An inverter degradation identification method and an inverter

Abstract

The application discloses an inverter degradation identification method and an inverter, and relates to the technical field of photovoltaic power generation operation and maintenance. The method comprises the following steps: obtaining an original operation data set of an inverter; extracting multi-dimensional degradation correlation features from the original operation data set; inputting the multi-dimensional degradation correlation features into N different degradation identification models respectively to obtain preliminary degradation identification results output by the N models, wherein the degradation types and time scales identified by the N models are not completely the same; and performing fusion analysis on the N preliminary degradation identification results to obtain a target degradation identification result. The application relies on the differentiated division of labor of multiple models to make up for the degradation omission and misjudgment defects caused by the fact that a single model only adapts to a single scale and part of degradation types, integrates multi-dimensional judgment evidence through the fusion of multiple model results, balances the identification advantages of each model, and improves the accuracy and robustness of the limit identification.

Description

Technical Field

[0001] This application relates to the field of photovoltaic power generation operation and maintenance technology, specifically to an inverter derating identification method and an inverter. Background Technology

[0002] In photovoltaic power plants, inverters may be derating due to multiple factors, including internal component temperature rise, dust accumulation and blockage in air ducts, MPPT branch abnormalities, and grid connection limitations. While an inverter can still output power during derating, the actual output power is significantly lower than its theoretical maximum power output under current irradiance, ambient temperature, and equipment conditions. Derating operation of inverters lacks specific alarms, making it far more difficult to detect than shutdown or offline faults. Therefore, accurately identifying inverter derating is crucial for improving photovoltaic power plant power generation and reducing operation and maintenance losses.

[0003] However, due to the variety of causes of inverter derating, the inverter derating operating parameters corresponding to different causes exhibit different characteristics. Conventional threshold determination schemes or single identification models are difficult to take into account all types of derating, and are prone to misjudgment and omission. Summary of the Invention

[0004] In view of this, this application provides an inverter derating identification method and an inverter, which can solve the problem of low accuracy in inverter derating identification.

[0005] To solve the above problems, the technical solution provided in this application is as follows:

[0006] The first aspect of this application provides a method for identifying inverter derating, including:

[0007] Obtain the raw operating dataset of the inverter;

[0008] Extract multidimensional depreciation correlation features from the original running dataset;

[0009] The multidimensional deduction correlation features are input into N different deduction identification models to obtain preliminary deduction identification results output by N deduction identification models; wherein, N>1, the preliminary deduction identification results include at least deduction type, the deduction types identified by the N deduction identification models are not completely the same, and the time scales of the deduction features identified by the N deduction identification models are not completely the same.

[0010] The target reduction identification result is obtained by fusing and analyzing the N preliminary reduction identification results.

[0011] In one possible implementation, multidimensional depreciation correlation features are extracted from the original running dataset, including:

[0012] The original running dataset is preprocessed to extract multidimensional parameter features, which include: basic running features, parameter trend features, parameter fluctuation features, and parameter correlation features.

[0013] If at least one of the parameter trend feature, the parameter fluctuation feature, and the parameter correlation feature exceeds the corresponding threshold, the multidimensional parameter feature is determined as a multidimensional depreciation correlation feature.

[0014] In one possible implementation, the N de-rating identification models include at least two of the following: a short-scale convolutional model, a long-scale temporal fusion model, a multi-scale adaptive convolutional model, and a branch-level anomaly identification model.

[0015] The short-scale convolution model is used to identify transient, abrupt derating features.

[0016] The long-scale temporal fusion model is used to identify the temporal evolution characteristics of continuously gradual depreciation.

[0017] The multi-scale adaptive convolution model is used to identify multi-scale derating features;

[0018] The branch-level anomaly identification model is used to identify MPPT branch-level derating characteristics.

[0019] In one possible implementation, the multidimensional devaluation correlation features are input into the short-scale convolutional model to obtain the preliminary devaluation recognition result, including:

[0020] The data input layer of the short-scale convolutional model maps the multidimensional derating correlation features into an enhanced input matrix corresponding to the first short time window. The enhanced input matrix includes: a power parameter group, a voltage parameter group, a current parameter group, and a temperature parameter group. The power parameter group includes inverter power and power change rate. The voltage parameter group includes the voltage of each MPPT branch and voltage change rate. The current parameter group includes the current of each MPPT branch and current change rate. The temperature parameter group includes the inverter internal temperature and temperature change rate.

[0021] The enhanced input matrix is ​​input into the grouped temporal convolutional layer, and one-dimensional convolution is performed according to the power parameter group, the voltage parameter group, the current parameter group and the temperature parameter group respectively to obtain the grouped convolution result;

[0022] The grouped convolution results are input into a multi-scale temporal convolutional layer to extract multi-scale restricted features;

[0023] After inputting the multi-scale restricted features into the residual convolutional layer and the temporal attention layer respectively, they are then input into the local peak pooling layer and the multi-task output layer in sequence to output the preliminary derated identification results. The preliminary derated identification results include the probability of short-term abrupt derated type, the probability of power plateau type derated type, the probability of temperature-triggered restriction, the probability of non-derated short-term perturbation, and the suspected trigger time points corresponding to various derated types.

[0024] In one possible implementation, the multidimensional reduction correlation features are input into the long-scale time-series fusion model to obtain the preliminary reduction identification result, including:

[0025] The data input layer of the long-scale time series fusion model maps the multi-dimensional derating correlation features into a long-time matrix corresponding to the first long-time window. The long-time matrix includes: inverter power, power ratio relative to rated value, power offset relative to historical baseline of similar operating conditions, power change rate, inverter internal temperature, temperature change rate, AC / DC conversion efficiency, voltage sum of each MPPT branch, and current of each MPPT branch.

[0026] The long temporal matrix is ​​input into the LSTM branch and the fully convolutional network branch, respectively;

[0027] The long-term state features output from the LSTM branch and the local morphological features output from the fully convolutional network branch are input into the trend feature fusion layer for feature concatenation to obtain the concatenated features.

[0028] The spliced ​​features are sequentially input into the attention layer and the multi-task output layer to output the preliminary deduction identification results. The preliminary deduction identification results include the probability of long-term trend deduction, the probability of slow deduction, the probability of recovery lag, the probability of atypical deduction, and the suspected deduction time period corresponding to each type of deduction.

[0029] In one possible implementation, the multidimensional devaluation correlation features are input into the multi-scale adaptive convolutional model to obtain the preliminary devaluation recognition result, including:

[0030] The multi-scale adaptive convolutional model's data input layer maps the multi-dimensional devaluation correlation features into input matrices corresponding to the second short-time window, the medium-scale time window, and the second long-time window.

[0031] Input the input matrix corresponding to the second short time window into the short-scale convolution branch, input the input matrix corresponding to the medium-scale time window into the medium-scale convolution branch, and input the input matrix corresponding to the second long time window into the long-scale convolution branch to obtain the reduction features at each scale.

[0032] Input the de-rating features at each scale into the scale attention layer to obtain the attention weights at each scale;

[0033] The multidimensional deduction correlation features are input into the channel attention layer and the time segment attention layer respectively to obtain the attention weights of each channel and the attention weights of each time segment.

[0034] The reduction features at each scale, the attention weights at each scale, the attention weights at each channel, and the attention weights at each time period are input into the cross-scale feature fusion layer to obtain multi-scale fused features.

[0035] The multi-scale fusion features are input into the multi-task output layer to output the preliminary devaluation identification results. The preliminary devaluation identification results include the probability of multi-scale continuous devaluation, the probability of short-term limitation, the probability of medium-term temperature-related limitation, the probability of long-term continuous inefficiency, and the suspected devaluation time periods corresponding to various devaluation types.

[0036] In one possible implementation, the multidimensional depreciation correlation features are input into the branch-level anomaly identification model to obtain the preliminary depreciation identification result, including:

[0037] The multidimensional derating correlation features are divided into power groups, temperature groups, and each MPPT branch group using the data input layer of the branch-level anomaly identification model.

[0038] The power group, the temperature group, and each MPPT branch group are respectively input into the corresponding intra-group convolutional branch in the intra-group convolutional branch layer to extract the local variation of each group.

[0039] The local variation patterns of each group are input into the branch difference feature construction layer. Based on the deviation of the multidimensional parameters of each MPPT branch group from the mean of the parameters of each MPPT branch, branch difference features are generated.

[0040] The branch difference characteristics are input into the cross-group relationship fusion layer to learn the correspondence between the inverter's total power decrease and branch anomalies, temperature increases, and grid connection constraints, and obtain the fusion result;

[0041] The fusion result is input into the physical consistency constraint layer for logical constraint verification to obtain the constraint result.

[0042] The constraint results are input into the multi-task output layer, which outputs the preliminary derating identification results. The preliminary derating identification results include the derating probability of a single MPPT branch abnormality, the derating probability of multiple MPPT branches consistent constraint, the derating probability of temperature linkage, the derating probability of grid-connected side constraint, and the suspected derating time period corresponding to each derating type.

[0043] In one possible implementation, the fusion analysis of the N preliminary rate reduction identification results to obtain the target rate reduction identification result includes:

[0044] Extract the probability, suspected reduction time period, and confidence level of each reduction type from the N preliminary reduction identification results;

[0045] The target time period for deduction is determined based on the overlap between the suspected time periods for deduction output by the N deduction identification models.

[0046] For the same target reduction period, the reduction probabilities output by N preliminary reduction identification results are weighted and summed to obtain the comprehensive reduction probability;

[0047] Based on the reduction type corresponding to the target reduction time period in the N preliminary reduction identification results, determine the reason for the reduction and the corresponding handling opinion;

[0048] Output the target credit reduction identification result, which includes: the reason for the credit reduction, the target credit reduction time period, the comprehensive credit reduction probability, the processing opinion, and the preliminary credit reduction identification result used for judgment.

[0049] In one possible implementation, the target deduction time period is determined based on the overlap between the suspected deduction time periods output by the N deduction identification models.

[0050] Calculate the overlap between the suspected deduction time periods output by any two of the deduction identification models;

[0051] For the same suspected period of credit limit reduction, the weighted sum of the confidence scores of N preliminary credit limit reduction identification results is used to obtain the weighted temporal support of the suspected period of credit limit reduction.

[0052] For each suspected depreciation period, it is marked as a candidate depreciation period if any of the following conditions are met: the overlap between the suspected depreciation periods output by at least two depreciation identification models is greater than a preset overlap threshold; the suspected depreciation period output by the short-scale model completely falls within the suspected depreciation period output by the long-scale model; the weighted temporal support of the suspected depreciation period is greater than a preset support threshold; wherein, the short-scale model is the depreciation identification model used to identify abrupt depreciation features, and the long-scale model is the depreciation identification model used to identify the temporal evolution features of continuous gradual depreciation.

