Preferably, in step S3, the residual sequence is calculated according to the following formula: ; In the formula, Sampling time The corresponding power residual, Sampling time The corresponding target photovoltaic string actual output power, Sampling time The corresponding baseline power prediction value.
[0015] Preferably, in step S3, the trend feature is the cumulative negative deviation of the residuals within a 7-day evaluation window, and only the negative residuals are accumulated, calculated according to the following formula: ; In the formula, As a trend feature, This is the set of sampling times within a continuous 7-day evaluation window. Sampling time, Sampling time The corresponding power residual, Indicates the power residual Take the smaller value between 0 and 0. This represents the time interval between adjacent sampling times.
[0016] Preferably, in step S3, the fluctuation characteristic is the coefficient of variation of the residual sequence, calculated according to the following formula: ; In the formula, It exhibits fluctuation characteristics. To evaluate the standard deviation of the power residuals within the window, To evaluate the mean of the power residual within the window, for The absolute value of.
[0017] Preferably, in step S3, the relative residual is calculated first, and then the persistence feature is calculated. The relative residual and the persistence feature are calculated according to the following formula: ; ; In the formula, Sampling time The corresponding relative residuals, Sampling time The corresponding power residual, Sampling time The corresponding baseline power prediction value, As a persistent feature, To evaluate the relative residuals within the window The cumulative duration of less than -3%, To assess the total duration of the window.
[0018] Preferably, in step S4, the joint determination of the three-dimensional residuals includes: when trend characteristics are simultaneously satisfied... Less than -5kWh, fluctuation characteristics Less than 10% and persistent characteristics When the rate is greater than 70%, the target photovoltaic string is determined to have permanent implicit degradation; when the trend characteristics Less than -5kWh and fluctuating characteristics When the shading rate is greater than 20%, the target photovoltaic string is determined to be affected by temporary shading; when the fluctuation characteristics are... When it is in the range of 10% to 20%, combine the trend characteristics. and persistent characteristics A comprehensive judgment will be made.
[0019] Preferably, in step S4, the health grading model is trained using the XGBoost algorithm, and the input feature vector of the health grading model is constructed according to the following formula: ; In the formula, This is the input feature vector for the health grading model. As a trend feature, It exhibits fluctuation characteristics. This is a persistent characteristic.
[0020] Preferably, the XGBoost algorithm has a learning rate of 0.1, a number of trees of 300, a maximum depth of 4, a minimum leaf node weight of 3, a subsampling ratio of 0.8, and a column sampling ratio of 0.8. The training parameters of the XGBoost algorithm include the learning rate, the number of trees, the maximum depth, the minimum leaf node weight, the subsampling ratio, the column sampling ratio, and the loss function, wherein the loss function adopts logarithmic loss. The attenuation levels include normal, mild attenuation, moderate attenuation, and severe attenuation, wherein normal corresponds to an attenuation amplitude of less than 3%, mild attenuation corresponds to an attenuation amplitude of 3% to 5%, moderate attenuation corresponds to an attenuation amplitude of 5% to 10%, and severe attenuation corresponds to an attenuation amplitude of greater than 10%. When the target photovoltaic string is determined to be at a mild attenuation level or above for 7 consecutive days, an early warning mechanism is triggered and the corresponding string number is located.
[0021] Preferably, in step S5, the periodic online self-updating includes: extracting healthy photovoltaic string data confirmed to be fault-free by operation and maintenance within the past week during the low-power period of the power plant each week, and incrementally training the healthy power benchmark model; retraining the health classification model quarterly using labeled data from the past 3 months; after incremental training of the healthy power benchmark model, if the validation set error satisfies the following formula, then rolling back to the previous version of the model and triggering manual review: ; In the formula, This represents the mean squared error of the model on the validation set after incremental training. This represents the mean squared error of the previous version of the model on the validation set.
[0022] An early warning system for latent degradation of photovoltaic strings based on time-series feature decoupling is disclosed. This system employs the aforementioned method for early warning of latent degradation of photovoltaic strings based on time-series feature decoupling, comprising a data acquisition module, a data processing module, a dynamic peer reference group construction module, a baseline model construction module, a residual feature extraction module, a health grading early warning module, and a model self-updating module, which are sequentially and interactively linked. Specifically: the data acquisition module acquires runtime time-series data and corresponding environmental time-series data of the photovoltaic strings; the data processing module performs timestamp alignment, outlier removal, invalid period removal, and missing data processing on the runtime time-series data and environmental time-series data; the dynamic peer reference group construction module is used to... The system selects adjacent healthy photovoltaic (PV) strings based on spatial location, commissioning time, equipment model, and historical operating status, and constructs a dynamic peer reference group. A benchmark model construction module trains a healthy power benchmark model based on the dynamic peer reference group and outputs the benchmark power prediction value. A residual feature extraction module calculates the residual sequence between the actual output power of the target PV string and the benchmark power prediction value, and extracts trend features, fluctuation features, and persistence features. A health grading and early warning module outputs the attenuation level based on the three-dimensional residual joint judgment result and the health grading model, and outputs early warning information when the early warning trigger conditions are met. A model self-updating module periodically updates the healthy power benchmark model and the health grading model.
