Intelligent estimation method for rated capacity of photovoltaic panel under multi-scale meteorological disturbance

Abstract

The invention discloses an intelligent estimation method for the rated capacity of a photovoltaic panel under multi-scale meteorological disturbance, and relates to the technical field of photovoltaic power generation and artificial intelligence. Active immunity to inherent interference of the multi-scale meteorological disturbance is realized by constructing a dual-channel-antagonism deep learning model; therefore, the real rated capacity of the photovoltaic panel is directly separated out in a complex meteorological environment, the attribute channel generates a feature representation capable of accurately returning to the rated capacity, the adversarial discriminator identifies a specific meteorological disturbance type from the feature representation, and a feature extraction encoder of the attribute channel is forced to extract the real rated capacity of the photovoltaic panel in the optimization process through a gradient back propagation mechanism. All information related to meteorological disturbance in the features is gradually filtered out, and the capacity regression layer carries out estimation based on a pure feature space in which external meteorological noise is eliminated, so that the estimation result can stably reflect the performance state of the component itself instead of instantaneous fluctuation of meteorological conditions.

Description

Technical Field

[0001] This invention relates to the fields of photovoltaic power generation and artificial intelligence technology, specifically to an intelligent estimation method for the rated capacity of photovoltaic panels under multi-scale meteorological disturbances. Background Technology

[0002] The rated capacity of photovoltaic panels is a key parameter for measuring their performance, evaluating power generation efficiency, diagnosing faults, and formulating operation and maintenance strategies. However, in actual operating environments, the output power of photovoltaic panels is continuously affected by various spatiotemporal meteorological disturbances, such as instantaneous changes in sunlight, temperature fluctuations, cloud cover, and sandstorms. These disturbances make it extremely difficult to directly and accurately separate the inherent rated capacity of photovoltaic panels from operating data.

[0003] Currently, existing technologies generally suffer from weak anti-interference capabilities. Most data-driven models simply mix meteorological data and operational data as input, making it difficult for the models to distinguish whether power changes originate from disturbances or changes in the component's own capabilities. This results in large fluctuations and low accuracy in the estimation results. Furthermore, traditional methods typically provide a definite estimate but cannot inform users of the reliability of that estimate under the current severe disturbance. Decision-makers cannot distinguish whether a result is obtained under clear and stable weather conditions or under rapidly changing cloudy and rainy weather, which poses risks to subsequent applications.

[0004] Therefore, there is an urgent need for an innovative method that can stably and accurately estimate the rated capacity under complex weather conditions and can also self-assess the reliability of the estimation results. This has become a technical problem that needs to be solved in the current photovoltaic operation and maintenance field. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an intelligent method for estimating the rated capacity of photovoltaic panels under multi-scale meteorological disturbances. This method achieves active immunity to the inherent interference of multi-scale meteorological disturbances by constructing a dual-channel adversarial deep learning model, thereby directly extracting the true rated capacity of photovoltaic panels under complex meteorological conditions. The attribute channel generates a feature representation that can accurately regress the rated capacity, while the adversarial discriminator identifies the specific type of meteorological disturbance from this feature representation. This contradiction, through a gradient backpropagation mechanism, forces the feature extraction encoder of the attribute channel to gradually filter out all meteorological disturbance-related information in the features during the optimization process, causing the attribute channel to fail in the adversarial process. After training, the feature vector output by the attribute channel becomes a highly purified, disturbance-invariant representation strongly correlated only with the inherent physical properties of the photovoltaic panel itself. The capacity regression layer performs estimation based on a pure feature space that excludes external meteorological noise, ensuring that the estimation result stably reflects the performance state of the module itself, rather than the instantaneous fluctuations of meteorological conditions.

[0006] To solve the above-mentioned technical problems, this invention provides the following technical solution: a smart estimation method for the rated capacity of photovoltaic panels under multi-scale meteorological disturbances, the specific steps of which are as follows:

[0007] S100. Collect inherent attribute data of photovoltaic panels and multi-scale meteorological disturbance data. Perform outlier removal, missing value filling and standardization on the collected data to obtain a preprocessed dataset.

