Cooperative variational multi-view clustering and high-order correlation fusion power load prediction method

Abstract

The invention relates to the technical field of power load prediction, in particular to a collaborative variational multi-view clustering and high-order correlation fusion power load prediction method, which comprises the steps of constructing multi-view data, designing a collaborative variational model for joint modeling, mining high-order correlation, fusing multi-view information, carrying out dimension reduction processing and carrying out deep learning model prediction. According to the method, complementarity and consistency of multi-view data are balanced through the collaborative variational model, high-order correlation is quantized by using the covariance matrix, multi-view features are integrated, and prediction precision is improved in combination with principal component analysis and the GRU model. According to the method, the defects of a traditional method in multi-view data processing can be effectively overcome, the accuracy and generalization ability of power load prediction are remarkably improved, and the efficient load prediction requirement of a power system is met.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power system load prediction, and specifically relates to a power load prediction method combining variational multi-view clustering and high-order correlation fusion. BACKGROUND

[0002] At present, the demand for load prediction of power systems is increasing, and the importance of multi-dimensional data processing in improving prediction accuracy is increasingly prominent. However, the existing power load prediction methods still have certain limitations in processing multi-view data, especially in balancing the complementarity and consistency of data and mining high-order correlations. These problems affect the generalization ability and accuracy of the prediction model in complex scenarios.

[0003] After searching, a power load prediction method with publication number CN109934375B proposes a technical solution based on DBSCAN clustering and pre-post LSTM neural network. This method first uses the DBSCAN algorithm to cluster power load data to obtain class centers, then inputs the load feature data of the previous N days before prediction into the pre-LSTM neural network to extract features, and selects the best training data by calculating the distance from the class center, and finally outputs the prediction result by the post-LSTM neural network. However, this technical solution mainly focuses on the data features of a single view, and lacks effective integration of the complementarity and consistency between different views when processing multi-view data, which may lead to insufficient utilization of information. In addition, this method does not consider the high-order correlation between views, ignoring the potential deep connections in multi-dimensional data, thus limiting the prediction accuracy of the model.

[0004] Another patent with publication number CN118780449B proposes a power load prediction method based on variational modal decomposition, gated recurrent unit (GRU) and attention mechanism. This method decomposes the original time series into sub-sequences and optimizes model parameters using an improved sparrow search algorithm to improve prediction accuracy. Although this method has improved in multi-source data fusion and model parameter optimization, its processing of multi-view data still stays at the shallow feature level and fails to deeply mine the high-order correlation between views. In addition, this technical solution lacks a collaborative mechanism in handling the consistency and complementarity of multi-view data, which may cause the loss of some important information and affect the overall performance of the prediction model.

[0005] The above problems show that the existing power load prediction methods have the following shortcomings in processing multi-view data: first, the balance ability of complementarity and consistency of multi-view data is limited; second, the high-order correlation between views is not fully mined, resulting in that useful information cannot be fully utilized. Therefore, the present application proposes a power load prediction method based on collaborative variational multi-view clustering and high-order correlation fusion, aiming to capture the correlation of multi-view data through a collaborative variational model, and extract deep features by combining a high-order correlation fusion strategy, so as to improve the accuracy and generalization ability of the prediction model, and meet the demand of the power system for efficient load prediction. SUMMARY

[0006] The present application provides a power load prediction method based on collaborative variational multi-view clustering and high-order correlation fusion to solve the problems that the prior art is difficult to balance complementarity and consistency when processing multi-view data, ignores high-order correlation, and information utilization is insufficient.