[0053] Based on the time boundaries output by the short-scale model and the long-scale model that output the same candidate reduction time period, the start and end times of the candidate reduction time period are corrected to obtain the target reduction time period.

[0054] In one possible implementation, the step of combining the reduction type corresponding to the target reduction time period from N preliminary reduction identification results to determine the reduction reason and the corresponding processing opinion includes:

[0055] Obtain the combination of reduction types consisting of the reduction types corresponding to the target reduction time period;

[0056] Based on the pre-defined mapping relationship between the combination of reduction types and the reasons for reduction, the reasons for reduction corresponding to the target reduction time period are determined;

[0057] Based on the pre-defined mapping relationship between the reasons for the reduction and the corresponding processing opinions, the processing opinions corresponding to the reasons for the reduction are determined.

[0058] In a second aspect of this application, an inverter is provided, the inverter including a controller, the controller being configured to perform the inverter derating identification method of the first aspect or any implementation thereof described above;

[0059] or,

[0060] The inverter is communicatively connected to the controller, and the inverter is configured to send the original operating dataset to the controller and receive the target derating identification result sent by the controller. The controller is configured to execute the inverter derating identification method of the first aspect or any implementation thereof.

[0061] This application provides an inverter derating identification method. By acquiring the original operating dataset of the inverter and extracting multi-dimensional derating correlation features, these features are input in parallel into N derating identification models with different identification targets and time scales. Each model can capture derating features at different time scales, and the preliminary derating identification results output by the N derating identification models are fused and analyzed to obtain the target derating identification result. On the one hand, the differentiated division of labor among multiple models compensates for the defects of derating omissions and misjudgments caused by a single model only adapting to a single scale and some derating types. It can fully cover various differentiated inverter derating scenarios such as short-term sudden changes, long-term gradual changes, multi-scale continuous changes, and branch anomalies. On the other hand, by fusing the results of multiple models, multi-dimensional judgment evidence is integrated and the identification advantages of each model are balanced to improve the accuracy of derating identification. Attached Figure Description

[0062] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the originals and elements are not necessarily drawn to scale.

[0063] Figure 1 A flowchart illustrating an inverter derating identification method provided in an embodiment of this application;

[0064] Figure 2 A schematic diagram of a short-scale convolution model provided in an embodiment of this application;

[0065] Figure 3 A schematic diagram of a long-scale temporal fusion model provided in an embodiment of this application;

[0066] Figure 4 This is a schematic diagram of a multi-scale adaptive convolution model provided in an embodiment of this application;

[0067] Figure 5 This is a schematic diagram of a branch-level anomaly identification model provided in an embodiment of this application;

[0068] Figure 6 This is a schematic diagram of a controller provided in an embodiment of this application. Detailed Implementation

[0069] The embodiments of this application are described below with reference to the accompanying drawings. The terminology used in the implementation section of this application is for explaining specific embodiments only and is not intended to limit the scope of this application.

[0070] As will be known to those skilled in the art, with the development of technology and the emergence of new scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0071] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms are interchangeable where appropriate; this is merely a way of distinguishing objects with the same attributes in the embodiments of this application. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, so that a process, method, system, product, or apparatus that comprises a series of elements is not necessarily limited to those elements, but may include other elements not explicitly listed or inherent to those processes, methods, products, or apparatuses.

[0072] In photovoltaic power plants, when an inverter experiences a shutdown, offline, or serious fault, the operation and maintenance platform can usually detect it in a timely manner through alarms. However, when an inverter is operating under derating conditions, it can still generate electricity, but the equipment-side alarms may not be triggered, making it easy for the derating problem to be overlooked for a long time.

[0073] Inverter derating can be caused by poor heat dissipation, blocked air ducts, excessively high device temperatures, grid connection limitations, MPPT branch anomalies, triggering of internal inverter protection strategies, or changes in the external environment. After derating occurs, the inverter does not completely stop generating power; rather, its output power is lower than its expected power generation capacity under current irradiance, temperature, and equipment conditions. This often manifests on the monitoring curve as an inefficient state where power is still generated but limited. This state is more insidious than a shutdown fault and is more easily confused with weather changes, module shading, communication anomalies, or grid-side limitations.

[0074] In practical applications, the shape of derating curves is not fixed. Some show a rapid increase in inverter power followed by a sudden plateau, while others show a rapid decrease in power after the inverter temperature reaches a certain range. Some only show slight jumps in voltage, current, or power within a local time period. These problems have obvious local morphological characteristics but are short-lived and have unclear boundaries, making them easy to miss when using daily power generation thresholds or single-point power thresholds. Some derating has long-term trend characteristics rather than obvious abrupt changes. For example, derating caused by poor heat dissipation may gradually deviate from the normal power curve as the inverter temperature gradually increases. A single sampling point may not show any abnormality, but over a long time window, the power will be continuously suppressed. In addition, the duration and scale of derating are unstable and affected by weather, equipment capacity, and installation environment. Derating may last for tens of minutes or span multiple power generation periods. Fixed windows or single-scale features can only cover part of the samples. At the same time, inverter data has the characteristics of multiple physical channels and multiple MPPT branches. If all channels are directly mixed for processing, it is easy to mask branch-level anomalies or misjudge non-derating factors as derating.

[0075] This application provides a method for identifying inverter derating. The method can be implemented by either a locally built-in controller of the inverter or a remote controller deployed outside the inverter and capable of establishing data communication with it. The inverter derating identification method of this application embodiment will be described in detail below with reference to the accompanying drawings.

[0076] Reference Figure 1 , Figure 1 This is a flowchart illustrating an inverter derating identification method provided in an embodiment of this application, as shown below. Figure 1 As shown in the embodiment of this application, an inverter derating identification method may include steps 101 to 104, which are described in detail below.

[0077] 101: Obtain the raw operating dataset of the inverter;

[0078] The raw operational dataset refers to the collection of multi-dimensional time-series data generated by the inverter during operation. Its sources include the inverter's local data acquisition unit, the historical database of the cloud monitoring platform, or the real-time cache of the edge computing gateway. The raw operational dataset provides fundamental data support for subsequent derating identification and comprehensively reflects the inverter's electrical performance and thermal state under different operating conditions.

[0079] For example, the original operating dataset includes real-time operating data and installation information of the inverter. The real-time operating data covers parameters such as MPPT (Maximum Power Point Tracking) branch voltage, MPPT branch current, inverter output power, internal air temperature, DC input power, AC output power, AC-DC conversion efficiency, grid connection voltage, and grid connection frequency; while the installation information includes the inverter's rated power, the number of MPPT branches, etc.

[0080] 102: Extract multidimensional depreciation correlation features from the original running dataset.

[0081] Multidimensional derating correlation features refer to a set of key data indicators that can characterize the derating operation status of inverters after data preprocessing and screening.

[0082] By extracting multidimensional depreciation correlation features, interference data from obviously non-depreciation scenarios such as nighttime, offline, and communication anomalies are filtered out, and multidimensional features highly correlated with depreciation are focused on, thereby reducing the input noise of subsequent models.

[0083] For example, the multidimensional reduction correlation features include: basic operation features, parameter trend features, parameter fluctuation features, and parameter correlation features.

[0084] 103: Input the multidimensional reduction correlation features into N different reduction identification models to obtain the preliminary reduction identification results output by the N reduction identification models.

[0085] Where N > 1, that is, at least two different derating identification models are included. Each model is specifically trained for derating characteristics of a specific time scale or a specific physical mechanism. Therefore, the derating types identified by the N derating identification models are not completely the same.

[0086] For example, the N derating identification models may include at least two of the following: a short-scale convolutional model, a long-scale time-series fusion model, a multi-scale adaptive convolutional model, and a branch-level anomaly identification model. When the inverter experiences a sudden drop in power due to short-term cloud cover, the short-scale convolutional model can quickly capture the abrupt change in the slope of the power curve and output a high-probability short-term abrupt derating result. Conversely, when the inverter experiences a cumulative increase in temperature due to fan failure, leading to power limitation, the long-scale time-series fusion model can identify the negative correlation between temperature and power over a long time window and output a long-term trend-based derating result. Since the different models identify different types of derating and different time scales, they can perform complementary derating identification for the same operating state from different dimensions, avoiding missed or false judgments caused by a single perspective.

[0087] The preliminary reduction identification results output by the reduction identification model include at least the reduction type, and may also include the probability of the reduction type, the suspected reduction period, and the confidence level.

[0088] 104: Perform fusion analysis on N preliminary reduction identification results to obtain the target reduction identification result.

[0089] The N preliminary reduction identification results reflect the reduction identification results of N reduction identification models from different dimensions. Therefore, the N preliminary reduction identification results are not completely the same and have redundancy and complementarity. By fusing and analyzing the N preliminary reduction identification results, we can make up for the defects of missed or misjudged reduction caused by a single model only adapting to a single scale and some reduction types, and finally generate accurate target reduction identification results.

[0090] This embodiment provides an inverter derating identification method. By acquiring the original operating dataset of the inverter and extracting multi-dimensional derating correlation features, these features are input in parallel into N derating identification models with different identification targets and time scales. Each model can capture derating features at different time scales, and the preliminary derating identification results output by the N derating identification models are fused and analyzed to obtain the target derating identification result. On the one hand, the differentiated division of labor among multiple models compensates for the defects of derating omissions and misjudgments caused by a single model only adapting to a single scale and some derating types. It can fully cover various differentiated inverter derating scenarios such as short-term sudden changes, long-term gradual changes, multi-scale continuous changes, and branch anomalies. On the other hand, by fusing the results of multiple models, multi-dimensional judgment evidence is integrated and the identification advantages of each model are balanced to improve the accuracy and robustness of derating identification.

[0091] Step 102 in the above embodiments can be implemented in various ways.

[0092] In one possible implementation, extracting multidimensional depreciation correlation features from the original running dataset includes the following steps 1021-1022:

[0093] 1021: Perform data preprocessing on the original running dataset to extract multidimensional parameter features.

[0094] Data preprocessing refers to a series of data processing steps performed before feature extraction from the original inverter operating data to eliminate noise interference, standardize data format, and remove invalid operating conditions. For example, data preprocessing includes: removing data records from nighttime, offline, and communication failure times to eliminate zero or null value interference in non-power generation states; resampling and time-aligning data at different sampling frequencies to ensure strict synchronization of multi-dimensional parameters such as power, voltage, current, and temperature on the time axis; performing linear interpolation to complete missing data or filling with nearest-neighbor values, and correcting or removing obvious out-of-bounds values, duplicate values, and spike values; finally, normalizing or standardizing the data after the above processing, and generating time-series samples according to a preset window length and sliding step size.

