Power System Current Fluctuation Prediction Method Based on Multivariate Data and Feature Selection

Abstract

The present invention relates to the technical field of power system data modeling and intelligent prediction, and discloses a power system current fluctuation prediction method based on multi-source data and feature selection, including the following steps: S1. Data perturbation enhancement processing: Collect the original data of the power system with multi-source heterogeneity including current, voltage, load, and meteorology, construct a feature matrix, introduce a perturbation term into the data to form a perturbed sample matrix, which is used to enhance the robustness of the model to data missing and noise interference; S2. Perturbed sample repair and reconstruction: Input the perturbed sample matrix into an auto-encoder structure, and obtain the repaired feature matrix by minimizing the reconstruction error and the coding sparsity constraint, which is used for subsequent modeling. By introducing perturbation enhancement, auto-encoder reconstruction, graph structure embedding, and feature scoring mechanisms, an end-to-end current fluctuation prediction framework is constructed, effectively enhancing the adaptability of the model to data anomalies and system structure information.

Description

Technical Field

[0001] The present invention relates to the technical field of power system data modeling and intelligent prediction, and specifically to a method for predicting current fluctuations in a power system based on multivariate data and feature selection. Background Art

[0002] With the continuous expansion of the scale of power systems and the rapid increase in the proportion of new energy access, the monitoring and prediction of current fluctuations have become an important part of ensuring the stable operation of the power grid and assisting dispatching decisions. Especially in the context of a high proportion of intermittent power access, the short-term fluctuations of current present more complex nonlinear and random characteristics, which puts higher requirements on the stability and responsiveness of the prediction method.

[0003] Traditional current fluctuation prediction methods mainly rely on time series modeling of historical data or regression analysis of single statistical features. Such methods usually assume that the input data is complete and stable, and lack the ability to systematically model measurement errors, data missing, and multi-source disturbances in actual operation. In real applications, the monitoring data collected by the power system often has inevitable noise interference, packet loss, and multi-source feature heterogeneity, which will have a serious impact on the reliability and generalization ability of the prediction model.

[0004] On the other hand, existing methods also have certain limitations in dealing with the structural relationship between multi-dimensional features. Due to the failure to fully consider the complex dependencies between features formed by the power grid topology, physical coupling relationship or operation logic, the prediction model often lacks structural perception ability and is difficult to effectively capture the potential spatial correlation and information propagation path between high-dimensional features, resulting in expression bias in the model when dealing with multi-variable inputs.

[0005] In addition, in high-dimensional input scenarios, feature selection mechanisms usually rely on static statistical evaluation or model-built-in weight analysis, lacking a systematic feature evaluation method for disturbance responses. This results in the model containing a large number of redundant or low-value features, further increasing the model's computational burden and reducing its stability and interpretability in different operating scenarios. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present invention provides a method for predicting current fluctuations in power systems based on multivariate data and feature selection, which solves the problem of achieving efficient, stable and structure-aware prediction of current fluctuation trends under the conditions of disturbance, incompleteness and high-dimensional redundant features in multi-source data of power systems in the prior art.

[0007] To achieve the above objectives, the present invention is implemented by the following technical scheme: a method for predicting current fluctuations in a power system based on multivariate data and feature selection, comprising the following steps:

[0008] S1. Data perturbation enhancement processing: Collect the original data of the power system with multi-source heterogeneity including current, voltage, load, meteorology, etc., construct a feature matrix, introduce a perturbation term to the data to form a perturbed sample matrix, which is used to enhance the robustness of the model to data loss and noise interference;

[0009] S2. Perturbed sample repair and reconstruction: Input the perturbed sample matrix into the autoencoder structure, and obtain the repaired feature matrix by minimizing the reconstruction error and the coding sparsity constraint for subsequent modeling;

[0010] S3. Graph structure embedding modeling: Based on the power grid nodes and their connection relationships, construct an adjacency matrix, perform graph convolution processing on the repaired feature matrix to generate an embedded feature representation, so as to introduce the ability of the power grid topology structure to perceive the propagation of current fluctuations;

[0011] S4. Feature importance scoring and screening: Use the collaborative filtering mechanism to construct a feature-sample relationship matrix based on the prediction error, extract the feature scores through non-negative matrix factorization, and select the feature subset according to the set threshold. This subset will be used as the input features for the final prediction model training;

[0012] S5. Multi-objective loss modeling and prediction training: Based on the feature subset, construct a current fluctuation prediction model, and optimize the multi-objective loss function including the prediction error term, the perturbation repair stability term, and the graph structure consistency term during the training process to improve the accuracy of the model and the system coordination;

[0013] S6. Module feedback linkage adjustment: During the training process, feedback the feature scoring results to the perturbation reconstruction module to focus on enhancing the repair accuracy of highly sensitive features. At the same time, use the structure indicators in the graph embedding to optimize the feature screening weights to achieve the information closed-loop adjustment between each functional module;

[0014] S7. Prediction output: Use the trained prediction model to predict the current fluctuation of the new input data and output the prediction result.

