A photovoltaic implicit fault identification and power prediction correction method

Abstract

The present application belongs to the technical field of photovoltaic power generation monitoring, and provides a photovoltaic hidden fault identification and power prediction correction method, which comprises: fusing an infrared micro-temperature difference time sequence characteristic vector and a residual low-frequency drift characteristic vector, identifying the hidden fault type and fault position of a photovoltaic module according to a pre-constructed bimodal fault discrimination model; constructing a decay correction coefficient to generate a corrected photovoltaic power prediction value; updating a parameter to be corrected according to the difference between the corrected photovoltaic power prediction value and the actual power output of a power station at the latest time; constructing a power correction input characteristic vector and performing time sequence stacking, inputting a pre-constructed power correction model to obtain a low-frequency residual compensation sequence in a future period, and inputting the low-frequency residual compensation sequence into a fault decay layer to output a corrected photovoltaic power prediction sequence. The present application can not only identify early hidden faults, but also feed back the fault decay degree to the power prediction correction process, thereby reducing the persistent prediction deviation caused by component hidden faults.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation monitoring technology, specifically to a method for identifying hidden photovoltaic faults and correcting power prediction. Background Technology

[0002] After long-term operation, photovoltaic power plants are subject to various factors such as manufacturing defects, thermal cycling fatigue, mechanical loads, humid and hot environments, deterioration of terminal contacts, bypass diode malfunctions, and localized shading contamination. Over time, these factors can lead to hidden faults such as microcracks in cells, poor solder joints, early hot spots, potential induced degradation, and string mismatch. In the early stages, these faults may only manifest as slight localized temperature differences, a slow decrease in output, or low-frequency drift in the prediction residual. Furthermore, the temperature anomalies and power deviations are not obvious, making them easily mistaken for normal fluctuations by conventional threshold methods.

[0003] Currently, photovoltaic module fault detection mainly relies on periodic infrared inspection, EL (Electroluminescence) graph detection, IV (Infrared) curve testing, and power deviation analysis using monitoring data. Infrared inspection is useful for identifying obvious hot spots, but its stability is insufficient when encountering minor thermal anomalies caused by the combined effects of ambient temperature, irradiance, and wind speed. EL and IV testing mostly require shutdown or offline operation, making them unsuitable for continuous online diagnostics of large-scale power plants. When relying solely on power prediction residuals to identify anomalies, meteorological random errors and module system attenuation errors become intertwined, making it difficult to distinguish whether the cause is rapid weather changes or the gradual aggravation of hidden module faults.

[0004] Furthermore, photovoltaic ultra-short-term power prediction models are primarily built using real-time meteorological data and historical power output data. If a hidden fault in the modules causes the amount of electricity that can be generated to gradually decrease, the basic output parameters set in the original prediction model will find it difficult to keep up with the health status of the modules, potentially leading to continuous overestimation or underestimation. If the fault is discovered and the prediction error cannot be resolved simultaneously, it will affect how the power plant operates, how the grid allocates power, and the accuracy of reporting.

[0005] Therefore, a method is needed to extract minute infrared anomalies under complex environmental interference, remove the bias of the component failure system from the prediction residual, and feed back the failure attenuation to the ultra-short-term power prediction model to identify and correct photovoltaic hidden faults. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method for identifying photovoltaic latent faults and correcting power prediction, thereby overcoming the difficulties such as weak infrared anomalies being easily masked by weather conditions, the difficulty in distinguishing between random weather errors and component fault system errors in prediction residuals, and the inability of fault diagnosis results to reverse-correct the power prediction model.

[0007] This invention provides a method for identifying latent photovoltaic faults and correcting power prediction, comprising: S1. Collect time-series infrared temperature data of photovoltaic modules, actual power output data of power plants, ultra-short-term raw power data and environmental meteorological data to form a multi-source observation vector after confidence weighting. S2. Extract the cell-level region from the infrared temperature image to form an infrared micro-temperature difference time-series feature vector; among them, perform meteorological disturbance attenuation processing on the infrared micro-temperature difference time-series feature vector with a large degree of environmental impact to reduce the degree of environmental impact; S3. Based on the actual power output of the power plant and the ultra-short-term original power prediction value, establish the power prediction residual time series. Use the multi-scale wavelet decomposition algorithm to decompose the power prediction residual time series into low-frequency approximate components and multiple high-frequency detail components at different scales to obtain the residual low-frequency drift feature vector. S4. By fusing the infrared micro-temperature difference time-series feature vector and the residual low-frequency drift feature vector, and based on the pre-constructed dual-mode fault discrimination model, the latent fault type and fault location of the photovoltaic module are identified, and the module fault attenuation index is output. S5. Construct a degradation correction coefficient based on the component failure degradation index and generate a corrected photovoltaic power prediction value; update the parameters to be corrected based on the difference between the corrected photovoltaic power prediction value and the actual power output of the power station at the latest moment. S6. Construct the power correction input feature vector and perform time-series stacking. Input the pre-constructed infrared-residual joint modulation power correction model to obtain the low-frequency residual compensation sequence for more than two future time periods. Input the low-frequency residual compensation sequence into the fault attenuation layer to output the corrected photovoltaic power prediction sequence. S7. Construct a joint optimization loss to update the trainable parameters, calculate the comprehensive latent fault score, and output the corresponding warning level and associated warning information.

[0008] Optionally, S1 includes: Construct a multi-source observation vector, including infrared pixel temperature matrix, actual power plant output, ultra-short-term raw power, irradiance, ambient temperature, wind speed, relative humidity, and module tilt angle; Unify the components of the multi-source observation vector to the same time axis; Confidence weights are established for different data sources, and the corresponding components of each data source after being unified to the same time axis are weighted according to the confidence weights to form a weighted multi-source observation vector. Abnormal samples were removed according to physical consistency rules. These rules included power output constraints when nighttime irradiance was zero, power response range constraints under step changes in irradiance, constraints on the negative correlation between component surface temperature rise and wind speed, and constraints on the non-negativity of predicted power.

[0009] Optionally, S2 includes: Extract the cell-level regions from the infrared temperature image, and calculate the pixel-level infrared micro-temperature difference after environmental compensation for each cell-level region; Calculate the mean, maximum, and drift slope of the infrared micro-temperature difference within a sliding window W. Local temperature gradient Thermal diffusion asymmetry and the proportion of abnormal duration To form an infrared micro-temperature difference time-series feature vector ; Calculate the correlation between the infrared micro-temperature difference time series feature vector and the environmental meteorological variable series. The elements in the environmental meteorological variable series include irradiance, ambient temperature, wind speed, and relative humidity. When the correlation is not greater than the preset meteorological disturbance threshold, the unattenuated infrared micro-temperature difference feature is output. At this time, if the drift duration of the corresponding infrared micro-temperature difference feature reaches or exceeds the preset duration window, the corresponding cell-level area is marked as a suspected area of ​​latent fault. If the drift duration does not reach the preset duration window, the corresponding cell-level area is marked as an area to be observed, and the drift duration continues to accumulate in the subsequent sliding window.

[0010] Optionally, weather synchronization disturbance attenuation includes: The corresponding meteorological disturbance attenuation weight is calculated based on the meteorological correlation. ; , For the preset attenuation intensity coefficient, The meteorological correlation of the k-th infrared micro-temperature difference feature is given. Preset meteorological disturbance threshold; The k-th infrared micro-temperature difference feature is attenuated using the meteorological disturbance attenuation weight; the k-th infrared micro-temperature difference feature after weighted attenuation... .