[0008] S200. Construct a deep learning model that includes an attribute channel, a disturbance channel, a feature fusion layer, a capacity regression layer, and an adversarial discriminator, wherein the attribute channel is used to extract feature vectors of inherent attribute data of photovoltaic panels, and the disturbance channel is used to extract feature vectors of meteorological disturbance data.

[0009] S300. The deep learning model is jointly trained using the preprocessed dataset, wherein the training objective of the attribute channel is to work in conjunction with the feature fusion layer and the capacity regression layer to minimize the rated capacity estimation error, and at the same time to compete with the adversarial discriminator so that the feature vector output by the adversarial discriminator cannot be accurately identified by the adversarial discriminator as the corresponding meteorological disturbance type.

[0010] S400. For the photovoltaic panel data to be estimated, input it into the deep learning model trained in S300, and perform multiple random forward propagations on the same input data to obtain a set of estimated values ​​for the rated capacity.

[0011] S500. Calculate the mean and standard deviation of the estimated value set. The mean is used as the final estimated value, and the standard deviation is used as an uncertainty measure. A preset uncertainty threshold is set. The estimated value and confidence interval are output together. When the standard deviation exceeds the uncertainty threshold, an alarm is automatically triggered.

[0012] Furthermore, in S100, the inherent attribute data includes photovoltaic panel factory parameters, operating time, historical degradation data, and component temperature distribution. The factory parameters cover rated power, open-circuit voltage, and short-circuit current.

[0013] The multi-scale meteorological disturbance data includes light intensity, ambient temperature, daily wind speed, and weather type, and the weather type is converted into numerical features through unique thermal coding.

[0014] Furthermore, in S200, both the attribute channel and the perturbation channel are composed of multi-layer fully connected neural networks, wherein:

[0015] The input of the attribute channel is the intrinsic attribute data. The number of neurons in its input layer is the same as the dimension of the intrinsic attribute data. The intrinsic attribute data is propagated forward through at least one fully connected layer and a non-linear activation function in sequence, and the first feature vector is output in the hidden layer.

[0016] The input to the disturbance channel is multi-scale meteorological disturbance data. The number of neurons in the input layer is consistent with the dimension of the multi-scale meteorological disturbance data. The multi-scale meteorological disturbance data is propagated forward through at least one fully connected layer and a nonlinear activation function in sequence, and a second feature vector is output in the hidden layer.

[0017] The feature fusion layer receives a first feature vector from the attribute channel output and a second feature vector from the perturbation channel output, and fuses them through a vector concatenation operation to generate a comprehensive feature vector.

[0018] The capacity regression layer consists of at least one fully connected layer that performs nonlinear transformation. Its input is the comprehensive feature vector output by the feature fusion layer, and it outputs a single estimate of the rated capacity of the photovoltaic panel through a linearly activated output layer.

[0019] Furthermore, the adversarial discriminator in S200 is a multi-class neural network, specifically configured as follows:

[0020] The input to the adversarial discriminator is taken from the first feature vector output by the attribute channel;

[0021] The first feature vector is sequentially passed through multiple fully connected layers of the adversarial discriminator and a nonlinear activation function;

[0022] Output a probability distribution vector in the Softmax output layer. ,in This indicates that the model determines the input data to belong to the first... The probability of each type of disturbance.

