A photovoltaic array anomaly feature perception method, device and medium

Abstract

The application provides a photovoltaic array abnormal feature perception method, device and medium, and belongs to the technical field of power systems, and the method comprises the following steps: acquiring photovoltaic array operation data; performing distribution correction on the photovoltaic array operation data to obtain photovoltaic array corrected data; inputting the photovoltaic array corrected data into a pre-constructed abnormal feature perception model to output a photovoltaic array abnormal feature perception result; wherein the construction of the abnormal feature perception model comprises the following steps: embedding a physical quantity constraint for correcting weakly associated physical quantities in the hidden space of an original self-encoder to obtain a PCAE structure; sequentially adding a CNN structure and an LSTM structure after the PCAE structure to obtain a constructed abnormal feature perception model. The application can ensure that the coding features conform to the working principle of the photovoltaic array, improve the comprehensiveness of abnormal identification and the accuracy of data state estimation, and solve the problem of parameter distribution deviation.

Description

Technical Field

[0001] This invention relates to the field of power system technology, and in particular to a method, device and medium for sensing abnormal features of photovoltaic arrays. Background Technology

[0002] Photovoltaic arrays, with their advantages of being clean, pollution-free, and resource-renewable, have become a core component of the new energy power generation field. Their safe and stable operation is directly related to the power generation efficiency and return on investment of photovoltaic power plants. With the large-scale development of the photovoltaic industry, photovoltaic arrays face various operational risks in complex outdoor environments. Common abnormal phenomena such as hot spots on modules, loose or burned junction boxes, inverter malfunctions, power attenuation due to partial shading, and module aging failure can not only lead to a significant decrease in the output power of the photovoltaic array and a reduction in power generation efficiency, but may also cause safety accidents such as module overheating damage and fires, resulting in huge economic losses.

[0003] Currently, photovoltaic (PV) array anomaly detection technology faces numerous challenges, failing to meet the demands of online monitoring in complex environments. On one hand, PV array operating data is susceptible to factors such as sudden changes in irradiance, ambient temperature fluctuations, sensor noise, and transmission interference, resulting in significant redundancy and outliers. Traditional data-driven identification methods lack constraints on the physical working principles of PV arrays, often learning noise characteristics and even outputting results that do not conform to reality, such as output power exceeding theoretical limits or voltage deviating from normal operating ranges. This leads to a persistently high false positive rate. On the other hand, different types of PV array anomaly patterns exhibit different characteristics during evolution. Some anomalies manifest as localized sudden features (such as junction box short circuits), while others exhibit long-term time-dependent features (such as progressive module aging). When a single model is used alone, the anomaly identification accuracy is generally low, failing to comprehensively cover all types of anomalies. Furthermore, existing identification methods often neglect the impact of data state estimation accuracy on anomaly judgment. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method, device and medium for sensing abnormal features of photovoltaic arrays. By constructing an abnormal feature sensing model, it is possible to ensure that the encoded features conform to the working principle of photovoltaic arrays, improve the comprehensiveness of anomaly identification and the accuracy of data state estimation, and solve the problem of parameter distribution offset.

[0005] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0006] On one hand, the present invention provides a method for sensing abnormal features of a photovoltaic array, comprising:

[0007] Acquire photovoltaic array operation data;

[0008] The photovoltaic array operation data is distributed and corrected to obtain photovoltaic array correction data;

[0009] The photovoltaic array correction data is input into a pre-built anomaly feature perception model, and the photovoltaic array anomaly feature perception result is output.

[0010] The construction of the anomaly feature perception model includes:

[0011] The PCAE structure is obtained by embedding physical quantity constraints for correcting weakly correlated physical quantities in the latent space of the original autoencoder.

[0012] After the PCAE structure, CNN and LSTM structures are added sequentially to obtain the constructed anomaly feature perception model. The PCAE structure is represented as follows: ; ; ; ; in, This represents the output of the encoder linear transform layer at sampling time k; , , These represent the weight matrices of the encoder linear transformation layer, the encoder activation layer, and the decoder, respectively. This represents the encoder's input vector at sampling time k; , , These represent the bias terms of the encoder linear transformation layer, encoder activation layer, and decoder, respectively. This represents the output of the encoder's active layer at sampling time k; , These represent the activation functions of the encoder and decoder, respectively. , These represent the sequence prediction state and the secondary sequence prediction state of the photovoltaic array, respectively. , These represent the sequential prediction output voltage value and the quadratic prediction output voltage value of the photovoltaic array, respectively. This represents the decoder output at sampling time k; This represents the latent space sampling point at sampling time k.

