Day-ahead market decision sample processing method and system based on unsupervised feature extraction

Abstract

The invention belongs to the technical field of big data analysis of a power system and a power market, and provides a day-ahead market decision sample processing method and system based on unsupervised feature extraction, which introduces a function orthogonalization, residual vector and alternative covariance matrix construction mechanism under a unified modeling framework of a spot day-ahead market decision sample matrix. On the premise of not depending on any supervision label, a complex nonlinear redundancy relation formed between factors such as multi-element loads and uncertain power output is recursively stripped in a function space by using function reduction and component analysis, and only an effective structure which still has high variance and high uncertainty in the sense of information theory is reserved; unsupervised linear feature extraction of the high-dimensional sample matrix is realized; the method can explicitly utilize a function family and an orthogonalization process to describe a redundant structure while keeping an overall linear projection form and relatively low calculation complexity, and is suitable for engineering application of large-scale spot day-ahead market samples.

Description

Technical Field

[0001] This invention belongs to the field of big data analysis technology for power systems and power markets, and particularly relates to a method and system for processing day-ahead market decision samples based on unsupervised feature extraction. Background Technology

[0002] In the day-ahead spot market of the power grid, dispatching agencies and market operators need to consider a large number of time-varying factors for clearing and decision-making. For various time-varying factors, this study examines the data characteristics of sample matrix sequences from the perspectives of data structure, magnitude, dimensions, and dimensionality. Based on statistical indicators, feature selection and dimensionality compression methods are designed to reduce feature redundancy and noise while preserving key information, thereby improving the computational efficiency and stability of subsequent clustering, prediction, and optimization.

[0003] The day-ahead market decision samples are often in a high-dimensional nonlinear correlation state, making it difficult for traditional principal component analysis techniques to effectively identify and eliminate nonlinear redundant features. While nonlinear methods such as kernel principal component analysis (KPCA) and deep autoencoders have stronger expressive power, they suffer from poor interpretability, hyperparameter sensitivity, and high training costs. Furthermore, if high-value features are not fully extracted and redundant features are not eliminated under unsupervised conditions before performing kernel norm clustering and typical scenario extraction on the sample matrix, the clustering results will be sensitive to noisy features, leading to the curse of dimensionality in the high-dimensional feature space, deteriorating the clustering effect, and significantly increasing model training time and storage overhead. Summary of the Invention

[0004] To address the aforementioned problems, this invention proposes a method and system for processing day-ahead market decision samples based on unsupervised feature extraction. This invention first utilizes function reduction and component analysis to recursively remove complex nonlinear redundant relationships formed by factors such as multivariate loads, uncertain power output and its stochastic characteristics, climate factors, unit maintenance, and grid power construction and decommissioning in the function space. Only the effective structure, which still possesses high variance and high uncertainty in an information theory sense, is retained, achieving unsupervised linear feature extraction of high-dimensional sample matrices. While maintaining the overall linear projection form and low computational complexity, it can explicitly characterize redundant structures using function families and orthogonalization processes, providing theoretical guarantees such as upper bounds on conditional entropy, making the dimensionality reduction results more physically interpretable and controllable, and suitable for engineering applications of large-scale day-ahead spot market samples.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution: In a first aspect, the present invention provides a method for processing day-ahead market decision samples based on unsupervised feature extraction, comprising: Obtain a sample of market decisions made in the previous day, and construct a random vector based on the sample of market decisions made in the previous day; Based on the random vector and the pre-defined linearly independent family of functions, a family of random functions is obtained; The random function family is orthogonalized in the function space to generate a set of orthogonal normalized functions, and the residual random vector is determined based on the set of orthogonal normalized functions to obtain the alternative covariance matrix; Based on the alternative covariance matrix, iterative feature extraction is performed using a function reduction algorithm to obtain new linear features under unsupervised conditions, which are used to characterize the key information of the spot market decision sample matrix. Component analysis is used to identify and remove redundant principal components that can be nonlinearly represented by other principal components. Unsupervised selection of original features is performed based on residual information analysis to obtain key original features. Based on feature analysis, the functional relationship between the original features is analyzed to identify features in the multi-source data of the spot market that can be approximately represented by combinations of other variables, thereby determining the final set of retained features. The features and feature subsets obtained by function reduction algorithm, component analysis method, residual information analysis and feature analysis method are fused with the preset spot day-ahead market decision sample matrix to obtain the dimension-reduced feature matrix; Clustering is performed on the dimensionality-reduced feature matrix.

[0006] Furthermore, the construction of the sample vector and random vector includes: constructing each day's clearing day as a sample, and uniformly representing the input of multi-element loads, uncertain power output and random characteristics, probabilistic correlations between uncertain factors, independent energy storage, virtual power plants, climate, unit maintenance plans, grid power construction and power decommissioning, as well as the output of thermal power, wind power, photovoltaic, independent energy storage and virtual power plants as random vectors.

