CNN-BiLSTM-AM model-based photovoltaic power generation power missing data completion method and device

Abstract

The invention relates to a CNN-BiLSTM-AM model-based photovoltaic power generation power missing data completion method. The method comprises the steps of defining a missing matrix with the same size as photovoltaic power generation power data X; a CNN-BiLSTM-AM model is constructed, and the CNN- The CNN-BiLSTM-AM model is optimized, and the CNN-BiLSTM And inputting to-be-complemented photovoltaic power generation power missing data into the optimized model to obtain a missing data complementation result, and evaluating an error of the missing data complementation result. Supervised training is carried out based on a predefined missing matrix M, the distinguishing precision of an original observation value and a generated complementation value is greatly improved, reliable feedback is provided for a generator, and complementation data is pushed to be close to reality in distribution and rationality; the method is suitable for photovoltaic power data scenes with different missing rates, the complemented complete data can be directly used for power prediction, operation and maintenance optimization and other services, and an accurate basis is provided for power grid dispatching and energy management; compared with a traditional method, the complemented data can reflect the power fluctuation rule more truly, and decision errors caused by data missing are reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of photovoltaic power generation data processing, and particularly relates to a photovoltaic power generation power missing data completion method and device based on a CNN-BiLSTM-AM model. BACKGROUND

[0002] In the actual operation process of a photovoltaic power station, due to the influence of the natural environment uncertainty of the region, such as extreme rainstorms, strong winds and sand, high temperature and sun exposure, and other adverse weather, the system data acquisition equipment is prone to operation abnormity, data transmission communication interruption or sensor failure, thereby causing the SCADA system to have photovoltaic power generation power data missing problems. Such data missing not only significantly reduces the real-time monitoring accuracy of the SCADA system on the operation state of the photovoltaic power station, but also causes the complete photovoltaic power station operation data set to be unable to be obtained: on the one hand, the number of effective samples in the data set is reduced, and on the other hand, the time continuity and overall integrity of the data are destroyed, causing the data accuracy to decrease, and ultimately directly affecting the training effect of the subsequent photovoltaic power generation power prediction model, and even possibly causing the model to have overfitting risk due to data distribution distortion. The above data missing problem has become a key bottleneck restricting the improvement of the photovoltaic power generation power prediction accuracy, and therefore, developing an efficient and reliable photovoltaic power generation power missing data completion technology has become a core technical problem to be solved in the current photovoltaic energy field.

[0003] At present, the data completion methods proposed by the academic and industrial circles are mainly divided into two categories: a classical statistical method and a data-driven method. In the classical statistical method, an autoregressive integrated moving average model (ARIMA) is widely used in the missing value completion scene of power, voltage, current and other parameters in the power system due to its mature linear time series modeling capability. However, this kind of method relies on a fixed mathematical model structure, and is difficult to effectively depict the high-dimensional nonlinear characteristics unique to photovoltaic power generation data; and a large number of experiments show that as the data missing rate increases, the completion accuracy and stability of the classical statistical method will significantly decrease, and cannot meet the actual application requirements of the photovoltaic power station in a high missing rate scene. Based on the deficiencies of the above existing technologies, the future research on the photovoltaic power generation power missing data completion technology should focus on optimizing the model structure design to adapt to the time dependence and multi-feature correlation of the data, and improving the model training strategy to enhance the learning ability of the nonlinear and non-stationary data, thereby improving the depiction accuracy of the model on the complex distribution of the photovoltaic power generation power data, and ultimately realizing the further improvement of the photovoltaic power generation power missing data completion accuracy and stability, and providing reliable data support for the subsequent power prediction and power grid dispatching. SUMMARY

[0004] To address the problem that existing technologies struggle to effectively capture potential correlations among multiple variables, resulting in incomplete data that fails to meet practical application requirements, the primary objective of this invention is to provide a photovoltaic power generation missing data completion method based on a CNN-BiLSTM-AM model. This method significantly improves the accuracy of distinguishing between original observations and generated complete values, more accurately reflects power fluctuation patterns, and reduces decision-making errors caused by missing data.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for completing missing photovoltaic power generation data based on a CNN-BiLSTM-AM model, the method comprising the following sequential steps:

[0006] (1) Define a missing matrix of the same size as the photovoltaic power generation data X. ;

[0007] (2) Based on the missing matrix Construct a CNN-BiLSTM-AM model, which includes a generator and a discriminator, and design a hinting mechanism for the discriminator.

