Pre-training method for photovoltaic system equipment operation data based on self-learning

Abstract

This invention relates to the field of photovoltaic system technology, and an embodiment provides a pre-training method for photovoltaic system equipment operation data based on self-learning. The method includes: acquiring temporal features of photovoltaic system equipment operation information; constructing positive and negative sample data based on the temporal features, and combining the positive and negative sample data into pairwise sample pairs; inputting the sample pairs into a deep Siamese network model to obtain the features and classification prediction values ​​of the sample pairs; and constructing data features of the equipment operation information based on the classification prediction values ​​and features. The embodiments provided by this invention improve the learning effect of photovoltaic equipment operation data representation, and further enhance the robustness and generalization of the data representation learning.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic system technology, and more specifically to a self-learning-based pre-training method for photovoltaic system equipment operation data, a self-learning-based pre-training device for photovoltaic system equipment operation data, an electronic device, and a corresponding storage medium. Background Technology

[0002] As a leader in renewable energy, photovoltaic (PV) power generation costs have gradually approached or even fallen below those of traditional energy sources, making it popular worldwide. Consequently, the safe and reliable operation of PV systems has become a key focus of PV system operation and maintenance. Currently, utilizing information technology to manage and monitor key equipment in PV systems, and to detect abnormal operation, has become an indispensable means to ensure the safe and reliable operation of PV systems.

[0003] In recent years, advancements in sensor technology have made it possible to collect real-time, fine-grained monitoring data from new energy systems. Supervisory Control and Data Acquisition (SCADA) systems based on standard photovoltaic (PV) power generation system installations can monitor the voltage and current of PV strings and the temperature of PV modules at the minute level. Simultaneously, significant progress has been made in the development of artificial intelligence systems. This data collection and technological advancements have spurred numerous downstream services for data-driven monitoring and analysis of PV system equipment, such as anomaly detection. To achieve these analytical tasks, it is often necessary to train a task-related feature extraction model to support these downstream objectives.

[0004] Training task-specific models is typically based on data-driven machine learning methods. These methods usually involve two steps: task-related feature learning and task implementation. To learn task-specific feature representations, a large amount of supervised information (labeled data) is often needed to train high-performance models. However, the amount of labeled data in the daily operation of photovoltaic systems is extremely insufficient. This is because SCADA systems themselves do not collect system operational status data; SCADA data itself is unlabeled, and manual labeling is costly and error-prone. To address the challenge of training task-specific feature extraction models, research has also been conducted based on unsupervised techniques. Unsupervised techniques do not require supervisory information and can detect abnormally operating equipment based solely on reasonable assumptions. For example, based on the assumption that most equipment should be operating normally at the same time, effective clustering methods can group most normally operating equipment into one category, thereby identifying abnormally operating equipment. However, unsupervised methods are based on ideal assumptions, which may lead to some false alarms or missed alarms during application, resulting in low model performance. Summary of the Invention

[0005] The purpose of this invention is to provide a self-learning-based pre-training method for photovoltaic system equipment operation data. Based on the self-learning algorithm, it learns a better data representation form without knowing the specific task, thereby better supporting downstream specific tasks, such as anomaly detection.

[0006] To achieve the above objectives, the first aspect of the present invention provides a pre-training method for photovoltaic system equipment operation data based on self-learning, the method comprising:

[0007] The process involves: acquiring the time-series characteristics of the photovoltaic system equipment's operation information; constructing positive and negative sample data based on the time-series characteristics, and combining the positive and negative sample data into pairs; inputting the sample pairs into a deep Siamese network model to obtain the classification prediction values ​​and features of the sample pairs; and constructing the data features of the equipment operation information based on the classification prediction values ​​and features.

[0008] Preferably, the method further includes: preprocessing the equipment operation information before acquiring the time-series characteristics of the equipment operation information of the photovoltaic system equipment, wherein the preprocessing includes at least one of outlier handling, missing value filling, secondary filtering, and dead number repair.

[0009] Preferably, the secondary filtering includes: performing median filtering on the device operation information, and performing mean filtering on the median-filtered data.