[0023] Furthermore, in step S300, during the joint training of the deep learning model, a comprehensive total loss function is used. As an optimization objective, the loss function is derived from capacity regression loss. and adversarial losses Linear weighted construction, i.e. ,in, For capacity regression loss and , The number of samples in the preprocessed dataset, The output of the capacity regression layer, This is the actual rated capacity value. For the purpose of combating losses and , It is a sample In the True labels on the type of disturbance It is the probability distribution vector output by the adversarial discriminator. medium sample In the Predicted probabilities for different types of disturbances and To balance the hyperparameters used to weigh the two loss terms, an alternating optimization strategy is employed during training:

[0024] With the parameters of the attribute channel, perturbation channel, feature fusion layer, and capacity regression layer fixed, the parameters of the adversarial discriminator are updated to minimize ;

[0025] With the parameters of the adversarial discriminator fixed, the parameters of the attribute channel, perturbation channel, feature fusion layer, and capacity regression layer are updated to minimize... ;

[0026] The alternating optimization strategy is used to achieve adversarial training between the attribute channels and the adversarial discriminator.

[0027] Furthermore, the multiple random forward propagations in S400 are implemented using Monte Carlo Dropout technology, specifically as follows:

[0028] A Dropout layer is introduced after each fully connected layer of the attribute channel and the capacity regression layer. This Dropout layer is trained with probability... Randomly set the output neurons of the previous layer to zero;

[0029] Keep all Dropout layers active and perform the following steps on the same set of input data: Each independent forward propagation, equivalent to performing a different subnetwork, yields a set of estimates. .

[0030] Furthermore, in S500, the confidence interval It is based on the set of estimated values. The statistical properties of the estimated values ​​are used to obtain an estimate of the uncertainty of the estimated values, which is obtained by calculating the sample mean of the set of estimated values. and sample standard deviation And combined with confidence level coefficient The above is obtained. , , Indicates the first The estimated rated capacity of the photovoltaic panel obtained from the first forward propagation. For the number of forward propagations, the , .

[0031] Furthermore, the automatic alarm triggering in S500 specifically involves: calculating the standard deviation... Compared with the preset uncertainty threshold When comparing, When that happens, a message containing the current estimated value is generated and sent. and its uncertainty confidence interval The alarm information is sent to the user.

[0032] Compared with existing technologies, this intelligent estimation method for the rated capacity of photovoltaic panels under multi-scale meteorological disturbances has the following advantages:

[0033] This invention constructs a dual-channel adversarial deep learning model to achieve active immunity to the inherent interference of multi-scale meteorological disturbances, thereby directly extracting the true rated capacity of photovoltaic panels under complex meteorological environments. The attribute channel generates a feature representation that can accurately regress the rated capacity, while the adversarial discriminator identifies the specific type of meteorological disturbance from this feature representation. This contradiction is addressed through a gradient backpropagation mechanism, which forces the feature extraction encoder of the attribute channel to gradually filter out all information related to meteorological disturbances in the features during the optimization process, thus causing the attribute channel to fail in the adversarial process. After training, the feature vector output by the attribute channel becomes a highly purified disturbance-invariant representation that is strongly correlated only with the inherent physical properties of the photovoltaic panel itself. The capacity regression layer estimates based on a pure feature space that excludes external meteorological noise, so that the estimation result can stably reflect the performance status of the module itself, rather than the instantaneous fluctuations of meteorological conditions.

[0034] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0035] Figure 1 A flowchart of a method for intelligent estimation of the rated capacity of photovoltaic panels under multi-scale meteorological disturbances;

[0036] Figure 2 This is a flowchart illustrating the steps of an intelligent estimation method for the rated capacity of photovoltaic panels under multi-scale meteorological disturbances. Detailed Implementation

[0037] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0038] Example 1

[0039] This embodiment provides the working principle of an intelligent estimation method for the rated capacity of photovoltaic panels under multi-scale meteorological disturbances. It aims to solve the problems of weak anti-interference capability and unreliable results in estimating the rated capacity of photovoltaic panels in complex meteorological environments. Figure 2 As shown, this method achieves proactive immunity to multi-scale meteorological disturbances through five stages: data preprocessing, construction of a dual-channel adversarial deep learning model, joint training, multiple random forward propagation estimations, and uncertainty analysis and alarms. It can accurately output the estimated value of the rated capacity of photovoltaic panels and quantify the uncertainty of the results through standard deviation and confidence interval, providing reliable data support for photovoltaic operation and maintenance decisions.