[0023] Preferably, the data acquisition module connects to the string inverter via the Modbus TCP protocol and to the weather station via the HTTP protocol; the system is deployed on the power plant's local edge gateway or cloud operation and maintenance platform, and pushes early warning information through at least one of the following methods: WeChat Work, DingTalk robot, SMS gateway, or platform messages.
[0024] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the method for early warning of latent degradation of photovoltaic strings based on time-series feature decoupling.
[0025] The present invention can achieve the following beneficial effects: 1. This invention acquires runtime timing data such as string current, string voltage, and DC power of the target photovoltaic string, as well as environmental timing data such as irradiance, module backsheet temperature, ambient wind speed, and installation tilt angle. It then performs timestamp alignment, outlier removal, invalid time period removal, and missing data processing to form cleaned minute-level timing data, providing a unified and continuous data foundation for subsequent power benchmark modeling and residual analysis.
[0026] 2. By removing data corresponding to downtime, inverter failure, and extreme weather conditions such as low irradiance and high wind speed, this invention can reduce the interference of non-string attenuation factors on model training and early warning judgment, and avoid mistaking equipment downtime, inverter failure, or abnormal environmental conditions as string latent attenuation for analysis.
[0027] 3. Based on the spatial location, commissioning time, module model, inverter model and historical operating status of the target photovoltaic string, the present invention constructs a dynamic peer reference group from adjacent healthy photovoltaic strings, so that the target photovoltaic string and the reference group are consistent or similar in terms of installation area, equipment configuration, commissioning stage and operating status, thereby reducing the comparison error caused by differences in equipment, installation and natural aging stages between different strings.
[0028] 4. This invention trains a healthy power benchmark model based on a dynamic peer reference group, and uses irradiance, module backsheet temperature, ambient wind speed and installation tilt angle as input features to output the benchmark power prediction value of the target photovoltaic string under the current environmental conditions. This allows the string's healthy output benchmark to change with actual environmental conditions, avoiding the problem that fixed power thresholds or fixed current thresholds are difficult to adapt to meteorological fluctuations.
[0029] 5. This invention uses the LightGBM algorithm to construct a healthy power benchmark model, which can fit the relationship between nonlinear environmental factors such as irradiance, temperature, wind speed and installation tilt angle and healthy output power, so that the difference between the actual output power and the benchmark power prediction value can better reflect the operating status of the target photovoltaic string itself.
[0030] 6. This invention separates string output anomalies from the original power sequence by calculating the residual sequence between the actual output power of the target photovoltaic string and the predicted baseline power. This allows subsequent analysis to no longer directly rely on the magnitude of the original output power, but instead use the degree of deviation under the same healthy operating conditions as the basis for judgment.
[0031] 7. This invention extracts trend features, fluctuation features, and persistence features from the residual sequence. The trend features are used to characterize the cumulative negative deviation of the residual, the fluctuation features are used to characterize the dispersion of the residual change, and the persistence features are used to characterize the proportion of time that the relative residual is continuously lower than a set proportion, so that the latent decay identification can simultaneously consider the deviation magnitude, stability, and duration.
[0032] 8. This invention performs a three-dimensional residual joint determination based on trend characteristics, fluctuation characteristics, and persistence characteristics, which can distinguish between temporary occlusion and permanent latent decay. For cases where the trend deviates significantly but the fluctuation is large, it can be determined as temporary occlusion or short-term disturbance; for cases where the trend deviates significantly but the fluctuation is small and the duration is long, it can be determined as permanent latent decay, thereby reducing false alarms caused by a single threshold judgment.
[0033] 9. This invention inputs three-dimensional residual features into the health grading model and outputs attenuation levels such as normal, mild attenuation, moderate attenuation, and severe attenuation. This makes the early warning results no longer just a binary judgment of whether it is abnormal, but can reflect the attenuation degree of the target photovoltaic string, providing a grading basis for operation and maintenance personnel to formulate measures such as cleaning, inspection, testing, or component replacement.
[0034] 10. The present invention sets up a mechanism to output early warning information only after the early warning triggering conditions are met continuously. When the target photovoltaic string is judged to be at the level of mild degradation or above for several consecutive days, the string-level early warning is triggered, which can reduce occasional alarms caused by single-day abnormalities, short-term weather disturbances or temporary shading.
[0035] 11. The early warning information output by this invention includes string number, string location, attenuation level and operation and maintenance suggestions. It can locate the early warning result to a specific target photovoltaic string, which makes it easier for operation and maintenance personnel to quickly identify the target on site and reduce the workload of string-by-string inspection in large-scale photovoltaic power plants.
[0036] 12. This invention uses a periodic online self-updating mechanism to incrementally train the healthy power benchmark model weekly using healthy photovoltaic string data that has been confirmed to be fault-free by operation and maintenance, and retrains the health grading model quarterly using labeled data, so that the model can adapt to seasonal changes, natural aging of components, and slow changes in the operating conditions of the power plant.
[0037] 13. This invention sets up a validation set error judgment and model rollback mechanism after incremental training of the health power benchmark model. When the model error after incremental training is significantly higher than the error of the previous version model, it automatically rolls back to the previous version model and triggers manual review, which can reduce the risk of model benchmark drift caused by abnormal samples entering the training set during online update.
[0038] 14. This invention can be deployed on the local edge gateway of the power station or the cloud operation and maintenance platform, and can connect to string inverters via Modbus TCP protocol and power station weather station via HTTP protocol. No additional string-level dedicated detection hardware is required, which facilitates software deployment on the basis of existing photovoltaic power station monitoring system.