[0011] Optionally, S3 includes: Based on the actual power output of the power plant and the ultra-short-term raw power forecast, a power forecast residual sequence is established. ; The power prediction residual sequence is decomposed into low-frequency approximate components and multiple high-frequency detail components at different scales. Calculate the correlation between the high-frequency detail components at each scale and each meteorological variable in the environmental meteorological variable series; When the correlation is greater than the preset meteorological correlation threshold, the corresponding high-frequency detail component is identified as a high-frequency meteorological random error component; when the correlation is not greater than the preset meteorological correlation threshold, if the high-frequency detail component or low-frequency approximate component exhibits monotonic drift, periodic slow drift, or persistent negative deviation, the corresponding component is identified as a low-frequency component failure systematic error component; when the correlation is not greater than the preset meteorological correlation threshold, but the high-frequency detail component or low-frequency approximate component does not exhibit monotonic drift, periodic slow drift, or persistent negative deviation, the corresponding scale component is identified as other model error components. Based on the systematic error components of the low-frequency component faults, a residual low-frequency drift feature vector is constructed. .

[0012] Optionally, S4 includes: Establish infrared micro-temperature difference characteristics or With residual low-frequency drift eigenvector The relationship between them; For the m-th cell-level region, calculate the infrared micro-temperature difference abnormal energy based on the environmentally compensated infrared micro-temperature difference. ; Single-scale inverse wavelet reconstruction is performed on each scale component of the systematic error component of low-frequency component faults to obtain the low-frequency fault residual sequence. The ratio of the energy of the low-frequency fault residual sequence to the total energy of the original power prediction residual sequence is calculated to obtain the proportion of low-frequency residual drift energy. ; Based on a pre-established dual-modal fault discrimination model, infrared micro-temperature difference characteristics are analyzed. With residual low-frequency drift eigenvector The data is encoded and then fused with contextual information to output the latent fault type, fault confidence, fault location, and component fault attenuation index. 。

[0013] Optionally, the dual-modal fault discrimination model includes an infrared micro-feature encoding channel, a residual low-frequency drift encoding channel, a gated attention fusion module, and a fault output module; The infrared micro-feature encoding channel and the residual low-frequency drift encoding channel respectively encode infrared micro-temperature difference features. With residual low-frequency drift eigenvector Encode to obtain infrared latent representation and residual latent representation ; Gated attention fusion module based on infrared latent representation Residual latent representation In addition to operating context information, the gating fusion vector is calculated. g ( t According to the gated fusion vector g ( t ), for infrared potential representation and residual latent representation Adaptive weighted fusion is performed to obtain the fused feature representation. ; Fault output module fusion feature representation Output the latent fault types respectively. c ( t Fault confidence p ( t ), location of the fault area loc ( t ) and component failure degradation index DI ( t ).

[0014] Optionally, S5 includes: According to component failure attenuation index DI ( t Establish component failure attenuation correction coefficients. ; Attenuation correction factor Component failure attenuation index DI(t), future low-frequency residual compensation amount for systematic errors in low-frequency component failures. Online bias compensation item and ultra-short-term raw power prediction values Input the fault attenuation correction layer to obtain the corrected photovoltaic power prediction value. ; Based on the deviation between the latest actual power output of the power plant and the corrected predicted power, the parameters to be corrected are... Online updates will be performed; parameters to be corrected include environmental compensation coefficients and online bias compensation terms.

[0015] Optionally, S6 includes: Constructing power-corrected input feature vectors Including infrared micro-temperature difference time-series feature vectors Low-frequency drift eigenvectors of residuals Component failure degradation index DI ( t ), attenuation correction factor Environmental meteorological data vector M ( t Ultra-short-term raw predicted power and historical efforts t represents time. Stack the power-corrected input feature vectors from time t-L+1 to time t to obtain the multi-source time-series input sequence at time t. ; Multi-source time-series input sequence By inputting a pre-constructed infrared-residual joint modulation power correction model, low-frequency residual compensation sequences for two or more future time periods are obtained. ;include: Multi-source time-series input sequence The input feature embedding layer is then superimposed with positional encoding to obtain the embedded input sequence. ; Based on infrared micro-temperature difference anomalous energy Low-frequency residual drift energy ratio Using the fault confidence level p(t) and the component fault attenuation index DI(t), a latent fault influence factor is constructed at time step t. ; The embedded input sequence The projection consists of a query matrix Q, a key matrix K, and a value matrix V; The sparse attention mechanism is used to compute the i-th query vector. Original information content score Utilizing latent fault influencing factors Information content score Modulation is performed to obtain the modulated information content fraction. Information content fraction after modulation Used to filter active query vectors to participate in sparse attention computation; Future generated based on time t to The ultra-short-term raw power prediction values ​​and corresponding time position codes for each time period are used to construct the decoder prediction query sequence. After self-attention, cross-attention, and linear output layer calculations, low-frequency residual compensation sequences for two or more future time periods are obtained. T represents the prediction step size. Low-frequency residual compensation sequence By inputting the fault attenuation correction layer, we obtain corrected photovoltaic power prediction sequences for two or more future time periods. .

[0016] Optionally, S7 includes: Compare the corrected photovoltaic power forecast values ​​for two or more future time periods with the corresponding measured power values ​​to calculate the power forecast correction loss. ; N represents the number of training samples, and T represents the prediction step size. This represents the predicted power of the nth sample after the k-th step correction. This indicates the corresponding measured power; Construct joint optimization loss The trainable parameters are updated based on the joint optimization loss; the joint optimization loss , Indicates the loss by fault type classification. Indicates the loss in fault location. Represents physical constraint loss. This indicates the correction magnitude for smoothing loss. to For loss weights; Component failure attenuation index DI ( t Infrared micro-temperature differential energy Low-frequency residual drift energy ratio Fault confidence p ( t The absolute difference between the corrected predicted power and the original ultra-short-term predicted power. The weighted summation yields the comprehensive latent fault score. S ( t Based on the comprehensive hidden fault score, the system outputs the warning level and the corresponding warning information.

[0017] By adopting the above technical solution, this application has the following beneficial effects: First, this invention does not rely solely on the absolute value threshold of infrared temperature for judgment, but rather performs early detection by observing temporal abnormalities such as the micro-temperature difference drift slope, local temperature gradient, thermal diffusion asymmetry, and abnormal duration. Therefore, it can discover problems such as microcracks, poor soldering, bypass diode abnormalities, and potential induced degradation in battery cells that have not yet generated obvious hot spots. Second, this invention uses multi-scale wavelet decomposition combined with meteorological correlation to judge residual power prediction. This can separate short-period, high-frequency, and highly correlated meteorological random errors from low-frequency, long-term, and less correlated systematic component errors, thereby reducing the occurrence of false prompts and missed alarms caused by rapid weather changes. Third, this invention performs gated attention fusion of infrared local anomaly information and residual global drift information to determine the fault location, fault type and how the fault decays, making it easier for maintenance personnel to handle early faults. Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0019] Figure 1 A flowchart of a photovoltaic latent fault identification and power prediction correction method provided by an embodiment of the present invention is shown; Figure 2 A flowchart of S2 provided in an embodiment of the present invention is shown; Figure 3 A flowchart of S3 provided in an embodiment of the present invention is shown; Figure 4 The flowcharts for S4-S7 provided in the embodiments of the present invention are shown; Figure 5 The diagram shows the effect of photovoltaic power prediction correction according to an embodiment of the present invention; Figure 6 The diagram shows the multi-scale decoupling results of power prediction residuals according to an embodiment of the present invention. Figure 7 The diagram shows the latent fault scoring and early warning results of an embodiment of the present invention. Detailed Implementation

[0020] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are only used to more clearly illustrate the technical solution of the present invention and are therefore merely examples, and should not be construed as limiting the scope of protection of the present invention.

[0021] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0022] In one embodiment, such as Figure 1 As shown, a method for identifying photovoltaic latent faults and correcting power prediction is provided, including: S1. Collect time-series infrared temperature data of photovoltaic modules, actual power output data of power plants, ultra-short-term raw power data and environmental meteorological data to form a multi-source observation vector after confidence weighting.