[0015] Preferably, the perturbation enhancement processing in step S1 is: adding a noise term that follows a Gaussian distribution to the original feature matrix to construct a perturbed sample

[0016] ;

[0017] Among them, is the original feature matrix, is the input sample matrix, is a noise matrix that follows a normal distribution with a mean of 0 and a covariance of and is the identity matrix.

[0018] Preferably, the reconstruction loss function of the auto-encoding structure in step S2 is defined as:

[0019] ;

[0020] where, is the input perturbation sample matrix, is the encoder function, is the decoder function, is the weight factor of the sparse regularization term.

[0021] Preferably, the graph structure embedding in step S3 adopts a normalized adjacency matrix

[0022] ;

[0023] where, is the original adjacency matrix, is the degree matrix, which is a diagonal matrix;

[0024] And calculate the graph embedding feature representation

[0025] ;

[0026] where, is the repaired feature matrix, is the trainable weight matrix, is the activation function.

[0027] Preferably, the element of the feature-sample error matrix constructed in step S4 is defined as:

[0028] ;

[0029] where, is the true value of the th sample, is the prediction model after removing the th feature, is the corresponding input.

[0030] Preferably, in step S4, the non-negative matrix factorization M≈UV is performed on the feature-sample relationship matrix, and the feature score is the L2 norm of each row of the matrix U And obtain the feature subset Sf={j|sj≥τ} according to the set threshold τ.

[0031] Preferably, the feedback mechanism in step S6 includes feeding back the high-scoring features identified in step S4 to the auto-encoding structure in step S2, and performing key weighting processing on its reconstruction process.

[0032] Preferably, the feedback mechanism in step S6 further includes participating in the weighted operation of feature scoring using the node centrality index obtained in graph embedding.

[0033] Preferably, the parameter optimization process of the prediction model in step S5 adopts a Bayesian optimization strategy to minimize the multi-objective loss function.

[0034] The present invention provides a method for predicting current fluctuations in a power system based on multivariate data and feature selection, having the following beneficial effects:

[0035] 1. By introducing a perturbation enhancement mechanism to simulate data uncertainty in the operation of the power system and combining with an autoencoder to jointly reconstruct missing and noisy data, the present invention effectively improves the model's adaptability to input anomalies, provides a more complete and reliable feature representation for prediction modeling, and has significant engineering applicability in the current fluctuation prediction scenario.

[0036] 2. By combining the implicit dependence relationship between the physical structure or features of the power grid, constructing a graph structure embedding module, and introducing graph convolution to perform structure-aware re-representation of multi-dimensional features, the present invention effectively models the spatial topology and logical connection between features, provides an input with both semantic and structural consistency for downstream tasks, and helps to improve the model's ability to depict system-level association patterns.

[0037] 3. By quantitatively analyzing the impact of perturbation samples on prediction performance and constructing a global feature importance evaluation mechanism through non-negative matrix factorization, the present invention solves the problems of lack of robustness and interpretability in traditional feature selection methods and ensures the effectiveness of the selected features in actual prediction tasks.

[0038] 4. By designing a multi-objective loss function including prediction error, perturbation stability, and graph structure consistency and establishing a feedback linkage mechanism, the present invention realizes the system optimization from feature construction to model training, makes the parameters of each sub-module depend on each other and dynamically adjust, constructs a unified and highly collaborative end-to-end prediction process, and improves the comprehensive adaptability of the method under complex working conditions of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a flowchart of the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0040] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0041] Please refer to the attached Figure 1 , the embodiment of the present invention provides a power system current fluctuation prediction method based on multi-source data and feature selection, including the following steps:

[0042] S1. Data perturbation enhancement processing: Collect the original data of the power system including multi-source heterogeneous data such as current, voltage, load, and meteorology, construct a feature matrix, introduce a perturbation term to the data, and form a perturbation sample matrix to enhance the robustness of the model to data missing and noise interference;

[0043] In this embodiment, aiming at the actual engineering problems such as unstable quality of original data, data missing, and noise pollution faced in the current fluctuation prediction problem of the power system, a data processing strategy based on perturbation enhancement is proposed. As the first step of the overall method of the present invention, it lays a reliable data foundation for subsequent feature reconstruction, structure modeling, and prediction training.

[0044] In an actually operating power system, the collected multi-source data often comes from various sensors and measurement systems, such as smart meters, substation monitoring equipment, environmental monitoring modules, etc. These data usually include but are not limited to observed values with time series and multi-dimensional attributes such as current, voltage, active power, reactive power, load, frequency, temperature, humidity, wind speed, etc. Due to factors such as sensing accuracy, electromagnetic interference, and communication link failures, these original data often have different degrees of missing, inconsistent, or noise perturbation phenomena.