[0039] 15. This invention, through the collaborative work of a data acquisition module, a data processing module, a dynamic peer reference group construction module, a benchmark model construction module, a residual feature extraction module, a health grading and early warning module, and a model self-updating module, forms a closed-loop process from minute-level data acquisition, health benchmark modeling, residual feature analysis, attenuation grading to early warning push, which is suitable for long-term online monitoring of implicit attenuation at the string level in photovoltaic power plants. Attached Figure Description
[0040] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 The flowchart shows the photovoltaic string latent attenuation early warning method based on time-series feature decoupling as described in this invention. Figure 2 This is a schematic diagram illustrating the principle of dynamic peer reference group construction and health benchmark model training described in this invention. Figure 3 This is a schematic diagram of the logic for residual feature extraction and attenuation classification early warning as described in this invention; Figure 4 This is a diagram of the system deployment architecture described in this invention. Detailed Implementation
[0041] like Figures 1 to 4 As shown, this invention provides an early warning method for the implicit degradation of photovoltaic strings based on time-series feature decoupling. It is applicable to centralized photovoltaic power plants, distributed photovoltaic power plants, and photovoltaic power generation systems using string inverters for operational monitoring. This method uses the operational time-series data of the target photovoltaic string and the corresponding environmental time-series data as a basis. By constructing a dynamic peer reference group, it trains a healthy power benchmark model (hereinafter referred to as the healthy power benchmark model) decoupled from environmental factors to obtain the benchmark power prediction value of the target photovoltaic string under the current environmental conditions. Then, it extracts trend features, fluctuation features, and persistence features based on the residual sequence between the actual output power of the target photovoltaic string and the benchmark power prediction value. Finally, it outputs the degradation level through a three-dimensional residual joint judgment and a health grading model, thereby providing an early warning of the implicit degradation of the photovoltaic string.
[0042] In this embodiment, the target photovoltaic string refers to the photovoltaic string to be identified and warned of implicit degradation; the dynamic peer reference group refers to the healthy reference group dynamically selected from adjacent healthy photovoltaic strings based on the spatial location, commissioning time, module model, inverter model and historical operating status of the target photovoltaic string; the healthy power benchmark model refers to the model trained based on the dynamic peer reference group, used to output the benchmark power prediction value of the target photovoltaic string under the current environmental conditions; the residual sequence refers to the difference sequence between the actual output power of the target photovoltaic string and the benchmark power prediction value; the trend feature, fluctuation feature and persistence feature together constitute the three-dimensional residual feature, which is used to distinguish between temporary shading and permanent implicit degradation.
[0043] S1. Acquisition and Preprocessing of Multi-Source Time Series Data The runtime timing data and corresponding environmental timing data of the target photovoltaic string are obtained. The runtime timing data and environmental timing data are then subjected to timestamp alignment, outlier removal, invalid time period removal, and missing data processing to form cleaned minute-level timing data.
[0044] The runtime timing data includes string current, string voltage, and DC power. As one implementation, the runtime timing data is acquired from the string inverter or power plant data acquisition gateway via the Modbus TCP protocol, with a sampling frequency of 1 minute / time. The environmental timing data includes irradiance, module backsheet temperature, ambient wind speed, and installation tilt angle. As one implementation, the environmental timing data is acquired from the power plant weather station via the HTTP protocol, with a sampling frequency of 1 minute / time.
[0045] During data acquisition, the string current is denoted as... The string voltage is denoted as The actual output power of the target photovoltaic string is denoted as Irradiance is denoted as The component backsheet temperature is denoted as The ambient wind speed is recorded as The installation tilt angle is recorded as ,in Indicates the sampling time. For scenarios where the installation tilt angle remains constant during the monitoring period, As a fixed input feature, it participates in the subsequent calculation of the health power benchmark model; for scenarios with tracking brackets, It can be updated synchronously according to the sampling time.
[0046] When aligning runtime timing data and environmental timing data with timestamps, minute-level timestamps are used as a unified index to merge string current, string voltage, DC power, irradiance, module backplane temperature, ambient wind speed, and installation tilt angle at the same sampling time into a single timing sample. For timestamp offsets caused by communication delays, alignment can be performed using the nearest-neighbor timestamp matching method; when the time offset exceeds a preset allowable range, the corresponding data is marked as invalid.
[0047] Outlier removal is performed using the 3σ criterion. For runtime or environmental time-series data within the same monitoring window, the mean and standard deviation are calculated, and data points with amplitudes exceeding the mean ± 3 times the standard deviation are removed. Invalid time period removal includes removing data corresponding to downtime, inverter failure periods, and extreme weather periods. Extreme weather periods include periods with irradiance below 50 W / m² or ambient wind speeds exceeding 25 m / s.
[0048] Missing data processing includes: for data segments with consecutive missing durations not exceeding 5 minutes, forward imputation is used to complete them; for data segments with consecutive missing durations exceeding 5 minutes, they are marked as invalid and removed. After the above processing, cleaned minute-level time-series data is formed, which serves as the data foundation for subsequent dynamic peer reference group construction, health power benchmark model training, and residual feature extraction.