[0023] S1 includes: S11. Constructing multi-source observation vectors The multi-source observation vector includes the infrared pixel temperature matrix. Actual power output of the power station Ultra-short-term raw power Irradiance Ambient temperature Wind speed relative humidity and component tilt angle t represents the current time. The lead time for prediction is represented as: (1) S12. Based on the timestamp, sampling period and communication delay, perform dynamic time warping on each data source to unify the components in the multi-source observation vector to the same time axis.

[0024] S13. Establish confidence weights for different data sources. The corresponding components of each data source after being unified to the same time axis are weighted according to the confidence weight to form a weighted multi-source observation vector. (2) in, Let represent the confidence weight of the i-th data source at time t, where i and j represent the data source indices, and j iterates over all data sources. Let be the moving average of the recent prediction errors of the i-th data source. For scale parameters, This indicates a summation.

[0025] S14. Remove abnormal samples according to the physical consistency rules; the physical consistency rules include power output constraints when the nighttime irradiance is zero, power response range constraints under step changes in irradiance, negative correlation constraints between component surface temperature rise and wind speed, and non-negative constraints on predicted power.

[0026] The physical consistency rules specifically include: when the irradiance is lower than a preset nighttime irradiance threshold, the actual output and predicted power of the power plant are not higher than a preset nighttime power threshold; when the change in irradiance at adjacent times reaches or exceeds a preset step threshold, the change in the actual output of the power plant is within a preset power response range corresponding to the change in irradiance; when the irradiance and ambient temperature are within a preset similar range, the surface temperature rise of the modules and the wind speed meet a preset negative correlation condition; the ultra-short-term original power prediction value is not less than zero; samples that do not meet any of the physical consistency rules are judged as abnormal samples and removed. The nighttime irradiance threshold, nighttime power threshold, irradiance step threshold, power response range, and negative correlation judgment threshold are determined based on historical data statistics or validation sets during the healthy operation phase of the photovoltaic power plant.

[0027] S2. Extract the cell-level region from the infrared temperature image to form an infrared micro-temperature difference time-series feature vector; among them, perform meteorological disturbance attenuation processing on the infrared micro-temperature difference time-series feature vector with a large degree of environmental impact to reduce the degree of environmental impact.

[0028] like Figure 2 As shown, S2 includes: S21. Extract the cell-level regions from the infrared temperature image, and calculate the pixel-level infrared micro-temperature difference after environmental compensation for each cell-level region. ; (3) in, This represents the infrared pixel temperature at coordinates (x, y) at time t. Let represent the neighborhood set of the region containing pixel (x, y), and median represent the median function. , , These represent the changes in ambient temperature, irradiance, and wind speed at adjacent time points, respectively. , , This is the environmental compensation factor, where m represents the cell-level number. Represents the coordinates of neighboring pixels.

[0029] S22. Calculate the mean, maximum, and drift slope of the infrared micro-temperature difference within the sliding window W. Local temperature gradient Thermal diffusion asymmetry and the proportion of abnormal duration To form an infrared micro-temperature difference time-series feature vector ; (4) Infrared micro-temperature difference average The arithmetic mean of the micro-temperature differences within the sliding window region; Maximum infrared micro-temperature difference The maximum value of the micro-temperature difference among all pixels within the window; Drift slope The slope obtained by least-squares linear fitting of the average micro-temperature difference within the window region; Local temperature gradient : The average value of the horizontal and vertical temperature gradient amplitudes; Thermal diffusion asymmetry : The normalized value of the difference in average micro-temperature difference between the left and right sides or the top and bottom sides of the region; abnormal duration ratio The ratio of the number of sampling points with a micro-temperature difference exceeding the abnormal threshold within the window to the total number of sampling points in the window.

[0030] Calculate the correlation between the infrared micro-temperature difference time series feature vector and the environmental meteorological variable series. The elements in the environmental meteorological variable series include irradiance, ambient temperature, wind speed, and relative humidity. S23. When the relevance Greater than the preset meteorological disturbance threshold At the same time, meteorological synchronization disturbance attenuation is performed on the corresponding infrared micro-temperature difference features, and the attenuated infrared micro-temperature difference features are output. ; When the correlation is not greater than the preset meteorological disturbance threshold At that time, the output is an unattenuated infrared micro-temperature difference feature. At this time, if the corresponding infrared micro-temperature difference characteristics If the drift duration is not less than a preset duration window, the corresponding cell-level area will be marked as a suspected latent fault area; if the corresponding infrared micro-temperature difference feature If the drift duration is less than the preset duration window, the corresponding cell-level area will be marked as the area to be observed, and the drift duration will continue to accumulate in subsequent sliding windows.

[0031] Attenuation of meteorological synchronization disturbances includes, in particular: According to meteorological correlation Calculate the corresponding meteorological disturbance attenuation weights ; , , For the preset attenuation intensity coefficient, The meteorological correlation of the k-th infrared micro-temperature difference feature is given. A preset meteorological disturbance threshold is set; the greater the meteorological correlation, the smaller the weight of the meteorological disturbance attenuation, thereby reducing the influence of infrared features caused by synchronous changes in irradiance, ambient temperature or wind speed on subsequent latent fault identification.

[0032] The k-th infrared micro-temperature difference feature is weighted and attenuated using meteorological disturbance attenuation weights; the k-th infrared micro-temperature difference feature after weighted attenuation. .

[0033] S3. Based on the actual power output of the power plant and the ultra-short-term original power prediction value, establish the power prediction residual time series. Use the multi-scale wavelet decomposition algorithm to decompose the power prediction residual time series into low-frequency approximate components and multiple high-frequency detail components of different scales to obtain the residual low-frequency drift feature vector.

[0034] like Figure 3 As shown, S3 includes: S31. Based on the actual power output of the power plant and the ultra-short-term raw power forecast, establish a power forecast residual sequence. .

[0035] For the target time t, obtain the actual power output of the power station at that time. And the power plant's existing ultra-short-term power prediction system at the time of prediction release. Ultra-short-term raw power predictions generated for the target time t The difference between the two is calculated to obtain the power prediction residual sequence. This is used to characterize the deviation between the actual output of a power plant and the original predicted output. Its calculation formula is: (5) in, This represents the power prediction residual at time t; This represents the actual power output of the power station at time t; This indicates that the power plant already has an ultra-short-term power prediction system at time [time value missing]. The original power prediction value output at the target time t; This indicates the forecast lead time. The ultra-short-term raw power forecast value is output data from an existing power forecasting system, not actual sensor-collected data.

[0036] S32. Perform multi-scale decomposition on the power prediction residual sequence, decomposing it into low-frequency approximate components and multiple high-frequency detail components at different scales.

[0037] (6) in, This represents the J-th layer approximation component, used to characterize the low-frequency, slowly varying trend in the residual sequence; The j-th level detail component is used to characterize the fluctuation characteristics of the residual sequence at the j-th scale; J represents the wavelet decomposition level; and j represents the wavelet decomposition scale number.

[0038] S33. Calculate the high-frequency detail components at each scale. With environmental meteorological variable series The correlation between various meteorological variables.

[0039] Environmental meteorological variable series: (7) Single-scale reconstruction is performed on the j-th level wavelet detail components to obtain time-domain scale components with the same sampling length as the original power prediction residual sequence. In a sliding time window ending at the current time t Within, calculate the time-domain scale components respectively. With environmental meteorological variable series The Pearson correlation coefficients among various environmental meteorological variables were calculated, and the maximum absolute value of the correlation coefficients was taken as the meteorological correlation index for the j-th scale component. The calculation formula is: (8) in, This represents the meteorological correlation index of the j-th scale component within the current sliding time window; This represents the temporal scale component obtained after the j-th level wavelet detail component is reconstructed by a single-scale inverse wavelet; Represents the series of environmental meteorological variables The k-th environmental meteorological variable in the equation; K represents the number of environmental meteorological variables; This represents a preset sliding time window ending at the current time t; corr represents the Pearson correlation coefficient calculation function.