[0045] To improve the generalization ability and anti-noise performance of the subsequent model, this embodiment preferably introduces a perturbation enhancement mechanism to perform perturbation sampling processing on the original feature matrix to construct a robust input sample and simulate the data perturbation behavior in the operation of the power system.

[0046] Let the collected original feature data set be a real number matrix , where represents the number of samples, represents the feature dimension. Apply a perturbation function based on a probability distribution to this original matrix to construct a perturbation sample matrix . The form of this perturbation function is preferably a Gaussian distribution, that is:

[0047] ;

[0048] Among them, is the embedded feature matrix output by the graph convolution, represents a noise matrix that follows a normal distribution with zero mean and covariance of , is an identity matrix with a dimension of , is an adjustable noise intensity parameter, usually tuned as a model hyperparameter during training.

[0049] The purpose of this perturbation operation is not only to simulate the possible perturbation patterns in the actual acquisition and transmission process of input data, but more importantly, to guide the subsequent model to learn the implicit redundant structure and dependence relationship between input features through a random perturbation mechanism, thereby enhancing the model's adaptability to incomplete data and offset data.

[0050] It should be noted that the perturbation process should not be simply understood as a data augmentation operation. It plays an important role as a "detector" in stimulating the structure recovery and feature selection process in the overall framework of the present invention. In the subsequent reconstruction stage, the model needs to perform adaptive reconstruction based on the perturbed samples to identify the internal structure between data; at the same time, in the feature selection stage, this perturbation also provides a premise for evaluating the sensitivity of features to prediction errors.

[0051] In addition, to ensure that the perturbed samples still retain the statistical distribution characteristics of the original data and can be used for training and optimization, in the embodiment, the perturbed samples are further normalized or standardized to keep them consistent with the original data distribution.

[0052] Preferably, this step can also be combined with a time sliding window strategy to batch process multi-time point data to construct a data perturbation cluster within the time series segment, further enhancing the model's ability to identify local time perturbations.

[0053] The perturbed sample matrix will be input as an input into the autoencoder structure in the next step to participate in the subsequent feature reconstruction and missing repair process, constituting the first important link in the data processing chain.

[0054] In summary, the data perturbation enhancement process in this embodiment injects noise into the original feature matrix by simulating the actual system noise, missing, and abnormal fluctuations, thereby providing basic support for the model's robustness and generalization ability. It has a clear engineering background and theoretical feasibility, and can effectively support the subsequent deep reconstruction, structure modeling, and feature evaluation processes.

[0055] S2. Repair and reconstruction of perturbed samples: Input the perturbed sample matrix into the autoencoder structure, and obtain the repaired feature matrix by minimizing the reconstruction error and the coding sparsity constraint for subsequent modeling;

[0056] In this embodiment, based on the perturbed sample matrix constructed in step S1 , a self-encoding reconstruction method that combines compressive learning and sparse constraints is proposed for jointly modeling and repairing the potential missing information and noise pollution in the perturbed data.

[0057] To effectively restore the expression integrity of the original features and avoid information redundancy or structural deformation introduced by perturbations, in this embodiment, an autoencoder structure is preferably used as a repair tool for the input perturbation samples. The autoencoder is a non-linear dimensionality reduction and reconstruction structure that can learn the low-dimensional latent representation in the sample data and adaptively reconstruct the input data based on this.

[0058] The autoencoder includes a set of encoder functions and decoder functions , where:

[0059] The encoder function is responsible for mapping the input samples to the low-dimensional latent space;

[0060] The decoder function is responsible for restoring the latent representation to the original feature space.

[0061] Let the perturbation sample matrix be , then the encoder output is the latent variable matrix , corresponding to the low-dimensional latent representation of each sample. Subsequently, the decoder reconstructs into , that is, the restored matrix , with the goal of making this reconstruction result as close as possible to the original unperturbed sample.

[0062] To ensure that the model can not only restore the surface features but also extract potential correlation information during the reconstruction process, and at the same time suppress the interference of the input perturbation term on the output, in this embodiment, a dual loss constraint is introduced into the training process, including a reconstruction error term and a sparse regularization term. The overall loss function is in the following form:

[0063] ;

[0064] where:

[0065] The first term is the average reconstruction error (in the form of the square of the L2 norm), which measures the difference between the decoded output and the original sample;

[0066] The second term is the L1 regularization term of the latent coding vector, which introduces a sparsity constraint and encourages the model to extract concise and representative feature expressions;

[0067] is the regularization weight coefficient, which is used to balance the relative importance between reconstruction accuracy and sparse expression.