[0007] According to one aspect of the present application, a power load prediction method based on collaborative variational multi-view clustering and high-order correlation fusion is provided, comprising:

[0008] Based on historical load data, environmental parameters and energy price data, a plurality of data views are constructed to generate a first view matrix, a second view matrix and a third view matrix, and first statistical eigenvalues of the first view matrix, second statistical eigenvalues of the second view matrix, third statistical eigenvalues of the third view matrix and initial association weights between views are obtained;

[0009] A collaborative variational model is designed to jointly model the plurality of view matrices, and shared latent variables are introduced in the model to capture the correlation between different views, and a variational inference process is used to estimate the distribution of the latent variables;

[0010] Based on the output results of the collaborative variational model, an optimization algorithm is used to solve the clustering centers and the attribution probability of each sample to obtain a plurality of clustering label sets, and the class center vector of each clustering label set is calculated;

[0011] For each view matrix, the high-order correlation between different samples is calculated by a covariance matrix formula, which is expressed as:

[0012]

[0013] where ∑ ij represents the covariance between the i-th sample and the j-th sample, x ik and x jk represent the values of the k-th sample in the i-th dimension and the j-th dimension, respectively, and represent the mean of the i-th dimension and the j-th dimension, respectively, and N represents the total number of samples;

[0014] Based on the high-order correlation matrix of multiple views, a fusion strategy is designed to integrate the correlation information of different views, and a weighted average method is used to calculate the fused feature vector, which is expressed as:

[0015]

[0016] where F represents the fused feature vector, V represents the number of views, Wv represents the weight of the vth view, and Rv represents the high-order correlation matrix of the vth view;

[0017] The fused feature vector is processed by dimension reduction, and the principal component analysis method is used to extract the main features, calculate the feature contribution rate, and select the top K principal components as the final input features;

[0018] Based on the extracted feature vector, a deep learning model is selected for training, and the model structure includes an input layer, a hidden layer, and an output layer. The input layer receives the feature vector, the hidden layer uses the gated recurrent unit (GRU) structure to extract the time series features, and the output layer generates the prediction results;

[0019] The performance of the model is evaluated by cross-validation method, and the model hyperparameters are adjusted to optimize the prediction accuracy, and finally the power load prediction results are output.

[0020] In the scheme realized by the above-mentioned power load prediction method based on collaborative variational multi-view clustering and high-order correlation fusion, multiple data views are constructed and a collaborative variational model is introduced to realize joint modeling of multi-view data, which can effectively balance the complementarity and consistency of data. At the same time, by mining the high-order correlation between different views, deep data features are extracted, avoiding the information loss problem caused by traditional methods only focusing on shallow features. Specifically, the calculation of covariance matrix is used to quantify the high-order correlation between samples, and the weighted average fusion strategy further integrates the correlation information of multiple views, providing high-quality input features for principal component analysis. In addition, the fused features are processed by dimension reduction through principal component analysis, removing redundant information and improving the generalization ability of the prediction model. Finally, combined with the time series feature extraction ability of the deep learning model, accurate prediction of power load is realized.

[0021] In the above manner, the present application solves the deficiencies of the prior art in multi-view data processing, especially the limitations in balancing complementarity and consistency and mining high-order correlation. At the same time, this method significantly improves the accuracy and generalization ability of the prediction model, meeting the demand of power system for efficient load prediction. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1This is a flowchart illustrating the method of the present invention, showing the overall process from constructing multi-view data to outputting power load prediction results.

[0023] Figure 2 This is a schematic diagram of the structure of a collaborative variational model, including the core modules of the introduction of shared latent variables and the variational inference process.

[0024] Figure 3 This diagram illustrates the calculation and fusion strategy for higher-order correlations, including the calculation of the covariance matrix and the specific steps of weighted average fusion.

[0025] Figure 4 This is a schematic diagram of the structure of a deep learning model, showing the connection relationship and functional division of the input layer, hidden layer (GRU structure), and output layer.