[0095] After data preprocessing, multidimensional parameter features are extracted, including: basic operational features, parameter trend features, parameter fluctuation features, and parameter correlation features.

[0096] Basic operating characteristics refer to physical quantities that reflect the instantaneous operating state of the inverter, including but not limited to: inverter power, voltage of each MPPT branch, current of each MPPT branch, internal temperature, AC / DC conversion efficiency, and the ratio of real-time power to rated value.

[0097] The parameter trend characteristics represent the changing trends of the inverter's multidimensional operating parameters over time. They can be obtained by calculating the first-order and second-order differences of the inverter power, voltage of each MPPT branch, current of each MPPT branch, internal temperature, and AC / DC conversion efficiency over time. They can reflect the monotonic changes in power ramp-up or temperature rise during inverter derating operation.

[0098] The parameter fluctuation characteristics represent the degree of drastic change of the inverter's multidimensional operating parameters within a local time window. They can be obtained by calculating the sliding standard deviation, range, and mean of parameters such as inverter power, voltage of each MPPT branch, current of each MPPT branch, and internal temperature within a preset time step (such as 6 time steps). This characteristic is used to capture short-term steps or oscillations in inverter derating operation scenarios.

[0099] Parameter correlation characteristics represent the coupling relationship between the multidimensional operating parameters of the inverter. They can be obtained by calculating the Pearson correlation coefficient or mutual information between temperature and power, temperature change rate and power change rate, and temperature and MPPT voltage. They are used to identify abnormal correlations that violate normal physical laws, such as temperature increase but power not increase, in inverter derating operation scenarios.

[0100] For example, within a certain sampling window, if the first-order difference of inverter power is continuously negative and its absolute value exceeds a preset slope threshold, while the first-order difference of internal temperature is continuously positive, then the parameter trend characteristics show anomalies. If the correlation coefficient between temperature and power changes from positive to negative or is significantly lower than the historical baseline, then the parameter correlation characteristics also show anomalies. Through this multi-dimensional feature extraction method, various derating correlation characteristics can be comprehensively covered, ranging from instantaneous states to long-term evolution, and from single-point values ​​to multivariate coupling.

[0101] 1022: If at least one of the parameter trend characteristics, parameter fluctuation characteristics, and parameter correlation characteristics exceeds the corresponding threshold, the multidimensional parameter characteristics shall be determined as multidimensional reduction correlation characteristics.

[0102] The process of determining multidimensional derating correlation characteristics is a threshold-based pre-judgment mechanism used to filter samples that clearly do not meet the derating judgment criteria. Specifically, corresponding thresholds are pre-set for parameter trend characteristics, parameter fluctuation characteristics, and parameter correlation characteristics. These thresholds can be dynamic baselines derived from historical normal operating data or fixed limits set based on experimental data from inverter operation. When at least one of the following multidimensional parameter characteristics exceeds its corresponding preset threshold: parameter trend characteristic (e.g., power decrease slope), parameter fluctuation characteristic (e.g., standard deviation of sharp voltage jumps), or parameter correlation characteristic (e.g., negative correlation coefficient between temperature and power), the current multidimensional parameter characteristic is determined to meet the derating judgment criteria and is thus marked as a multidimensional derating correlation characteristic. Conversely, if none of the above three types of characteristics exceed their corresponding preset thresholds, the current multidimensional parameter characteristic is considered to be within the normal fluctuation range or not in an abnormal derating state and is filtered out.

[0103] For example, the threshold for parameter trend characteristics is set as power decreasing by more than 5% of rated power within 10 minutes, and the threshold for parameter correlation characteristics is the correlation coefficient between temperature and power being less than -0.8. If the current multidimensional parameter characteristics indicate that the power has decreased by 6% within 10 minutes (exceeding the trend threshold), even if its parameter fluctuation characteristics and parameter correlation characteristics do not exceed the limits, the current multidimensional parameter characteristics will be immediately identified as multidimensional derating correlation characteristics and passed into the subsequent identification model.

[0104] Pre-screening significantly reduces the amount of data input to subsequent complex models, thereby improving the overall system's response speed and computational efficiency.

[0105] This embodiment, through the coordinated efforts of data preprocessing and multi-dimensional feature extraction, eliminates interference from non-derating scenarios such as nighttime operation, shutdown, and communication anomalies. Based on this, it comprehensively characterizes the inverter's operating status by constructing four complementary feature dimensions: basic operating features, parameter trend features, parameter fluctuation features, and parameter correlation features. Furthermore, by utilizing a threshold pre-judgment mechanism based on parameter trend features, parameter fluctuation features, and parameter correlation features, samples that clearly do not meet the derating determination criteria are filtered out, significantly reducing the amount of data input to the complex downstream derating identification model and improving the overall system's response speed and computational efficiency.

[0106] In another possible implementation, the extracted multidimensional parameter features can be directly determined as multidimensional devaluation correlation features. This embodiment does not impose any specific limitations.

[0107] In the above embodiments, the N de-rating identification models include at least two of the following: short-scale convolution model, long-scale temporal fusion model, multi-scale adaptive convolution model, and branch-level anomaly identification model.

[0108] N derating identification models constitute a parallel decision-making model array, which solves the technical problems that a single model cannot simultaneously capture the short-term sudden change and long-term gradual change characteristics during the inverter derating process, and that it is difficult to distinguish the branch-level subtle faults in multi-channel mixed data.

[0109] Among them, the short-scale convolutional model can adopt the Temp Convolutional Network (TempCNN) architecture, which focuses on capturing local morphological differences within a time window of microseconds to minutes by stacking causal convolutional layers with different dilation rates, and identifies instantaneous abrupt devaluation features.

[0110] Long-scale time series fusion models can employ a dual-branch architecture that integrates Long Short-Term Memory (LSTM) networks and Fully Convolutional Networks (FCNs) to identify the evolution characteristics of continuously gradual derating time series, with time scales ranging from several hours to even spanning multiple power generation periods.

[0111] Multi-scale adaptive convolutional models can employ an attention-based multi-scale convolutional neural network architecture (MACNN) containing short, medium, and long parallel convolutional branches to identify multi-scale devaluation features.

[0112] The branch-level anomaly identification model can adopt a disjoint-CNN architecture and use a physical grouping strategy to decouple the total power, temperature, grid-connected parameters and parameters of each MPPT branch from the input. It can identify MPPT branch-level derating characteristics and perform root cause analysis through a cross-group relationship fusion layer.

[0113] When the system is configured in high-precision mode, all four models mentioned above can be deployed simultaneously; however, in edge computing scenarios with limited computing resources, at least two models can be selected. This combination of at least two models ensures that the system can cover key derating characteristics under different hardware conditions, avoiding missed detections of specific types of derating due to missing models.

[0114] In the case of N de-rating recognition models, including the aforementioned short-scale convolutional models, please refer to... Figure 2 The diagram shows a short-scale convolutional model, which includes: a data input layer, a grouped temporal convolutional layer, a multi-scale temporal convolutional layer, a residual convolutional layer, a temporal attention layer, a local peak pooling layer, and a multi-task output layer. Inputting multi-dimensional devaluation correlation features into the short-scale convolutional model yields preliminary devaluation recognition results, including the following steps A1-A4:

[0115] A1: The data input layer of the short-scale convolutional model maps the multidimensional reduction correlation features to the enhanced input matrix corresponding to the first short time window.

[0116] The first short time window refers to the time window used to capture the instantaneous changes in the inverter's operating state. Its length is shorter than the time window corresponding to the multi-dimensional derating correlation feature. For example, the value range of the first short time window can be from adjacent sampling intervals to tens of minutes, so as to focus on the local morphological differences near the derating trigger point.

[0117] The multidimensional derating correlation features are mapped to the enhanced input matrix corresponding to the first short time window. Specifically, a multi-channel matrix for the first short time window is constructed based on the multidimensional derating correlation features. The multi-channel matrix includes: inverter power, voltage of each MPPT branch, current of each MPPT branch, and internal temperature of the inverter. Then, the rate of change dimension is added to the multi-channel matrix to obtain the enhanced input matrix.

[0118] The enhanced input matrix includes: a power parameter group, a voltage parameter group, a current parameter group, and a temperature parameter group. The power parameter group includes inverter power and power change rate; the voltage parameter group includes the voltage of each MPPT branch and voltage change rate; the current parameter group includes the current of each MPPT branch and current change rate; and the temperature parameter group includes the inverter internal temperature and temperature change rate. Specifically, the power change rate characterizes abrupt changes in the slope of the power curve, the voltage change rate of each MPPT branch reflects the step behavior of the DC-side voltage, the current change rate captures transient current fluctuations, and the temperature change rate correlates the temperature rise rate with the timing relationship of power limitation.

[0119] A2: Input the enhanced input matrix into the grouped temporal convolutional layer, and perform one-dimensional convolution according to the power parameter group, voltage parameter group, current parameter group and temperature parameter group respectively to obtain the grouped convolution result.

[0120] The grouped temporal convolutional layer comprises multiple independent convolutional kernels, each corresponding to a physical parameter group for feature extraction. This avoids premature mixing of different physical channels in the shallow layers of the model and prevents interference between signals with different dimensions and physical meanings. Specifically, the convolutional kernels for the power parameter group focus on learning power plateaus and power cutoffs; those for the voltage parameter group focus on learning voltage drops or oscillations in the MPPT branch; those for the current parameter group focus on learning sudden current changes; and those for the temperature parameter group focus on learning temperature rise curves. For example, if a component shading in an MPPT branch causes a step drop in voltage, but the total power does not fluctuate significantly due to compensation from other branches, traditional hybrid convolution might ignore this voltage feature. However, grouped convolution can independently extract the local variation of this voltage parameter group, generating independent grouped convolution results.

[0121] A3: Input the grouped convolution results into a multi-scale temporal convolutional layer to extract multi-scale restricted features;

[0122] Multi-scale temporal convolutional layers refer to convolutional structures with different receptive fields, set up in parallel or cascaded, to simultaneously extract derating features across different time spans. Because short-term abrupt derating exhibits diverse forms—ranging from dramatic jumps between adjacent sampling points to plateaus lasting tens of minutes—a single-scale convolutional kernel cannot comprehensively cover all of them. Therefore, this layer employs short, medium, and long convolutional kernels with dilation rates to process the input data.

[0123] Among them, short convolutional kernels are used to extract step features between adjacent sampling points, such as instantaneous power drops; medium convolutional kernels are used to extract plateau features over a range of tens of minutes, such as the stable state after power enters the confined region; and long convolutional kernels with dilation are used to extract the confined trend within a longer short time window. For example, short convolutional kernels can produce a high response to millisecond-level power cut-offs caused by instantaneous grid voltage exceeding limits; while for the gradual entry into a power-limited state caused by device overheating, medium and long convolutional kernels can capture the continuous pattern of power peaks being gradually suppressed. Through parallel extraction of multi-scale features, the model can adaptively match derating events of different durations, avoiding missed detections caused by fixed windows.