[0010] Preferably, acquiring the time-series characteristics of the photovoltaic system equipment operation information includes: discretizing the acquired equipment operation information according to a preset sampling frequency to obtain the time-series characteristics; the formula is as follows:

[0011]

[0012] in Norm represents the difference feature of the i-th string in the j-th combiner at time t. j (t) represents the parameter value of the junction box j in its normal state at time t. This represents the actual parameter value of the i-th substring in combiner j at time t.

[0013] Preferably, constructing positive and negative sample data based on the time-series characteristics, and combining the positive and negative sample data into pairs, includes: constructing an evaluation criterion based on the mean and deviation of the time-series characteristics, and dividing the time-series characteristics into positive and negative sample data according to the evaluation criterion; each positive sample data and its closest positive sample data form a positive sample pair; each positive sample data and several randomly selected negative sample data form a negative sample pair; the sample pair includes the positive sample pair and the negative sample pair.

[0014] Preferably, inputting the sample pair into a deep Siamese network model to obtain the classification prediction value and features of the sample pair includes: using the activation function of one layer in the deep Siamese network model as a boundary, inputting the sample pair into the deep Siamese network model to obtain the features output by the activation function at the boundary; concatenating the obtained features with the aggregate features and relational features of the sample pair to obtain concatenated features; replacing the features output by the activation function with the concatenated features and inputting them into the part after the boundary of the deep Siamese network model; determining the features of the sample pair using the output of the deep Siamese network model; and obtaining the classification prediction value of the sample pair by applying the features of the sample pair to an activation function.

[0015] Preferably, the loss function of the deep Siamese network model includes: the sum of the value of the binary cross-entropy loss function after being corrected by the balance factor and the contrastive loss function.

[0016] In a second aspect of the invention, a pre-training device for photovoltaic system equipment operation data based on self-learning is also provided, the device comprising:

[0017] The system comprises a time-series feature module for acquiring time-series features of the photovoltaic system equipment operation information; a sample pair module for constructing positive and negative sample data based on the time-series features, and combining the positive and negative sample data into sample pairs; a Siamese network module for inputting the sample pairs into a deep Siamese network model to obtain the classification prediction value and features of the sample pairs; and a feature construction module for constructing data features of the equipment operation information based on the classification prediction value and features.

[0018] A third aspect of the present invention provides an electronic device, comprising: at least one processor; and a memory connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the at least one processor implements the steps of the aforementioned self-learning-based pre-training method for photovoltaic system equipment operation data by executing the instructions stored in the memory.

[0019] A fourth aspect of the present invention provides a machine-readable storage medium storing instructions that, when executed by a processor, configure the processor to perform the aforementioned self-learning-based pre-training method for photovoltaic system equipment operation data.

[0020] The fifth aspect of the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the above-described self-learning-based pre-training method for photovoltaic system equipment operation data.

[0021] The above technical solution has the following beneficial effects:

[0022] (1) This invention proposes a self-learning-based pre-training method for daily operation data of photovoltaic power generation system equipment. This method requires no labeled information; it learns the basic representation of daily operation data of photovoltaic power generation system equipment solely through a large amount of unlabeled data. This is achieved through reasonable data preprocessing, the construction of positive and negative sample pairs, the introduction of expert information, network design, and the similarity between positive and negative data during training. This learned representation can be used to support the performance of many downstream tasks (such as anomaly detection) that suffer from low performance due to a lack of labeled data.

[0023] (2) Based on the understanding of the field data, the designed secondary filtering algorithm achieves more effective data cleaning and improves the effect of data representation learning.

[0024] (3) Based on the understanding of the physical mechanism of photovoltaic power generation system equipment operation, the handcrafted features based on expert experience are designed, which not only improves the effect of data representation learning, but also makes the data representation learning more robust and generalizable.