[0040] First, the data acquisition and preprocessing stage (S100) is initiated. This stage forms the foundation for subsequent model training and estimation. The core objective is to acquire high-quality, standardized data, eliminating anomalies and missing data in the raw data to provide a reliable data source for model input. During data acquisition, two key types of data are collected simultaneously: inherent photovoltaic (PV) panel attribute data and multi-scale meteorological disturbance data. Inherent PV panel attribute data reflects the PV panel's performance and historical state as static and quasi-static data, specifically including factory parameters, operating time, historical degradation data, and module temperature distribution. Factory parameters, as fundamental performance indicators of PV panels, cover rated power, open-circuit voltage, and short-circuit current, directly determining the theoretical upper limit of the PV panel's power generation capacity. Operating time characterizes the aging degree of the PV panel. Historical degradation data records the performance changes of the PV panel over time during past operation, assisting the model in correcting the impact of long-term use on rated capacity. Module temperature distribution reflects the heating state of different areas of the PV panel. Multi-scale meteorological disturbance data represents dynamic external factors affecting the real-time output power of the PV panel, including irradiance, ambient temperature, daily wind speed, and weather type. Sunlight intensity is the core energy source for photovoltaic (PV) power generation, and its strength directly determines the actual output power of the PV panel. Ambient temperature affects the electron migration efficiency of the PV panel. Daily wind speed mainly affects the temperature of the PV panel through heat dissipation, indirectly affecting the output power. Weather type comprehensively reflects the overall state of various meteorological factors. Since weather type data is non-numerical, it is converted into numerical features through one-heat encoding to ensure that it can be recognized and processed by the model. After data collection, outlier removal, missing value imputation, and standardization are performed to obtain a preprocessed dataset with a uniform format and qualified quality, which is ready for model training.

[0041] Then, the deep learning model construction stage (S200) is entered. This stage requires building a dual-channel adversarial model architecture that includes an attribute channel, a perturbation channel, a feature fusion layer, a capacity regression layer, and an adversarial discriminator to jointly achieve the function of anti-interference rated capacity estimation. Among them, the attribute channel is used to extract feature vectors of the inherent attribute data of the photovoltaic panel. Its structure consists of a multi-layer fully connected neural network. The input is the preprocessed inherent attribute data. The number of neurons in the input layer is strictly consistent with the dimension of the inherent attribute data. In this embodiment, if the inherent attribute data includes six dimensions: rated power, open-circuit voltage, short-circuit current, operating time, historical decay rate, and average component temperature, then the number of neurons in the input layer is set to 6. The inherent attribute data is propagated forward through at least one fully connected layer and a nonlinear activation function. The fully connected layer is responsible for mapping the input data to a higher-dimensional feature space, while the nonlinear activation function introduces nonlinear transformation capability, enabling the model to fit complex attribute-capacity relationships. Finally, the first feature vector, which is only related to the attributes of the photovoltaic panel itself, is output in the hidden layer. The perturbation channel extracts feature vectors from multi-scale meteorological perturbation data, providing external environmental information for subsequent feature fusion. Its structure also employs a multi-layer fully connected neural network. The input is preprocessed multi-scale meteorological perturbation data, with the number of neurons in the input layer matching the dimensionality of the perturbation data. Features are extracted from the perturbation data through fully connected layers and nonlinear activation functions, with a second feature vector reflecting the state of the meteorological perturbation output in the hidden layer. The feature fusion layer effectively fuses the first feature vector output from the attribute channel with the second feature vector output from the perturbation channel, generating a comprehensive feature vector containing both the photovoltaic panel's inherent properties and external meteorological environment information. The fusion method uses vector concatenation, directly connecting the first and second feature vectors in dimensional order. This vector retains both the inherent properties of the photovoltaic panel and the meteorological perturbation features, providing comprehensive input information for the capacity regression layer. The capacity regression layer aims to output a single estimate of the rated capacity of photovoltaic panels based on a comprehensive feature vector. Its structure consists of at least one fully connected layer. It fits the mapping relationship between comprehensive features and rated capacity through nonlinear transformation. The input is the comprehensive feature vector output by the feature fusion layer, and the output layer uses a linear activation function to ensure that the output value covers a reasonable range of rated capacity values, ultimately yielding the single-time rated capacity estimate. The adversarial discriminator identifies the corresponding meteorological disturbance type from the first feature vector output by the attribute channel, thereby forcing the attribute channel to filter out meteorological disturbance information from the features. Its structure is a multi-class neural network. The input is taken from the first feature vector output by the attribute channel. The data undergoes feature transformation through multiple fully connected layers and nonlinear activation functions, finally outputting a probability distribution vector at the Softmax output layer. ,in The number of categories for meteorological disturbance types. This indicates that the first feature vector of the model's input corresponds to the first... The probability of a weather disturbance type is considered. The higher the probability value, the stronger the discriminator's confidence in identifying that disturbance type.