[0049] S2. Construction of Dynamic Peer Reference Groups and Training of Health Power Benchmark Model Based on the spatial location, commissioning time, module model, inverter model, and historical operating status of the target photovoltaic string, a dynamic peer reference group is constructed from adjacent healthy photovoltaic strings, and a healthy power benchmark model is trained based on the dynamic peer reference group.
[0050] Healthy photovoltaic (PV) strings in the dynamic peer reference group must simultaneously meet the following conditions: located in the same power station and area as the target PV string; spatial distance from the target PV string not exceeding 50m; commissioning time difference from the target PV string not exceeding 30 days; component and inverter models identical to the target PV string; no degradation or fault records detected in historical operating data; and at least three healthy PV strings in the dynamic peer reference group. These constraints ensure that the dynamic peer reference group and the target PV string share the same or nearly identical environmental conditions, equipment configuration, and natural aging stage, thereby reducing the interference of environmental and equipment differences on degradation identification.
[0051] The healthy power benchmark model is trained using the LightGBM algorithm, taking irradiance, module backsheet temperature, ambient wind speed, and installation tilt angle as input features, and outputting a benchmark power prediction value for the target photovoltaic string under the same operating conditions. The benchmark power prediction value sequence is denoted as... ,in Expressed as follows:
[0052] In the formula, The reference power prediction sequence at sampling time The value of ; The mapping function for the healthy power benchmark model; Sampling time The corresponding irradiance; Sampling time Corresponding component backplane temperature; Sampling time Corresponding ambient wind speed; For installation tilt angle.
[0053] To ensure the model has stable environmental factor decoupling capabilities, the LightGBM algorithm's learning rate was set to 0.05, the number of trees to 500, the maximum depth to 6, the minimum number of samples per leaf node to 20, the L2 regularization parameter to 1.0, and the feature sampling ratio to 0.8. The mean squared error loss function was used, and training terminated when the validation set loss no longer decreased after 10 consecutive rounds. Training samples consisted of at least 30 consecutive days of fault-free historical data.
[0054] After training, the healthy power baseline model is used to output a baseline power prediction value that the target photovoltaic string should achieve under the current environmental conditions, based on the current irradiance, module backsheet temperature, ambient wind speed, and installation tilt angle. This baseline power prediction value is not the rated power of the target photovoltaic string, but rather a healthy output baseline under the same operating conditions and health reference group status.
[0055] S3. Residual sequence calculation and extraction of trend characteristics, volatility characteristics, and persistence characteristics. The target photovoltaic string is output as a baseline power prediction value under the current environmental conditions using a healthy power baseline model. The residual sequence between the actual output power of the target photovoltaic string and the baseline power prediction value is calculated, and trend features, fluctuation features, and persistence features are extracted from the residual sequence.
[0056] The residual sequence is calculated according to the following formula:
[0057] In the formula, Sampling time The corresponding power residual; Sampling time The corresponding target photovoltaic string's actual output power; Sampling time The corresponding baseline power prediction value.
[0058] when When the value is consistently negative and its absolute value gradually increases, it indicates that the actual output power of the target photovoltaic string is consistently lower than the predicted baseline power under the same operating conditions, suggesting potential hidden degradation. When there are short-term, drastic fluctuations, it usually corresponds to temporary obstruction, cloud disturbance, or instantaneous environmental changes.
[0059] The trend characteristic is the cumulative negative deviation of the residuals within a 7-day evaluation window, and only negative residuals are accumulated, calculated according to the following formula:
[0060] In the formula, As a trend feature; This is a set of sampling times within a continuous 7-day evaluation window; Sampling time; Sampling time The corresponding power residual; Indicates the power residual Take the smaller value between 0 and 0; This represents the time interval between adjacent sampling times.
[0061] The volatility characteristic is the coefficient of variation of the residual sequence, calculated according to the following formula:
[0062] In the formula, It exhibits fluctuation characteristics; To evaluate the standard deviation of the power residuals within the window; To evaluate the mean of the power residual within the window; for The absolute value. When participating in threshold comparison, Expressed as a percentage.
[0063] First, calculate the relative residuals, then calculate the persistence characteristics. The relative residuals and persistence characteristics are calculated according to the following formulas:
[0064]
[0065] In the formula, Sampling time The corresponding relative residuals; Sampling time The corresponding power residual; Sampling time The corresponding baseline power prediction value; It is a persistent feature; To evaluate the relative residuals within the window The cumulative duration of being below -3%; To assess the total duration of the window.
[0066] Trend features, volatility features, and persistence features are used to characterize the cumulative negative trend, volatility, and persistence of the residuals, respectively. Using trend features alone makes it difficult to distinguish between persistent occlusion and permanent implicit decay; using volatility features alone makes it difficult to identify small, slow decays; and using persistence features alone makes it difficult to eliminate occasional interference. Therefore, this invention combines these three types of residual features for subsequent determination.
[0067] S4, Output of Three-Dimensional Residual Joint Judgment and Health Classification Model Based on the trend characteristics, fluctuation characteristics, and persistence characteristics, a three-dimensional residual joint determination is performed to distinguish between temporary shading and permanent latent attenuation. The three-dimensional residual characteristics are then input into the health grading model to output the attenuation level of the target photovoltaic string.