[0013] Optionally, an inverse Gaussian distribution is used to correct the distribution of the photovoltaic array operating data, resulting in corrected photovoltaic array data. The probability density formula for the inverse Gaussian distribution is:

[0014] ;

[0015] ;

[0016] ;

[0017] ;

[0018] in, This represents the output voltage of the i-th photovoltaic module in the photovoltaic array. The inverse Gaussian probability density function; This represents the output voltage of the i-th photovoltaic module in the photovoltaic array. The average value; The shape parameter represents the inverse Gaussian distribution; This represents the operation of exponential functions; This represents the output voltage of the i-th photovoltaic module in the photovoltaic array. The variance; This represents the output voltage correction value of the i-th photovoltaic module in the photovoltaic array at sampling time k; Indicates the sampling time Output voltage of the i-th photovoltaic module in the lower photovoltaic array; Let represent the inverse Gaussian probability density function of the i-th photovoltaic module in the photovoltaic array at sampling time k; This represents the difference between the output voltage of the i-th photovoltaic module in the photovoltaic array and the output voltage correction value at sampling time k.

[0019] Optionally, the processing steps of the anomaly feature perception model include:

[0020] In the PCAE structure, the photovoltaic array correction data is subjected to data dimensionality reduction and feature extraction to obtain a low-dimensional feature vector, and abnormal pattern features are screened out from the low-dimensional feature vector.

[0021] In the CNN structure, the abnormal detail features within a short time window and the abnormal association features within a long time window are extracted from the abnormal pattern features respectively. The abnormal detail features and the abnormal association features are then concatenated to obtain the local abnormal features.

[0022] In the LSTM structure, long-term temporal dependency information of the local anomaly features is captured to obtain the photovoltaic array anomaly feature perception results.

[0023] Optionally, the LSTM structure is represented as:

[0024] ;

[0025] ;

[0026] ;

[0027] in, , , , Let represent the output functions of the forget gate, input gate, candidate memory cell function, and output gate at time t, respectively. express The weight matrix; express The weight matrix; express The weight matrix; express The weight matrix; This represents the input at time t; , Let represent the hidden states at time t-1 and time t, respectively; , , , They represent bias vector, bias vector, bias vector, The bias vector; , These represent memory cell units at time t-1 and time t, respectively. This represents the hyperbolic tangent activation function.

[0028] Optionally, the training of the anomaly feature perception model includes:

[0029] The hyperparameters of the anomaly perception model are trained by hierarchical iterative collaborative mean square error loss and root mean square error loss, resulting in a well-trained anomaly perception model.

[0030] Optional, also includes:

[0031] The anomaly feature perception model is used as a state space model, and the STKF algorithm is used to track the state changes of the photovoltaic array anomaly feature perception results in real time.

[0032] When the state of the photovoltaic array abnormal feature sensing result changes abruptly, increase the Kalman gain in the STKF algorithm;

[0033] When the state of the photovoltaic array abnormal feature sensing result is stable, reduce the Kalman gain in the STKF algorithm.

[0034] Optionally, the state-space model is represented as:

[0035] ;

[0036] ;

[0037] in, This represents the state vector of the photovoltaic array at time t. This represents the state transition matrix at time t-1. This represents the control input matrix at time t-1. This represents the process noise at time t; Let represent the observation vector at time t; Let represent the observation matrix at time t; This represents the measurement noise at time t.

[0038] In a second aspect, the present invention provides a photovoltaic array anomaly detection device, comprising:

[0039] The data acquisition module is used to acquire photovoltaic array operating data.

[0040] The data correction module is used to: perform distribution correction on the photovoltaic array operating data to obtain photovoltaic array correction data;

[0041] An anomaly perception module is used to: input the photovoltaic array correction data into a pre-constructed anomaly feature perception model and output the photovoltaic array anomaly feature perception result;

[0042] The model building module is used to build an anomaly feature perception model. The building process includes:

[0043] The PCAE structure is obtained by embedding physical quantity constraints for correcting weakly correlated physical quantities in the latent space of the original autoencoder.

[0044] After the PCAE structure, CNN and LSTM structures are added sequentially to obtain the constructed anomaly feature perception model. The PCAE structure is represented as follows: ; ; ; ; in, This represents the output of the encoder linear transform layer at sampling time k; , , These represent the weight matrices of the encoder linear transformation layer, the encoder activation layer, and the decoder, respectively. This represents the encoder's input vector at sampling time k; , , These represent the bias terms of the encoder linear transformation layer, encoder activation layer, and decoder, respectively. This represents the output of the encoder's active layer at sampling time k; , These represent the activation functions of the encoder and decoder, respectively. , These represent the sequence prediction state and the secondary sequence prediction state of the photovoltaic array, respectively. , These represent the sequential prediction output voltage value and the quadratic prediction output voltage value of the photovoltaic array, respectively. This represents the decoder output at sampling time k; This represents the latent space sampling point at sampling time k.