[0124] A4: After inputting the multi-scale restricted features into the residual convolutional layer and the temporal attention layer respectively, they are then input into the local peak pooling layer and the multi-task output layer in sequence to output the preliminary derated identification results. The preliminary derated identification results include the probability of short-term abrupt derated type, the probability of power plateau type derated type, the probability of temperature-triggered restriction, the probability of non-derated short-term perturbation, and the suspected trigger time points corresponding to various derated types.

[0125] Among them, the residual convolutional layer is used to preserve the original local morphology of multi-scale restricted features and prevent the loss of key abrupt boundary information during the process of network deepening; the temporal attention layer is used to calculate the importance weight of each time step in the time series of multi-scale restricted features, automatically increase the weight of key segments such as power slope drop, power plateau, and power limitation after high temperature, while suppressing noise interference such as short-term fluctuations or sampling spikes caused by cloud cover.

[0126] Local peak pooling layers are used to retain the maximum value of each type of local anomaly response and its corresponding time index, thereby achieving a mapping from feature space to time space and accurately outputting the suspected trigger time point.

[0127] The multi-task output layer, based on the fused features of the input, outputs classification probabilities and temporal location information for various derating types in parallel. For example, when a sustained increase in temperature is detected followed by a power plateau, the temporal attention layer assigns high weight to this time period, the residual connection ensures that the steepness of the plateau's starting point is not smoothed, and the local peak pooling layer pinpoints the exact moment when power begins to be limited. Finally, the multi-task output layer provides a high probability of temperature-triggered limitation and a specific suspected trigger time point. This series of processes not only improves recognition accuracy but also provides direct quantitative evidence for determining the start time of the target derating period by outputting specific trigger time points, effectively solving the problem of ambiguous trigger point location in traditional methods.

[0128] For issues such as power plateaus, short-term steps, and rapid limiting after temperature triggering in inverter derating, the abnormal information is usually concentrated within a short time range, manifested as sudden changes in curve slope, truncation of the rising process, or a change in local waveform from continuous change to plateau change. This embodiment constructs an enhanced input matrix containing the rate of change of multiple physical quantities, and combines a short-scale convolution model with grouped temporal convolution, multi-scale temporal convolution, residual connections, temporal attention mechanisms, and local peak pooling layers. Through its shallow grouped temporal convolution layers, independent one-dimensional convolution operations are performed on the power parameter group, voltage parameter group, current parameter group, and temperature parameter group, respectively, to avoid premature mixing of channels with different physical meanings in the early stage of feature extraction, thereby preserving the independent abrupt change signals of each channel. Subsequently, using short convolution kernels and convolution kernels with dilation rates in the multi-scale temporal convolution layers, step features between adjacent sampling points and plateau features within tens of minutes are extracted. Through local peak pooling layers, the abstract feature response is transformed into specific suspected trigger time points, which can quickly respond to sudden changes in operating conditions and effectively distinguish between real derating events and transient noise interference.

[0129] In the case of N devaluation identification models, including the aforementioned long-scale temporal fusion model, please refer to... Figure 3The diagram shows a long-scale temporal fusion model, which includes: a data input layer, an LSTM branch, a fully convolutional network branch, a trend feature fusion layer, an attention layer, and a multi-task output layer. Inputting multi-dimensional reduction-of-quotient correlation features into the long-scale temporal fusion model yields preliminary reduction-of-quotient recognition results, including the following steps B1-B4:

[0130] B1: The data input layer of the long-scale time series fusion model maps the multidimensional devaluation correlation features into a long-time series matrix corresponding to the first long time window.

[0131] The first long time window can refer to a continuous time series covering several hours to multiple power generation periods, used to capture the slow derating process caused by poor heat dissipation, blocked air ducts, or device temperature rise.

[0132] The long-term time-series matrix includes: inverter power, power ratio relative to rated value, power offset relative to historical baseline under similar operating conditions, power change rate, inverter internal temperature, temperature change rate, AC / DC conversion efficiency, voltage sum of each MPPT branch, and current of each MPPT branch. The power ratio relative to rated value is the ratio of real-time power to the inverter's rated power, used to eliminate dimensional differences between devices of different capacities. The power offset relative to historical baseline under similar operating conditions is the difference between the real-time power and the standard power curve in the historical database under the same irradiance and ambient temperature; this feature is used to identify implicit derating where the power has not dropped but is significantly lower than historical levels under the same conditions. The AC / DC conversion efficiency can be the ratio of AC output power to DC input power, used to reflect changes in internal inverter losses. For example, when the inverter's heat dissipation efficiency decreases due to dust accumulation, although the instantaneous power may not trigger a low-limit alarm, its offset relative to the historical baseline will remain negative and its absolute value will gradually increase, accompanied by an abnormal increase in the temperature change rate. By constructing a long time series matrix containing the aforementioned multidimensional parameters, it is possible to transform insignificant anomalies at single points into feature sequences with obvious evolutionary trends.

[0133] B2: Input the long time series matrix into the LSTM branch and the fully convolutional network branch respectively.

[0134] The long-term matrix is ​​simultaneously fed into the LSTM branch and the fully convolutional network branch for parallel processing. The LSTM branch, through its internal input, forget, and output gate mechanisms, transmits state information between multiple sampling times, focusing on remembering the continuous evolutionary dependencies from normal power generation to derating, sustained derating, and then recovery, thus addressing the weakness of pure convolutional networks in modeling long-term dependencies. The fully convolutional network branch, on the other hand, uses convolutional kernels with different receptive fields to scan the long-term matrix in parallel, focusing on extracting waveform details such as local power plateaus, peak derating, curve inflections, short-term fluctuations, and recovery hysteresis, thus addressing the insensitivity of pure LSTM networks to waveform boundaries. For example, for a slow derating caused by device temperature rise, the LSTM branch can record the causal timeline between the continuous temperature increase and the gradual power derating, while the fully convolutional network branch can accurately locate minute steps or slope changes on the power curve.

[0135] B3: Input the long-term state features output from the LSTM branch and the local morphological features output from the fully convolutional network branch into the trend feature fusion layer for feature concatenation to obtain the concatenated features.

[0136] Feature concatenation refers to connecting the feature vectors output by two branches along the channel dimension to form a high-dimensional fused feature that includes state evolution and local morphology. This process involves not only simple concatenation but also deep fusion incorporating temperature-power hysteresis, power deviation from historical baselines, and relative temperature change rates. For example, when the LSTM branch detects a high temperature while the fully convolutional network branch detects a flat-top feature in the power curve, the fusion layer strongly correlates these two features, significantly improving its ability to represent temperature-triggered derating in this segment. This fusion mechanism ensures that the model can grasp both the macroscopic trend of derating and the microscopic morphological changes, avoiding missed detections caused by a single perspective.

[0137] B4: Input the spliced ​​features into the attention layer and the multi-task output layer in sequence, and output the preliminary reduction recognition results. The preliminary reduction recognition results include the probability of long-term trend reduction, the probability of slow reduction, the probability of recovery lag, the probability of atypical reduction, and the suspected reduction time period corresponding to each reduction type.

[0138] The attention layer is used to dynamically adjust feature weights, increasing the weight of key segments and suppressing noise interference. Specifically, the attention layer calculates weight coefficients based on splicing features, assigning high weights to segments with continuously rising temperatures, gradually decreasing power peaks, and power recovery slower than irradiation or reference power recovery, while reducing the weights of short-term cloud cover, occasional spikes, or low-quality sampling segments.

[0139] The multi-task output layer, based on the fused features of the input, outputs in parallel the classification probabilities of various derating types and their corresponding suspected derating time periods. Long-term trend derating probability represents the possibility that derating will gradually develop over time; slow derating probability represents the possibility that the derating rate is low but the duration is long; recovery lag probability represents the possibility that power fails to recover in time after external conditions improve; atypical derating probability represents the possibility that it does not conform to common derating patterns but exhibits anomalies. For example, if the attention layer detects a high negative correlation between the temperature rise curve and the power fall curve within a certain time period and that this correlation is long-lasting, it will output a high long-term trend derating probability and mark this time period as a suspected derating time period.

[0140] The long-scale time-series fusion model provided in this embodiment utilizes the memory gating mechanism of LSTM branches to transfer state information between multiple sampling times, capturing the entire process dependency from normal power generation to power limitation and then to continuous power limitation. Simultaneously, it uses fully convolutional network branches to extract local power plateaus, curve inflections, and recovery lag patterns within long windows in parallel, and uses a trend feature fusion layer to stitch together long-term state features and local morphological features, achieving complementary advantages in state evolution and waveform details. Furthermore, it strengthens the weights of key segments such as temperature rise, power limitation, and recovery lag through an attention layer, effectively reducing interference from short-term cloud cover or sampling noise. This enables accurate identification of atypical faults such as slow derating and recovery lag caused by poor heat dissipation, airflow blockage, or device temperature rise, significantly improving the accuracy of derating diagnosis for residential inverters under complex operating conditions.

[0141] In the case of N devaluation recognition models, including the aforementioned multi-scale adaptive convolutional model, please refer to... Figure 4 The diagram illustrates a multi-scale adaptive convolutional model, which includes: a data input layer, short-scale convolutional branches, medium-scale convolutional branches, long-scale convolutional branches, a scale attention layer, a channel attention layer, a temporal segment attention layer, a cross-scale feature fusion layer, and a multi-task output layer. Inputting multi-dimensional reduction-related features into the multi-scale adaptive convolutional model to obtain preliminary reduction-based recognition results involves the following steps C1-C6:

[0142] C1: The data input layer of the multi-scale adaptive convolutional model maps the multi-dimensional reduced-rate correlation features into the input matrices corresponding to the second short time window, the medium-scale time window, and the second long time window.

[0143] The second short-term window, the mesoscale window, and the second long-term window are three sliding windows of different lengths pre-defined based on the uncertainty of the inverter derating duration. For example, the second short-term window can be a time range of tens of minutes to capture rapid limiting or short-term step features; the mesoscale window can be a time range of several hours to capture power limiting features caused by sustained temperature increases; and the second long-term window can be half a day, a day, or multiple power generation periods to capture sustained inefficiency features across time periods. The data input layer maps the same set of multi-dimensional derating-related features to these three time windows of different lengths, forming three independent input matrices. By constructing multi-scale input matrices, it is ensured that subsequent model branches can simultaneously perceive local abrupt changes and long-term trends, avoiding feature omissions caused by a single fixed window.