[0068] During the training process, the reconstruction mechanism adjusts the parameters through backpropagation and gradient descent optimization. Preferably, activation functions (such as ReLU, LeakyReLU, etc.) can be used to improve the non-linear modeling ability, and combined with batch normalization operations to accelerate the convergence speed and enhance the model stability.

[0069] It should be noted that the design goal of the autoencoder structure in this embodiment is not limited to reconstructing the surface features of the input data. Instead, it aims to jointly model the perturbed samples to identify the relevant structures and internal mapping relationships between the input features, so as to achieve automatic completion of possible missing items, active correction of outliers, and provide a higher-quality input feature matrix for subsequent graph structure modeling and feature selection.

[0070] The reconstructed output data matrix is denoted as in this embodiment. As the structure recovery result in the method of the present invention, this matrix has removed certain perturbation interferences in structure and filled in some missing or non-ideal observed features through latent space learning.

[0071] This step is not only a preliminary repair of the input samples, but also a fundamental link in the feature reconstruction and context association capture in the prediction framework of the present invention. In subsequent steps, the repaired data will be used as the direct input of the graph embedding module and the feature scoring mechanism, affecting the final feature screening and prediction performance.

[0072] In summary, in this embodiment, by introducing a structured sparse autoencoder, the perturbed data is jointly reconstructed to establish a non-linear mapping relationship between the perturbed samples and the original data, thereby effectively completing the data missing repair and feature denoising processing, and forming a reconstructed feature matrix with structural integrity and embeddability. This provides a solid foundation for the data-driven modeling in the overall method of the present invention.

[0073] S3. Graph structure embedding modeling: Based on the power grid nodes and their connection relationships, construct an adjacency matrix, and perform graph convolution processing on the repaired feature matrix to generate an embedded feature representation, so as to introduce the ability of the power grid topology structure to perceive the propagation of current fluctuations.

[0074] In this embodiment, in order to more accurately depict the structural correlation relationships between the various feature variables in the power system and make full use of their potential physical constraint information, before subsequent modeling of the feature matrix repaired in step S2, preferably introduce a graph neural network modeling mechanism to perform structure-aware feature embedding processing on it.

[0075] In an actual power system, the characteristics of each electrical quantity are not independent of each other, and there are often physical transmission paths and logical dependencies. For example, the current fluctuations of adjacent substation nodes have a topological connection in space, and there is a coupled evolution characteristic between voltage and load in time. This complex structural relationship is difficult to directly model through traditional feature engineering. Therefore, in this embodiment, a graph structure modeling method is introduced to express and learn the mutual dependencies between features from the structural level.

[0076] Specifically, first, a graph structure is constructed based on the physical or logical connection relationships between features , where: represents the feature set; represents the edge set between features, reflecting the correlation or adjacency relationship between each feature variable.

[0077] According to the constructed graph structure, the adjacency matrix of the feature graph is defined , where represents the feature and the feature are connected, represents no connection relationship.

[0078] To eliminate the imbalance caused by the difference in feature connection degrees, preferably, the adjacency matrix is normalized to obtain the symmetric normalized adjacency matrix , and its calculation method is as follows:

[0079] ;

[0080] where, is the original adjacency matrix, is the degree matrix, which is a diagonal matrix;

[0081] where, is the degree matrix, , and the remaining elements are 0, indicating the number of connections of each node in the graph. This normalization operation helps to maintain the weight balance of each node during the information propagation process and prevent the high-degree features from having a disproportionate impact on the embedding results.

[0082] After obtaining the normalized graph structure, a Graph Convolutional Network (GCN) is used to perform a graph embedding operation on the repaired feature matrix . The graph convolution process is essentially an information aggregation process based on the node adjacency structure, and its embedding expression form in the present invention is:

[0083] ;

[0084] where: represents the feature representation matrix after graph embedding; is the repaired feature matrix of the input; is the graph convolution weight matrix, which needs to be optimized through training; is the activation function, and preferably non-linear functions such as ReLU and tanh can be selected; is the aforementioned normalized adjacency matrix, which ensures the effective transmission of the graph structure in the convolution propagation.

[0085] The core idea of the above graph convolution operation is that each feature representation is not only composed of its own information, but also incorporates the information of its adjacent feature nodes in the graph structure. This mechanism enables the feature expression to have the ability of structural context awareness, and is particularly suitable for modeling the power linkage changes caused by the network topology in the power system.

[0086] In the method of the present invention, the feature matrix after graph embedding will be used in the subsequent feature importance evaluation and selection process. The embedded features retain the semantic information of the original features, and at the same time introduce the dependency constraints of the structural hierarchy, which helps to more accurately evaluate the contribution of each feature in the entire network structure during the feature screening stage.