[0042] Next, the deep learning model joint training phase (S300) begins. This phase, through the construction of a comprehensive loss function and alternating optimization strategies, achieves collaborative training of the attribute channel, perturbation channel, feature fusion layer, capacity regression layer, and adversarial discriminator. This ensures that the model can accurately estimate the rated capacity while also achieving immunity of the attribute channel to meteorological perturbation information. The core of model training is defining the comprehensive total loss function. This function is derived from the capacity regression loss. and adversarial losses Linear weighted construction, the expression is: ,in and These are hyperparameters used to balance the weights of the two losses: volumetric regression loss. The error between the estimated value output by the capacity regression layer and the actual rated capacity value is measured using the mean squared error. ,in The number of samples in the preprocessed dataset, This is the single-time rated capacity estimate output by the capacity regression layer. This represents the actual rated capacity value corresponding to the sample. The smaller the value, the higher the accuracy of the model's nominal capacity estimation and the better the resistance loss. To measure the accuracy of the adversarial discriminator in identifying meteorological disturbance types, the cross-entropy loss is used for calculation. ,in It is a sample In the The true labels for each type of perturbation are encoded using one-hot encoding, such as the sample. For the second type of disturbance, then ,the remaining , The sample output by the adversarial discriminator Corresponding to the Predicted probability of each type of disturbance The smaller the value, the stronger the discriminator's ability to identify the type of disturbance; conversely, the larger the value, the stronger the discriminator's ability to identify the type of disturbance. The larger the value, the lower the discriminator's recognition accuracy. The training process employs an alternating optimization strategy, consisting of two cyclical steps: First, the parameters of the attribute channel, perturbation channel, feature fusion layer, and capacity regression layer are fixed, while only the parameters of the adversarial discriminator are updated, with the optimization objective being to minimize... The first step aims to improve the discriminator's ability to identify meteorological disturbance types from the first feature vector. The second step involves fixing the parameters of the adversarial discriminator and updating the parameters of the attribute channel, disturbance channel, feature fusion layer, and capacity regression layer, with the optimization objective being to minimize... At this point, the attribute channel needs to reduce the capacity regression error while minimizing the discriminator's accuracy in identifying disturbance types. This adversarial approach forces the first feature vector output by the attribute channel to gradually filter out meteorological disturbance-related information, becoming a pure feature that is only related to the inherent properties of the photovoltaic panel. The above two steps are repeated until the model converges and the model training is completed.