[0144] C2: Input the input matrix corresponding to the second short time window into the short-scale convolution branch, input the input matrix corresponding to the medium-scale time window into the medium-scale convolution branch, and input the input matrix corresponding to the second long time window into the long-scale convolution branch to obtain the reduction features at each scale.

[0145] The short-scale, medium-scale, and long-scale convolutional branches are structurally similar but have different receptive fields within a convolutional neural network, each processing the input matrix at its corresponding time scale. The short-scale branch uses a smaller kernel or fewer layers to extract features such as step changes between adjacent sampling points, short-term plateaus, and local peak-valley variations, identifying rapid limiting conditions within tens of minutes. The medium-scale branch uses a medium-sized kernel or an appropriate expansion rate to extract features such as continuously rising temperature, persistent power limiting, delayed power recovery, and medium-term plateauing, identifying temperature-related limiting conditions over several hours. The long-scale branch uses a larger kernel or a higher expansion rate to extract features such as persistent inefficiency across power generation periods, intraday peak limiting, long-term underperformance relative to historical baselines, and recurring derating features across multiple windows, identifying persistent inefficiencies across days or multiple time periods. Each scale's convolutional branch operates independently and in parallel, outputting derating feature vectors at its respective scale. For example, when the inverter's power gradually decreases over 3 hours due to poor heat dissipation, the mesoscale convolutional branch can extract obvious temperature-power negative correlation evolution features, while the short-scale branch may only detect small local fluctuations, and the long-scale branch may detect the overall inefficiency trend. This parallel extraction mechanism ensures that derating features of different durations can be effectively captured.

[0146] C3: Input the de-rating features of each scale into the scale attention layer to obtain the attention weights of each scale.

[0147] The scale attention layer is a neural network layer used to dynamically evaluate the importance of features at different time scales, automatically calculating the weight coefficients of the output features of short-scale, medium-scale, and long-scale branches. When the derating features at each scale exhibit short-term abrupt derating (such as rapid recovery after cloud cover or transient failure), the scale attention layer assigns higher weights to short-scale derating features; when the derating features at each scale exhibit temperature-related progressive derating, it assigns higher weights to medium-scale derating features; and when the derating features at each scale exhibit inefficient operation throughout the day, it assigns higher weights to long-scale derating features. The generation process of these weights is based on global pooling and fully connected mapping of the feature vectors at each scale, and is finally normalized to a probability distribution using the Softmax function. By introducing the scale attention mechanism, the model can adaptively select the best-matching decision scale, avoiding subjective bias caused by manually setting fixed thresholds or rules, and improving the robustness of identifying derating events with uncertain durations.

[0148] C4: Input the multidimensional reduction correlation features into the channel attention layer and the time segment attention layer respectively to obtain the attention weights of each channel and the attention weights of each time segment.

[0149] The channel attention layer is used to evaluate the contribution of different physical quantity channels (such as power, temperature, voltage, current, etc.) to the derating determination in multidimensional derating correlation characteristics. Since different types of derating are often triggered by specific physical quantity anomalies (for example, temperature-triggered derating mainly depends on the temperature channel, and branch anomaly-triggered derating mainly depends on the MPPT branch channel), the channel attention layer can automatically focus on key physical quantities and suppress the influence of noise channels.

[0150] The temporal segment attention layer is used to locate key anomalous segments in the time dimension, such as power plateaus, high-temperature periods, or specific time periods when branch anomalies occur. This layer scores the time series, highlighting time segments strongly correlated with derating and reducing the weight of segments with cloud cover, occasional spikes, or missing data. For example, in samples with brief communication loss or sampling spikes, the temporal segment attention layer reduces the weight of these anomalies while increasing the weight of segments with persistently limited power. In multi-MPPT branch scenarios, if only a single branch experiences a voltage step, the channel attention layer increases the weight of that branch channel.

[0151] Through a dual attention mechanism, the model can further extract the most discriminative channel information and time segments before feature fusion, thereby enhancing its noise resistance to complex interference environments.

[0152] C5: Input the reduction features at each scale, the attention weights at each scale, the attention weights at each channel, and the attention weights at each time period into the cross-scale feature fusion layer to obtain multi-scale fused features.

[0153] The cross-scale feature fusion layer receives de-rating features at each scale, attention weights at each scale, attention weights at each channel, and attention weights at each time period. First, it uses the scale attention weights to perform a weighted summation of the de-rating features at each scale to achieve adaptive aggregation in the time dimension. Then, it combines the channel attention weights and the time period attention weights to perform fine-grained correction and enhancement on the aggregated features to ensure that the feature representation of key physical quantities is maximized in key time segments.

[0154] The cross-scale feature fusion layer can be composed of fully connected layers or element-wise multiplication operations to generate a comprehensive feature vector containing multi-scale temporal evolution information, key physical quantity orientation, and key time segment positioning information, i.e., multi-scale fused features.

[0155] For example, when multidimensional dereasing correlation features simultaneously exhibit short-term power step (short-scale feature) and long-term temperature accumulation effect (long-scale feature), and the temperature channel is marked with high weight, the cross-scale feature fusion layer organically combines these two types of evidence to form a comprehensive representation that reflects both abrupt changes and trends. This fusion method avoids information redundancy caused by simple splicing, achieves reliable aggregation of multi-source evidence, and lays the foundation for subsequent accurate classification.

[0156] C6: Input the multi-scale fusion features into the multi-task output layer and output the preliminary de-rating identification results. The preliminary de-rating identification results include the probability of multi-scale continuous de-rating, the probability of short-term limitation, the probability of medium-term temperature-related limitation, the probability of long-term continuous inefficiency, and the suspected de-rating time periods corresponding to various de-rating types.

[0157] The multi-task output layer, based on the fusion features of the input, outputs in parallel the classification probabilities of various depreciation types and their corresponding suspected depreciation time periods. The multi-scale persistent depreciation probability represents the overall probability that a sample belongs to persistent depreciation at any scale; the short-term restricted probability specifically refers to the possibility of depreciation caused by short-term abrupt changes; the medium-term temperature-related restricted probability specifically refers to the possibility of depreciation caused by the cumulative effect of temperature; and the long-term persistent inefficiency probability specifically refers to the possibility of inefficient operation across time periods.

[0158] The multi-scale adaptive convolutional model provided in this embodiment constructs three parallel time window input matrices—short-scale, medium-scale, and long-scale—corresponding to different time spans from adjacent sampling points to tens of minutes, several hours, and more than half a day, respectively. Each scale convolutional branch independently extracts fast-limited, temperature-related limited, and persistently inefficient features across time periods within its corresponding time range. A scale attention layer dynamically allocates the weights of each scale branch. Based on this, a channel attention layer and a time segment attention layer are introduced to further focus on key physical quantities (such as temperature and branch parameters) and key time segments (such as power plateau periods), effectively suppressing interference from non-derating factors such as cloud cover, noise, and missing data. This adaptive mechanism of multi-scale input, dynamic weighting, and cross-dimensional fusion significantly improves the model's recognition accuracy and generalization ability for complex derating scenarios such as short-term abrupt changes, long-term gradual changes, temperature linkages, and branch anomalies, providing high-quality feature input for subsequent multi-model fusion analysis.

[0159] In the case where N derating identification models include the aforementioned branch-level anomaly identification models, please refer to [link / reference]. Figure 5 The diagram shows a branch-level anomaly detection model. This model includes: a data input layer, an intra-group convolutional branch layer, a branch difference feature construction layer, a cross-group relationship fusion layer, a physical consistency constraint layer, and a multi-task output layer. Inputting multi-dimensional deduction correlation features into the branch-level anomaly detection model to obtain preliminary deduction detection results includes the following steps D1-D6:

[0160] D1: The data input layer of the branch-level anomaly identification model is used to divide the multidimensional derating correlation features into power groups, temperature groups, and each MPPT branch group.

[0161] The power group includes parameters directly related to the overall output capability, such as the inverter's total power, power change rate, and power ratio relative to the rated value. The temperature group includes the inverter's internal temperature parameters. Each MPPT branch group contains the voltage, current, power, and their change rates for the corresponding branch. This division method is based on the independence and correlation of each physical channel of the inverter, avoiding premature mixing of parameters with different physical meanings at the shallow level of the model, thus laying the foundation for subsequent branch-level anomaly decoupling analysis. For example, when the inverter contains two MPPTs, the data input layer assigns the voltage, current, and power parameters belonging to MPPT1 in the original feature matrix to the MPPT1 branch group, and the corresponding parameters belonging to MPPT2 to the MPPT2 branch group, while global parameters such as total power and internal temperature are assigned to the power group and temperature group, respectively. Through this grouping, the model can focus on the shape of the total power platform, the trend of temperature evolution, and the local changes of each branch, providing a structured input for accurate identification of branch-level anomalies.

[0162] D2: Input the power group, temperature group, and each MPPT branch group into the corresponding intra-group convolutional branch in the intra-group convolutional branch layer to extract the local variation pattern of each group.

[0163] Intra-group convolutional branch layers refer to feature extraction modules composed of multiple parallel one-dimensional convolutional neural network branches, with each branch specifically processing data from a physical group. Each intra-group convolutional branch extracts the local morphological features of its corresponding group parameters in the time dimension through multiple layers of one-dimensional convolution operations. For example, power plateaus, steps, or abrupt slope changes in the power group; continuous temperature rises or high-level plateaus in the temperature group; and local changes such as voltage drops, current fluctuations, or power limitations in a certain MPPT branch group.

[0164] By using independent convolutions within physically grouped modules, the unique anomalous patterns within each channel are preserved and enhanced, preventing the total power or signals from other channels from masking subtle faults at the branch level. For example, for the MPPT1 branch group, its intra-group convolutional branches may extract features where the voltage drops continuously while the current remains essentially constant over a short period, indicating potential component obstruction or connection faults in that branch. By independently extracting local morphologies from each group, the model provides a clean and physically meaningful feature representation for subsequent branch difference analysis and cross-group relationship fusion.

[0165] D3: Input the local variation patterns of each group into the branch difference feature construction layer, and generate branch difference features based on the deviation of the multidimensional parameters of each MPPT branch group from the mean of the parameters of each MPPT branch.

[0166] The branch difference feature construction layer is used to quantify the differences in operating status between different MPPT branches. Specific branch difference features include: the difference between the voltage of each MPPT branch and the average voltage of all branches, the deviation ratio of the current of each branch from the average current, the power proportion of each branch, the maximum and minimum power difference between branches, and the coefficient of variation of branch parameters. These features directly reflect the degree of imbalance between branches and are key evidence for identifying single-branch anomalies or consistent limitation across multiple branches.