[0045] Thirdly, the present invention provides a computer-readable storage medium having a computer program / instruction stored thereon, which, when executed by a processor, implements the steps of the photovoltaic array anomaly feature sensing method described in the first aspect.

[0046] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0047] This invention adaptively adjusts the accuracy of data state estimation and corrects the distribution of photovoltaic array operating data, solving the problem of parameter distribution offset caused by factors such as differences in component production batches, uneven array arrangement, and environmental interference. During the design of the PCAE structure, the overall model state of the PCAE structure is compensated and corrected based on the overall predicted state. Furthermore, by superimposing the overall model state of the PCAE structure with the deviation model state of the PCAE structure, the coupling relationship of mapping the state of individual photovoltaic modules can be realized. The PCAE structure ensures that the encoded features conform to the working principle of the photovoltaic array, providing high-quality input for subsequent feature extraction. The CNN structure and LSTM structure are used in synergy to capture local abnormal features, improving the comprehensiveness of anomaly recognition. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the overall process of the photovoltaic array anomaly feature sensing method provided in the embodiments of the present invention;

[0049] Figure 2 This is a partial flowchart of the photovoltaic array anomaly feature sensing method provided in an embodiment of the present invention;

[0050] Figure 3 This is another schematic flowchart of the photovoltaic array anomaly feature sensing method provided in the embodiments of the present invention;

[0051] Figure 4 A schematic diagram illustrating the construction framework of the anomaly feature perception model provided in an embodiment of the present invention;

[0052] Figure 5 This is a schematic diagram of the CNN and LSTM structures provided in the embodiments of the present invention. Detailed Implementation

[0053] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0054] The term "and / or" simply describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0055] Example 1

[0056] like Figure 1 As shown in the figure, this embodiment introduces a method for sensing abnormal features of a photovoltaic array, including:

[0057] Acquire photovoltaic array operation data;

[0058] The photovoltaic array operation data is distributed and corrected to obtain photovoltaic array correction data;

[0059] The photovoltaic array correction data is input into a pre-built anomaly feature perception model, and the photovoltaic array anomaly feature perception result is output.

[0060] The construction of the anomaly feature perception model includes:

[0061] The PCAE structure is obtained by embedding physical quantity constraints for correcting weakly correlated physical quantities in the latent space of the original autoencoder.

[0062] After the PCAE structure, CNN and LSTM structures are added sequentially to obtain the constructed anomaly feature perception model. The PCAE structure is represented as follows: ; ; ; ; in, This represents the output of the encoder linear transform layer at sampling time k; , , These represent the weight matrices of the encoder linear transformation layer, encoder activation layer, and decoder, respectively. This represents the encoder's input vector at sampling time k; , , These represent the bias terms of the encoder linear transformation layer, encoder activation layer, and decoder, respectively. This represents the output of the encoder's active layer at sampling time k; , These represent the activation functions of the encoder and decoder, respectively. , These represent the sequence prediction state and the secondary sequence prediction state of the photovoltaic array, respectively. , These represent the sequential prediction output voltage value and the quadratic prediction output voltage value of the photovoltaic array, respectively. This represents the decoder output at sampling time k; This represents the latent space sampling point at sampling time k.

[0063] This embodiment can solve the problem of parameter distribution offset caused by factors such as component production batch differences, uneven array arrangement, and environmental interference. When designing the PCAE structure, the overall model state of the PCAE structure is compensated and corrected based on the overall prediction state. By superimposing the overall model state of the PCAE structure and the deviation model state of the PCAE structure, the coupling relationship of mapping the state of individual photovoltaic modules can be realized. The PCAE structure ensures that the encoded features conform to the working principle of the photovoltaic array, providing high-quality input for subsequent feature extraction. The CNN structure and LSTM structure are used to capture local abnormal features in a coordinated manner, improving the comprehensiveness of anomaly recognition.

[0064] Example 2

[0065] like Figure 2 As shown in the figure, this embodiment introduces a method for sensing abnormal features of a photovoltaic array, including the following steps:

[0066] Step 1: Obtain photovoltaic array operating data, specifically:

[0067] like Figure 3As shown, firstly, photovoltaic array operating data is acquired. Abnormal photovoltaic array data is affected by various factors such as the number of sensors, data quality, and random operating conditions. Traditional identification methods struggle to fully capture these multi-source anomalies, resulting in low prediction accuracy. In this embodiment, abnormal data is first removed from the photovoltaic array operating data based on the physical working boundary of the photovoltaic array. Then, a mean-filling method is used to address missing and duplicate values ​​that may occur during the transmission and storage of the photovoltaic array operating data. This eliminates noise interference and filters effective features, providing high-quality data support for subsequent model input.