[0026] The attached figures are labeled as follows:

[0027] 1. First view matrix; 2. Second view matrix; 3. Third view matrix; 4. Covariance model; 5. Shared latent variables; 6. Covariance matrix; 7. Weighted average fusion module; 8. Principal component analysis module; 9. GRU hidden layer; 10. Output layer. Detailed Implementation

[0028] This invention provides a power load forecasting method based on the fusion of collaborative variational multi-view clustering and higher-order correlation. The following is a detailed explanation in conjunction with the appendix. Figure 1 To be continued Figure 4 The accompanying drawings will be described in detail, including the specific component markings shown. Firstly... Figure 1 The document illustrates the overall process, from constructing multi-view data to the final output of power load forecast results. The first view matrix 1, second view matrix 2, and third view matrix 3 correspond to historical load data, environmental parameters, and energy price data, respectively. These data are preprocessed to generate multiple view matrices before proceeding to subsequent steps. A collaborative variational model 4 is used for joint modeling. Shared latent variables 5 are introduced as a core module to capture the correlations between different views. The covariance matrix 6 is used to calculate higher-order correlations between samples. The weighted average fusion module 7 integrates the correlation information from each view into a fused feature vector. The principal component analysis module 8 performs dimensionality reduction on the fused feature vector, extracts the main features, and inputs them into a deep learning model. This model includes a GRU hidden layer 9 and an output layer 10.

[0029] In actual implementation, first, a plurality of data views need to be constructed based on historical load data, environmental parameters and energy price data to form a first view matrix 1, a second view matrix 2 and a third view matrix 3. The first view matrix 1 is derived from historical load data and contains power consumption records in the past period of time; the second view matrix 2 is derived from environmental parameters, including temperature, humidity, wind speed and other meteorological data; and the third view matrix 3 is derived from energy price data, such as electricity price fluctuation. Each view matrix needs to be standardized to ensure that the data distribution is within the same scale range. Then, the first statistical characteristic value of the first view matrix 1, the second statistical characteristic value of the second view matrix 2 and the third statistical characteristic value of the third view matrix 3 are calculated, and the initial correlation weight between views is obtained. The initial correlation weight is calculated using the cosine similarity formula, which measures the correlation between two views by calculating their similarity.

[0030] Next, a collaborative variational model 4 is designed to jointly model the plurality of view matrices. According to Figure 2 As shown in the figure, the core module of the collaborative variational model 4 is a shared latent variable 5, which is used to capture the correlation between different views. The shared latent variable 5 is represented in the form of a probability distribution, and its distribution parameters are estimated through a variational inference process. The specific implementation of the variational inference process includes defining an approximate distribution function, and optimizing the difference between the approximate distribution and the true distribution by maximizing the evidence lower bound (ELBO). In this process, the model updates the parameters using the gradient descent algorithm until convergence. The output of the collaborative variational model 4 includes the latent variable distribution of each sample and the consistency and complementarity information between views.

[0031] Based on the output of the collaborative variational model 4, the K-means optimization algorithm is used to solve the clustering center and the attribution probability of each sample to obtain a plurality of clustering label sets. Each clustering label set corresponds to a class center vector, and the class center vector is calculated as the mean of the feature values of all samples belonging to the class. Then, the high-order correlation between different samples is calculated for each view matrix. According to Figure 3 As shown in the figure, the calculation formula of the covariance matrix 6 is:

[0032]

[0033] where ∑ ij represents the covariance between the i-th sample and the j-th sample, x ik and x jk represent the values of the k-th sample in the i-th dimension and the j-th dimension, respectively, and represent the mean of the i-th dimension and the j-th dimension, respectively, and N represents the total number of samples. Through the calculation of the covariance matrix 6, the high-order correlation between different samples can be quantified.

[0034] After the calculation of the covariance matrix 6 is completed, the weighted average fusion module 7 is used to integrate the correlation information of different views. According to Figure 3 The specific formula of weighted average fusion is shown in the following formula (2):

[0035]

[0036] where F represents the fused feature vector, V represents the number of views, Wv represents the weight of the vth view, and Rv represents the high-order correlation matrix of the vth view.

[0037] The calculation of the view weight Wv uses a normalization method, and the specific formula is as follows:

[0038]

[0039] where α v represents the importance score of the vth view, which is obtained by evaluating the consistency and complementarity between views. The fused feature vector F contains the correlation information from multiple views, providing high-quality input for further feature extraction.