[0068] Joint determination of three-dimensional residuals includes: when trend characteristics are simultaneously satisfied Less than -5kWh, fluctuation characteristics Less than 10% and persistent characteristics When the rate is greater than 70%, the target photovoltaic string is determined to have permanent implicit degradation; when the trend characteristics Less than -5kWh and fluctuating characteristics When the shading rate is greater than 20%, the target photovoltaic string is determined to be affected by temporary shading; when the fluctuation characteristics are... When it is in the range of 10% to 20%, combine the trend characteristics. and persistent characteristics A comprehensive judgment will be made.
[0069] ; In the formula, This is the input feature vector for the health grading model. As a trend feature, It exhibits fluctuation characteristics. This is a persistent characteristic.
[0070] The XGBoost algorithm uses a learning rate of 0.1, 300 trees, a maximum depth of 4, a minimum leaf node weight of 3, a subsampling ratio of 0.8, a column sampling ratio of 0.8, and a logarithmic loss function. The health grading model outputs attenuation levels of normal, mild attenuation, moderate attenuation, and severe attenuation. Normal attenuation corresponds to an attenuation of less than 3%, mild attenuation to 3% to 5%, moderate attenuation to 5% to 10%, and severe attenuation to greater than 10%. When a target photovoltaic string is classified as mildly attenuated or worse for seven consecutive days, an early warning mechanism is triggered, and the corresponding string number is located.
[0071] As one implementation method, the health grading model can construct a training set using operation and maintenance confirmation data, historical fault work order data, and manual review labels. For each training sample, trend features, fluctuation features, and persistence features are extracted as input features, and the corresponding attenuation level is used as the output label for training. After training, the health grading model is used to perform string-level health grading of target photovoltaic strings.
[0072] S5, Early Warning Information Output and Periodic Online Self-Update of the Model When the target photovoltaic string continuously meets the early warning triggering conditions, an early warning message containing the string number, string location, attenuation level and operation and maintenance suggestions is output, and the healthy power benchmark model and health classification model are periodically updated online.
[0073] When a target photovoltaic (PV) string is assessed as experiencing mild degradation or higher for seven consecutive days, an early warning mechanism is triggered, and the corresponding string number is located. The early warning information includes the string number, string location, degradation level, current trend characteristics, current fluctuation characteristics, current ongoing characteristics, and maintenance recommendations. The early warning information can be pushed to maintenance personnel via at least one of the following methods: WeChat Work, DingTalk robot, SMS gateway, or platform message.
[0074] The periodic online self-updating includes: extracting healthy photovoltaic string data that has been confirmed to be fault-free by operation and maintenance within the past week during the low power period of the power plant every week, and incrementally training the healthy power benchmark model; and retraining the health classification model every quarter using labeled data from the past 3 months.
[0075] After incremental training of the healthy power baseline model, if the validation set error satisfies the following formula, the model will be rolled back to the previous version and manual review will be triggered:
[0076] In the formula, This represents the mean squared error of the model on the validation set after incremental training. This represents the mean squared error of the previous version of the model on the validation set.
[0077] The above-mentioned online self-updating mechanism can adapt the healthy power benchmark model to seasonal changes, natural aging of components, and slow drift of power plant operating conditions. At the same time, the error rollback condition avoids the risk of benchmark drift caused by model self-updating.
[0078] like Figure 4 As shown, the present invention also provides an early warning system for latent degradation of photovoltaic strings based on time-series feature decoupling. The method for early warning of latent degradation of photovoltaic strings based on time-series feature decoupling includes a data acquisition module, a data processing module, a dynamic peer reference group construction module, a benchmark model construction module, a residual feature extraction module, a health classification early warning module, and a model self-updating module that are sequentially and interactively linked.
[0079] The data acquisition module is used to acquire the runtime timing data and corresponding environmental timing data of the photovoltaic string. In one implementation, the data acquisition module connects to the string inverter via the Modbus TCP protocol to collect string current, string voltage, and DC power; and connects to a weather station via the HTTP protocol to collect irradiance, module backsheet temperature, ambient wind speed, and installation tilt angle.
[0080] The data processing module performs timestamp alignment, outlier removal, invalid period removal, and missing data processing on runtime and environmental time-series data. The dynamic peer reference group construction module filters adjacent healthy PV strings based on spatial location, commissioning time, equipment model, and historical operating status, and constructs a dynamic peer reference group. The benchmark model construction module trains a healthy power benchmark model based on the dynamic peer reference group and outputs the benchmark power prediction value.
[0081] The residual feature extraction module calculates the residual sequence between the actual output power of the target photovoltaic string and the predicted baseline power, and extracts trend features, fluctuation features, and persistence features. The health grading early warning module outputs the attenuation level based on the joint judgment results of the three-dimensional residuals and the health grading model, and outputs early warning information when the early warning trigger conditions are met. The model self-updating module periodically updates the healthy power baseline model and the health grading model.
[0082] The system can be deployed on a local edge gateway at the power plant or on a cloud-based operations and maintenance platform. When deployed on a local edge gateway at the power plant, data acquisition, data cleaning, model inference, and early warning triggering are all completed locally. When deployed on a cloud-based operations and maintenance platform, data from the power plant side is uploaded to the cloud via the edge gateway, and model inference and early warning push are performed by the cloud.
[0083] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the aforementioned method for early warning of latent degradation of photovoltaic strings based on time-series feature decoupling. The computer-readable storage medium may be a hard disk, solid-state drive, read-only memory, random access memory, removable storage device, or other medium capable of storing a computer program.