[0040] S34. When the relevance When the value exceeds a preset meteorological threshold, the corresponding high-frequency detail components will be... It was determined to be a high-frequency meteorological random error component. When relevance When the value is not greater than the preset meteorological threshold, in the high-frequency detail component or low-frequency approximate components When a component exhibits monotonic drift, periodic slow drift, or persistent negative deviation, it is identified as a systematic error component due to low-frequency component failure; when the correlation... When the value is not greater than the preset meteorological threshold, but the high-frequency detail components or low-frequency approximate components If no monotonic drift, periodic slow drift, or persistent negative bias is observed, the corresponding scale component is identified as another model error component.

[0041] S35. Based on the systematic error components of low-frequency component failures, extract the low-frequency residual mean, residual drift slope, cumulative negative deviation area, low-frequency energy proportion, and the estimated time to reach the preset residual threshold to construct a residual low-frequency drift feature vector. .

[0042] (9) in, This represents the residual low-frequency drift eigenvector at time t; This represents the mean of low-frequency residuals; Indicates the residual drift slope; Indicates the cumulative negative deviation area; Indicates the proportion of low-frequency energy; This indicates the estimated time to reach the preset residual threshold.

[0043] Low-frequency drift eigenvectors of residuals The specific components are as follows: The components identified as systematic error components of component failure in S34 are reconstructed using single-scale inverse wavelet reconstruction, and the reconstructed components are summed to obtain the low-frequency fault residual sequence. , is represented as: ; in, This represents the set of scale numbers that are identified as systematic error components of component failure. This represents the time-domain component obtained after the j-th scale component is reconstructed by a single-scale inverse wavelet.

[0044] Low-frequency residual mean: ; in, L represents the arithmetic mean of the residuals of low-frequency faults within the sliding time window. The number of sampling points within.

[0045] Residual drift slope: ; in, This represents the drift slope of the low-frequency fault residual obtained by least-squares linear fitting; This represents the average value at each sampling time point within the sliding time window.

[0046] Cumulative negative deviation area: ; in, This represents the cumulative area of ​​the low-frequency fault residual that is below zero within the sliding time window; This indicates the sampling time interval between adjacent sampling points.

[0047] Low-frequency energy percentage: ; in, This represents the proportion of low-frequency fault residual energy to the total residual energy of the original power prediction. ε represents the original power prediction residual obtained by formula (5); ε represents a preset positive number used to prevent the denominator from being zero.

[0048] Estimated time to reach the residual threshold: ; in, This indicates the estimated time required to reach the preset residual threshold based on the current residual drift trend; This represents a preset negative residual threshold. When the low-frequency fault residual has reached this threshold, let... When the residual does not drift towards the threshold direction, let Or preset the maximum time.

[0049] S4. By fusing the infrared micro-temperature difference time-series feature vector and the residual low-frequency drift feature vector, and based on the pre-constructed dual-mode fault discrimination model, the latent fault type and fault location of the photovoltaic module are identified, and the module fault attenuation index is output.

[0050] S4 includes: S41. Establishing Infrared Micro-Temperature Difference Characteristics or With residual low-frequency drift eigenvector The relationship between them.

[0051] Relevance in S23 This will affect the output of the infrared micro-temperature difference feature. If the correlation is greater than the preset meteorological disturbance threshold, it indicates that the environment has a significant impact on the infrared micro-temperature difference feature, and meteorological synchronization disturbance attenuation needs to be performed to reduce the degree of environmental influence. In this case, the infrared micro-temperature difference feature used in this step has been perturbed. .

[0052] Based on the spatial and electrical topology mapping relationship between the infrared image region and photovoltaic modules, strings and inverters, the equipment level to which the infrared micro-temperature difference feature belongs is determined, and it is paired and associated with the residual low-frequency drift feature vector corresponding to the same equipment level.

[0053] S42. For the m-th cell-level region, calculate the infrared micro-temperature difference abnormal energy based on the environmentally compensated infrared micro-temperature difference. .

[0054] For the m-th cell-level region Based on the infrared micro-temperature difference after environmental compensation Calculate the abnormal energy of infrared microtemperature differences The calculation formula is: (10) in, This represents the infrared micro-temperature difference anomalous energy of the m-th component region at time t; This represents the m-th component region; This represents the number of pixels within the m-th component region; (x,y) represents the pixel coordinates in the infrared image. represents the infrared micro-temperature difference at coordinates (x, y) at time t after environmental compensation; max represents the function that takes the maximum value.

[0055] S43. Perform single-scale inverse wavelet reconstruction on each scale component of the systematic error component of low-frequency component faults to obtain the low-frequency fault residual sequence. The ratio of the energy of the low-frequency fault residual sequence to the total energy of the original power prediction residual sequence is calculated to obtain the proportion of low-frequency residual drift energy. The proportion of low-frequency residual drift energy is used to subsequently construct the latent fault influence factor. and comprehensive hidden fault score S ( t ).

[0056] In a sliding time window ending at the current time t Within this range, the ratio of the energy of the low-frequency fault residual sequence to the total energy of the original power prediction residual sequence is calculated to obtain the proportion of low-frequency residual drift energy. This is used to characterize the proportion of persistent low-frequency deviation caused by latent component faults in the power prediction residual, and the calculation formula is: (11) in, This indicates the proportion of low-frequency residual drift energy within the current sliding time window.

[0057] S44. Based on a pre-established dual-modal fault discrimination model, for infrared micro-temperature difference characteristics... With residual low-frequency drift eigenvector The data is encoded and then fused with contextual information to output the latent fault type, fault confidence, fault area location, and component fault attenuation index.

[0058] In this step, the dual-modal fault discrimination model includes an infrared micro-feature encoding channel, a residual low-frequency drift encoding channel, a gated attention fusion module, and a fault output module. The infrared micro-feature encoding channel and the residual low-frequency drift encoding channel respectively encode infrared micro-temperature difference features. With residual low-frequency drift eigenvector Encode to obtain infrared latent representation and residual latent representation ; Infrared micro-feature encoding channel for infrared micro-temperature difference time-series feature vector Encoding is performed specifically through a fully connected layer, a ReLU activation function, and layer normalization to map the infrared micro-temperature difference time-series feature vector to a latent feature space of a preset dimension, thereby obtaining the infrared latent representation. , is represented as: (12) The residual low-frequency drift coding channel for residual low-frequency drift feature vectors Encoding is performed specifically through a fully connected layer, a ReLU activation function, and layer normalization to map the residual low-frequency drift feature vector to a latent feature space of the same dimension as the infrared latent representation, thus obtaining the residual latent representation. , is represented as: (13) in, and These represent the trainable weight matrices for the two encoding channels, respectively. and These represent the corresponding bias vectors; ReLU represents the activation function; LN represents the layer normalization operation. The output dimensions of the two encoding channels are set to be the same so that they can be input into the subsequent gated attention fusion module.

[0059] Gated attention fusion module based on infrared latent representation Residual latent representation In addition to operating context information, the gating fusion vector is calculated. Wherein, Sigmoid represents the sigmoid activation function. and For learnable parameters, Indicates the potential infrared representation, Let C(t) represent the residual latent representation, and let C(t) represent the operating condition context vector. This indicates feature splicing.

[0060] Based on the gated fusion vector g ( t ), for infrared potential representation and residual latent representation Adaptive weighted fusion is performed to obtain the fused feature representation. ; This represents the dual-modal fusion feature representation at time t. This represents element-wise multiplication; This represents the adaptive weights corresponding to the low-frequency drift coding channels of the residual.

[0061] The fault output module represents the fusion features. Output the latent fault types respectively. c ( t Fault confidence p ( t ), location of the fault area loc ( t ) and component failure degradation index DI ( t ).