[0040] In order to remove redundant information, the principal component analysis module 8 is used to reduce the dimension of the fused feature vector F. The specific steps of principal component analysis include calculating the covariance matrix of the feature vector F, solving the eigenvalues and eigenvectors of the covariance matrix, and selecting the first K principal components as the final input features according to the size of the eigenvalues. The formula for calculating the feature contribution rate is as follows: where λ k represents the eigenvalue corresponding to the kth principal component. Through the processing of the principal component analysis module 8, the feature dimension can be effectively reduced while retaining the main information, thereby improving the generalization ability of the subsequent deep learning model.

[0041] After feature extraction is completed, a deep learning model is selected for training. According to Figure 4 The structure of the deep learning model includes an input layer, a GRU hidden layer 9, and an output layer 10. The input layer receives the feature vector processed by the principal component analysis module 8, the GRU hidden layer 9 uses a gated recurrent unit structure to extract time series features, and the output layer 10 generates power load prediction results. The specific implementation of the GRU hidden layer 9 includes the design of the update gate and the reset gate, the update gate controls how much historical information is retained in the current state, and the reset gate determines how much historical information is ignored. During model training, the mean square error is used as the loss function, and the model parameters are updated through the backpropagation algorithm.

[0042] To evaluate the model performance, a cross-validation method is adopted for testing. The specific steps include dividing the dataset into training and validation sets, randomly allocating samples according to a certain proportion. After the model training is completed, the prediction accuracy is evaluated on the validation set, and the hyperparameters are adjusted to optimize the model performance. Hyperparameter adjustment includes learning rate, hidden layer node number and batch size selection, and finally the power load prediction result is output.

[0043] The complete flow from constructing multi-view data to the final prediction result output is described in the above specific embodiment, covering the implementation details and operation methods of each step. Through the introduction of the collaborative variational model 4 and the weighted average fusion module 7, joint modeling of multi-view data and mining of high-order correlation are realized, providing high-quality input features for the principal component analysis module 8. Combined with the time series feature extraction capability of the deep learning model, the accurate prediction of power load is finally realized.

[0044] In order to better enable relevant persons in the art to fully understand and implement the present application, the specific implementation principles of the present application are further supplemented below in conjunction with a specific application scenario.

[0045] In actual application, assuming that a certain power company needs to predict the power load situation in the next week to optimize power grid scheduling and energy distribution. First, based on historical load data, environmental parameters and energy price data, multiple data views are constructed. The first view matrix 1 is derived from the power consumption records in the past year, the second view matrix 2 contains temperature, humidity and wind speed meteorological data, and the third view matrix 3 reflects the electricity price fluctuation. After standardization processing, these data generate multiple view matrices and enter the collaborative variational model 4 for joint modeling. The core module of the collaborative variational model 4 is the shared latent variable 5, which is used to capture the correlation between different views. For example, when the temperature rises, the power load may increase due to the increased use of air conditioners, and this correlation is reflected through the shared latent variable 5.

[0046] Subsequently, the first statistical eigenvalue of the first view matrix 1, the second statistical eigenvalue of the second view matrix 2 and the third statistical eigenvalue of the third view matrix 3 are calculated, and the initial correlation weight between views is obtained through the cosine similarity formula. For example, the cosine similarity between the first view matrix 1 and the second view matrix 2 is high, indicating that the historical load data and meteorological data have strong correlation. This correlation weight provides a basis for subsequent high-order correlation mining.