[0084] The computer program may include a data acquisition program, a data cleaning program, a dynamic peer reference group construction program, a healthy power benchmark model training program, a residual feature extraction program, a health grading and early warning program, and a model self-updating program. When the processor executes the computer program, it can complete minute-level time series data processing, benchmark power prediction output, residual sequence calculation, attenuation level determination, and early warning information push.
[0085] Example 1: Local Edge Gateway Deployment Example for Photovoltaic Power Plants This embodiment uses a centralized photovoltaic power station as an example. The power station is equipped with string inverters. The runtime timing data of the target photovoltaic string is collected from the inverters via the Modbus TCP protocol, while environmental timing data is collected from the power station's weather station via the HTTP protocol. The data sampling frequency is 1 minute / time.
[0086] The system is deployed on the local edge gateway of the power plant. The edge gateway is configured with data acquisition service, data processing service, LightGBM health power benchmark model service, XGBoost health classification model service, and early warning push service. The data processing service performs timestamp alignment, 3σ outlier removal, outage period removal, inverter fault period removal, extreme weather period removal, and missing data processing on the acquired data.
[0087] For each target photovoltaic string, the system selects a dynamic peer reference group based on spatial distance, commissioning time, module model, inverter model, and historical operating status. The dynamic peer reference group must contain at least three healthy photovoltaic strings. The benchmark model building module trains a healthy power benchmark model using at least 30 consecutive days of fault-free historical operating data and outputs the benchmark power prediction value for each sampling moment during real-time operation.
[0088] During the real-time early warning phase, the residual feature extraction module calculates the power residual and extracts trend features, fluctuation features, and persistence features over a continuous 7-day evaluation window. The health grading early warning module outputs the attenuation level based on the joint judgment results of the three-dimensional residuals and the health grading model. When the target photovoltaic string is judged to be at a mild attenuation level or above for 7 consecutive days, the system outputs the string number, string location, attenuation level, and operation and maintenance suggestions, and pushes them to the operation and maintenance personnel via platform messages.
[0089] Example 2: Residual Feature Calculation and Occlusion Detection Example This embodiment uses data from a target photovoltaic (PV) string within a 7-day evaluation window as an example to illustrate the residual characteristic calculation and discrimination process. The residual sequence of a target PV string within the same diurnal sampling period for 7 consecutive days is represented as follows:
[0090] In the formula, It is a residual sequence composed of the power residuals corresponding to several sampling times within a continuous 7-day evaluation window; Sampling time; Sampling time The corresponding power residual.
[0091] Trend characteristics are calculated from the residual sequence. ,satisfy Conditions. Calculate the mean and standard deviation of the power residuals within the evaluation window to obtain... , Thus, the wave characteristics are obtained. ,satisfy Condition. The proportion of the cumulative duration of relative residuals below -3% to the total duration of the assessment window is [condition]. Then the continuous feature satisfies condition.
[0092] According to the three-dimensional residual joint judgment rule, although both the trend feature and the persistence feature meet the condition for suspected attenuation, the fluctuation feature... A residual value greater than 20% indicates severe fluctuations in the residual value. Therefore, the target photovoltaic string is determined to be affected by temporary shading, rather than being directly classified as a permanent, latent degradation. This determination avoids false alarms caused by cloud cover, bird droppings, or temporary shadows.
[0093] Example 3: Application Example of Early Warning of Latent Attenuation This embodiment uses a 100MW centralized photovoltaic power station as an example. The power station is equipped with 200 string inverters, each connected to 20 photovoltaic strings, for a total of 4000 photovoltaic strings. The system collects runtime timing data of each photovoltaic string and environmental timing data from a weather station at a sampling frequency of 1 minute per instance. For each photovoltaic string, dynamic peer reference group construction, healthy power benchmark model prediction, residual feature extraction, and health classification early warning are performed.
[0094] During continuous operation, the system generates a corresponding baseline power prediction value and residual sequence for each target photovoltaic string, and outputs a normal, mildly degraded, moderately degraded, or severely degraded result through a health grading model. For target photovoltaic strings that are judged to be mildly degraded or worse for 7 consecutive days, the system automatically triggers an early warning and pushes the warning information to the operation and maintenance platform.
[0095] Without shutting down the system, it can identify early-stage, latently degraded photovoltaic strings with a degradation rate of 3% to 5%. For photovoltaic strings identified as having mild degradation, maintenance personnel can further confirm the cause of degradation by combining on-site infrared detection, module appearance inspection, electroluminescence detection, or insulation testing, and perform maintenance measures such as cleaning, wiring troubleshooting, repairing bypass diodes, or replacing modules based on the warning information.
[0096] The above embodiments and examples are used to illustrate the technical solution of the present invention. For photovoltaic power plants of different capacity levels, different component types, different inverter manufacturers, or different operation and maintenance platforms, the dynamic peer reference group screening conditions, healthy power benchmark model training cycle, healthy grading model training samples, and early warning push methods can be adaptively adjusted according to the power plant data interface, string arrangement density, and operation and maintenance management requirements.