[0087] Preferably, in the implementation process, the graph structure can be constructed based on the physical topology of the power grid (such as node connection graph, electrical distance matrix, etc.), or a similarity graph can be constructed according to the statistical correlation between historical data, so as to achieve the integration between physical prior and data-driven.

[0088] In summary, in this embodiment, by introducing the graph structure embedding method, the connection relationship between each feature variable in the power system is incorporated into the modeling process. Through graph convolution operations, the effective transmission and integration of structural information are realized, and a feature embedding result with structural expression ability is generated, providing structural support for the subsequent feature evaluation, screening and prediction modeling in the method of the present invention.

[0089] S4. Feature importance scoring and screening: Use the collaborative filtering mechanism to construct a feature-sample relationship matrix based on the prediction error, extract the feature scores through non-negative matrix factorization, and select the feature subset according to the set threshold. This subset will be used as the input features for the final prediction model training;

[0090] In this embodiment, based on the graph structure embedding feature result obtained in step S3 and the repaired feature matrix in step S2 , a feature importance scoring and screening method that combines the collaborative filtering mechanism and matrix factorization technology is proposed, which is used to identify the feature subset that contributes significantly to the current fluctuation prediction task from the original high-dimensional features, so as to reduce input redundancy, improve the generalization ability of the model, and provide an optimized feature representation for the subsequent prediction modeling.

[0091] In the actual operation data of power systems, there are obvious differences in the influence degrees of different features on the prediction target. Some features may have a weak dynamic correlation with current fluctuations and may even introduce noise interference. Traditional feature selection methods based on single statistical indicators or model weights often cannot fully consider the sensitivity of features to prediction errors in a perturbed environment. Therefore, in the solution of the present invention, the feature selection step not only focuses on the static correlation between features and the prediction target, but also introduces a feature scoring idea based on a "perturbation-response" mechanism, that is, by analyzing the impact of removing a certain feature on the prediction performance under perturbed samples to measure the prediction contribution degree of this feature.

[0092] Specifically, let the total number of samples be n and the total feature dimension be d. For each sample i and each feature j, a feature sensitivity evaluation matrix is constructed , where each element represents the error change generated by the -th sample in the prediction stage after removing the -th feature, and is defined as follows:

[0093] ;

[0094] where: is the true label value corresponding to the -th sample (i.e., the true current fluctuation amount); represents the sample feature vector obtained after removing the -th feature; represents the prediction model retrained in the sample space without the -th feature.

[0095] This error matrix M reflects the influence degree of each feature on the prediction result under perturbed samples. If a significant error increase occurs after a certain feature is removed, it can be considered that this feature has strong prediction value in the model.

[0096] To further extract the global feature importance score from this error matrix, in this embodiment, the non-negative matrix factorization (Non-negative Matrix Factorization, NMF) method is preferably used to factorize the error matrix M:

[0097] ;

[0098] where: is the feature factor matrix; is the sample factor matrix; is the set latent factor dimension, which is usually much smaller than and ;

[0099] All matrix elements satisfy the non - negative constraint to maintain interpretability and sparsity.

[0100] Through the above - mentioned decomposition process, the feature scores can be calculated based on the row norms (or row averages) of matrix U. Preferably, in this embodiment, the L2 norm is adopted as the feature - scoring metric, that is, for each feature j, its score is defined as:

[0101] ;

[0102] where, represents the th row of the matrix . This score represents the degree of factor contribution of feature in the interpretation error matrix. The higher the score, the more significant the role of this feature in the error reconstruction, and thus it can be considered more predictive.

[0103] After obtaining the scoring results of all features , screening can be performed by setting a threshold , and retaining the features with scores not lower than this threshold to form a feature subset , that is:

[0104] ;

[0105] Preferably, this threshold can be determined through cross - validation, stability analysis, or a feedback mechanism based on structural metrics, and the feature scores can also be weighted by combining the node - centrality metrics in the graph - embedding stage to improve the structural consistency and system interpretability of the screening results.

[0106] This feature subset will ultimately serve as the effective input dimension in the training stage of the prediction model of the present invention, directly affecting the subsequent learning efficiency and performance of the model. It should be noted that this screening method not only considers the influence degree of features on the prediction error, but also models the contribution structure of each feature to the prediction residual space under the perturbation environment through NMF factorization, and has strong global adaptability and module interpretability.

[0107] In summary, in this embodiment, by constructing a perturbation error matrix and introducing a non - negative matrix factorization mechanism, the importance measurement and effective screening of input features in the prediction task are realized, forming a feature subset that has both data - driven basis and fuses system - structure perception, providing refined and high - quality input support for subsequent model training.

[0108] S5. Multi-objective Loss Modeling and Prediction Training: Based on the feature subset, a current fluctuation prediction model is constructed. During the training process, a multi-objective loss function including a prediction error term, a disturbance repair stability term, and a graph structure consistency term is optimized to improve the accuracy of the model and the system coordination.