[0047] On the basis of the output results of the collaborative variational model 4, a K-means optimization algorithm is used to solve the clustering center and the attribution probability of each sample, and a plurality of clustering label sets are obtained. For example, a certain clustering label set can correspond to a high-load scenario in a summer high-temperature weather, and a category center vector is obtained by taking the average of the sample feature values belonging to the category. Then, the high-order correlation between different samples is calculated for each view matrix. According to Figure 3 As shown in the formula (6), the calculation formula of the covariance matrix 6 is:

[0048]

[0049] wherein ∑ ij represents the covariance between the i-th sample and the j-th sample. For example, in the first view matrix 1, the fluctuation of the power load in a certain time period and the fluctuation in another time period can have strong correlation, and this high-order correlation is quantified by the covariance matrix 6 and used for subsequent analysis.

[0050] After the calculation of the covariance matrix 6 is completed, the weighted average fusion module 7 is used to integrate the correlation information of different views. According to Figure 3 As shown in the formula (7), the specific formula of the weighted average fusion is:

[0051]

[0052] wherein F represents the fused feature vector, V represents the number of views, Wv represents the weight of the v-th view, and Rv represents the high-order correlation matrix of the v-th view. For example, the weight of the first view matrix 1 can be high because the historical load data has a greater impact on the prediction result. Through the weighted average fusion module 7, the correlation information from multiple views is integrated into a high-quality feature vector F.

[0053] In order to remove redundant information, the principal component analysis module 8 is used to perform dimension reduction processing on the fused feature vector F. The specific steps of the principal component analysis include calculating the covariance matrix of the feature vector F, solving the eigenvalues and eigenvectors of the covariance matrix, and selecting the first K principal components as the final input features according to the size of the eigenvalues. For example, if the contribution rate of the eigenvalue corresponding to the first principal component is 80%, only the first principal component can be retained to reduce the dimension while retaining the main information. Through the processing of the principal component analysis module 8, the dimension of the feature vector is significantly reduced, thereby improving the generalization ability of the subsequent deep learning model.

[0054] After the feature extraction is completed, a deep learning model is selected for training. According to Figure 4As shown, the structure of the deep learning model includes an input layer, a GRU hidden layer 9, and an output layer 10. The input layer receives the feature vectors processed by the principal component analysis module 8, and the GRU hidden layer 9 uses a gated recurrent unit structure to extract time series features. For example, the update gate controls how much historical information is retained in the current state, and the reset gate determines how much historical information to ignore. During model training, the mean square error is used as the loss function, and the model parameters are updated through the backpropagation algorithm. For example, when there is a deviation between the predicted result and the actual load value, the model parameters will be adjusted according to the deviation until convergence.

[0055] To evaluate the performance of the model, the cross-validation method is used for testing. The specific steps include dividing the data set into training and validation sets, and randomly allocating samples according to a certain proportion. For example, 80% of the data is used for training, and 20% of the data is used for validation. After the model is trained, the prediction accuracy is evaluated on the validation set, and the hyperparameters are adjusted to optimize the performance of the model. For example, by adjusting the learning rate, the number of hidden layer nodes, and the batch size, the final power load prediction result is output.

[0056] The above specific application scenarios describe how to build from multi-view data to the final prediction result output, covering the implementation details and operation methods of each step. Through the introduction of the collaborative variational model 4 and the weighted average fusion module 7, joint modeling of multi-view data and mining of high-order correlation are realized, providing high-quality input features for the principal component analysis module 8. Combined with the time series feature extraction capability of the deep learning model, the final realization of the accurate prediction of the power load is realized.