[0008] (3) The CNN-BiLSTM-AM model is optimized according to the game optimization strategy between the generator and the discriminator to obtain the optimized model;

[0009] (4) Input the missing photovoltaic power generation data to be completed into the optimized model to obtain the missing data completion result and evaluate the error of the missing data completion result.

[0010] Step (1) specifically refers to: defining a photovoltaic power generation data set containing T time points. ,in S is the number of feature dimensions for each time step; a missing matrix with the same shape as X is introduced. In the missing matrix middle, express It has been observed, and Indicates missing values. Indicates photovoltaic data in the first The first moment 1 eigenvalue, Representing the missing matrix In the The first moment One characteristic value; observation data ,in, For observation data In the The first moment 1 eigenvalue, From input and represents, as shown in the following formula:

[0011] ;

[0012] In the formula, is the Hadamard product operation of the matrix;

[0013] The final output of the completed data is defined as , so , wherein is a matrix with the same dimension as and all elements are 1, is the generator output matrix, is the first feature value of the generator output matrix at the moment;

[0014] The missing matrix is represented as:

[0015]

[0016] In the formula, the data missing item is represented as 0, and the data normal item is represented as 1.

[0017] In step (2), the generator includes a feature attention module, a time correlation module and a linear layer, the linear layer is used to extract features, the time correlation module captures the time dependence of the time series, and the feature attention module reveals the potential correlation between multiple variables by giving different weights to the sequence. The specific calculation formula is as follows:

[0018] ;

[0019] ;

[0020] ;

[0021] ;

[0022] In the formula, and are the query vector and the key vector for calculating the similarity respectively, is the value vector, is the dimension of the key, , , are the weight matrices of , , respectively; Attention is the attention mechanism; and softmax is the activation function.

[0023] The loss function of the generator The calculation formula is as follows:

[0024] ;

[0025] ;

[0026] ;

[0027] In the formula, is the output of the generator; is the noise; is the authenticity discrimination result of the discriminator on the generated data; is the discrimination loss, is the reconstruction loss, is a hyperparameter that controls the proportion between the discrimination loss and the reconstruction loss .

[0028] In step (2), the discriminator includes a bidirectional gated recurrent unit, a linear layer and a Sigmoid activation function; the bidirectional gated recurrent unit effectively captures the long-term dependence relationship existing in the time series data by fusing and integrating the forward and backward time series information of the time series data; the linear layer performs deep processing on the feature information extracted by the bidirectional gated recurrent unit to strengthen the effectiveness of the feature representation; finally, the Sigmoid activation function is used to output the probability evaluation result of the authenticity of the input data.

[0029] The prompting mechanism of the discriminator is as follows: in order to improve the ability of the discriminator to distinguish between original data and generated data, a prompt matrix H is designed, and the calculation formula of the prompt matrix H is as follows:

[0030] ;

[0031] In the formula, W is a random binary matrix, and the dimension is the same as that of the missing matrix M;

[0032] The loss function of the discriminator is as follows:

[0033]

[0034] In the formula, is the evaluation value of the discriminator for the corresponding position element; is the element of the missing matrix at the corresponding position; indicates that the missing matrix is , the expectation of the evaluation value output by the discriminator for the i-th moment and the j-th feature value is Missing matrix In When, the discriminator outputs the expectation of the evaluation value of the i-th moment and the j-th characteristic value .