[0062] Representing fusion features Input the fault classification header, location header, and attenuation regression header, and output the latent fault type c(t), fault confidence p(t), fault area location loc(t), and component fault attenuation index, respectively. DI ( t ); (14) Where c(t) represents the type of latent fault identified at time t; p(t) represents the fault discrimination confidence, which is obtained by the fault classification head based on the predicted probability of each fault type; and loc(t) represents the location of the fault area. DI ( t This represents the component failure degradation index, which is determined by the degradation regression head based on the fused potential representation. The calculations yielded a result used to quantify the impact of latent faults on the component's power generation capability and subsequent power prediction results. This represents the output mapping function of the fault output module.

[0063] Latent fault type c(t) includes one or more of the following: microcracks in the cell, deterioration of the solder strip or busbar contact, abnormal bypass diode, potential induced degradation, early hot spots, latent mismatch caused by local shading contamination, and abnormal string connection contact.

[0064] S5. Construct a degradation correction coefficient based on the component failure degradation index and generate a corrected photovoltaic power prediction value; update the parameters to be corrected based on the difference between the corrected photovoltaic power prediction value and the actual power output of the power station at the latest moment.

[0065] S5 includes: S51. Based on component failure attenuation index DI ( t Establish component failure attenuation correction coefficients. , used to characterize the degree of reduction in the power generation capacity of photovoltaic modules due to latent faults, is calculated using the following formula; (15) in, This represents the attenuation correction factor at time t, based on the component failure attenuation index at the current time. Calculated in real time; This indicates the lower limit of the attenuation correction factor; This represents the upper limit of the attenuation correction coefficient; DI(t) represents the component failure attenuation index at time t. and For learnable parameters, based on joint optimization loss It is updated and kept fixed after training; Sigmoid represents the sigmoid activation function.

[0066] S52. Adjust the attenuation correction factor. Component failure attenuation index DI(t), future low-frequency residual compensation amount for systematic errors in low-frequency component failures. Online bias compensation item and ultra-short-term raw power prediction values Input the fault attenuation correction layer to obtain the corrected photovoltaic power prediction value. .

[0067] Specifically, the historical low-frequency fault residual sequence of length L up to the current time t is input into the low-frequency residual time series prediction module, and the future fault residual sequence is output. The low-frequency residual compensation amount at time is expressed as (16) (17) The low-frequency residual prediction submodule is a component of the infrared-residual joint modulation power correction model. It is used to extract time-series features from historical low-frequency fault residual sequences and corresponding multi-source power correction input features, and to predict the low-frequency residual compensation amount at future times. The low-frequency residual prediction submodule adopts the encoder-decoder structure described in step S6. This indicates that the time-series prediction mapping established by the low-frequency residual prediction submodule is used to calculate future... Future low-frequency residual compensation amount .

[0068] The differentiable fault attenuation correction layer is not an independent neural network model, but rather a differentiable computation layer located at the output of the power correction model. This differentiable fault attenuation correction layer receives attenuation correction coefficients. Component failure degradation index Ultra-short-term raw power prediction values Future low-frequency residual compensation and online bias compensation items First utilize The fault attenuation correction is applied to the ultra-short-term raw power forecast, and then the future low-frequency residual compensation and online bias compensation are added to obtain the corrected photovoltaic power forecast. The fault attenuation correction layer is specifically implemented according to formula (18): (18) in, This indicates the prediction at time t. Photovoltaic power after time correction; This indicates that it was generated at time t. Predicted raw power values ​​for ultra-short time periods; This indicates the prediction at time t. Future low-frequency residual compensation amount; This indicates the online bias compensation term.

[0069] To avoid outliers in the power prediction correction results that do not conform to the operating rules of photovoltaic power plants, constraints are imposed on the corrected photovoltaic power prediction value, attenuation correction coefficient and power correction change range respectively; The non-negativity constraint for predicted power is as follows: ; The boundary constraints for the attenuation coefficient are: ; The continuity constraint for the correction magnitude is: ; Where, Δ τ Indicates the time interval between adjacent prediction times; This indicates the maximum allowable power correction amount determined based on the photovoltaic power plant's installed capacity, ramp-up rate limits, or historical operating data.

[0070] The differentiable fault attenuation correction layer retains the fault attenuation correction structure during backpropagation, ensuring that the module fault attenuation index DI(t) is consistent with the corrected photovoltaic power prediction. The influence of satisfies a monotonically decreasing relationship, and its gradient expression is: (19) in, The symbol represents a partial differential equation; this formula is used to constrain the model to maintain physical consistency during training, namely, "the greater the component failure attenuation index, the stronger the reduction in baseline output."

[0071] S53. Based on the deviation between the actual power output of the power plant at the latest time and the corrected predicted power, adjust the parameters. Perform online updates; correct parameters include environmental compensation coefficients. , , and online bias compensation items .

[0072] (20) Where θ(t) represents the correction parameter at time t; θ(t-1) represents the correction parameter at the previous time; and K(t) represents the adaptive gain. This represents the actual power output of the power station at time t; This represents the photovoltaic power predicted at time t-1 and corrected at time t.

[0073] S6. Construct the power correction input feature vector and perform time-series stacking. Input the pre-constructed infrared-residual joint modulation power correction model to obtain the low-frequency residual compensation sequence for more than two future time periods. Input the low-frequency residual compensation sequence into the fault attenuation layer to output the corrected photovoltaic power prediction sequence.

[0074] S6 includes: S61. Constructing the power-corrected input feature vector Including infrared micro-temperature difference time-series feature vectors Low-frequency drift eigenvectors of residuals Component failure degradation index DI ( t ), attenuation correction factor Environmental meteorological data vector M ( t Ultra-short-term raw predicted power and historical efforts t represents time.

[0075] (twenty one) The environmental meteorological data vector M(t) is the irradiance obtained from S11. Ambient temperature Wind speed relative humidity and component tilt angle Composed of the above, and after timestamp alignment, missing value completion, and outlier handling in S12, it is represented as: ; This represents the ultra-short-term raw prediction power generated at time t+τ, which is generated at time t. This represents the actual power output of the power station at time t.

[0076] S62. Stack the power-corrected input feature vectors from time t-L+1 to time t to obtain the multi-source time-series input sequence at time t. .

[0077] The historical lookback window length is set to L. The power correction input feature vectors of L consecutive time steps are stacked to obtain the multi-source time-series input sequence for power correction prediction at time t. Its expression is: (twenty two) in, Represents the multi-source time-series input sequence at time t; L represents the length of the history lookback window; to This represents the continuous input feature vector from time t-L+1 to time t.

[0078] S63. Input multi-source time-series sequences By inputting a pre-constructed infrared-residual joint modulation power correction model, low-frequency residual compensation sequences for two or more future time periods are obtained. .

[0079] S63 includes: S631. Input multi-source time-series sequences The input feature embedding layer is then superimposed with positional encoding to obtain the embedded input sequence. .

[0080] Multi-source time-series input sequence The input feature embedding layer is then superimposed with positional encoding to obtain the embedded input sequence. The calculation formula is: (twenty three) in, The input sequence is represented by the embedded part; Embed represents the feature embedding function. This represents the position code corresponding to time t, used to preserve the sequential relationship between different time steps in the time-series input sequence.

[0081] S632. Based on infrared micro-temperature difference constant energy Low-frequency residual drift energy ratio Using the fault confidence level p(t) and the component fault attenuation index DI(t), a latent fault influence factor is constructed at time step t. .

[0082] Latent Fault Influence Factors This is used to characterize the overall impact of latent faults on the power correction process, and the calculation formula is: (twenty four) in, This represents the latent fault impact factor at time step t; Indicates the abnormal energy of infrared micro-temperature difference; represents the proportion of low-frequency residual drift energy; p(t) represents the fault confidence level; DI(t) represents the component fault attenuation index; norm represents the normalization function; α , β , γ and δ The weights are non-negative and satisfy the following constraints: .