[0043] Secondly, the process moves to the stage of calculating the rated capacity of the photovoltaic panels to be estimated (S400). This stage involves multiple random forward propagations of the same data to be estimated, obtaining a set of estimated rated capacity values. This provides a data foundation for subsequent uncertainty analysis. The core technology is Monte Carlo Dropout. The principle of Monte Carlo Dropout is that during the model training phase, a Dropout layer is introduced after each fully connected layer in the attribute channel and capacity regression layers. During training, the Dropout layer calculates the estimated capacity using probability. By randomly setting the output neurons of the previous layer to zero and randomly shutting down some neurons, the training process of multiple different sub-networks is simulated to enhance the model's generalization ability and uncertainty quantification capability. When processing the photovoltaic panel data to be estimated, the data to be estimated, including the inherent attribute data of the photovoltaic panel and the current multi-scale meteorological disturbance data, is preprocessed. The preprocessed data to be estimated is then input into the deep learning model trained on the S300, while keeping all Dropout layers in an active state. The same set of input data is then processed. Each independent forward propagation, because the neurons in the Dropout layer that are randomly set to zero are different each time, is equivalent to performing estimations on different subnetworks, ultimately resulting in a set of nominal capacity estimates. ,in Indicates the first The estimated value obtained from the first forward propagation.

[0044] Finally, the estimation result analysis and alarm stage (S500) begins. This stage calculates statistical indicators from the estimated value set to determine the final rated capacity estimate, uncertainty measure, and confidence interval. Based on the uncertainty threshold, it determines whether to trigger an alarm, providing users with estimation results that combine accuracy and reliability. The sample mean of the estimated value set is also considered. with sample standard deviation Its sample mean As the final estimated rated capacity, the calculation formula is as follows: ,in Forward propagation count, For the first Second estimate, mean It can comprehensively reflect the average level of multiple estimates, reduce the error of a single random estimate, and the sample standard deviation. As a measure of uncertainty, the calculation formula is as follows: , The larger the value, the higher the dispersion of the multiple estimation results, and the stronger the uncertainty of the estimation result; conversely, the smaller the value, the higher the dispersion of the multiple estimation results. The smaller the value, the more stable the estimation result and the higher its reliability. Next, the confidence interval is calculated. Based on the mean Standard deviation With confidence level coefficient The range of the obtained estimated values ​​is calculated using the following formula: Preset an uncertainty threshold The calculated standard deviation and When comparing, When this occurs, it indicates that the uncertainty of the current estimation result exceeds the acceptable range, and an alarm message will be automatically generated and sent to the user. The alarm message must include the current final estimated value. and its confidence interval This helps users quickly understand the risk level of the estimation results. When this indicates that the uncertainty of the estimation result is within an acceptable range, the final estimated value can be directly expressed. confidence interval Output to the user.

[0045] In summary, this embodiment achieves intelligent estimation of the rated capacity of photovoltaic panels under multi-scale meteorological disturbances through the complete process from S100 to S500. The data preprocessing stage ensures the quality of the input data. The dual-channel-adversarial model architecture actively filters out meteorological disturbance information through the adversarial interaction between the attribute channel and the adversarial discriminator, thereby improving the estimation accuracy. Monte Carlo Dropout technology and statistical analysis quantify the uncertainty of the estimation results, solving the problem that traditional methods cannot assess the reliability of the results. The entire method balances estimation accuracy and result reliability, providing effective technical support for photovoltaic operation and maintenance decisions under complex meteorological environments.

[0046] Example 2

[0047] Based on Example 1, this example provides a method for intelligently estimating the rated capacity of photovoltaic panels under multi-scale meteorological disturbances. The specific steps for this method are as follows: Figure 1 As shown, the specific steps are as follows:

[0048] (1) Data preparation and preprocessing

[0049] Data Acquisition: Collect data on two types of photovoltaic panels.

[0050] Inherent attribute data: such as factory rated parameters, cumulative running time, historical performance degradation records, component surface temperature, etc.

[0051] Multi-scale meteorological disturbance data: such as light intensity and ambient temperature at the second / minute level, and average wind speed and weather type at the daily level.

[0052] Data cleaning: Perform quality checks on the collected raw data, remove obviously abnormal values, and fill in any missing data appropriately.