Claims
1. A method for early warning of implicit degradation of photovoltaic strings based on time-series feature decoupling, characterized in that, Includes the following steps: S1. Obtain the runtime timing data and corresponding environmental timing data of the target photovoltaic string, and perform timestamp alignment, outlier removal, invalid time period removal and missing data processing on the runtime timing data and environmental timing data to form cleaned minute-level timing data; S2. Based on the spatial location, commissioning time, module model, inverter model and historical operating status of the target photovoltaic string, construct a dynamic peer reference group from adjacent healthy photovoltaic strings, and train a healthy power benchmark model decoupled from environmental factors based on the dynamic peer reference group, hereinafter referred to as the healthy power benchmark model. S3. Using the healthy power benchmark model, output the benchmark power prediction value of the target photovoltaic string under the current environmental conditions, calculate the residual sequence between the actual output power of the target photovoltaic string and the benchmark power prediction value, and extract trend features, fluctuation features and persistence features from the residual sequence. S4. Perform a three-dimensional residual joint determination based on the trend characteristics, fluctuation characteristics and persistence characteristics to distinguish between temporary shading and permanent hidden attenuation, and input the three-dimensional residual characteristics into the health grading model to output the attenuation level of the target photovoltaic string; S5. When the target photovoltaic string continuously meets the early warning triggering conditions, output early warning information including string number, string location, attenuation level and operation and maintenance suggestions, and perform periodic online self-updates on the healthy power benchmark model and health classification model.
2. The method for early warning of latent degradation of photovoltaic strings based on time-series feature decoupling according to claim 1, characterized in that, In step S1, the runtime timing data includes string current, string voltage, and DC power; the environmental timing data includes irradiance, module backplane temperature, ambient wind speed, and installation tilt angle; the runtime timing data is collected via Modbus TCP protocol at a sampling frequency of 1 minute / time, and the environmental timing data is collected from the power plant meteorological station via HTTP protocol at a sampling frequency of 1 minute / time.
3. The method for early warning of implicit degradation of photovoltaic strings based on time-series feature decoupling according to claim 2, characterized in that, In step S1, the data cleaning includes: using the 3σ criterion to remove outliers with amplitudes exceeding the mean ± 3 times the standard deviation; removing invalid data corresponding to downtime, inverter failure periods, and extreme weather periods; the extreme weather periods include periods with irradiance below 50W / m² or wind speeds exceeding 25m / s; using forward filling to complete data segments with consecutive missing durations not exceeding 5 minutes, and marking data segments with consecutive missing durations exceeding 5 minutes as invalid and removing them.
4. The method for early warning of implicit degradation of photovoltaic strings based on time-series feature decoupling according to claim 1, characterized in that, In step S2, the healthy photovoltaic strings in the dynamic peer reference group simultaneously meet the following conditions: they are located in the same power station and the same area as the target photovoltaic string; the spatial distance between them and the target photovoltaic string is no more than 50m; the difference in commissioning time between them and the target photovoltaic string is no more than 30 days; the component model and inverter model are the same as those of the target photovoltaic string; no attenuation or fault records are detected in the historical operating data; and the number of healthy photovoltaic strings in the dynamic peer reference group is no less than 3.
5. The method for early warning of implicit degradation of photovoltaic strings based on time-series feature decoupling according to claim 2, characterized in that, In step S2, the healthy power benchmark model is trained using the LightGBM algorithm, with irradiance, module backsheet temperature, ambient wind speed, and installation tilt angle as input features. The output is a benchmark power prediction value for the target photovoltaic string under the same operating conditions, expressed by the following formula: ; In the formula, Represents a series of baseline power prediction values. This indicates the reference power prediction sequence at the sampling time. The value of , The mapping function for the healthy power benchmark model. Sampling time The corresponding irradiance, Sampling time The corresponding component backplane temperature, Sampling time Corresponding ambient wind speed, For installation tilt angle.
6. The method for early warning of implicit degradation of photovoltaic strings based on time-series feature decoupling according to claim 5, characterized in that, The LightGBM algorithm has a learning rate of 0.05, a number of trees of 500, a maximum depth of 6, a minimum number of samples per leaf node of 20, an L2 regularization parameter of 1.0, and a feature sampling ratio of 0.
8. The loss function is mean squared error, and the training termination condition is that the validation set loss no longer decreases after 10 consecutive rounds. The training samples consist of no less than 30 consecutive days of fault-free historical operating data.
7. The method for early warning of implicit degradation of photovoltaic strings based on time-series feature decoupling according to claim 1, characterized in that, In step S3, the residual sequence is calculated according to the following formula: ; In the formula, Sampling time The corresponding power residual, Sampling time The corresponding target photovoltaic string actual output power, Sampling time The corresponding baseline power prediction value.
8. The method for early warning of implicit degradation of photovoltaic strings based on time-series feature decoupling according to claim 7, characterized in that, In step S3, the trend feature is the cumulative negative deviation of the residuals within a 7-day evaluation window, and only the negative residuals are accumulated, calculated according to the following formula: ; In the formula, As a trend feature, This is the set of sampling times within a continuous 7-day evaluation window. Sampling time, Sampling time The corresponding power residual, Indicates the power residual Take the smaller value between 0 and 0. This represents the time interval between adjacent sampling times.
9. The method for early warning of latent degradation of photovoltaic strings based on time-series feature decoupling according to claim 7, characterized in that, In step S3, the fluctuation characteristic is the coefficient of variation of the residual sequence, calculated according to the following formula: ; In the formula, It exhibits fluctuation characteristics. To evaluate the standard deviation of the power residuals within the window, To evaluate the mean of the power residual within the window, for The absolute value of.