[0109] In this embodiment, after completing the selection of the feature subset in step S4, the embedded features corresponding to the feature subset are used as the main input of the prediction model of the present invention to construct a learning model for predicting the current fluctuation amount in the power system, and a multi-objective loss function is proposed to jointly optimize the prediction accuracy, disturbance robustness, and structural consistency, thereby improving the robustness and generalization ability of the model in practical applications.

[0110] First, let the input feature matrix finally used for training be , which is composed of the features selected by step S4 in the original feature matrix and can be part of the structure representation after graph embedding or the original reconstructed feature representation. The specific selection depends on the application strategy.

[0111] In the present invention, the prediction target is the current fluctuation amount at a certain time point or a certain time period in the power system, denoted as , where represents the prediction model, is the sample input feature vector, is the prediction output result, is the true label corresponding to the sample.

[0112] To achieve the systematic optimization of the prediction model, this embodiment designs a multi-objective joint loss function to simultaneously constrain the prediction error, disturbance sample stability, and graph structure preservation ability of the model. The form of the multi-objective loss function is as follows:

[0113] ;

[0114] where: is the prediction error term, which is used to measure the distance between the model output and the true value. Preferably, it can be expressed in the form of mean squared error (MSE) as:

[0115] ;

[0116] This term is the basic error index in supervised learning and directly reflects the fitting degree of the model to the target variable.

[0117] is the disturbance stability constraint term, which is used to ensure the robustness of the model to input disturbances and is defined as the difference in the prediction output between the original sample and the disturbance sample. Let the disturbance sample input be , whose predicted value is , then this item can be expressed as:

[0118] ;

[0119] This item ensures that the model remains stable in output when facing input noise, missing or perturbations, thereby enhancing the anti-interference ability during system operation.

[0120] is a structural consistency constraint term, mainly used to maintain the consistency of structural features before and after graph embedding. Preferably, it can be achieved through graph Laplacian constraint, feature reconstruction error or graph contrast loss. Let the feature before embedding be , and the graph embedding feature be , then the consistency loss term can be simplified as:

[0121] ;

[0122] Among them, is a structure-preserving mapping matrix, represents the Frobenius norm, which measures the deviation of the embedded feature from the original feature in the overall representation.

[0123] In the above loss function, α and β are adjustable weight coefficients, used to balance the importance of each subtask objective. This loss function can be jointly optimized through the backpropagation mechanism to drive the model to achieve coordination and balance among different tasks.

[0124] Preferably, the model structure can be a deep feedforward neural network (such as a multi-layer perceptron), a time series network (such as an LSTM), or a graph neural network (such as a GCN / GAT). The specific architecture can be flexibly selected according to the target prediction period and task characteristics. The model training process can adopt optimization algorithms such as gradient descent, such as Adam or RMSProp, combined with an appropriate learning rate adjustment strategy to achieve fast convergence.

[0125] After training, the obtained predicted model parameters will be used in the subsequent online prediction process to estimate the current fluctuation trend of new input samples. The prediction results can be used as auxiliary decision-making information in the power grid dispatching system, or for business modules such as load anomaly detection and system stability assessment.

[0126] In summary, in the training process of the current fluctuation prediction model in this embodiment, a structured multi-objective loss optimization mechanism is introduced. By jointly considering prediction accuracy, perturbation robustness, and graph structure preservation, the adaptability of the model in the complex environment of the power system is improved, providing a complete and implementable training strategy and optimization framework for achieving high-reliability prediction tasks.

[0127] S6, module feedback linkage adjustment: During the training process, the feature scoring results are fed back to the perturbation reconstruction module to focus on strengthening the restoration accuracy of highly sensitive features. At the same time, the structural indicators in the graph embedding are used to optimize the feature screening weights to achieve information closed-loop adjustment between the functional modules;

[0128] In this embodiment, during the process of completing the training and construction of the prediction model, a module-to-module feedback linkage mechanism is further introduced to open up the information path between the functional sub-modules of the present invention, and realize the coordinated adjustment between feature evaluation, graph structure modeling, disturbance repair and prediction optimization, thereby improving the stability and learning consistency of the overall model in the scenario of systematic data disturbance.

[0129] In the overall process of the present invention, although each module has a clear division of labor in terms of function, there is a close logical dependency between its input and output. For example, the reconstruction accuracy of the perturbed sample repair module in step S2 directly affects the subsequent graph embedding expression effect and feature scoring stability; and the result of feature selection in step S4 determines the final input dimension of the model, which further affects the model prediction performance and the consistency of graph structure embedding.