[0025] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0026] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings:

[0027] Figure 1 The illustration shows a schematic diagram of the steps of a self-learning-based pre-training method for photovoltaic system equipment operation data according to an embodiment of the present invention;

[0028] Figure 2 The illustration shows a schematic diagram of an implementation of a self-learning-based pre-training method for photovoltaic system equipment operation data according to an embodiment of the present invention;

[0029] Figure 3 The diagram illustrates the structure of a pre-training device for photovoltaic system equipment operation data based on self-learning according to an embodiment of the present invention. Detailed Implementation

[0030] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0031] Figure 1 The illustration schematically depicts the steps of a self-learning-based pre-training method for photovoltaic system equipment operation data according to an embodiment of the present invention. Figure 3 As shown, a self-learning-based pre-training method for photovoltaic system equipment operation data includes:

[0032] S01. Obtaining the temporal characteristics of equipment operation information of photovoltaic system equipment;

[0033] S02. Construct positive example sample data and negative example sample data according to the time series characteristics, and combine the positive example sample data and negative example sample data into sample pairs;

[0034] S03. Input the sample pairs into the deep Siamese network model to obtain the classification prediction values ​​and features of the sample pairs;

[0035] S04. Construct the data features of the equipment operation information based on the classification prediction values ​​and features.

[0036] Through the above implementation methods, the basic characteristics of the daily operation data of photovoltaic power generation system equipment can be learned from massive and unlabeled equipment operation information, thereby improving the utilization efficiency of equipment operation information, increasing the application scenarios of equipment operation information, and indirectly improving the processing efficiency of downstream tasks.

[0037] Preferably, the method further includes: preprocessing the equipment operation information before acquiring the time-series characteristics of the photovoltaic system equipment operation information; the preprocessing includes at least one of outlier handling, missing value imputation, secondary filtering, and dead value repair. Preprocessing is necessary to improve the data quality of the equipment operation information and to perform necessary data standardization.

[0038] In some optional implementations, the secondary filtering includes: performing median filtering on the device operation information, and performing mean filtering on the median-filtered data. Specifically, firstly, median filtering is performed on the data collected by the system to remove abnormal noise data. Secondly, mean filtering is applied to the data after the first filtering to smooth the data while removing data noise caused by issues such as sensor sensitivity, as shown in the following formula:

[0039]

[0040]

[0041] Where w represents the size of the filtering window, x i This represents the raw data at time i. and These represent the data after one and two filters, respectively, with `median{·}` representing the median of a set. The filtering strength can be adjusted by changing the size of the filtering window `w`; a larger window results in smoother filtered data.

[0042] In some optional embodiments of the present invention, equipment operation information can be divided into external information of the photovoltaic power station meteorological station and internal information of equipment operation. Meteorological information databases and equipment information databases are established for the photovoltaic power station meteorological station and photovoltaic system equipment at different granularities. Obtaining the time-series characteristics of the photovoltaic system equipment operation information includes: discretizing the acquired equipment operation information according to a preset sampling frequency to obtain the time-series characteristics; the formula is as follows:

[0043]

[0044] in Norm represents the difference feature of the i-th string in the j-th combiner at time t. j (t) represents the parameter value of the junction box j in its normal state at time t. This represents the actual operating parameter value of the i-th string in combiner box j at time t. This parameter is used as the string current data. For example, the data sampling frequency is 5 minutes, and the data dimension is 1×432. The parameter value under normal conditions can be taken as the median of the current under normal conditions after filtering. t is empirically taken as 3 days of time-series data. Irradiance data from the weather station. Alternatively, the current can be calculated using the method described above, which will not be repeated here.

[0045] Preferably, constructing positive and negative sample data based on the time-series characteristics, and combining the positive and negative sample data into pairs, includes: constructing an evaluation criterion based on the mean and deviation of the time-series characteristics, and dividing the time-series characteristics into positive and negative sample data according to the evaluation criterion; each positive sample data and its closest positive sample data form a positive sample pair; each positive sample data and several randomly selected negative sample data form a negative sample pair; the sample pair includes the positive sample pair and the negative sample pair.

[0046] Preferably, inputting the sample pair into a deep Siamese network model to obtain the classification prediction value and features of the sample pair includes: using the activation function of one layer in the deep Siamese network model as a boundary, inputting the sample pair into the deep Siamese network model to obtain the features output by the activation function at the boundary; concatenating the obtained features with the aggregated features and relational features of the sample pair to obtain concatenated features; replacing the features output by the activation function with the concatenated features and inputting them into the portion of the deep Siamese network model after the boundary; determining the features of the sample pair using the output of the deep Siamese network model; and obtaining the classification prediction value of the sample pair by applying the features of the sample pair to an activation function. Specifically, the aggregated features of the sample pair are defined as:

[0047]

[0048] Where Mean(), Median(), Std(), and Max() are the mean, median, standard deviation, and maximum value of the RD features of a certain string, respectively;

[0049] The relationship feature of the sample pairs is defined as follows: the relationship between the operating power of the equipment and the solar irradiance is regressed using the Random Sample Consensus (RANSAC) algorithm, and the slope of the regression is taken as the relationship feature.