[0167] For example, when the MPPT1 branch experiences a power decrease due to local shading, its voltage, current, and power parameters will be significantly lower than those of the MPPT2 branch. In this case, the calculated branch difference characteristics, such as the branch voltage difference and the deviation in power proportion, will show obvious peaks. By generating such difference characteristics, the model can transform potentially inconspicuous branch anomalies on the total power curve into quantifiable and comparable feature vectors, thereby providing direct evidence for root cause analysis.

[0168] D4: Input the branch difference characteristics into the cross-group relationship fusion layer, learn the correspondence between the inverter's total power decrease and branch anomalies, temperature increases, and grid connection constraints, and obtain the fusion result.

[0169] The cross-group relationship fusion layer integrates information from multiple sources, including power groups, temperature groups, and branch difference features, and learns their complex mapping relationships. Through fully connected layers or attention mechanisms, the cross-group relationship fusion layer learns the causal relationship between the global phenomenon of total power reduction and local or related factors such as specific branch anomalies, temperature increases, and grid voltage anomalies. For example, when branch difference features show a significant deviation in a single branch, while power group features show a step decrease in total power, and temperature group features show no significant change, the cross-group relationship fusion layer may learn a mapping relationship where a single branch anomaly leads to limited total power. Conversely, if temperature group features show a continuous temperature increase that highly overlaps with the power reduction period, while branch difference features show synchronous limitation across all branches, it may learn a mapping relationship where poor heat dissipation leads to consistent limitation across multiple branches. Through this fusion, the model can transcend the limitations of a single channel, understand the physical mechanism of derating at the system level, and thus improve the accuracy of root cause identification.

[0170] D5: Input the fusion result into the physical consistency constraint layer for logical constraint verification to obtain the constraint result.

[0171] The constraints used for logical constraint verification in the physical consistency constraint layer conform to the physical laws governing inverter operation. For example, for determining abnormal derating in a single MPPT branch, the constraint requires that the parameters of at least one branch deviate significantly from those of other branches; for consistent derating in multiple MPPT branches, the parameters of multiple branches must decrease synchronously and deviate in the same direction; for temperature-linked derating, there must be a clear temporal correlation between the period of high temperature or rapid rise and the period of power limitation; for grid-connected constraint derating, abnormal grid-connected voltage, frequency, or power limiting parameters must occur simultaneously with power limitation.

[0172] The physical consistency constraint layer can filter out model illusions that do not conform to physical laws, ensuring that the final identification results are interpretable and credible. For example, if the model outputs a high probability of a single-branch anomaly only due to noise, but the branch difference features do not show a significant deviation, the physical consistency constraint layer will suppress this output to avoid false alarms.

[0173] D6: Input the constraint results into the multi-task output layer and output the preliminary derating identification results. The preliminary derating identification results include the derating probability of single MPPT branch abnormal type, the derating probability of multiple MPPT branch consistent constraint type, the derating probability of temperature linkage type, the derating probability of grid-connected side constraint type, and the suspected derating time period corresponding to each derating type.

[0174] The multi-task output layer, based on the input fusion features, outputs in parallel the classification probabilities of various derating types and their corresponding suspected derating time periods. This layer, based on the constrained fusion features, calculates the occurrence probability of four typical derating root causes (single branch anomaly, multiple branch consistent limitation, temperature linkage, and grid connection constraint) and identifies the possible start and end time periods for each type. For example, the output results might show: a single MPPT branch anomaly derating probability of 0.85, with a suspected time period of 10:15-10:45; and a temperature linkage derating probability of 0.3, with a suspected time period of 11:00-12:30.

[0175] The branch-level anomaly identification model provided in this embodiment first uses the data input layer to divide the multidimensional derating correlation features into power groups, temperature groups, and each MPPT branch group, ensuring that each physical channel remains independent during the input stage. Then, through intra-group convolutional branch layers, the local variation patterns within each group are extracted, and a branch difference feature construction layer is used to calculate the deviation of each MPPT branch parameter from the mean, the coefficient of variation, and the maximum and minimum differences, generating explicit branch difference features. A cross-group relationship fusion layer learns the correspondence between total power reduction and specific branch anomalies, and a physical consistency constraint layer verifies the logical conditions for significant deviations in a single branch, thereby accurately outputting the probability of anomaly-type derating in a single MPPT branch. This layer-by-layer, mutually verifying mechanism enables the branch-level anomaly identification model to accurately distinguish derating scenarios with similar total power performance but different physical causes, effectively avoiding the averaging effect of the total power curve from masking subtle branch-level faults.

[0176] In one possible implementation, step 103 in the above embodiment performs a fusion analysis on N preliminary rate reduction identification results to obtain the target rate reduction identification result, specifically including the following steps 1031-1035:

[0177] 1031: Extract the probability, suspected reduction time period, and confidence level of each reduction type from N preliminary reduction identification results.

[0178] The probability of derating type represents the likelihood that the derating identification model considers the current sample to belong to a specific derating pattern (such as short-term mutation, long-term trend, branch anomaly, etc.); the suspected derating time period indicates the start and end time range or trigger point of the time when the model judges the anomaly to have occurred; and the confidence level reflects the reliability of the model's output results based on the quality of the input data and the feature matching degree.

[0179] 1032: Determine the target time period for deduction based on the overlap between the suspected time periods output by N deduction identification models.

[0180] Overlap is an indicator that quantifies the degree to which the suspected reduction time periods output by any two models overlap on the time axis. It can be calculated as the ratio of the intersection length to the union length, or the ratio of the intersection length to the shorter time period length, to determine whether the shorter time period falls within the longer time period.

[0181] 1033: For the same target reduction period, the reduction probabilities output by N preliminary reduction identification results are weighted and summed to obtain the comprehensive reduction probability.

[0182] The weight allocation strategy for the de-indexing identification model can be statically preset or dynamically adjusted based on data completeness, time period overlap, and prediction results. For example, when MPPT branch data is complete, the weight of the branch-level anomaly identification model is increased; when samples have short-term missing data, the weight of the short-scale convolution model is reduced.

[0183] Specifically, after selecting the output probabilities of all models covering the target depreciation period, they are multiplied by their respective dynamic weights and then summed to obtain the comprehensive depreciation probability. For example, if the short-scale model output probability is 0.9 (weight 0.3), the long-scale model output probability is 0.8 (weight 0.4), and the multi-scale model output probability is 0.85 (weight 0.3) within the target period, then the comprehensive depreciation probability is 0.9 × 0.3 + 0.8 × 0.4 + 0.85 × 0.3 = 0.845. Through weighted fusion, the misjudgment fluctuations of individual models under noise interference can be effectively smoothed out, making the final probability value more accurately reflect the actual operating state of the inverter.

[0184] 1034: Based on the reduction type corresponding to the target reduction time period in the N preliminary reduction identification results, determine the reason for the reduction and the corresponding handling opinion.

[0185] The corresponding handling opinions for the reasons for the reduction realize the semantic mapping from abstract probability output to specific operation and maintenance decisions, and transform the AI ​​recognition results into executable handling opinions through logical reasoning.

[0186] The derating cause refers to the physical root cause of the inverter's power limitation, such as poor heat dissipation, branch circuit faults, or grid constraints; the handling opinion is a specific maintenance or adjustment suggestion given for the specific cause. First, obtain the type combination composed of the derating types output by each model with high probability within the target derating time period. Then, determine the derating cause according to the pre-set mapping relationship between the derating type combination and the derating cause. Finally, generate the handling opinion based on the mapping relationship between the derating cause and the handling opinion.

[0187] For example, if the short-scale model outputs a temperature-triggered limitation and the long-scale model outputs a long-term trend-type derating, the type combination points to a heat dissipation-limited derating, indicating that the derating is caused by airflow blockage or fan failure. The corresponding handling suggestions are to check the inverter's airflow unobstructedness, clean the accumulated dust, and test the fan speed. If the branch-level anomaly identification model outputs a single MPPT branch anomaly and the total power decrease is limited, the derating is determined to be caused by component or DC cable failure. The corresponding handling suggestions are to check the component strings and connectors of the corresponding MPPT branch.

[0188] This attribution mechanism, which is based on complementary derating types from multiple model outputs, can distinguish between derating scenarios with similar total power curves but different physical mechanisms, ensuring that the proposed solutions are targeted and actionable.

[0189] 1035: Output the target reduction identification result, which includes: the reason for the reduction, the target reduction time period, the comprehensive reduction probability, the processing opinion, and the preliminary reduction identification result used in the judgment.

[0190] The target deduction identification result can be a well-structured and information-rich diagnostic report. The target deduction identification result not only includes the final conclusive data (cause, time, probability, opinion), but also retains traceability information (preliminary results involved in the judgment) to facilitate manual review or subsequent model optimization.

[0191] Specifically, the determined reasons for the rate reduction, the corrected target rate reduction period, the calculated comprehensive rate reduction probability, the generated processing opinions, and the original preliminary rate reduction identification results of each model (including the type probability, confidence level, etc. of each model) are packaged into a standard format data package or visual report for output.

[0192] For example, the output can be displayed as follows:

[0193] Reason for the throttling: Limited heat dissipation;

[0194] Target time period: 10:15-11:00;

[0195] Overall probability: 0.845;

[0196] Recommended solution: Inspect the air ducts and fans;

[0197] Participating models: short-scale model (probability 0.9), long-scale model (probability 0.8).

[0198] As a result, maintenance personnel can directly perform on-site troubleshooting based on the results, and at the same time, they can understand the contribution of each model by viewing the preliminary results of the judgment, thus realizing end-to-end intelligent diagnosis and decision support for inverter derating status.

[0199] This embodiment utilizes time-time overlap calculations to achieve precise fusion of short-scale abrupt changes and long-scale sustained intervals, resolving the issue of fragmented output time from multiple models. By weighted summation of the confidence levels of different models, it leverages the advantages of each model in specific scenarios, improving the robustness of the comprehensive judgment. More importantly, by mapping the derating types output by multiple models to the derating causes and handling suggestions in the domain knowledge base, it successfully transforms the black-box AI probabilistic output into transparent, executable operation and maintenance instructions. This mechanism not only overcomes the shortcomings of a single model in simultaneously considering local abrupt changes and long-term trends but also effectively reduces false alarms caused by weather fluctuations, communication noise, or non-derating factors. Ultimately, it outputs high-quality diagnostic results containing clear causes, precise times, and specific suggestions, significantly improving the operation and maintenance efficiency and fault response speed of residential photovoltaic power stations.

[0200] Furthermore, the target deduction identification result can also include an overall judgment result on whether a deduction has been reduced. For example, if the overall deduction probability is greater than a first threshold, the output is a target deduction identification result indicating a clear deduction; if the overall deduction probability is between the first and second thresholds, the output is a target deduction identification result indicating a suspected deduction; if the overall deduction probability is less than the second threshold, the output is a target deduction identification result indicating no deduction, wherein the first threshold is greater than the second threshold.