[0130] Therefore, in order to avoid local optimality or decision deviation caused by information fragmentation between modules, this embodiment introduces the following two types of feedback mechanisms during the training process:

[0131] In this embodiment, firstly based on the feature scoring result obtained in step S4 , identify the highly sensitive feature set, that is, the features that cause a significant increase in prediction error during feature elimination. Feed this set back to the autoencoder reconstruction module in step S2 to guide the feature weighting strategy in the reconstruction process. Specifically, the feature weighting term is introduced into the original reconstruction loss function, so that the model pays higher attention to the reconstruction accuracy of highly sensitive features during the reconstruction phase. The corrected loss function is as follows:

[0132] ;

[0133] in: Features The reconstruction weight of , which can preferably be determined according to its normalized score in the feature score; For sample No. The reconstructed value of the feature;

[0134] The remaining symbols remain the same as those defined in step S2.

[0135] In this way, the model can perform differentiated modeling based on the feedback of feature importance from the prediction task during the data repair phase, effectively improving the information consistency of the entire process.

[0136] Furthermore, in order to enhance the integration between graph structure modeling and feature selection, this embodiment also introduces the structural indicators obtained in the graph embedding stage as feedback signals to participate in the weighted calculation of feature importance scoring. Specifically, in step S3, the structural centrality metric of each feature node can be calculated after graph convolution, such as node degree, PageRank value or average attention weight based on embedding space, etc., which is recorded as This structural index can be compared with the feature score Jointly construct a weighted scoring function:

[0137] ;

[0138] in: , It is an adjustable weight coefficient, which is used to control the influence ratio of prediction error sensitivity and graph structure weight; It is the final comprehensive score used for feature screening.

[0139] The above structural feedback mechanism ensures that feature selection not only considers the contribution of prediction error, but also integrates the criticality of the features in the graph structure, which is more conducive to constructing a feature subset with reasonable structure and effective prediction.

[0140] Preferably, this embodiment can further optimize the graph embedding weight matrix W according to the final training error feedback, and realize parameter cascade update through an end-to-end optimization mechanism. If the graph neural network is used as the main model, the above graph embedding weights can be directly jointly trained under the prediction loss drive to form an overall unified optimization closed loop.

[0141] In summary, this embodiment, by constructing two types of feedback paths, namely “feature scoring → reconstruction weighting” and “graph structure index → feature weighting”, opens up the parameter dependency and information coupling between the processing modules, and constructs an adaptive linkage learning mechanism, thereby improving the overall modeling coordination and stability of the prediction model of the present invention in a complex disturbance environment.

[0142] S7, prediction output: use the trained prediction model to predict the current fluctuation of new input data and output the prediction result;

[0143] In this embodiment, after completing the model training and multi-objective joint optimization, based on the prediction model obtained in step S5 , and the feature subset obtained by screening in step S4 , carry out prediction operations on future current fluctuations in the power system and achieve the final prediction output results.

[0144] To ensure the consistency and stability of the input of the prediction model, in this step, it is preferred to preprocess the new sample data to be predicted according to the same feature processing flow as the training data. Specifically, it includes: according to the feature subset constrain the dimension of the input feature matrix; normalize or standardize the features to match the feature distribution in the training stage; in the case of structural embedding requirements, synchronously construct the graph structure information and calculate the embedding representation.

[0145] Let the sample set to be predicted be represented as , where each sample represents the feature input after feature screening and processing. Input it into the trained prediction model , and the corresponding prediction output can be obtained:

[0146] ;

[0147] The internal structure of the above model is preferably a feedforward regression model in this embodiment. Its training process has been completed in step S5, and the parameters have been optimized and learned based on the joint loss function. If the model used contains a graph neural network or a time series structure, the prediction function can also perform forward calculation through the corresponding graph convolution layer or sequence propagation mechanism.

[0148] In the case of graph embedding requirements, the graph representation of the prediction sample needs to be constructed synchronously through the graph structure generation method defined in step S3, and then the normalized adjacency matrix and the weight matrix W learned in the training stage are used to perform graph convolution operations on the input features. Its graph embedding representation can be expressed as:

[0149] ;

[0150] Among them, represents the original prediction sample feature matrix, is the graph embedding result, which can be used as the model input or input into the model after being concatenated with the original features.

[0151] The prediction result can be expressed as the current change amount of a certain node or multiple nodes in the power system at a future moment or the average fluctuation amount within the prediction interval range. This output form can be adjusted according to actual business needs, including but not limited to single-point prediction, sliding window prediction, or multi-step prediction, etc.

[0152] In the solution of the present invention, the prediction result can be directly used in the power dispatching center, the operation control system or the power grid load prediction module, and can also be linked with the alarm module or the load response control system to identify potential load abnormalities, system overloads or prediction deviations.

[0153] Preferably, to enhance the interpretability and traceability of the prediction output in the business system, this embodiment can also synchronously output intermediate quantity information such as feature response weights and node embedding activation intensities to assist the subsequent decision-making module in analyzing the prediction formation process.