[0083] S633. Embedded input sequence The projection consists of the query matrix Q, the key matrix K, and the value matrix V.

[0084] In the encoder of the infrared-residual joint modulation power correction model, the embedded input sequence The projection consists of the query matrix Q, the key matrix K, and the value matrix V, and its calculation formula is: (25) (26) (27) Where Q represents the query matrix; K represents the key matrix; and V represents the value matrix; , and It is a learnable projection matrix.

[0085] S634. Calculate the i-th query vector using a sparse attention mechanism. Original information content score Utilizing latent fault influencing factors Information content score Modulation is performed to obtain the modulated information content fraction. Information content fraction after modulation Used to filter active query vectors for attention calculation. Based on the modulated information content score. The query vectors are sorted, and a preset number of query vectors with higher information content scores are selected as active query vectors to participate in the sparse attention calculation. The remaining query vectors are not selected as active query vectors to participate in this round of sparse attention calculation.

[0086] (28) in, This represents the original information content score of the i-th query vector, based on the i-th query vector. The attention score is calculated by scaling the dot product between the query vector and all key vectors K. Specifically, the matching degree between the query vector and each key vector is first calculated, and then the information content is evaluated based on the difference between the maximum value and the average level of these matching scores. The greater the difference, the more the query vector can focus on a few key time steps, the higher its information content, and the more it should be retained for subsequent sparse attention calculation.

[0087] This represents the information content score after modulation of the latent fault influence factor. This indicates that a latent fault affects the modulation coefficient; This represents the latent fault influence factor at the i-th time step.

[0088] This is the modulation intensity coefficient of the latent fault influence factor, used to control the degree of influence of the latent fault influence factor on the original information content score. In this embodiment, As a non-negative learnable parameter, it is iteratively updated during model training based on power prediction loss and fault discrimination loss, and its final value is determined through the validation set. The larger the value, the stronger the modulatory effect of the latent fault influence factor on attention scores.

[0089] Based on the modulated information content fraction Active query vectors are selected to participate in attention calculation to enhance the model's ability to extract features during critical periods such as continuous drift of infrared micro-temperature difference, slow drift of low-frequency residual, aggravation of component attenuation, and continuous expansion of prediction bias.

[0090] S635. Future generated based on time t to The ultra-short-term raw power prediction values ​​and corresponding time position codes for each time period are used to construct the decoder prediction query sequence. After self-attention, cross-attention, and linear output layer calculations, low-frequency residual compensation sequences for two or more future time periods are obtained. T represents the prediction step size.

[0091] Future generated based on time t to The ultra-short-term raw power prediction value and the corresponding time position code of the time period are used to construct the decoder prediction query sequence; the decoder prediction query sequence is input into the self-attention layer of the decoder for temporal feature extraction to obtain the first decoding feature; the first decoding feature is used as the query, and the encoder output feature is used as the key and value to input into the cross-attention layer to obtain the second decoding feature.

[0092] The second decoded feature is used as the decoder output feature. The data is then input into a linear output layer for mapping to obtain low-frequency residual compensation sequences for two or more future time periods. .

[0093] (29) Where T represents the prediction step size; represents the hidden layer features output by the decoder; Linear represents the linear output layer.

[0094] S64. Low-frequency residual compensation sequence By inputting the fault attenuation correction layer, we obtain corrected photovoltaic power prediction sequences for two or more future time periods. .

[0095] Low-frequency residual compensation sequence The corrected photovoltaic power prediction sequence for two or more future time periods is calculated according to formula (18). .

[0096] This represents the corrected photovoltaic power prediction sequence from time t+1 to time t+T.

[0097] S7. Construct a joint optimization loss to update the trainable parameters, calculate a comprehensive latent fault score, and output the corresponding warning level and associated warning information.

[0098] S7 includes: S71. Compare the corrected photovoltaic power forecast values ​​with the corresponding measured power values ​​for two or more future time periods, and calculate the power forecast correction loss. ; (30) N represents the number of training samples, and T represents the prediction step size. This represents the corrected predicted power of the nth sample at the kth time step, calculated using the power correction model. This represents the power actually collected by the power plant's SCADA system at the corresponding future time for the nth sample. and The difference is used to calculate the prediction error.

[0099] S72. Constructing a joint optimization loss The trainable parameters are updated based on the joint optimization loss; the joint optimization loss , Indicates the loss by fault type classification. Indicates the loss in fault location. Represents physical constraint loss. To represent the correction magnitude smoothing loss, smoothed L1 loss can be used. to This is the loss weight.

[0100] The following provides a detailed explanation of each loss involved in the joint optimization loss.

[0101] a. Fault type classification loss Cross-entropy loss is used: (31) Where C represents the number of latent fault types; The true label indicating whether the nth sample belongs to the cth type of fault; This represents the predicted probability of the c-th type of fault output by the dual-modal fault discrimination model.

[0102] b. Fault location loss : (32) Where M represents the number of candidate cell-level regions; This represents the label of the actual fault region for the nth sample. This represents the probability that the m-th region in the model output is a fault region.

[0103] c. Physical constraint loss : (33) in, This indicates the installed capacity of the photovoltaic power plant or the maximum allowable power under current operating conditions. This loss is used to penalize prediction results that are less than zero or exceed the maximum allowable output.

[0104] d. Correcting the magnitude of the loss : First, define the correction range: (34) Smoothing loss is written as: (35) in, This represents the smoothed L1 loss function, used to limit abrupt changes in the correction magnitude between adjacent prediction times, while reducing the impact of outliers on model training.

[0105] Construct joint optimization loss During the model training phase, the trainable parameters are updated. These trainable parameters specifically include the weight matrices for the infrared micro-feature encoding channel and the residual low-frequency drift encoding channel. , and bias vector , Weight matrix of the gated attention fusion module and bias vector The weight matrix and bias vector of each output branch in the fault output module, used to output the latent fault type, fault confidence, fault area location, and component fault attenuation index; the projection matrix of the query matrix, key matrix, and value matrix. , and Trainable parameters for the attention layer, feedforward network layer, and normalization layer in the encoder and decoder; weight matrix and bias vector of the linear output layer; and decay correction mapping parameters in the fault decay correction layer. and .

[0106] It is important to note the meteorological disturbance threshold in the infrared micro-temperature difference-environmental disturbance decoupling process. and attenuation intensity coefficient Furthermore, the wavelet basis, decomposition level J, and meteorological related thresholds in the multi-scale decoupling processing of power prediction residuals are determined by preset or validation sets and are not used as gradient update objects for the joint optimization loss.

[0107] S73. Component Failure Attenuation Indicators DI ( t Infrared micro-temperature differential energy Low-frequency residual drift energy ratio Fault confidence p ( t The absolute difference between the corrected predicted power and the original ultra-short-term predicted power. The weighted summation yields the comprehensive latent fault score. S ( t Based on the comprehensive latent fault score S ( t Output the warning level and the corresponding warning information.

[0108] (36) in, This represents the weighting coefficient for each item. The overall hidden fault score is calculated using this formula. S( t ) and the first threshold Second threshold and the third threshold Comparison: when At that time, the output status was normal; when At that time, issue a warning at the level of concern; when At that time, a maintenance-level early warning will be issued; when At that time, it will issue a warning of shutdown or reduction in rating.

[0109] Output early warning information, including component number, string number, fault area coordinates, fault type, fault confidence level, component fault attenuation index, difference in predicted power before and after correction, and suggested maintenance measures; and update the infrared micro-temperature difference benchmark, residual decoupling threshold, fault discrimination model parameters, and power prediction correction parameters based on the maintenance verification results.

[0110] In summary, the steps S4-S7 can be summarized as follows: Figure 4 As shown.