[0201] For example, the aforementioned explicit rate reduction, suspected rate reduction, and no rate reduction can be represented by status identifiers. The explicit rate reduction identifier directly triggers the high-priority alarm mechanism of the operation and maintenance platform, automatically generating a fault work order and pushing it to the operation and maintenance personnel's terminal, recommending immediate on-site inspection or remote intervention. The business logic corresponding to the suspected rate reduction identifier does not immediately dispatch an emergency work order, but rather triggers an extended observation mechanism or manual review process. For example, the system can automatically retrieve irradiation data before and after the relevant time period, compare it with the operating data of adjacent inverters, or mark it as pending confirmation, waiting for the next data window update. The no-rate-reduction identifier is handled silently, without generating any alarms or work orders, only recording the current operating status in the historical database for subsequent trend analysis.

[0202] By introducing a three-level judgment mechanism based on the first and second thresholds, the probability of continuous numerical reduction is mapped to discrete business decision levels (clear reduction, suspected reduction, no reduction), thus achieving seamless integration between AI algorithm output and on-site operation and maintenance business processes.

[0203] In one possible implementation, determining the target deduction period based on the overlap of suspected deduction periods specifically includes the following steps E1-E4:

[0204] E1: Calculate the overlap of the suspected deduction time periods output by any two deduction identification models.

[0205] Specifically, select any two suspected time periods from the outputs of N rate reduction identification models, denoted as time period A and time period B. The overlap O can be expressed as:

[0206] O = len(A ∩ B) / len(A ∪ B);

[0207] Where len(A ∩ B) represents the length of the intersection of the two time intervals, and len(A ∪ B) represents the length of the union of the two time intervals.

[0208] To accommodate situations where TempCNN outputs short-term trigger points or shorter trigger intervals while LSTM-FCN and MACNN output longer sustained intervals, the following overlap can also be used:

[0209] O = len(A ∩ B) / min(len(A), len(B));

[0210] The ratio of the intersection length to the shorter time interval length indicates whether the shorter time interval falls within the longer time interval.

[0211] For example, when the suspected trigger point output by the short-scale convolutional model falls within the continuous restricted time period output by the long-scale temporal fusion model, if the ratio of the intersection to the shorter time period is used, the overlap is close to 1, indicating that the short-term abrupt change features are encompassed by the long-term trend features. The calculation process of this overlap excludes isolated anomalous segments caused by noise interference from a single model, ensuring that only time regions captured by multiple models are considered as high-confidence depreciation intervals.

[0212] E2: For the same suspected period of credit reduction, the weighted sum of the confidence scores of N preliminary credit reduction identification results is used to obtain the weighted time support of the suspected period of credit reduction.

[0213] Weighted temporal support integrates the confidence levels of each model for a specific time segment and their model weights, representing the overall strength of evidence for devaluation occurring within that time period. Specifically, for each sampling moment or time segment on the time axis, all N devaluation identification models are iterated. If the output of a model covers that moment, its confidence level is multiplied by its preset weight and then summed; if it does not cover that moment, its contribution is zero. For example, assuming that the short-scale convolutional model, the long-scale temporal fusion model, and the multi-scale adaptive convolutional model all cover a certain 15-minute time segment, and their confidence levels are 0.9, 0.85, and 0.8, respectively, with corresponding model weights of 0.4, 0.3, and 0.3, then the weighted temporal support for that time segment is 0.9 × 0.4 + 0.85 × 0.3 + 0.8 × 0.3 = 0.855. This indicator not only reflects how many models support the time period as a reduction range, but also reflects the degree of confidence of each model in this judgment, thus effectively distinguishing between occasional low-confidence fluctuations and real, continuous reduction events.

[0214] E3: For each suspected period of credit limit reduction, if any of the following conditions are met, it will be marked as a candidate period of credit limit reduction:

[0215] At least two deduction identification models output suspected deduction time periods with an overlap greater than a preset overlap threshold;

[0216] The suspected devaluation period output by the short-scale model falls entirely within the suspected devaluation period output by the long-scale model; the weighted time support of the suspected devaluation period is greater than the preset support threshold.

[0217] Among them, the short-scale model is a dereasing identification model used to identify abrupt dereasing characteristics, and the long-scale model is a dereasing identification model used to identify the temporal evolution characteristics of continuous gradual dereasing.

[0218] The labeling logic for candidate de-rating time periods employs a refined screening mechanism of multi-dimensional consensus and boundary calibration. It can be triggered by any one of the three judgment conditions to adapt to different de-rating scenarios. Specifically, the first condition requires that the overlap of at least two models exceeds a preset overlap threshold (e.g., 0.5) to ensure that the de-rating event has spatial consistency across models and exclude false positives from a single model. The second condition is specifically for short-to-long physical causal chains. That is, the instantaneous mutation point or short-term plateau period identified by the short-scale model (e.g., TempCNN) must be completely contained within the long-term trend decline interval identified by the long-scale model (e.g., LSTM-FCN). This simulates the physical process of a sudden temperature rise triggering power limitation for several hours in the following hours. Even if the overlap value is not high, it is considered valid as long as the inclusion relationship is met. The third condition utilizes the weighted time support calculated above. When it exceeds a preset support threshold (e.g., 0.6), even if the overlap between pairs does not meet the standard, the high confidence of multiple models is sufficient to prove the existence of de-rating. For example, in a derating event caused by poor heat dissipation, the short-scale model detects a power step, while the long-scale model detects a power reduction due to a continuous rise in temperature. Although the large difference in their time spans results in low conventional overlap, this period is still marked as a candidate derating period because it meets the short-to-long derating condition. This multi-criteria parallel strategy significantly improves the system's robustness to complex derating patterns.

[0219] E4: Based on the time boundaries output by the short-scale model and the long-scale model that output the same candidate reduction time period, the start and end times of the candidate reduction time period are corrected to obtain the target reduction time period.

[0220] Specifically, when a candidate dereasing time period is supported by both short-scale and long-scale models, the suspected trigger point or starting boundary output by the short-scale model is used as the start time of the target dereasing time period, because the short-scale model is more sensitive to abrupt signals. Simultaneously, the end boundary of the persistently restricted interval output by the long-scale model is used as the end time of the target dereasing time period, because the long-scale model is better able to accurately capture the recovery lag phenomenon of the dereasing state. For example, if the short-scale model determines that dereasing begins at 10:05 (power step point), while the long-scale model determines that the dereasing interval is from 10:00 to 11:30 (covering the heating and recovery process), then the corrected target dereasing time period will be determined as 10:05 to 11:30. This differentiated correction of start and end times avoids including normal fluctuations before dereasing in the statistics and prevents premature truncation of the dereasing tail that has not yet fully recovered, thus obtaining an accurate target dereasing time period and providing a reliable time benchmark for subsequent analysis of the causes of dereasing and the generation of treatment recommendations.

[0221] In one possible implementation, determining the reason for the rate reduction and the proposed handling includes: obtaining a combination of rate reduction types consisting of rate reduction types corresponding to the target rate reduction period; determining the rate reduction reason corresponding to the target rate reduction period based on a pre-defined mapping relationship between rate reduction type combinations and rate reduction reasons; and determining the proposed handling opinion corresponding to the rate reduction reason based on a pre-defined mapping relationship between rate reduction reasons and proposed handling opinions.

[0222] The depreciation type combination is composed of the depreciation types output by N depreciation identification models within the same target depreciation time period determined after the fusion analysis of the above embodiments. This depreciation type combination originates from the output results of the parallel determination of multiple models, including the short-scale convolutional model, the long-scale temporal fusion model, the multi-scale adaptive convolutional model, and the branch-level anomaly identification model. For example, when the short-scale convolutional model outputs a power plateau-type depreciation and the multi-scale adaptive convolutional model outputs time-temperature correlation limitation within the target depreciation time period, the generated depreciation type combination is {power plateau type, time-temperature correlation limitation}; similarly, when the branch-level anomaly identification model outputs a single MPPT branch anomaly-type depreciation while the long-scale temporal fusion model does not output a significant long-term trend anomaly, the generated depreciation type combination is {single MPPT branch anomaly type}.

[0223] The pre-defined mapping relationship between derating type combinations and derating causes can be a logical association established based on a rule base or lookup table built from expert knowledge in the photovoltaic operation and maintenance field. The mapping relationship defines several typical combinations and their corresponding derating causes: if the derating type combination includes power plateau type and short-term limited type, the mapping is determined as power limited derating (usually caused by grid power limitation or insufficient DC input); if the derating type combination includes long-term trend type and recovery lag probability, the mapping is determined as slow and inefficient derating (usually caused by module dust accumulation or aging); if the derating type combination includes temperature-triggered limited and long-term trend type, the mapping is determined as heat dissipation limited derating (usually caused by duct blockage or fan failure); if the derating type combination includes single MPPT branch anomaly type and the total power reduction does not reach the overall unit threshold, the mapping is determined as branch anomaly type derating (usually caused by a single module being blocked or junction box failure); if the derating type combination includes multiple MPPT branch consistent limited and grid-connected constraint type, the mapping is determined as grid-connected constraint type derating.

[0224] The pre-defined mapping relationship between derating reasons and corresponding solutions can be pre-stored in the operation and maintenance knowledge base in memory. Specifically, for different derating reasons, the mapping relationship is configured with differentiated solutions: For derating due to heat dissipation limitations, the corresponding solutions are to check whether the inverter air duct is blocked, whether the fan is operating normally, whether the installation spacing meets the heat dissipation requirements, and whether there are any obstructions in the surrounding area; for derating due to branch abnormalities, the corresponding solutions are to locate the abnormal MPPT branch number, check whether the photovoltaic modules corresponding to the branch are obstructed or damaged, check whether the DC connectors are loose or burned, and measure the insulation resistance of the DC cable; for derating due to grid connection constraints, the corresponding solutions are to monitor whether the voltage and frequency at the grid connection point are within the allowable range, confirm the grid dispatching limit instruction, and check the anti-islanding protection settings; for derating due to slow inefficiency, the corresponding solutions are to suggest cleaning the dust on the surface of the photovoltaic modules, assessing the module degradation rate, and comparing the power generation performance of power plants in the same area.

[0225] In one possible implementation, see Figure 6 This figure is a schematic diagram of a controller provided in an embodiment of this application.

[0226] The controller may include a memory 601 and a processor 602. For example... Figure 6 As shown, the memory can be random access memory (RAM), flash memory, read-only memory (ROM), EPROM, non-volatile read-only memory (Electronic Programmable ROM), registers, hard disks, removable disks, etc.