[0068] Z-score standardization is a method for normalizing filtered feature data. It can eliminate the differences in the units of measurement between different features, avoid the model from being overly biased towards features with a large numerical range, and ensure that each feature plays an equal role in model training.

[0069] Step 2: Correct the data using the inverse Gaussian distribution, specifically as follows:

[0070] The inverse Gaussian distribution is a commonly used distribution in statistics, mainly used to solve the problem of parameter distribution deviation caused by factors such as differences in component production batches, array layout characteristics, and uneven environmental shading during photovoltaic array modeling.

[0071] The photovoltaic array operating data after Z-score normalization is corrected using an inverse Gaussian distribution to obtain corrected photovoltaic array data. The probability density formula for the inverse Gaussian distribution is:

[0072] ;

[0073] ;

[0074] ;

[0075] ;

[0076] in, This represents the output voltage of the i-th photovoltaic module in the photovoltaic array. The inverse Gaussian probability density function; This represents the output voltage of the i-th photovoltaic module in the photovoltaic array. The average value; The shape parameter represents the inverse Gaussian distribution; This represents the operation of the exponential function; This represents the output voltage of the i-th photovoltaic module in the photovoltaic array. The variance; This represents the output voltage correction value of the i-th photovoltaic module in the photovoltaic array at sampling time k; Indicates the sampling time Output voltage of the i-th photovoltaic module in the lower photovoltaic array; Let represent the inverse Gaussian probability density function of the i-th photovoltaic module in the photovoltaic array at sampling time k; This represents the difference between the output voltage of the i-th photovoltaic module in the photovoltaic array and the output voltage correction value at sampling time k.

[0077] Step 3: Construct an anomaly feature perception model, specifically:

[0078] like Figure 4 As shown, physical quantity constraints for correcting weakly correlated physical quantities are embedded in the latent space of the original autoencoder to obtain the structure of the Physics-Constrained Autoencoder (PCAE).

[0079] After the PCAE structure, a Convolutional Neural Network (CNN) structure and a Long Short-Term Memory (LSTM) structure are added sequentially to obtain the constructed anomaly feature perception model.

[0080] In the design of the PCAE structure, the overall model state of the PCAE structure is compensated and corrected based on the overall predicted state. By superimposing the overall model state of the PCAE structure and the deviation model state of the PCAE structure, the coupling relationship of mapping the state of individual photovoltaic modules can be realized. It mainly includes two parts: encoder and decoder.

[0081] The encoder integrates the overall model state with the state prediction results, aiming to effectively avoid problems such as low model accuracy and poor robustness of neural networks, thereby improving prediction accuracy. Simultaneously, it compresses historical photovoltaic array operating data (such as module output power and terminal voltage) and environmental state data (such as irradiance and ambient temperature) as physical quantity constraints. These physical quantity constraints are used as correction terms in the encoder's latent space to correct weakly correlated physical quantities. The encoder represents:

[0082] ;

[0083] ;

[0084] ;

[0085] in, This represents the output of the encoder linear transform layer at sampling time k; , These represent the weight matrices of the encoder linear transformation layer and the encoder activation layer, respectively. This represents the encoder's input vector at sampling time k; , These represent the bias terms of the encoder linear transformation layer and the encoder activation layer, respectively; This represents the output of the encoder's active layer at sampling time k; This represents the encoder's activation function; , These represent the sequence prediction state and the secondary sequence prediction state of the photovoltaic array, respectively. , These represent the sequential prediction output voltage value and the quadratic prediction output voltage value of the photovoltaic array, respectively.

[0086] The core task of the decoder is to map the sampling points in the latent space to the data space of the future state of the photovoltaic array. The decoder is represented as:

[0087] ;

[0088] in, This represents the decoder output at sampling time k; The weight matrix represents the decoder; This represents the latent space sampling point at sampling time k; This represents the bias term of the decoder; It is an activation function designed based on the operating conditions of the photovoltaic array to prevent gradient explosion during training and ensure the stability of the model under the complex operating conditions of the photovoltaic array.

[0089] It is the overall predicted state and overall predicted voltage of the photovoltaic array. Includes the current state, via The essence of integration is to use the predicted state to compensate and correct the model state. Reflecting the overall model state, The activation function incorporates constraints. For the superposition result, the output features implicitly contain the deviation between the overall state and the individual state. The deviation model state captures the local differences, and the decoder uses... The output feature is then reconstructed into photovoltaic state data.