Claims

1. A method for power load forecasting that integrates collaborative variational multi-view clustering and higher-order correlation, characterized in that, Includes the following steps: Based on historical load data, environmental parameters and energy price data, a first view matrix (1), a second view matrix (2) and a third view matrix (3) are constructed. The first statistical feature value of the first view matrix (1), the second statistical feature value of the second view matrix (2), the third statistical feature value of the third view matrix (3) and the initial correlation weight between each view are obtained. The co-variant model (4) is used to jointly model multiple view matrices, and a shared latent variable (5) is introduced to capture the correlation between different views. The distribution of the latent variable is estimated by the variational inference process. Based on the output of the collaborative variational model (4), the cluster center and the probability of belonging to each sample are solved by the optimization algorithm to obtain multiple cluster label sets, and the category center vector of each cluster label set is calculated. For each view matrix, the higher-order correlation between different samples is calculated using the covariance matrix formula, which is expressed as: Where ∑ ij Let x represent the covariance between the i-th sample and the j-th sample. ik and x jk These represent the values ​​of the k-th sample in the i-th and j-th dimensions, respectively. and Let represent the mean of the i-th dimension and the j-th dimension, respectively, and N represent the total number of samples; Based on a high-order correlation matrix of multiple views, a weighted average fusion strategy is designed to integrate the correlation information of different views. The formula is as follows: Where F represents the fused feature vector, V represents the number of views, Wv represents the weight of the v-th view, and Rv represents the high-order correlation matrix of the v-th view. The dimensionality of the fused feature vector is reduced, and the main features are extracted using the principal component analysis module (8). The feature contribution rate is calculated and the top K principal components are selected as the final input features. Based on the extracted feature vectors, a deep learning model is selected for training. The model structure includes an input layer, a hidden layer (GRU hidden layer (9)) and an output layer (10). The input layer receives the feature vectors, the hidden layer uses a gated recurrent unit structure to extract time series features, and the output layer generates prediction results. The model performance is evaluated using cross-validation, and the model hyperparameters are adjusted to optimize prediction accuracy. Finally, the power load prediction results are output.

2. The method according to claim 1, characterized in that, In the step of constructing the first view matrix (1), the second view matrix (2), and the third view matrix (3) based on historical load data, environmental parameters, and energy price data, the first view matrix (1) is derived from historical load data and includes records of electricity consumption over a period of time. The second view matrix (2) is derived from environmental parameters, including temperature, humidity, and wind speed data. The third view matrix (3) is derived from energy price data, including electricity price fluctuations.

3. The method according to claim 1, characterized in that, The initial association weights are calculated using the cosine similarity formula, which measures the degree of association between pairs of views by calculating their similarity.

4. The method according to claim 1, characterized in that, The shared latent variable (5) in the collaborative variational model (4) is represented by a probability distribution and its distribution parameters are estimated through a variational inference process. The specific implementation of the variational inference process includes defining an approximate distribution function and optimizing the difference between the approximate distribution and the true distribution by maximizing the lower bound of evidence.

5. The method according to claim 1, characterized in that, The optimization algorithm is the K-means algorithm, which is used to solve for the cluster centers and the probability of belonging to each sample. The cluster center vector is calculated as the mean of the feature values ​​of all samples belonging to that category.

6. The method according to claim 1, characterized in that, The view weights wv in the weighted average fusion strategy are calculated using a normalization method, specifically: Where α v The importance score of the v-th view is determined by evaluating the consistency and complementarity among the views.

7. The method according to claim 1, characterized in that, The specific steps of the principal component analysis module (8) include calculating the covariance matrix of the eigenvector F, solving for the eigenvalues ​​and eigenvectors of the covariance matrix, and selecting the top K principal components as the final input features based on the magnitude of the eigenvalues. The formula for calculating the feature contribution rate is as follows: Where λ k This represents the eigenvalue corresponding to the k-th principal component.

8. The method according to claim 1, characterized in that, The hidden layer (GRU hidden layer (9)) of the deep learning model adopts a gated recurrent unit structure, including the design of update gate and reset gate. The update gate controls how much historical information is retained in the current state, and the reset gate determines how much historical information is ignored.

9. The method according to claim 1, characterized in that, The specific steps of the cross-validation method include dividing the dataset into a training set and a validation set, randomly allocating samples according to a certain ratio, evaluating the prediction accuracy on the validation set after the model is trained, and adjusting hyperparameters to optimize model performance. Hyperparameter adjustment includes the selection of learning rate, number of hidden layer nodes, and batch size.

10. The method according to claim 1, characterized in that, The loss function of the deep learning model is the mean squared error, and the model parameters are updated through the backpropagation algorithm until convergence.

Images