[0035] In step (3), the game optimization strategy specifically refers to: the generator takes the generation of the completed data conforming to the real photovoltaic data distribution as the core target, and the generation strategy of the generator needs to focus on making the generated completed data pass the authenticity evaluation of the discriminator; the discriminator takes the accurate distinction between the real photovoltaic data and the completed data of the generator as the core target, and the discrimination ability of the discriminator needs to be iteratively improved around the difference between the data generated by the generator and the real data; in the continuous iteration of the game process, the discriminator feeds back the difference between the generated data and the real data to the generator by outputting the data authenticity probability evaluation result; the generator adjusts the generation logic based on the feedback, and further reduces the distribution deviation between the generated completed data and the real data; at the same time, the output data of the generator after optimization will also be used as new training samples to drive the discriminator to further improve the recognition accuracy; the generator and the discriminator realize bidirectional dynamic optimization through the closed-loop game mechanism of generation, discrimination, feedback and optimization, so that the completed data generated by the generator is highly consistent with the real photovoltaic data distribution, and the discriminator has stable and accurate authenticity discrimination ability, thereby guaranteeing the effectiveness of the overall data completion scheme.

[0036] In step (4), the root mean square error is used to evaluate the error of the photovoltaic power generation power missing data completion result; the smaller the value of the root mean square error, the closer the generated value is to the real value.

[0037] Another object of the present application is to provide an electronic device comprising:

[0038] a processor; and

[0039] a memory having computer program instructions stored therein, the computer program instructions causing the processor to execute the CNN-BiLSTM-AM model-based photovoltaic power generation power missing data completion method as described above when executed by the processor.

[0040] The present application also provides a computer readable storage medium having computer program instructions stored thereon, the computer program instructions causing the processor to execute the CNN-BiLSTM-AM model-based photovoltaic power generation power missing data completion method as described above when executed by the processor.

[0041] From the above technical solution, the beneficial effects of the present application are: first, the generator in the present application precisely adapts to the photovoltaic data characteristics and efficiently mines data value, adopts a CNN-BiLSTM-AM fusion architecture, a convolutional neural network extracts potential correlations and local patterns among multiple features, and captures the periodicity and trend of photovoltaic power; a bidirectional long short-term memory network integrates bidirectional time sequence information and fully mines long-term dependencies; an attention mechanism focuses on key features and time steps and strengthens effective information extraction; the three work together to solve the problem that traditional models cannot balance multiple feature correlations and long-term sequence dependencies; second, the discriminator loss optimization improves the completion accuracy and model stability, introduces the Wasserstein distance to replace the traditional cross-entropy loss, accurately measures the distribution difference between generated data and real data, alleviates the gradient vanishing problem, and makes the training more stable; based on the pre-defined missing matrix M, supervised training is carried out, which greatly improves the discrimination accuracy of the original observed values and the generated completed values, provides reliable feedback for the generator, and promotes the completion data to be close to the real data in distribution and rationality; third, it is suitable for photovoltaic power data scenarios with different missing rates, and the completed complete data can be directly used for power prediction, operation optimization and other businesses, providing accurate basis for power grid dispatching and energy management; compared with traditional methods, the data completed by the present application more truly reflects the power fluctuation law and reduces the decision-making errors caused by data missing. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 is a structure diagram of the generator in the present application;

[0043] Figure 2 is a structure diagram of the discriminator in the present application;

[0044] Figure 3 is a method flowchart of the present application. DETAILED DESCRIPTION

[0045] As shown in Figure 3 , a photovoltaic power generation power missing data completion method based on a CNN-BiLSTM-AM model, the method comprising the following sequential steps:

[0046] (1) define a missing matrix M with the same size as the photovoltaic power generation power data X ;

[0047] (2) construct a CNN-BiLSTM-AM model according to the missing matrix M , the CNN-BiLSTM-AM model includes a generator and a discriminator, and a prompt mechanism of the discriminator is designed;

[0048] (3) according to the game optimization strategy of the generator and the discriminator, optimize the CNN-BiLSTM-AM model to obtain an optimized model;

[0049] (4) Input the missing photovoltaic power generation data to be completed into the optimized model to obtain the missing data completion result and evaluate the error of the missing data completion result.

[0050] Step (1) specifically refers to: defining a photovoltaic power generation data set containing T time points. ,in S is the number of feature dimensions for each time step; a missing matrix with the same shape as X is introduced. In the missing matrix middle, express It has been observed, and Indicates missing values. Indicates photovoltaic data in the first The first moment 1 eigenvalue, Representing the missing matrix In the The first moment One characteristic value; observation data ,in, For observation data In the The first moment 1 eigenvalue, From input and It is represented as shown in the following formula:

[0051] ;

[0052] In the formula, This refers to the Hadamard product operation of matrices;

[0053] Define the final output as the completed data. Then there is ,in, To and A matrix with the same dimensions and all elements being 1. Output matrix for generator, The generator output matrix is ​​at the th The first moment One eigenvalue;

[0054] Missing matrix Represented as:

[0055]

[0056] In the formula, missing data items are represented by 0, and normal data items are represented by 1.