[0090] like Figure 5 As shown, to further enhance the ability of CNN structure to capture local anomaly features, a multi-scale convolution kernel design can be adopted. By setting convolution kernels of different sizes (such as 3×1, 5×1, 7×1), local anomaly information at different time spans in the output data of PCAE structure can be extracted respectively.

[0091] like Figure 5As shown, the LSTM structure, as a special type of recurrent neural network, has the core advantage of accurately learning long-term dependencies in sequence data. It works in synergy with the cell state through three gates: the input gate, the forget gate, and the output gate. The forget gate filters and retains key historical features, the input gate controls the effective integration of new information, and the output gate regulates the transmission of the cell state to subsequent layers. Ultimately, relying on this dynamic mechanism, it efficiently processes sequence data, accurately captures the long-term correlation of data features, and outputs reliable fitting results for outlier data.

[0092] The LSTM structure is represented as follows:

[0093] ;

[0094] ;

[0095] ;

[0096] in, , , , Let represent the output functions of the forget gate, input gate, candidate memory cell function, and output gate at time t, respectively. express The weight matrix; express The weight matrix; express The weight matrix; express The weight matrix; This represents the input at time t; , Let represent the hidden states at time t-1 and time t, respectively; , , , They represent bias vector, bias vector, bias vector, The bias vector; , These represent memory cell units at time t-1 and time t, respectively. This represents the hyperbolic tangent activation function.

[0097] Step 4: Parameter optimization, specifically:

[0098] like Figure 4As shown, since the state of the photovoltaic array anomaly feature sensing results is prone to sudden changes, the Strong Tracking Kalman Filter (STKF) algorithm is introduced to make the optimal estimate of the state of the dynamic photovoltaic array anomaly feature sensing results.

[0099] The anomaly feature perception model is represented as a state-space model as follows:

[0100] ;

[0101] ;

[0102] in, This represents the state vector of the photovoltaic array at time t. Let f(t) represent the state transition matrix at time t-1, describing the state change pattern from time t-1 to time t. This represents the control input matrix at time t-1, corresponding to the external influencing factors of the photovoltaic array; The process noise at time t is represented by a Gaussian distribution with a mean of 0, reflecting the uncertainty in the state transition. Let represent the observation vector at time t; Let represent the observation matrix at time t, which describes the mapping relationship from the state vector to the observation vector; The measurement noise at time t is represented by a Gaussian distribution with a mean of 0, reflecting the error of the sensor measurement.

[0103] The STKF algorithm introduces a fading factor and dynamically adjusts the Kalman gain to track the state changes of the photovoltaic array anomaly sensing results. When the state of the photovoltaic array anomaly sensing results changes abruptly, the Kalman gain is increased to quickly track the state change; when the state of the photovoltaic array anomaly sensing results is stable, the Kalman gain is decreased to reduce the impact of noise, ensuring that the photovoltaic array anomaly sensing results can adapt to dynamic changes under random operating conditions.

[0104] The training steps for the anomaly feature perception model include:

[0105] First, it is necessary to identify the hyperparameters that need to be optimized in the anomaly feature perception model, including:

[0106] The PCAE structure includes output parameter distribution correction parameters, physical constraint loss weights, and the number of nodes in the fully connected layer.

[0107] The CNN architecture includes kernel size, number of convolutional layers, and number of filters;

[0108] The LSTM structure includes the number of hidden units and the number of iteration rounds.

[0109] Layered iteration prioritizes ensuring the physical rationality of photovoltaic array operation, avoids parameter interference between modules and reduces optimization space complexity, and further prevents deviations from the photovoltaic array's operating rules caused by parameter optimization. The specific steps are as follows:

[0110] First, an anomaly feature perception model is constructed based on the initial hyperparameter combination;

[0111] Then, the hyperparameters of the PCAE structure are iteratively optimized by first traversing the latent variable dimensions and the number of encoder nodes, and then fixing the selected structural hyperparameters.

[0112] Secondly, the hyperparameters of the CNN structure and the hyperparameters of the LSTM structure are jointly iteratively optimized. First, the number of convolutional layers is fixed at 3, and the number of convolutional kernels and filters is traversed. Then, the number of convolutional kernels and filters is fixed, and the number of convolutional layers is traversed. The performance of 2 convolutional layers and 3 convolutional layers is compared to obtain the pre-trained abnormal feature perception model.