[0053] Feature engineering: transforms non-numerical data into numerical features and standardizes all numerical data to facilitate model learning.

[0054] (2) Estimation model construction

[0055] Build a dual-channel model framework: Construct a deep learning model that includes two independent input channels.

[0056] Attribute channel: A neural network specifically responsible for receiving and processing inherent attribute data, from which the first feature vector representing the state of the photovoltaic panel itself is extracted.

[0057] Disturbance Channel: Another neural network, specifically responsible for receiving and processing meteorological disturbance data, from which a second feature vector describing the impact of the external environment is extracted.

[0058] Design Feature Fusion and Output Layer: The feature vectors output from the two channels above are concatenated and fused into a comprehensive feature vector. This comprehensive feature vector is then input into a regression layer, which ultimately outputs a single estimate of the rated capacity.

[0059] Introducing an adversarial discriminator: An additional neural network is connected to the end of the attribute channel to determine the type of current weather disturbance based on the features extracted from the attribute channel.

[0060] (3) Joint training of models

[0061] Initialize model parameters: Randomly initialize all parameters of the dual-channel fusion-regression network and the adversarial discriminator.

[0062] Perform alternating training until the model converges overall.

[0063] (4) Probabilistic estimation and uncertainty quantification

[0064] Enable stochastic inference: For a photovoltaic panel to be estimated, its data is input into a trained deep learning model. During model inference, all Dropout layers are activated, and forward propagation calculations are repeated multiple times on the same set of input data to obtain a set of estimated values.

[0065] Final estimate: The arithmetic mean of the estimated values ​​is taken as the final reported value of the rated capacity of the photovoltaic panel.

[0066] Uncertainty measure: Calculate the standard deviation of this set of estimates. The smaller the standard deviation, the more consistent the results of multiple estimates, and the higher the credibility of this estimate.

[0067] (5) Results output and decision support

[0068] Output and warning: The final estimated value and its uncertainty confidence interval are output as results, with a preset uncertainty threshold.

[0069] Intelligent decision linkage: The calculated uncertainty measure is compared with the preset threshold in real time. If the standard deviation exceeds the uncertainty threshold, an alarm is automatically triggered, and an alarm message containing the current estimated value and its uncertainty confidence interval is generated and sent to the user.

[0070] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for intelligent estimation of the rated capacity of a photovoltaic panel under multi-scale meteorological disturbances, characterized in that, The specific steps of the method are: S100, collecting inherent attribute data of photovoltaic panels and multi-scale meteorological disturbance data, performing outlier rejection, missing value filling and standardization processing on the collected data to obtain a preprocessed data set; S200, constructing a deep learning model comprising an attribute channel, a disturbance channel, a feature fusion layer, a capacity regression layer, and an adversarial discriminator, wherein the attribute channel is used to extract a feature vector of the inherent attribute data of the photovoltaic panel, and the disturbance channel is used to extract a feature vector of the meteorological disturbance data; S300, jointly training the deep learning model using the preprocessed data set, wherein the training target of the attribute channel is to work with the feature fusion layer and the capacity regression layer to minimize the rated capacity estimation error, and at the same time, to be in opposition with the adversarial discriminator, so that the output feature vector cannot be accurately identified by the adversarial discriminator as corresponding to the meteorological disturbance type; S400, for the photovoltaic panel data to be estimated, input into the deep learning model trained in S300, and perform multiple random forward propagations on the same input data to obtain a set of estimated values of the rated capacity; S500, calculating the mean and standard deviation of the estimated value set, the mean as the final estimated value, and the standard deviation as the uncertainty measure, a preset uncertainty threshold, outputting the estimated value and the confidence interval together, and automatically triggering an alarm when the standard deviation exceeds the uncertainty threshold.

2. The method of claim 1, wherein, In step S100, the inherent attribute data includes photovoltaic panel factory parameters, running time, historical attenuation data and component temperature distribution, and the factory parameters include rated power, open circuit voltage and short circuit current. The multi-scale meteorological disturbance data includes light intensity, environmental temperature, daily wind speed and weather type, and the weather type is converted into a numerical feature through one-hot encoding.