[0050] Deep twin networks share the same deep network structure, consisting of Linear1>Tanh1>Linear2>Tanh2>Linear3>Tanh3>Linear4>Tanh4 layers. A pair of samples (referred to as sample A and sample B) data X... i X j The input is fed into a deep Siamese network, and the feature representation is obtained after the Tanh3 layer. The aggregated features z extracted from samples A and B k1 z k2 Relationship feature z l1 z l2 With features The concatenated features are then processed through a Linear4>Tanh4 layer following Tanh3 for further feature extraction, yielding feature z. i z j Further compression is performed using a Linear layer, and the classification prediction probabilities p of samples A and B are obtained using the sigmoid function.

[0051] Preferably, the loss function of the deep Siamese network model includes: the sum of the binary cross-entropy loss function after balance factor correction and the contrastive loss function. Specifically, the method for updating the weight matrix of the neural network model according to the loss function is as follows:

[0052] Calculate the loss function of the neural network model:

[0053] L = L fcntrst +×Lf clss

[0054]

[0055]

[0056] Where: L is the loss function; The comparison loss function is used; Y is the true classification value of the current data. When Y = 0, the current data belongs to the same category; when Y = 1, the current data belongs to different categories; D W For the features z of the current data i , z j In the latent space, the Euclidean distance is used, where m is the current loss cutoff value. When the Euclidean distance between features of different classes exceeds m, no optimization is performed, which can reduce model training time and resources. L fclss Let be the cross-entropy loss function for binary classification, N be the number of data points in the batch, Y be the true classification value of the current data, and p be the probability that the current data is predicted as a positive sample pair. The weight matrix of each layer of the neural network model is updated using stochastic gradient descent.

[0057] Figure 2 The illustration schematically depicts an implementation diagram of a self-learning-based pre-training method for photovoltaic system equipment operation data according to an embodiment of the present invention. Figure 2 As shown, it includes the following steps:

[0058] Step 1: Obtain string operation data, including string current data, from the station operation database. Irradiance data from weather stations The data sampling frequency was 5 minutes, and the data dimension was 1×432.

[0059] Step 2: Data preprocessing mainly includes outlier handling, missing value imputation, noise reduction, and dead number repair to obtain preprocessed string current data. and irradiance data All data dimensions are 1×432.

[0060] Step 3: Based on expert experience and reasonable assumptions, construct positive and negative sample pairs, and perform manual feature extraction. The specific process is as follows:

[0061] Positive sample pair construction: Obtain the equipment current data of the photovoltaic power station for three consecutive days, average the current data for each string, and arrange them in descending order of mean. The top-k strings with the highest mean are then paired with the k strings whose mean values ​​are closest to each other. Here, the first 40% of the data is used, and the three strings with the closest mean values ​​are used to form positive sample pairs.

[0062] Negative sample pair construction: Obtain string current data from the photovoltaic power station for three consecutive days. Average the current data for each string, and sort them from largest to smallest mean. The bottom-k strings are considered abnormal strings, and the top-k strings are considered normal strings. For each normal string, randomly select k samples from the abnormal strings to form negative sample pairs. Here, the first 40% of the data is considered normal strings, and the bottom 40% is considered abnormal strings. For each normal string, randomly select 3 samples from the abnormal strings to form negative sample pairs.

[0063] Use current data from all strings and irradiance data Extract the temporal feature values ​​RD(T) of all strings. The feature value dimension is 1×432. Generate aggregated features based on the temporal feature values ​​RD(T), namely the mean, median, standard deviation, and maximum value.

[0064] z k= {Mean(()),Median(()),Std((T)),Max((T))}, with a feature dimension of 1×4. The Random Sample Consensus (RANSAC) algorithm is used to regress the relationship between the device's operating power and solar irradiance. The slope feature z of the regression is then taken. l The feature dimension is 1×1.