[0113] Then, the pre-trained anomaly perception model is evaluated using a validation set to obtain the mean squared error (MSE) loss, which is used as the fitness value.

[0114] Finally, the iteration continues until a stopping condition is met: the proportion of physically anomalous samples is <0.01 and the change in reconstructed MSE is <0.001, indicating no performance improvement or meeting performance requirements. When the hyperparameter optimization process meets the preset stopping condition, the optimal hyperparameter combination is selected from the global optimization history, corresponding to the globally best historical optimization position. Based on this optimal hyperparameter combination, a trained anomaly feature perception model is obtained, such as... Figure 2 As shown, by verifying the root mean square error (RMSE) of the state prediction loss, the accuracy of anomaly type identification, and the real-time inference time, the final performance evaluation of the model is completed, ensuring that the optimal hyperparameters are adapted to the intelligent perception requirements of photovoltaic array anomaly characteristics under random operating conditions.

[0115] Step 5: Obtain the results of photovoltaic array anomaly detection, specifically:

[0116] The photovoltaic array correction data obtained in step two is input into the pre-constructed anomaly feature perception model, which outputs the photovoltaic array anomaly feature perception result, namely:

[0117] To address the issue that pure time-series models are not sensitive enough to local sudden anomaly features, we can first use the PCAE structure to perform data dimensionality reduction and feature extraction on the photovoltaic array correction data to obtain low-dimensional feature vectors. From these low-dimensional feature vectors, we can then select anomaly pattern features that can reflect local anomalies.

[0118] These local anomaly features are then input into a CNN structure. Leveraging the local perception and weight sharing characteristics of the CNN structure, the spatial correlation of features is further explored. In the CNN structure, small-sized convolutional kernels focus on the sudden anomaly details within a short time window, while large-sized convolutional kernels capture the anomaly correlation features within a slightly longer time window. Subsequently, through feature concatenation, the anomaly details and anomaly correlation features output by multi-scale convolutions are merged into local anomaly features of a unified dimension. This preserves the subtle differences in local anomaly features while avoiding the feature omission problem caused by single-scale convolution. It can more comprehensively reflect the essential features of local sudden anomaly patterns, providing more discriminative feature inputs for subsequent anomaly feature perception models, thereby improving the model's sensitivity to the recognition of local sudden anomaly patterns.

[0119] Finally, the local anomaly features are input into the LSTM structure to capture the long-term temporal dependency information of the local anomaly features. The forget gate dynamically filters redundant historical information by adjusting the weight matrix and bias to avoid the accumulation of invalid features. The input gate works with the candidate memory cells to selectively integrate the key information of the current input into the cell state. The output gate generates a feature representation adapted to subsequent tasks based on the current cell state and the hidden state, thus obtaining the photovoltaic array anomaly feature perception result.

[0120] This mechanism not only enables the LSTM structure to stably capture long-term dependencies in sequence data, but also allows it to flexibly adapt to the temporal volatility of abnormal data from equipment. In scenarios such as photovoltaic array anomaly trend prediction and long-term anomaly pattern recognition, it effectively improves the model's fitting accuracy and generalization ability to anomaly evolution patterns.

[0121] Common anomaly detection results for photovoltaic arrays include three typical scenarios:

[0122] Open circuits between photovoltaic array strings and shading on the surface of photovoltaic array modules occur simultaneously, as do dust accumulation on the surface of photovoltaic array modules.

[0123] The coexistence of surface shading and photovoltaic array aging: photovoltaic array modules are shaded, photovoltaic array modules are damaged, photovoltaic array modules are disconnected, interfaces are aged and corroded, and photovoltaic arrays are abnormally aged.

[0124] Simultaneous short circuit within the photovoltaic array module string and surface shading of the photovoltaic array module: Electrolyte leakage caused by damage to the cell casing or failure of sealing will reduce the cell's ion conduction capacity, and at the same time cause corrosion of external circuits and reduced insulation performance, exhibiting continuous degradation characteristics.

[0125] In the intelligent perception of abnormal features of photovoltaic arrays, it is necessary to calculate indicators such as accuracy, recall, and F1 score for different types of anomalies (such as dust accumulation on modules, abnormal wiring between modules, etc.) to quantify the model's ability to identify and predict various anomaly patterns.

[0126] Accuracy refers to the proportion of abnormal samples (including all types of abnormal and normal samples) correctly identified by the model out of the total identified samples, reflecting the model's overall perception and identification accuracy of normal photovoltaic arrays and various types of abnormal photovoltaic arrays.