[0227] The memory 601 can store computer instructions. When the computer instructions stored in the memory 602 are executed by the processor 602, the processor 602 can be used to execute the inverter derating identification method. The memory 601 can also store data, such as threshold information involved in the above embodiments.

[0228] This application also provides an inverter that includes a controller, i.e., a built-in local controller, which is configured to execute any of the inverter derating identification methods provided in this application.

[0229] This application also provides an inverter that is communicatively connected to a controller, which is located outside the inverter, and the inverter and the controller establish a wired or wireless data communication link.

[0230] The inverter is configured to send the raw operating dataset to the controller and receive the target derating identification result sent by the controller. The controller is configured to execute any of the inverter derating identification methods provided in the embodiments of this application.

[0231] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape) or a semiconductor medium (e.g., solid-state disk (SSD)).

[0232] This application also provides a readable storage medium for storing the methods provided in the above embodiments. Examples include random access memory (RAM), flash memory, read-only memory (ROM), EPROM, non-volatile read-only memory (EPROM), registers, hard disks, removable disks, or any other form of storage medium in the art.

[0233] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. Regarding the methods disclosed in the embodiments, since they correspond to the product embodiments disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the description of the product embodiments.

[0234] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for identifying inverter derating, characterized in that, include: Obtain the raw operating dataset of the inverter; Extract multidimensional depreciation correlation features from the original running dataset; The multidimensional deduction correlation features are input into N different deduction identification models to obtain preliminary deduction identification results output by N deduction identification models; wherein, N>1, the preliminary deduction identification results include at least deduction type, the deduction types identified by the N deduction identification models are not completely the same, and the time scales of the deduction features identified by the N deduction identification models are not completely the same. The target reduction identification result is obtained by fusing and analyzing the N preliminary reduction identification results.

2. The inverter derating identification method according to claim 1, characterized in that, Multidimensional devaluation correlation features were extracted from the original running dataset, including: The original running dataset is preprocessed to extract multidimensional parameter features, which include: basic running features, parameter trend features, parameter fluctuation features, and parameter correlation features. If at least one of the parameter trend feature, the parameter fluctuation feature, and the parameter correlation feature exceeds the corresponding threshold, the multidimensional parameter feature is determined as a multidimensional depreciation correlation feature.

3. The inverter derating identification method according to claim 1, characterized in that, The N de-rating identification models include at least two of the following: short-scale convolution model, long-scale temporal fusion model, multi-scale adaptive convolution model, and branch-level anomaly identification model; The short-scale convolution model is used to identify transient, abrupt derating features. The long-scale temporal fusion model is used to identify the temporal evolution characteristics of continuously gradual depreciation. The multi-scale adaptive convolution model is used to identify multi-scale derating features; The branch-level anomaly identification model is used to identify MPPT branch-level derating characteristics.

4. The inverter derating identification method according to claim 3, characterized in that, The multidimensional reduction correlation features are input into the short-scale convolutional model to obtain the preliminary reduction recognition result, including: The data input layer of the short-scale convolutional model maps the multidimensional derating correlation features into an enhanced input matrix corresponding to the first short time window. The enhanced input matrix includes: a power parameter group, a voltage parameter group, a current parameter group, and a temperature parameter group. The power parameter group includes inverter power and power change rate. The voltage parameter group includes the voltage of each MPPT branch and voltage change rate. The current parameter group includes the current of each MPPT branch and current change rate. The temperature parameter group includes the inverter internal temperature and temperature change rate. The enhanced input matrix is ​​input into the grouped temporal convolutional layer, and one-dimensional convolution is performed according to the power parameter group, the voltage parameter group, the current parameter group and the temperature parameter group respectively to obtain the grouped convolution result; The grouped convolution results are input into a multi-scale temporal convolutional layer to extract multi-scale restricted features; After inputting the multi-scale restricted features into the residual convolutional layer and the temporal attention layer respectively, they are then input into the local peak pooling layer and the multi-task output layer in sequence to output the preliminary derated identification results. The preliminary derated identification results include the probability of short-term abrupt derated type, the probability of power plateau type derated type, the probability of temperature-triggered restriction, the probability of non-derated short-term perturbation, and the suspected trigger time points corresponding to various derated types.

5. The inverter derating identification method according to claim 3, characterized in that, The multidimensional reduction correlation features are input into the long-scale time-series fusion model to obtain the preliminary reduction identification result, including: The data input layer of the long-scale time series fusion model maps the multi-dimensional derating correlation features into a long-time matrix corresponding to the first long-time window. The long-time matrix includes: inverter power, power ratio relative to rated value, power offset relative to historical baseline of similar operating conditions, power change rate, inverter internal temperature, temperature change rate, AC / DC conversion efficiency, voltage sum of each MPPT branch, and current of each MPPT branch. The long temporal matrix is ​​input into the LSTM branch and the fully convolutional network branch, respectively; The long-term state features output from the LSTM branch and the local morphological features output from the fully convolutional network branch are input into the trend feature fusion layer for feature concatenation to obtain the concatenated features. The spliced ​​features are sequentially input into the attention layer and the multi-task output layer to output the preliminary deduction identification results. The preliminary deduction identification results include the probability of long-term trend deduction, the probability of slow deduction, the probability of recovery lag, the probability of atypical deduction, and the suspected deduction time period corresponding to each type of deduction.

6. The inverter derating identification method according to claim 3, characterized in that, The multidimensional reduction correlation features are input into the multi-scale adaptive convolutional model to obtain the preliminary reduction recognition result, including: The multi-scale adaptive convolutional model's data input layer maps the multi-dimensional devaluation correlation features into input matrices corresponding to the second short-time window, the medium-scale time window, and the second long-time window. Input the input matrix corresponding to the second short time window into the short-scale convolution branch, input the input matrix corresponding to the medium-scale time window into the medium-scale convolution branch, and input the input matrix corresponding to the second long time window into the long-scale convolution branch to obtain the reduction features at each scale. Input the de-rating features at each scale into the scale attention layer to obtain the attention weights at each scale; The multidimensional deduction correlation features are input into the channel attention layer and the time segment attention layer respectively to obtain the attention weights of each channel and the attention weights of each time segment. The reduction features at each scale, the attention weights at each scale, the attention weights at each channel, and the attention weights at each time period are input into the cross-scale feature fusion layer to obtain multi-scale fused features. The multi-scale fusion features are input into the multi-task output layer to output the preliminary devaluation identification results. The preliminary devaluation identification results include the probability of multi-scale continuous devaluation, the probability of short-term limitation, the probability of medium-term temperature-related limitation, the probability of long-term continuous inefficiency, and the suspected devaluation time periods corresponding to various devaluation types.

7. The inverter derating identification method according to claim 3, characterized in that, The multidimensional deduction correlation features are input into the branch-level anomaly identification model to obtain the preliminary deduction identification results, including: The multidimensional derating correlation features are divided into power groups, temperature groups, and each MPPT branch group using the data input layer of the branch-level anomaly identification model. The power group, the temperature group, and each MPPT branch group are respectively input into the corresponding intra-group convolutional branch in the intra-group convolutional branch layer to extract the local variation of each group. The local variation patterns of each group are input into the branch difference feature construction layer. Based on the deviation of the multidimensional parameters of each MPPT branch group from the mean of the parameters of each MPPT branch, branch difference features are generated. The branch difference characteristics are input into the cross-group relationship fusion layer to learn the correspondence between the inverter's total power decrease and branch anomalies, temperature increases, and grid connection constraints, and obtain the fusion result; The fusion result is input into the physical consistency constraint layer for logical constraint verification to obtain the constraint result. The constraint results are input into the multi-task output layer, which outputs the preliminary derating identification results. The preliminary derating identification results include the derating probability of a single MPPT branch abnormality, the derating probability of multiple MPPT branches consistent constraint, the derating probability of temperature linkage, the derating probability of grid-connected side constraint, and the suspected derating time period corresponding to each derating type.

8. The inverter derating identification method according to claim 1, characterized in that, The process of fusing and analyzing the N preliminary rate reduction identification results to obtain the target rate reduction identification result includes: Extract the probability, suspected reduction time period, and confidence level of each reduction type from the N preliminary reduction identification results; The target time period for deduction is determined based on the overlap between the suspected time periods for deduction output by the N deduction identification models. For the same target reduction period, the reduction probabilities output by N preliminary reduction identification results are weighted and summed to obtain the comprehensive reduction probability; Based on the reduction type corresponding to the target reduction time period in the N preliminary reduction identification results, determine the reason for the reduction and the corresponding handling opinion; Output the target credit reduction identification result, which includes: the reason for the credit reduction, the target credit reduction time period, the comprehensive credit reduction probability, the processing opinion, and the preliminary credit reduction identification result used for judgment.

9. The inverter derating identification method according to claim 8, characterized in that, The target time period for tax reduction is determined based on the overlap between the suspected time periods for tax reduction output by the N tax reduction identification models. Calculate the overlap between the suspected deduction time periods output by any two of the deduction identification models; For the same suspected period of credit limit reduction, the weighted sum of the confidence scores of N preliminary credit limit reduction identification results is used to obtain the weighted temporal support of the suspected period of credit limit reduction. For each suspected period of deduction, if any of the following conditions are met, it is marked as a candidate period of deduction: the overlap between the suspected periods of deduction output by at least two deduction identification models is greater than a preset overlap threshold. The suspected deflator period output by the short-scale model falls completely within the suspected deflator period output by the long-scale model; the weighted temporal support of the suspected deflator period is greater than a preset support threshold; wherein, the short-scale model is the deflator identification model used to identify abrupt deflator features, and the long-scale model is the deflator identification model used to identify the temporal evolution features of continuous gradual deflator. Based on the time boundaries output by the short-scale model and the long-scale model that output the same candidate reduction time period, the start and end times of the candidate reduction time period are corrected to obtain the target reduction time period.

10. The inverter derating identification method according to claim 8, characterized in that, The step involves combining the reduction types corresponding to the target reduction time period from the N preliminary reduction identification results to determine the reduction reason and the corresponding processing opinion, including: Obtain the combination of reduction types consisting of the reduction types corresponding to the target reduction time period; Based on the pre-defined mapping relationship between the combination of reduction types and the reasons for reduction, the reasons for reduction corresponding to the target reduction time period are determined; Based on the pre-defined mapping relationship between the reasons for the reduction and the corresponding processing opinions, the processing opinions corresponding to the reasons for the reduction are determined.

11. An inverter, characterized in that, The inverter includes a controller configured to perform the inverter derating identification method according to any one of claims 1-10; or, The inverter is communicatively connected to the controller, the inverter is configured to send the original operating dataset to the controller, and receive the target derating identification result sent by the controller, the controller is configured to execute the inverter derating identification method according to any one of claims 1-10.

Images