[0065] Step 4: Using a deep Siamese network architecture and a self-learning pre-training method, pre-train the positive and negative sample pairs constructed in Step 3 to obtain the predicted values ​​p for the positive and negative sample pairs. i and their respective characteristics z i .

[0066] Step 5: Repeat step 4 to train the data on other positive and negative samples in the constructed positive and negative sample sets one by one, and obtain the predicted value p and feature z for each pair of data.

[0067] Step Six: Calculate the loss function of the neural network model based on the predicted values ​​of all data, and update the weight matrix of the neural network model according to the loss function. The specific method is as follows:

[0068] Step 6.1: Train the data using sample A and sample B from the positive and negative sample pairs. The specific training process is as follows:

[0069] First, the original string data of sample A and sample B are... Each of these data points consists of batch_size × 1 × 432 dimensions, where batch_size represents the size of each batch. The feature representation of the original data is obtained through a Linear>Tanh>Linear>Tanh>Linear>Tanh>Linear>Tanh layer. Each of them consists of batch_size × 1 × 32 dimensional data.

[0070] Secondly, the artificially designed aggregation features z extracted from samples A and B will be used to further analyze these features. k1 z k2 Relationship feature z l1 z l2 , where z k For batch_size × 1 × 4 dimensional data, z l Given batch_size×1×1 dimensional data, the feature z is obtained by concatenating and merging the data. m1 z m2 Its dimensions are batch_size × 1 × 5.

[0071] Merge feature z m1 z m2 With features The data is concatenated and further feature extraction is performed through a Linear > Tanh layer to obtain feature z. i z j This makes its dimensions batch_size×1×16.

[0072] Finally, the obtained feature z i z j The data is concatenated and compressed using a Linear layer to obtain batch_size×1×1 dimensional data, which is then input into the sigmoid function to obtain the classification prediction values ​​p for samples A and B.

[0073] Step 6.2: Calculate the loss function and its corresponding loss gradient function of the neural network model by taking the positive and negative sample pairs, the current model weights, positive and negative sample features, and the classification prediction value p according to the calculation method of the loss function. Then, update the weights of each layer of the neural network using the stochastic gradient descent method.

[0074] The loss function is calculated as follows:

[0075] Where: L is the loss function; To compare loss functions; λ is the cross-entropy loss function for binary classification; λ is the balancing factor, a hyperparameter used to balance the two loss functions. In the experiment, it was set to 0.8.

[0076] Specifically, its detailed components are as follows:

[0077]

[0078] The comparison loss function is used; Y is the true classification value of the current data. When Y = 0, the current data belongs to the same category; when Y = 1, the current data belongs to different categories; D W denoted as Euclidean distance of the current data's feature values ​​in the latent space, and m as the current loss cutoff value, which is set to 2 in the experiment. When the Euclidean distance between features of different categories exceeds m, no optimization is performed, which can reduce model training time and resources.

[0079]

[0080] Let be the binary cross-entropy loss function, N be the number of data in the batch, Y be the true classification value of the current data, and p be the probability that the current data is predicted as a positive sample pair.

[0081] Step 7: Determine if the training of the neural network model has reached the training stopping condition. If it has, save the model and its weights; otherwise, return to Step 5 to continue the next round of training. The specific process for determining whether the training has converged is as follows:

[0082] Training neural networks typically involves setting a fixed number of epochs, stopping training after the set number of epochs is reached, or stopping training prematurely when the model loss no longer continues to decrease. In this experiment, the number of epochs was set to 500.

[0083] By using the above methods, we can obtain data representations of equipment operation data that have better robustness and generalizability.

[0084] Based on the same inventive concept, the embodiments of the present invention also provide a pre-training device for photovoltaic system equipment operation data based on self-learning. Figure 3 The schematic diagram illustrates the structure of a pre-training device based on self-learning for photovoltaic system equipment operation data according to an embodiment of the present invention, such as... Figure 3 As shown, the device includes:

[0085] The system comprises a time-series feature module for acquiring time-series features of the photovoltaic system equipment operation information; a sample pair module for constructing positive and negative sample data based on the time-series features, and combining the positive and negative sample data into sample pairs; a Siamese network module for inputting the sample pairs into a deep Siamese network model to obtain the classification prediction value and features of the sample pairs; and a feature construction module for constructing data features of the equipment operation information based on the classification prediction value and features.