10. The method for early warning of latent degradation of photovoltaic strings based on time-series feature decoupling according to claim 7, characterized in that, In step S3, the relative residual is calculated first, and then the persistence feature is calculated. The relative residual and persistence feature are calculated according to the following formula: ; ; In the formula, Sampling time The corresponding relative residuals, Sampling time The corresponding power residual, Sampling time The corresponding baseline power prediction value, As a persistent feature, To evaluate the relative residuals within the window The cumulative duration of less than -3%, To assess the total duration of the window.
11. The method for early warning of latent degradation of photovoltaic strings based on time-series feature decoupling according to any one of claims 8 to 10, characterized in that, In step S4, the joint determination of the three-dimensional residuals includes: when the trend characteristics are simultaneously satisfied... Less than -5kWh, fluctuation characteristics Less than 10% and persistent characteristics When the rate is greater than 70%, the target photovoltaic string is determined to have permanent implicit degradation; when the trend characteristics Less than -5kWh and fluctuating characteristics When the shading rate is greater than 20%, the target photovoltaic string is determined to be affected by temporary shading; when the fluctuation characteristics are... When it is in the range of 10% to 20%, combine the trend characteristics. and persistent characteristics A comprehensive judgment will be made.
12. The method for early warning of latent degradation of photovoltaic strings based on time-series feature decoupling according to claim 1, characterized in that, In step S4, the health grading model is trained using the XGBoost algorithm, and the input feature vector of the health grading model is constructed according to the following formula: ; In the formula, This is the input feature vector for the health grading model. As a trend feature, It exhibits fluctuation characteristics. This is a persistent characteristic.
13. The method for early warning of latent degradation of photovoltaic strings based on time-series feature decoupling according to claim 12, characterized in that, The XGBoost algorithm has a learning rate of 0.1, a number of trees of 300, a maximum depth of 4, a minimum leaf node weight of 3, a subsampling ratio of 0.8, and a column sampling ratio of 0.
8. The training parameters of the XGBoost algorithm include the learning rate, number of trees, maximum depth, minimum leaf node weight, subsampling ratio, column sampling ratio, and loss function, where the loss function is logarithmic loss. The attenuation levels include normal, mild attenuation, moderate attenuation, and severe attenuation. Normal attenuation corresponds to an attenuation of less than 3%, mild attenuation corresponds to an attenuation of 3% to 5%, moderate attenuation corresponds to an attenuation of 5% to 10%, and severe attenuation corresponds to an attenuation of greater than 10%. When a target photovoltaic string is determined to be at a mild attenuation level or higher for 7 consecutive days, an early warning mechanism is triggered and the corresponding string number is located.
14. The method for early warning of latent degradation of photovoltaic strings based on time-series feature decoupling according to claim 4, characterized in that, In step S5, the periodic online self-updating includes: extracting healthy photovoltaic string data confirmed to be fault-free by operation and maintenance within the past week during the low-power period of the power plant each week, and incrementally training the healthy power benchmark model; retraining the health classification model quarterly using labeled data from the past 3 months; after incremental training of the healthy power benchmark model, if the validation set error satisfies the following formula, then rolling back to the previous version of the model and triggering manual review: ; In the formula, This represents the mean squared error of the model on the validation set after incremental training. This represents the mean squared error of the previous version of the model on the validation set.
15. An early warning system for implicit degradation of photovoltaic strings based on time-series characteristic decoupling, characterized in that, The method for early warning of latent degradation of photovoltaic strings based on time-series feature decoupling, as described in any one of claims 1-14, includes a data acquisition module, a data processing module, a dynamic peer reference group construction module, a benchmark model construction module, a residual feature extraction module, a health grading early warning module, and a model self-updating module that are sequentially and interactively linked: wherein: The data acquisition module is used to acquire the runtime timing data and corresponding environmental timing data of the photovoltaic string; The data processing module is used to perform timestamp alignment, outlier removal, invalid time period removal, and missing data processing on the runtime sequence data and environmental time sequence data. The dynamic peer reference group construction module is used to filter adjacent healthy photovoltaic strings based on spatial location, commissioning time, equipment model and historical operating status, and to construct a dynamic peer reference group. The benchmark model building module is used to train a healthy power benchmark model based on a dynamic peer reference group and output the benchmark power prediction value. The residual feature extraction module is used to calculate the residual sequence between the actual output power of the target photovoltaic string and the predicted reference power, and to extract trend features, fluctuation features and persistence features. The health grading early warning module is used to output the attenuation level based on the joint judgment result of the three-dimensional residual and the health grading model, and to output early warning information when the early warning triggering conditions are met. The model self-updating module is used to periodically update the health power benchmark model and the health classification model.
16. The photovoltaic string latent degradation early warning system based on time-series feature decoupling according to claim 15, characterized in that, The data acquisition module connects to the string inverter via the Modbus TCP protocol and to the weather station via the HTTP protocol. The system is deployed on the power plant's local edge gateway or cloud operation and maintenance platform, and pushes early warning information through at least one of the following methods: WeChat, DingTalk robot, SMS gateway, or platform messages.
17. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the photovoltaic string latent degradation early warning method based on time-series feature decoupling as described in any one of claims 1 to 14.