[0057] In step (2), the generator includes a feature attention module, a time correlation module and a linear layer, the linear layer adopts a convolutional neural network (CNN) for feature extraction, the time correlation module adopts a bidirectional long short-term memory (Bilstm) to capture the time dependence of the time series, and the feature attention module adopts an attention mechanism (AM) to reveal the potential correlation between multiple variables by assigning different weights to the sequence, and the specific calculation formula is as follows:

[0058] ;

[0059] ;

[0060] ;

[0061] ;

[0062] In the formula, and are the query vector and the key vector for calculating the similarity respectively, is the value vector, is the dimension of the key, , , are the weight matrices of , , respectively; Attention is the attention mechanism; and softmax is the activation function;

[0063] The loss function of the generator is calculated according to the following formula:

[0064] ;

[0065] ;

[0066] ;

[0067] In the formula, is the output of the generator; is the noise; is the authenticity discrimination result of the generated data by the discriminator; is the discrimination loss, is the reconstruction loss, is a hyperparameter for controlling the proportion between the discrimination loss and the reconstruction loss .

[0068] In step (2), as Figure 2As shown, the discriminator includes a bidirectional gated recurrent unit, a linear layer and a Sigmoid activation function; the bidirectional gated recurrent unit effectively captures the long-term dependence relationship existing in the time series data by fusing and integrating the forward and backward time series information of the time series data; the linear layer performs deep processing on the feature information extracted by the bidirectional gated recurrent unit to strengthen the effectiveness of the feature representation; finally, the Sigmoid activation function is used to output the probability evaluation result of the authenticity of the input data;

[0069] The prompting mechanism of the discriminator is as follows: in order to improve the ability of the discriminator to distinguish between original data and generated data, a prompt matrix H is designed, and the calculation formula of the prompt matrix H is as follows:

[0070]

[0071] In the formula, W is a random binary matrix, and the dimension of W is the same as that of the missing matrix M;

[0072] The core task of the discriminator is to distinguish the source of the values in the completed data, that is, to distinguish which data are original observed values and which data are model generated values. The input of the discriminator includes two parts, which are the completed data and the prompt matrix; the output is a probability matrix consistent with the size of the original data, and the value of each element in the matrix corresponds to the probability size of the original observed value at the same position of the input data. In order to optimize the performance of the model, the Wasserstein distance is used as the core distinguishing measure to realize the difference discrimination between the generated value and the true value, i.e., the original observed value. Compared with the traditional generative adversarial network (GAN), the design based on the Wasserstein distance can effectively improve the stability of the model learning process, solve the mode collapse problem, and reduce the difficulty of model training. The loss function of the discriminator is as follows:

[0073]

[0074] In the formula, W is a random binary matrix, and the dimension of W is the same as that of the missing matrix M; is the evaluation value of the discriminator for the element at the corresponding position; is the element at the corresponding position of the missing matrix M; is the element at the corresponding position of the missing matrix M; When the missing matrix M is , the expectation of the evaluation value output by the discriminator for the i-th time and the j-th feature value is . When the missing matrix M is , the expectation of the evaluation value output by the discriminator for the i-th time and the j-th feature value is .

[0075] ​​In step (3), the game optimization strategy specifically refers to that the generator generates the completed data conforming to the real photovoltaic data distribution as the core target, and the generation strategy of the generator needs to focus on enabling the generated completed data to pass the authenticity evaluation of the discriminator; the discriminator focuses on accurately distinguishing the real photovoltaic data from the completed data generated by the generator, and the discrimination ability of the discriminator needs to be iteratively improved around the difference between the data generated by the generator and the real data; in the continuous iteration process of the game, the discriminator feeds back the difference between the generated data and the real data to the generator by outputting the data authenticity probability evaluation result; the generator adjusts the generation logic based on the feedback to further reduce the distribution deviation between the generated completed data and the real data; at the same time, the output data of the generator after optimization will also be used as new training samples to drive the discriminator to further improve the identification accuracy; through the closed-loop game mechanism of generation, discrimination, feedback and optimization, the generator and the discriminator realize bidirectional dynamic optimization, finally make the completed data generated by the generator highly consistent with the real photovoltaic data distribution, and the discriminator has stable and accurate authenticity discrimination ability, which guarantees the effectiveness of the overall data completion scheme.