[0086] In some optional embodiments, the apparatus further includes: preprocessing the equipment operation information before acquiring the time-series characteristics of the equipment operation information of the photovoltaic system equipment, wherein the preprocessing includes at least one of outlier handling, missing value filling, secondary filtering, and dead number repair.

[0087] In some optional implementations, the secondary filtering includes: performing median filtering on the device operation information, and performing mean filtering on the median-filtered data.

[0088] In some optional embodiments, obtaining the time-series characteristics of the photovoltaic system equipment operation information includes: discretizing the obtained equipment operation information according to a preset sampling frequency to obtain the time-series characteristics; the formula is shown below:

[0089]

[0090] in Norm represents the difference feature of the i-th string in the j-th combiner at time t. j (t) represents the parameter value of the junction box j in its normal state at time t. This represents the actual parameter value of the i-th substring in combiner j at time t.

[0091] In some optional implementations, constructing positive and negative sample data based on the time-series characteristics, and combining the positive and negative sample data into pairs, includes: constructing an evaluation criterion based on the mean and deviation of the time-series characteristics, and dividing the time-series characteristics into positive and negative sample data according to the evaluation criterion; each positive sample data and its nearest several positive sample data constitute a positive sample pair; each positive sample data and several randomly selected negative sample data constitute a negative sample pair; the sample pair includes the positive sample pair and the negative sample pair.

[0092] In some optional implementations, inputting the sample pair into a deep Siamese network model to obtain the classification prediction value and features of the sample pair includes: using the activation function of one layer in the deep Siamese network model as a boundary, inputting the sample pair into the deep Siamese network model to obtain the features output by the activation function at the boundary; concatenating the obtained features with the aggregate features and relational features of the sample pair to obtain concatenated features; replacing the features output by the activation function with the concatenated features and inputting them into the portion of the deep Siamese network model after the boundary; determining the features of the sample pair using the output of the deep Siamese network model; and obtaining the classification prediction value of the sample pair by applying the features of the sample pair to an activation function.

[0093] In some alternative implementations, the loss function of the deep Siamese network model includes the sum of the binary cross-entropy loss function after being corrected for a balance factor and the contrastive loss function.

[0094] The specific limitations of each functional module in the aforementioned pre-training device for photovoltaic system equipment operation data based on self-learning can be found in the limitations of the pre-training method for photovoltaic system equipment operation data based on self-learning described above, and will not be repeated here. Each module in the aforementioned device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0095] In some embodiments of the present invention, an electronic device is also provided, comprising: at least one processor; and a memory connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the at least one processor implements the aforementioned steps of the self-learning-based pre-training method for photovoltaic system equipment operation data by executing the instructions stored in the memory. The control module or processor here has numerical calculation and logical operation functions, and at least has a central processing unit (CPU) with data processing capabilities, random access memory (RAM), read-only memory (ROM), multiple I / O ports, and an interrupt system. The processor contains a kernel, which retrieves corresponding program units from the memory. One or more kernels can be configured, and the aforementioned method can be implemented by adjusting the kernel parameters. The memory may include non-permanent memory in computer-readable media, random access memory (RAM), and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0096] In one embodiment of the present invention, a machine-readable storage medium is provided, on which instructions are stored, which, when executed by a processor, cause the processor to be configured to perform the above-described self-learning-based pre-training method for photovoltaic system equipment operation data.

[0097] In one embodiment of the present invention, a computer program product is provided, including a computer program that, when executed by a processor, implements the above-described self-learning-based pre-training method for photovoltaic system equipment operation data.

[0098] 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.

[0099] 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.

[0100] 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.

[0101] 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.