[0111] To verify the effectiveness and applicability of the method of this invention, an application test was conducted on a centralized photovoltaic power station. The power station, with an installed capacity of 100MW, includes multiple inverters, combiner boxes, strings, and module arrays. To verify the feasibility of the method, a simulation dataset containing infrared temperature, actual power output, raw ultra-short-term power prediction values, and environmental meteorological variables was constructed in MATLAB based on the actual operating characteristics of the photovoltaic power station.

[0112] The simulation data includes irradiance sequences, ambient temperature sequences, wind speed sequences, theoretical power output sequences under healthy conditions, actual power output sequences with attenuation due to latent faults, original ultra-short-term power prediction sequences without considering latent faults, and infrared micro-temperature difference sequences used to characterize local infrared anomalies.

[0113] In the simulation, 288 sampling points were set, corresponding to 5-minute intervals of operational data within a day; the installed capacity of the power plant was set to 100MW; and the latent faults of the components were set to gradually increase from the 150th sampling point, causing the component fault attenuation index DI(t) to gradually increase from a small value. This simulated the process of early latent faults of the components gradually worsening from slight anomalies.

[0114] In this embodiment, MATLAB was used to reproduce the method of the present invention. Then, RMSE, MAE, MAPE, accuracy, recall, F1 score, positioning accuracy and warning lead time were calculated for different methods.

[0115] The power prediction errors of the original ultra-short-term power prediction, conventional power residual correction, ordinary data-driven fusion method, and the complete method of this invention were calculated using a MATLAB program. The results are shown in Table 1. The power prediction correction effect is as follows: Figure 5 As shown.

[0116] Table 1

[0117] As shown in Table 1, the original RMSE for ultra-short-term power prediction was 2.5679 MW, MAE was 1.9462 MW, and MAPE was 13.943%. After using conventional power residual correction, the RMSE decreased to 1.2195 MW; after using the ordinary data-driven fusion method, the RMSE was 1.4570 MW; after using the complete method of this invention, the RMSE decreased to 0.9522 MW, MAE decreased to 0.7302 MW, and MAPE decreased to 5.7285%, with an RMSE reduction rate of 62.921%.

[0118] Depend on Figure 5 It can be seen that the prediction curve corrected by the present invention is closer to the actual output curve as a whole. Especially after the latent faults of the components gradually worsen, the original ultra-short-term prediction is prone to continuous deviation. However, the present invention can reduce the deviation between the original predicted power and the actual output by correcting the low-frequency residual compensation and the component fault attenuation index.

[0119] Table 2

[0120] As shown in Table 2, although the traditional infrared threshold method has a high recall rate, its fault location accuracy is only 72.40%, and the early warning time is 18.50 min. The conventional power residual analysis method, due to the lack of infrared spatial positioning information, has an accuracy, recall rate, and F1 score of 58.015%, 69.725%, and 63.333%, respectively, with a positioning accuracy of only 45.60%. The general performance of the ordinary data-driven fusion method is improved, but its positioning accuracy and early warning time are still lower than those of the complete method of this invention.

[0121] The complete method of this invention achieves an accuracy, recall, and F1 score of 96.330%, a fault location accuracy of 91.70%, and an early warning time of 43.80 minutes. This demonstrates that this invention, through dual-modal fusion of infrared micro-features and low-frequency residual drift features, can simultaneously improve the accuracy of latent fault identification, fault location capability, and early warning capability.

[0122] Depend on Figure 6It can be seen that the original prediction residuals contain both short-term fluctuations and persistent slow drift. After multi-scale decoupling, the high-frequency component mainly exhibits short-term random fluctuations, which can be used to characterize meteorological random errors; the low-frequency component exhibits persistent deviations during the fault development stage, which can be used to characterize systematic errors caused by component latent faults or string attenuation.

[0123] Depend on Figure 7 It can be seen that the comprehensive latent fault score was at a low level before the fault occurred; as the latent fault gradually intensified, the comprehensive latent fault score continued to rise and exceeded the preset warning threshold. This result shows that the present invention can achieve graded early warning of latent faults by utilizing infrared micro-temperature difference anomalies, low-frequency residual drift, and component fault attenuation indicators.

[0124] The latent fault status was further determined based on the comprehensive latent fault score, and the identification results of different methods were compared with the fault labels. The results are shown in Table 2. The multi-scale decoupling results of the power prediction residual are as follows: Figure 6 As shown, the comprehensive latent fault score and early warning results are as follows: Figure 7 As shown.

[0125] As can be seen from the above embodiments, the method of the present invention can extract weak local infrared anomalies of the module under complex meteorological disturbances and normal power fluctuations, and separate low-frequency systematic errors related to latent faults of the module from the power prediction residuals. Compared with the traditional infrared threshold method and conventional power residual analysis method, the present invention can not only identify early latent faults, but also feed back the degree of fault attenuation into the power prediction correction process, thereby reducing the persistent prediction bias caused by latent faults of the module.

[0126] In summary, this invention forms a complete process of infrared micro-feature perception, prediction residual decoupling, dual-mode fault discrimination, power prediction adaptive correction, and hierarchical early warning closed-loop update, which can provide technical support for photovoltaic power plant power application, scheduling control, and operation and maintenance.

[0127] The above embodiments are only used to provide a detailed description of the technical solutions of this application. However, the descriptions of the above embodiments are only for the purpose of helping to understand the methods of the embodiments of the present invention and should not be construed as limiting the embodiments of the present invention. Any variations or substitutions that can be easily conceived by those skilled in the art should be covered within the protection scope of the embodiments of the present invention.

Claims

1. A method for identifying latent photovoltaic faults and correcting power prediction, characterized in that, include: S1. Collect time-series infrared temperature data of photovoltaic modules, actual power output data of power plants, ultra-short-term raw power data and environmental meteorological data to form a multi-source observation vector after confidence weighting. S2. Extract the cell-level region from the infrared temperature image to form an infrared micro-temperature difference time-series feature vector; among them, perform meteorological disturbance attenuation processing on the infrared micro-temperature difference time-series feature vector with a large degree of environmental impact to reduce the degree of environmental impact; S3. Based on the actual power output of the power plant and the ultra-short-term original power prediction value, establish the power prediction residual time series. Use the multi-scale wavelet decomposition algorithm to decompose the power prediction residual time series into low-frequency approximate components and multiple high-frequency detail components at different scales to obtain the residual low-frequency drift feature vector. S4. By fusing the infrared micro-temperature difference time-series feature vector and the residual low-frequency drift feature vector, and based on the pre-constructed dual-mode fault discrimination model, the latent fault type and fault location of the photovoltaic module are identified, and the module fault attenuation index is output. S5. Construct a degradation correction coefficient based on the component failure degradation index and generate a corrected photovoltaic power prediction value; update the parameters to be corrected based on the difference between the corrected photovoltaic power prediction value and the actual power output of the power station at the latest moment. S6. Construct the power correction input feature vector and perform time-series stacking. Input the pre-constructed infrared-residual joint modulation power correction model to obtain the low-frequency residual compensation sequence for more than two future time periods. Input the low-frequency residual compensation sequence into the fault attenuation layer to output the corrected photovoltaic power prediction sequence. S7. Construct a joint optimization loss to update the trainable parameters, calculate the comprehensive latent fault score, and output the corresponding warning level and associated warning information.

2. The method according to claim 1, characterized in that, S1 includes: Construct a multi-source observation vector, including infrared pixel temperature matrix, actual power plant output, ultra-short-term raw power, irradiance, ambient temperature, wind speed, relative humidity, and module tilt angle; Unify the components of the multi-source observation vector to the same time axis; Confidence weights are established for different data sources, and the corresponding components of each data source after being unified to the same time axis are weighted according to the confidence weights to form a weighted multi-source observation vector. Abnormal samples were removed according to physical consistency rules. These rules included power output constraints when nighttime irradiance was zero, power response range constraints under step changes in irradiance, constraints on the negative correlation between component surface temperature rise and wind speed, and constraints on the non-negativity of predicted power.