[0076] In step (4), the root mean square error is used to evaluate the error of the photovoltaic power generation power missing data completion result; the smaller the value of the root mean square error, the closer the generated value is to the real value.

[0077] To sum up, the generator in the application accurately adapts to the characteristics of photovoltaic data, efficiently mines data value, adopts a CNN-BiLSTM-AM fusion architecture, a convolutional neural network extracts potential correlations and local patterns among multiple features, and captures periodicity and trend of photovoltaic power; a bidirectional long short-term memory network integrates bidirectional time sequence information and fully mines long-term dependency; an attention mechanism focuses on key features and time steps and strengthens effective information extraction; the three work together to solve the problem that traditional models cannot balance multiple feature correlations and long-term sequence dependencies; second, the discriminator loss optimization improves the completion accuracy and model stability, introduces the Wasserstein distance to replace the traditional cross-entropy loss, accurately measures the distribution difference between the generated data and the real data, alleviates the gradient vanishing problem, and makes the training more stable; based on the predefined missing matrix M, supervised training is carried out, which greatly improves the discrimination accuracy of the original observed value and the generated completed value, provides reliable feedback for the generator, and promotes the completed data to be close to the real data in distribution and rationality; the application is suitable for photovoltaic power data scenes with different missing rates, and the completed complete data can be directly used for power prediction, operation optimization and other businesses, providing accurate basis for power grid dispatching and energy management; compared with traditional methods, the data completed by the application more truly reflects the power fluctuation law and reduces the decision-making errors caused by data missing.

[0078] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above-mentioned embodiments, and the above-mentioned embodiments and descriptions in the specification are only the principles of the present application. Various changes and improvements can be made without departing from the spirit and scope of the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for photovoltaic power generation power missing data completion based on a CNN-BiLSTM-AM model, characterized in that: The method comprises the following steps in sequence: (1) define a missing matrix with the same size as the photovoltaic power data X ; (2) according to the missing matrix A CNN-BiLSTM-AM model is constructed, the CNN-BiLSTM-AM model comprises a generator and a discriminator, and a prompt mechanism of the discriminator is designed; (3) optimizing the CNN-BiLSTM-AM model according to the game optimization strategy of the generator and the discriminator to obtain an optimized model; (4) inputting the missing photovoltaic power generation data to be completed into the optimized model to obtain a missing data completion result and evaluating the error of the missing data completion result.

2. The CNN-BiLSTM-AM model-based photovoltaic power generation power missing data completion method according to claim 1, characterized in that: Step (1) is specifically defined as: defining a photovoltaic power data containing T time photovoltaic power data , wherein , S is the characteristic dimension number of each time data; introducing a missing matrix with the same shape as X , in the missing matrix , , indicates has been observed, and indicates a missing value , indicates the jth characteristic value of the photovoltaic data at the ith time , indicates the jth characteristic value of the photovoltaic data at the ith time , indicates the jth characteristic value of the missing matrix at the ith time , indicates the jth characteristic value of the missing matrix at the ith time , indicates the observed data , wherein is the jth characteristic value of the observed data at the ith time , indicates the jth characteristic value of the observed data at the ith time , is represented by input and , as shown in the following formula: ； In the formula, is the Hadamard product operation of matrices; The post-completion data defining the final output is Then there is where is a matrix of the same dimension as with all elements being 1, is the generator output matrix, is the th eigenvalue of the generator output matrix at the th time instant; Missing matrix is represented as: ； In the formula, the data missing item is represented as 0, and the data normal item is represented as 1.