[0127] For anomalous features of photovoltaic arrays, recall refers to the proportion of anomalous samples correctly identified by the model out of the total number of such anomalous samples. For photovoltaic arrays, a low recall rate may result in key anomalous patterns not being discovered. Therefore, the recall rate for anomalous features such as dust accumulation on the surface of the module and disconnection of wiring between modules is usually required to be >90%.

[0128] When abnormal samples are imbalanced, F1-Score can avoid the bias of a single indicator and more objectively measure the model performance. It is the most commonly used comprehensive indicator for sensing multiple anomalies in photovoltaic arrays.

[0129] The anomaly perception model in this implementation can efficiently, quickly, and accurately identify and predict anomalies in photovoltaic arrays. It performs particularly well in dealing with power attenuation problems caused by hot spots, junction box failures, inverter anomalies, and shading. At the same time, this method provides strong technical support and a reference solution for improving the operational safety, anomaly early warning capabilities, and power generation economic benefits of photovoltaic arrays.

[0130] Example 3

[0131] This embodiment introduces a photovoltaic array anomaly detection device, including:

[0132] The data acquisition module is used to acquire photovoltaic array operating data.

[0133] The data correction module is used to: perform distribution correction on the photovoltaic array operation data to obtain photovoltaic array correction data;

[0134] Anomaly detection module is used to: input the photovoltaic array correction data into a pre-constructed anomaly feature detection model and output the photovoltaic array anomaly feature detection result;

[0135] The model building module is used to build an anomaly feature perception model. The building process includes:

[0136] Obtain the original autoencoder structure;

[0137] The PCAE structure is obtained by embedding physical quantity constraints for correcting weakly correlated physical quantities in the latent space of the original autoencoder.

[0138] After the PCAE structure, CNN and LSTM structures are added sequentially to obtain the constructed anomaly feature perception model. The PCAE structure is represented as follows: ; ; ; ; in, This represents the output of the encoder linear transform layer at sampling time k; , , These represent the weight matrices of the encoder linear transformation layer, the encoder activation layer, and the decoder, respectively. This represents the encoder's input vector at sampling time k; , , These represent the bias terms of the encoder linear transformation layer, encoder activation layer, and decoder, respectively. This represents the output of the encoder's active layer at sampling time k; , These represent the activation functions of the encoder and decoder, respectively. , These represent the sequence prediction state and the secondary sequence prediction state of the photovoltaic array, respectively. , These represent the sequential prediction output voltage value and the quadratic prediction output voltage value of the photovoltaic array, respectively. This represents the decoder output at sampling time k; This represents the latent space sampling point at sampling time k.

[0139] The specific functions of each module described above are explained in the relevant content of Embodiment 1 or 2, and will not be repeated here.

[0140] Example 4

[0141] This embodiment introduces a computer-readable storage medium storing a computer program / instruction thereon, which, when executed by a processor, implements the steps of the photovoltaic array anomaly feature sensing method described in Embodiment 1 or 2.

[0142] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0143] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0144] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0145] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0146] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for sensing abnormal features of a photovoltaic array, characterized in that, include: Acquire photovoltaic array operation data; The photovoltaic array operation data is distributed and corrected to obtain photovoltaic array correction data; The photovoltaic array correction data is input into a pre-built anomaly feature perception model, and the photovoltaic array anomaly feature perception result is output. The construction of the anomaly feature perception model includes: The PCAE structure is obtained by embedding physical quantity constraints for correcting weakly correlated physical quantities in the latent space of the original autoencoder. After the PCAE structure, CNN and LSTM structures are added sequentially to obtain the constructed anomaly feature perception model. The PCAE structure is represented as follows: ； ； ； ； in, This represents the output of the encoder linear transform layer at sampling time k; , , These represent the weight matrices of the encoder linear transformation layer, encoder activation layer, and decoder, respectively. This represents the encoder's input vector at sampling time k; , , These represent the bias terms of the encoder linear transformation layer, encoder activation layer, and decoder, respectively. This represents the output of the encoder's active layer at sampling time k; , These represent the activation functions of the encoder and decoder, respectively. , These represent the sequence prediction state and the secondary sequence prediction state of the photovoltaic array, respectively. , These represent the sequential prediction output voltage value and the quadratic prediction output voltage value of the photovoltaic array, respectively. This represents the decoder output at sampling time k; This represents the latent space sampling point at sampling time k.