3. The method according to claim 1, characterized in that, S2 include: Extract the cell-level regions from the infrared temperature image, and calculate the pixel-level infrared micro-temperature difference after environmental compensation for each cell-level region; Calculate the mean, maximum, and drift slope of the infrared micro-temperature difference within a sliding window W. Local temperature gradient Thermal diffusion asymmetry and the proportion of abnormal duration To form an infrared micro-temperature difference time-series feature vector ; Calculate the correlation between the infrared micro-temperature difference time series feature vector and the environmental meteorological variable series. The elements in the environmental meteorological variable series include irradiance, ambient temperature, wind speed, and relative humidity. When the correlation is not greater than the preset meteorological disturbance threshold, the unattenuated infrared micro-temperature difference feature is output. At this time, if the drift duration of the corresponding infrared micro-temperature difference feature reaches or exceeds the preset duration window, the corresponding cell-level area is marked as a suspected latent fault area; if the drift duration does not reach the preset duration window, the corresponding cell-level area is marked as an area to be observed, and the drift duration continues to accumulate in subsequent sliding windows.

4. The method according to claim 3, characterized in that, Attenuation of meteorological synchronization disturbances includes: The corresponding meteorological disturbance attenuation weight is calculated based on the meteorological correlation. ; , For the preset attenuation intensity coefficient, The meteorological correlation of the k-th infrared micro-temperature difference feature is given. Preset meteorological disturbance threshold; The k-th infrared micro-temperature difference feature is attenuated using the meteorological disturbance attenuation weight; the k-th infrared micro-temperature difference feature after weighted attenuation... .

5. The method according to claim 4, characterized in that, S3 include: Based on the actual power output of the power plant and the ultra-short-term raw power forecast, a power forecast residual sequence is established. ; The power prediction residual sequence is decomposed into low-frequency approximate components and multiple high-frequency detail components at different scales. Calculate the correlation between the high-frequency detail components at each scale and each meteorological variable in the environmental meteorological variable series; When the correlation is greater than the preset meteorological correlation threshold, the corresponding high-frequency detail component is identified as a high-frequency meteorological random error component; when the correlation is not greater than the preset meteorological correlation threshold, if the high-frequency detail component or low-frequency approximate component exhibits monotonic drift, periodic slow drift, or persistent negative deviation, the corresponding component is identified as a low-frequency component failure systematic error component; when the correlation is not greater than the preset meteorological correlation threshold, but the high-frequency detail component or low-frequency approximate component does not exhibit monotonic drift, periodic slow drift, or persistent negative deviation, the corresponding scale component is identified as other model error components. Based on the systematic error components of the low-frequency component faults, a residual low-frequency drift feature vector is constructed. .

6. The method according to claim 5, characterized in that, S4 include: Establish infrared micro-temperature difference characteristics or With residual low-frequency drift eigenvector The relationship between them; For the m-th cell-level region, calculate the infrared micro-temperature difference abnormal energy based on the environmentally compensated infrared micro-temperature difference. ; Single-scale inverse wavelet reconstruction is performed on each scale component of the systematic error component of low-frequency component faults to obtain the low-frequency fault residual sequence. The ratio of the energy of the low-frequency fault residual sequence to the total energy of the original power prediction residual sequence is calculated to obtain the proportion of low-frequency residual drift energy. ; Based on a pre-established dual-modal fault discrimination model, infrared micro-temperature difference characteristics are analyzed. With residual low-frequency drift eigenvector The data is encoded and then fused with contextual information to output the latent fault type, fault confidence, fault area location, and component fault attenuation index.

7. The method according to claim 1, characterized in that, The dual-modal fault discrimination model includes an infrared micro-feature encoding channel, a residual low-frequency drift encoding channel, a gated attention fusion module, and a fault output module; The infrared micro-feature encoding channel and the residual low-frequency drift encoding channel respectively encode infrared micro-temperature difference features. With residual low-frequency drift eigenvector Encode to obtain infrared latent representation and residual latent representation ; Gated attention fusion module based on infrared latent representation Residual latent representation In addition to operating context information, the gating fusion vector is calculated. g ( t According to the gated fusion vector g ( t ), for infrared potential representation and residual latent representation Adaptive weighted fusion is performed to obtain the fused feature representation. ; Fault output module fusion feature representation Output the latent fault types respectively. c ( t Fault confidence p ( t ), location of the fault area loc ( t ) and component failure degradation index DI ( t ).

8. The method according to claim 7, characterized in that, S5 include: According to component failure attenuation index DI ( t Establish component failure attenuation correction coefficients. ; Attenuation correction factor Component failure attenuation index DI(t), future low-frequency residual compensation amount for systematic errors in low-frequency component failures. Online bias compensation item and ultra-short-term raw power prediction values Input the fault attenuation correction layer to obtain the corrected photovoltaic power prediction value. ; Based on the deviation between the latest actual power output of the power plant and the corrected predicted power, the parameters to be corrected are... Online updates will be performed; parameters to be corrected include environmental compensation coefficients and online bias compensation terms.

9. The method according to claim 8, characterized in that, S6 include: Constructing power-corrected input feature vectors Including infrared micro-temperature difference time-series feature vectors Low-frequency drift eigenvectors of residuals Component failure degradation index DI ( t ), attenuation correction factor Environmental meteorological data vector M ( t Ultra-short-term raw predicted power and historical efforts t represents time. Stack the power-corrected input feature vectors from time t-L+1 to time t to obtain the multi-source time-series input sequence at time t. ; Multi-source time-series input sequence By inputting a pre-constructed infrared-residual joint modulation power correction model, low-frequency residual compensation sequences for two or more future time periods are obtained. ;include: Multi-source time-series input sequence The input feature embedding layer is then superimposed with positional encoding to obtain the embedded input sequence. ; Based on infrared micro-temperature difference anomalous energy Low-frequency residual drift energy ratio Using the fault confidence level p(t) and the component fault attenuation index DI(t), a latent fault influence factor is constructed at time step t. ; The embedded input sequence The projection consists of a query matrix Q, a key matrix K, and a value matrix V; The sparse attention mechanism is used to compute the i-th query vector. Original information content score Utilizing latent fault influencing factors Information content score Modulation is performed to obtain the modulated information content fraction. Information content fraction after modulation Used to filter active query vectors to participate in sparse attention computation; Future generated based on time t to The ultra-short-term raw power prediction values ​​and corresponding time position codes for each time period are used to construct the decoder prediction query sequence. After self-attention, cross-attention, and linear output layer calculations, low-frequency residual compensation sequences for two or more future time periods are obtained. T represents the prediction step size. Low-frequency residual compensation sequence By inputting the fault attenuation correction layer, we obtain corrected photovoltaic power prediction sequences for two or more future time periods. .

10. The method according to claim 9, characterized in that, S7 includes: Compare the corrected photovoltaic power forecast values ​​for two or more future time periods with the corresponding measured power values ​​to calculate the power forecast correction loss. ; N represents the number of training samples, and T represents the prediction step size. This represents the predicted power of the nth sample after the k-th step correction. This indicates the corresponding measured power; Construct joint optimization loss The trainable parameters are updated based on the joint optimization loss; the joint optimization loss , Indicates the loss by fault type classification. Indicates the loss in fault location. Represents physical constraint loss. This indicates the correction magnitude for smoothing loss. to For loss weights; Component failure attenuation index DI ( t Infrared micro-temperature differential energy Low-frequency residual drift energy ratio Fault confidence p ( t The absolute difference between the corrected predicted power and the original ultra-short-term predicted power. The weighted summation yields the comprehensive latent fault score. S ( t Based on the comprehensive hidden fault score, the system outputs the warning level and the corresponding warning information.

Images