3. The CNN-BiLSTM-AM model-based photovoltaic power generation power missing data completion method according to claim 1, characterized in that: In step (2), the generator comprises a feature attention module, a time correlation module and a linear layer, the linear layer is used to extract features, the time correlation module captures the time dependence of the time series, and the feature attention module reveals the potential correlation between the multiple variables by giving different weights to the sequence, and the specific calculation formula is as follows: ； ； ； ； wherein, and are a query vector and a key vector for calculating similarity, respectively, is a value vector, is a dimension of the key, , , are , , a weight matrix of ; Attention is an attention mechanism; Softmax is an activation function; The loss function of the generator The calculation formula is as follows: ； ； ； wherein is the output of the generator; is the noise; is the discriminator's authenticity discrimination result for the generated data; is the discrimination loss, is the reconstruction loss, is a hyperparameter that controls the ratio between the discrimination loss and the reconstruction loss .

4. The CNN-BiLSTM-AM model-based photovoltaic power generation power missing data completion method according to claim 1, characterized in that: In step (2), the discriminator comprises a bidirectional gated recurrent unit, a linear layer and a Sigmoid activation function; the bidirectional gated recurrent unit effectively captures the long-term dependence relationship existing in the time series data by fusing and integrating the forward and backward time series information of the time series data; The linear layer performs deep processing on the feature information extracted by the bidirectional gated recurrent unit to strengthen the effectiveness of the feature representation; finally, the Sigmoid activation function is used to output the probability evaluation result of the authenticity of the input data; The prompting mechanism of the discriminator is as follows: in order to improve the ability of the discriminator to distinguish between original data and generated data, a prompt matrix H is designed, and the calculation formula of the prompt matrix H is as follows: ； In the formula, W is a random binary matrix, and the dimension of W is the same as that of the missing matrix M; The loss function of the discriminator is: ； In the formula, is the evaluation value of the discriminator for the corresponding position element; is the missing matrix is the element of the corresponding position; denotes the missing matrix In the formula, is the expectation of the evaluation value output by the discriminator for the ith moment and the jth feature value In the formula, denotes the missing matrix In the formula, is the expectation of the evaluation value output by the discriminator for the ith moment and the jth feature value In the formula, 5. The CNN-BiLSTM-AM model-based photovoltaic power generation power missing data completion method according to claim 1, characterized in that: In step (3), the game optimization strategy specifically refers to: the generator takes generating the completed data conforming to the distribution of the real photovoltaic data as the core target, and the generation strategy of the generator needs to be around making the generated completed data pass the authenticity evaluation of the discriminator; the discriminator takes accurately distinguishing the real photovoltaic data from the completed data generated by the generator as the core target, and the discrimination ability of the discriminator needs to be iteratively improved around identifying the difference between the generated data and the real data; In the continuous iterative game process, the discriminator feeds back the difference between the generated data and the real data to the generator by outputting the data authenticity probability evaluation result; The generator adjusts the generation logic based on the feedback to further reduce the distribution deviation between the generated completed data and the real data; at the same time, the output data of the generator after optimization will be used as new training samples to drive the discriminator to further improve the recognition accuracy; through the closed-loop game mechanism of generation, discrimination, feedback and optimization, the generator and the discriminator realize bidirectional dynamic optimization, finally make the completed data generated by the generator highly consistent with the distribution of the real photovoltaic data, and the discriminator has stable and accurate authenticity discrimination ability, which guarantees the effectiveness of the overall data completion scheme.

6. The CNN-BiLSTM-AM model-based photovoltaic power generation power missing data completion method according to claim 1, characterized in that: In step (4), the root mean square error is used to evaluate the error of the photovoltaic power generation missing data completion result; the smaller the value of the root mean square error is, the closer the generated value is to the real value.

7. An electronic device comprising: a processor; and A memory, in which computer program instructions are stored, the computer program instructions, when executed by the processor, causing the processor to perform the method for photovoltaic power generation power missing data completion based on the CNN-BiLSTM-AM model according to any one of claims 1-6. 8.A computer readable storage medium having stored thereon computer program instructions, the computer program instructions, when executed by a processor, causing the processor to perform the method for photovoltaic power generation power missing data completion based on the CNN-BiLSTM-AM model according to any one of claims 1-6.

Images