3. The method of claim 1, wherein, In step S200, the attribute channel and the disturbance channel are both composed of multiple layers of fully connected neural networks, wherein: The attribute channel input is the inherent attribute data, the number of input layer neurons is consistent with the dimension of the inherent attribute data, and the inherent attribute data is sequentially forwarded through at least one fully connected layer and a nonlinear activation function to output a first feature vector at the hidden layer; The disturbance channel input is multi-scale meteorological disturbance data, the number of input layer neurons is consistent with the dimension of the multi-scale meteorological disturbance data, and the multi-scale meteorological disturbance data is sequentially forwarded through at least one fully connected layer and a nonlinear activation function to output a second feature vector at the hidden layer; The feature fusion layer receives the first feature vector output from the attribute channel and the second feature vector output from the disturbance channel, and fuses them through a vector splicing operation to generate a comprehensive feature vector; The capacity regression layer is composed of at least one fully connected layer for nonlinear transformation, the input of which is the comprehensive feature vector output by the feature fusion layer, and the output layer with linear activation outputs a single estimated value of the rated capacity of the photovoltaic panel.

4. The method of claim 1, wherein, The adversarial discriminator in step S200 is a multi-classification neural network, which is specifically composed of: The input of the adversarial discriminator is taken from the first feature vector output by the attribute channel; sequentially passing the first feature vector through multiple fully connected layers of the adversarial discriminator and a nonlinear activation function; The Softmax output layer outputs a probability distribution vector where denotes the probability that the model judges the input data to belong to the type of perturbation.

5. The method of claim 1, wherein, In step S300, during the joint training of the deep learning models, a comprehensive total loss function is used. As an optimization objective, the loss function is derived from capacity regression loss. and adversarial losses Linear weighted construction, i.e. ,in, For capacity regression loss and , The number of samples in the preprocessed dataset, The output of the capacity regression layer, This is the actual rated capacity value. For the purpose of combating losses and , It is a sample In the True labels on the type of disturbance It is the probability distribution vector output by the adversarial discriminator. medium sample In the Predicted probabilities for different types of disturbances and To balance the hyperparameters used to weigh the two loss terms, an alternating optimization strategy is employed during training: fixing parameters of the attribute channel, the perturbation channel, the feature fusion layer, and the capacity regression layer, updating parameters of the adversarial discriminator to minimize ; fixing parameters of the adversarial discriminator, updating parameters of the attribute channel, the perturbation channel, the feature fusion layer, and the capacity regression layer to minimize ; implementing adversarial training between the attribute channel and the adversarial discriminator through the alternating optimization strategy.

6. The method of claim 1, wherein, The multiple random forward propagations in step S400 are implemented through a Monte Carlo Dropout technology, and the specific process is as follows: Dropout layers are introduced after each fully connected layer of the attribute channel and the capacity regression layer, which randomly zeroes out neurons of the previous layer output with probability during training; and during training. Keep all Dropout layers active, perform a separate forward pass for the same set of input data, each pass equivalent to being on a different subnetwork, resulting in a set of estimates .

7. The method of claim 6, wherein the method further comprises: In step S500, the confidence interval It is based on the set of estimated values. The statistical properties of the estimated values ​​are used to obtain an estimate of the uncertainty of the estimated values, which is obtained by calculating the sample mean of the set of estimated values. and sample standard deviation And combined with confidence level coefficient The above is obtained. , , Indicates the first The estimated rated capacity of the photovoltaic panel obtained from the first forward propagation. For the number of forward propagations, the , .

8. The method of claim 7, wherein, The automatic triggering of the alarm in step S500 is specifically implemented by comparing the calculated standard deviation with a preset uncertainty threshold , and when , generating and sending an alarm message containing the current estimated value and its uncertainty confidence interval to the user.

Images