[0154] In summary, through the forward inference process based on the final trained model in this embodiment the prediction of the current fluctuation amount of the newly input sample is completed, forming the prediction output mechanism of the last link in the method of the present invention. On the basis of maintaining consistency with the training stage, this process ensures the rationality, stability and practicability of the output by introducing the collaborative mechanism of structure embedding and feature screening, and constitutes a deployable and scalable prediction execution scheme for the actual scenario.

[0155] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principle and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting current fluctuations in a power system based on multi-source data and feature selection, characterized in that It includes the following steps: S1. Data perturbation enhancement processing: Collect the original data of the power system with multi-source heterogeneous data including current, voltage, load, and meteorology, construct a feature matrix, introduce a perturbation term to the data to form a perturbed sample matrix, which is used to enhance the robustness of the model to data missing and noise interference; S2. Perturbed sample repair and reconstruction: Input the perturbed sample matrix into an auto-encoder structure, and obtain the repaired feature matrix by minimizing the reconstruction error and the coding sparsity constraint for subsequent modeling; S3. Graph structure embedding modeling: Based on the power grid nodes and their connection relationships, construct an adjacency matrix, perform graph convolution processing on the repaired feature matrix to generate an embedded feature representation, so as to introduce the ability of the power grid topology structure to perceive the propagation of current fluctuations; S4. Feature importance scoring and screening: Use a collaborative filtering mechanism to construct a feature-sample relationship matrix based on the prediction error, extract feature scores through non-negative matrix factorization, and select a feature subset according to a set threshold. This subset will be used as the input features for the final prediction model training; S5. Multi-objective loss modeling and prediction training: Based on the feature subset, construct a current fluctuation prediction model, and optimize the multi-objective loss function including the prediction error term, the perturbation repair stability term, and the graph structure consistency term during the training process to improve the accuracy of the model and the system coordination; S6. Module feedback linkage adjustment: During the training process, feedback the feature scoring results to the perturbation reconstruction module to focus on strengthening the repair accuracy of highly sensitive features. At the same time, use the structure indicators in the graph embedding to optimize the feature screening weights to achieve the information closed-loop adjustment between each functional module; S7. Prediction output: Use the trained prediction model to predict the current fluctuation of new input data and output the prediction result.

2. The method for predicting current fluctuations in a power system based on multi-source data and feature selection according to claim 1, wherein The perturbation enhancement processing in step S1 is: adding a noise term that follows a Gaussian distribution to the original feature matrix to construct a perturbed sample ； Among them, is the original feature matrix, is the input sample matrix, is the noise matrix that follows a normal distribution with a mean of 0 and a covariance of and is the identity matrix.

3. The power system current fluctuation prediction method based on multi-source data and feature selection according to claim 1, wherein, The reconstruction loss function of the auto-encoder structure in step S2 is defined as: ； Among them, is the input perturbation sample matrix, is the encoder function, is the decoder function, is the weight factor of the sparse regularization term.

4. The method for predicting current fluctuations in a power system based on multi-source data and feature selection according to claim 1, wherein The graph structure embedding in step S3 uses a normalized adjacency matrix ； Among them, is the original adjacency matrix, is the degree matrix, which is a diagonal matrix; and calculates the graph embedding feature representation ； Among them, is the embedded feature matrix output by graph convolution, is the repaired feature matrix, is the trainable weight matrix, is the activation function.

5. The method for predicting current fluctuations in a power system based on multi-source data and feature selection according to claim 1, wherein The feature-sample error matrix constructed in the step S4 is defined as follows: ； Among them, is the true value of the th sample, is the prediction model after removing the th feature, is the corresponding input.

6. The method for predicting current fluctuations in a power system based on multi-source data and feature selection according to claim 1, wherein In the step S4, non-negative matrix factorization M≈UV is performed on the feature-sample relationship matrix, and the feature score is the L2 norm of each row of the matrix U. And a feature subset Sf={j|sj≥τ} is obtained according to the set threshold τ.

7. The method for predicting current fluctuations in a power system based on multi-source data and feature selection according to claim 1, characterized in that The feedback mechanism in step S6 includes feeding back the high-scoring features identified in step S4 to the auto-encoder structure in step S2 and performing key weighted processing on its reconstruction process.

8. The method for predicting current fluctuations in a power system based on multi-source data and feature selection according to claim 1, characterized in that The feedback mechanism in step S6 also includes using the node centrality index obtained in the graph embedding to participate in the weighted operation of feature scoring.

9. The method for predicting current fluctuations in a power system based on multi-source data and feature selection according to claim 1, wherein The parameter optimization process of the prediction model in step S5 adopts a Bayesian optimization strategy to minimize the multi-objective loss function.

Images