[0102] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0103] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, like read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0104] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0105] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0106] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A self-learning-based pre-training method for photovoltaic system equipment operation data, characterized in that, The method includes: The temporal characteristics of acquiring equipment operation information of photovoltaic system equipment; Based on the time-series characteristics, positive example sample data and negative example sample data are constructed, and the positive example sample data and negative example sample data are combined in pairs to form sample pairs; The process of inputting sample pairs into a deep Siamese network model to obtain classification prediction values ​​and features for the sample pairs includes: using the activation function of one layer in the deep Siamese network model as a boundary, inputting the sample pairs into the deep Siamese network model to obtain the features output by the activation function at the boundary; concatenating the obtained features with the aggregate features and relational features of the sample pairs to obtain concatenated features; replacing the features output by the activation function with the concatenated features and inputting them into the portion of the deep Siamese network model after the boundary; determining the features of the sample pairs based on the output of the deep Siamese network model; and obtaining the classification prediction value of the sample pairs by applying the features of the sample pairs to an activation function. The data features of the equipment operation information are constructed based on the classification prediction values ​​and features; Among them, the aggregation feature is: in, , , , Temporal feature values ​​of a certain string The mean, median, standard deviation, and maximum value; The relationship feature is as follows: the relationship between the operating power of the equipment and the solar irradiance is regressed using the random sampling consensus algorithm, and the slope of the regression is taken as the relationship feature.

2. The method according to claim 1, characterized in that, The method further includes: preprocessing the equipment operation information before acquiring the time-series characteristics of the equipment operation information of the photovoltaic system equipment, wherein the preprocessing includes at least one of outlier handling, missing value filling, secondary filtering, and dead number repair.

3. The method according to claim 2, characterized in that, The secondary filtering includes: performing median filtering on the equipment operation information, and performing mean filtering on the median-filtered data.

4. The method according to claim 1, characterized in that, The time-series characteristics of acquiring equipment operation information for photovoltaic system equipment include: The acquired device operation information is discretized according to a preset sampling frequency to obtain the time-series characteristics; the formula is as follows: in Indicates the first The first junction box Each string at time Difference characteristics over time This indicates the junction box exist The parameter values ​​for the normal state at all times. Indicates junction box The Middle Each string in The actual parameter values ​​at any given time.

5. The method according to claim 1, characterized in that, Based on the aforementioned time-series characteristics, positive and negative sample data are constructed, and the positive and negative sample data are paired to form sample pairs, including: An evaluation criterion is constructed based on the mean and deviation of the time series features, and the time series features are divided into positive sample data and negative sample data according to the evaluation criterion. Each positive sample data and its nearest several positive sample data constitute a positive sample pair; Each positive example sample data and several randomly selected negative example sample data constitute a negative sample pair; The sample pairs include the positive sample pairs and the negative sample pairs.

6. The method according to claim 1, characterized in that, The loss function of the deep twin network model includes the sum of the binary cross-entropy loss function after being corrected by the balance factor and the contrastive loss function.

7. A pre-training device based on self-learning for photovoltaic system equipment operation data, characterized in that, The device includes: The time-series feature module is used to acquire the time-series features of the equipment operation information of photovoltaic system equipment; The sample pair module is used to construct positive sample data and negative sample data according to the time series features, and to combine the positive sample data and negative sample data into sample pairs; A Siamese network module is used to input sample pairs into a deep Siamese network model to obtain classification prediction values ​​and features for the sample pairs. This includes: inputting the sample pairs into the deep Siamese network model with the activation function of one layer as a boundary, and obtaining the features output by the activation function at the boundary; concatenating the obtained features with the aggregate features and relational features of the sample pairs to obtain concatenated features; replacing the features output by the activation function with the concatenated features and inputting them into the portion of the deep Siamese network model after the boundary; determining the features of the sample pairs based on the output of the deep Siamese network model; obtaining the classification prediction value of the sample pairs by applying the features of the sample pairs to an activation function; and a feature construction module for constructing data features of the device operation information based on the classification prediction value and features. Among them, the aggregation feature is: in, , , , Temporal feature values ​​of a certain string The mean, median, standard deviation, and maximum value; The relationship feature is as follows: the relationship between the operating power of the equipment and the solar irradiance is regressed using the random sampling consensus algorithm, and the slope of the regression is taken as the relationship feature.

8. An electronic device, characterized in that, include: At least one processor; A memory connected to the at least one processor; The memory stores instructions executable by the at least one processor, which executes the instructions stored in the memory to implement the steps of the self-learning-based pre-training method for photovoltaic system equipment operation data as described in any one of claims 1 to 6.

9. A machine-readable storage medium, characterized in that, The machine-readable storage medium stores instructions that, when executed by a processor, cause the processor to be configured to implement the self-learning-based pre-training method for photovoltaic system equipment operating data as described in any one of claims 1 to 6.

Images