2. The photovoltaic array anomaly detection method according to claim 1, characterized in that, The photovoltaic array operating data is corrected using an inverse Gaussian distribution to obtain corrected photovoltaic array data. The probability density formula for the inverse Gaussian distribution is: ； ； ； ； in, This represents the output voltage of the i-th photovoltaic module in the photovoltaic array. The inverse Gaussian probability density function; This represents the output voltage of the i-th photovoltaic module in the photovoltaic array. The average value; The shape parameter represents the inverse Gaussian distribution; This represents the operation of the exponential function; This represents the output voltage of the i-th photovoltaic module in the photovoltaic array. The variance; This represents the output voltage correction value of the i-th photovoltaic module in the photovoltaic array at sampling time k; Indicates the sampling time Output voltage of the i-th photovoltaic module in the lower photovoltaic array; Let represent the inverse Gaussian probability density function of the i-th photovoltaic module in the photovoltaic array at sampling time k; This represents the difference between the output voltage of the i-th photovoltaic module in the photovoltaic array and the output voltage correction value at sampling time k.

3. The photovoltaic array anomaly detection method according to claim 1, characterized in that, The processing steps of the anomaly feature perception model include: In the PCAE structure, the photovoltaic array correction data is subjected to data dimensionality reduction and feature extraction to obtain a low-dimensional feature vector, and abnormal pattern features are selected from the low-dimensional feature vector. In the CNN structure, the abnormal detail features within a short time window and the abnormal association features within a long time window are extracted from the abnormal pattern features, and the abnormal detail features and abnormal association features are concatenated to obtain local abnormal features. In the LSTM structure, long-term temporal dependency information of the local anomaly features is captured to obtain the photovoltaic array anomaly feature perception results.

4. The photovoltaic array anomaly detection method according to claim 1, characterized in that, The LSTM structure is represented as follows: ； ； ； in, , , , Let represent the output functions of the forget gate, input gate, candidate memory cell function, and output gate at time t, respectively. express The weight matrix; express The weight matrix; express The weight matrix; express The weight matrix; This represents the input at time t; , Let represent the hidden states at time t-1 and time t, respectively; , , , They represent bias vector, bias vector, bias vector, The bias vector; , These represent memory cell units at time t-1 and time t, respectively. This represents the hyperbolic tangent activation function.

5. The photovoltaic array anomaly detection method according to claim 1, characterized in that, The training of the anomaly feature perception model includes: The hyperparameters of the anomaly perception model are trained by hierarchical iterative collaborative mean square error loss and root mean square error loss, resulting in a well-trained anomaly perception model.

6. The photovoltaic array anomaly detection method according to claim 1, characterized in that, Also includes: The anomaly feature perception model is used as a state space model, and the STKF algorithm is used to track the state changes of the photovoltaic array anomaly feature perception results in real time. When the state of the photovoltaic array abnormal feature sensing result changes abruptly, increase the Kalman gain in the STKF algorithm; When the state of the photovoltaic array abnormal feature sensing result is stable, reduce the Kalman gain in the STKF algorithm.

7. The photovoltaic array anomaly detection method according to claim 6, characterized in that, The state-space model is represented as follows: ； ； in, This represents the state vector of the photovoltaic array at time t. This represents the state transition matrix at time t-1. This represents the control input matrix at time t-1. This represents the process noise at time t; Let represent the observation vector at time t; Let represent the observation matrix at time t; This represents the measurement noise at time t.

8. A photovoltaic array anomaly detection device, characterized in that, include: The data acquisition module is used to acquire photovoltaic array operating data. The data correction module is used to: perform distribution correction on the photovoltaic array operation data to obtain photovoltaic array correction data; Anomaly detection module is used to: input the photovoltaic array correction data into a pre-constructed anomaly feature detection model and output the photovoltaic array anomaly feature detection result; The model building module is used to build an anomaly feature perception model. The building process includes: The PCAE structure is obtained by embedding physical quantity constraints for correcting weakly correlated physical quantities in the latent space of the original autoencoder. After the PCAE structure, CNN and LSTM structures are added sequentially to obtain the constructed anomaly feature perception model. The PCAE structure is represented as follows: ； ； ； ； in, This represents the output of the encoder linear transform layer at sampling time k; , , These represent the weight matrices of the encoder linear transformation layer, encoder activation layer, and decoder, respectively. This represents the encoder's input vector at sampling time k; , , These represent the bias terms of the encoder linear transformation layer, encoder activation layer, and decoder, respectively. This represents the output of the encoder's active layer at sampling time k; , These represent the activation functions of the encoder and decoder, respectively. , These represent the sequence prediction state and the secondary sequence prediction state of the photovoltaic array, respectively. , These represent the sequential prediction output voltage value and the quadratic prediction output voltage value of the photovoltaic array, respectively. This represents the decoder output at sampling time k; This represents the latent space sampling point at sampling time k.

9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps of the photovoltaic array anomaly feature sensing method according to any one of claims 1 to 7.

Images