[0007] Furthermore, the step of obtaining the random function family based on the random vector and a preset linearly independent function family includes: defining a formal variable vector: ; in, Represents a set of formal variables; Indicates the first A formal variable, which is represented by its corresponding random variable in the function. replace; Indicates the total dimension of the features; Define a family of linearly independent functions: ; in, Represents a set of functions; Indicates the first One function; Indicates the number of functions; Define an arbitrary set of indices as: ; in, Represented by formal variables The middle belongs to the index set The subvector composed of the components; Representing a family of functions China only depends on A subset of functions, i.e., all functions that satisfy The function; When using random vectors Replacement of formal variables When this happens, we obtain a family of random functions: ; in, Represents a random vector The middle belongs to the index set The subvector composed of the components; Indicates random variables Substitute into function The random variable obtained afterwards; This represents the set of these random functions.

[0008] Furthermore, in random vector distributions In the corresponding function space, the inner product is defined as follows: ; in, Represents two random functions; Representation function AND function exist Inner product in space; Represents random variables and The mathematical expectation of the product of .

[0009] Furthermore, the determination of the alternative covariance matrix includes: selecting several random variables; based on the several random variables, obtaining a set of orthogonal normalized functions by performing Gram-Schmidt orthogonalization on the family of functions; for the new function, constructing a random function under the current set of independent variables; then removing the projection from the obtained set of orthogonal functions; subsequently performing normalization; adding the normalized random function to the set of orthogonal normalized functions to form a new set of orthogonal functions.

[0010] Furthermore, the alternative covariance matrix is: ; ; in, Indicates the first The residual vector in the step state; Represents a random vector With random functions The covariance vector; In function space China Regression in the least squares sense; Indicates the first The alternative covariance matrix during step iteration; Represents the outer product of the residual vectors; It represents the term-by-term mathematical expectation of the product matrix.

[0011] Furthermore, iterative feature extraction using a function reduction algorithm includes: defining an initial covariance matrix; in the first... In the next iteration, the largest eigenvalue is determined; ; in, Represents a candidate direction vector; express The Euclidean norm is 1; Represents a random vector In direction variance on; Indicates the first The optimal direction vector obtained from the next iteration; Set variance threshold ,like If no direction with variance greater than the threshold exists in the current alternative covariance matrix, the feature extraction process is stopped; where, Indicates along direction The residual variance; The square of the variance threshold controls the amount of remaining unexplained information; otherwise, a new linear feature is defined; the new linear feature is added to the variable set, the orthogonal function set and residual vector are updated, the covariance matrix is ​​replaced, and the search step for the maximum feature direction is returned to continue iterating.

[0012] Furthermore, the component analysis method is used to identify and remove redundant principal components that can be nonlinearly represented by other principal components. This includes: performing eigenvalue decomposition on the initial covariance matrix to obtain eigenvectors; defining principal components and redundant functions in the traditional sense; if there is a redundant relationship that allows a certain principal component to be represented by other principal component functions, the principal component is considered redundant with respect to the selected principal component set, and thus the principal component is skipped in the principal component screening process, retaining only the principal components that cannot be reconstructed on other principal components through a family of functions.

[0013] Furthermore, based on residual information analysis, unsupervised selection of original features is performed, resulting in key original features including: defining the sum of squared components of the residual vector as a substitute for the variance vector; defining the infinite norm. ; in, This represents the maximum value of a vector component. Indicates the first The alternative variance vector during step iteration; The components of the residual vector are represented. The variance of all features in the current residual space is determined by the infinite norm. If the variance does not exceed the threshold, the feature selection process ends. Otherwise, the original feature corresponding to the index is selected and the unsupervised selection is repeated.

[0014] Furthermore, the analysis of the functional relationships between the original features based on feature analysis methods includes: Let: ; in, Represents an abstract high-dimensional random vector; Indicates the first One random variable; Represent the total dimension of the features; assume there exists a measurable function: ; and parameters ,like Then it is assumed that there exists in the original feature space. - Redundancy ,in, Representation function In random vectors The value of ; This represents the second moment of the redundant function on the data; An upper bound on the degree of redundancy; if an index exists. with subset and functions , so that: ; And a family of functions exists. Projection: ; satisfy: ; Then it is considered a feature In the mean square sense, it can be determined by the characteristics The functional representation of is a redundant feature; among which, Indicates index by feature Subvectors formed; Indicates from arrive The function; Indicates in variable The set of functions after orthogonalization; This indicates the use of a family of functions to analyze features. The mean square error after approximation; This indicates the allowable error threshold.

[0015] Secondly, the present invention also provides a day-ahead market decision sample processing system based on unsupervised feature extraction, comprising: The data acquisition module is configured to: acquire current day market decision samples and construct random vectors based on the current day market decision samples; The random function family building module is configured to: obtain a random function family based on a random vector and a preset linearly independent function family; The alternative covariance matrix establishment module is configured to: orthogonalize the family of random functions in the function space, generate a set of orthogonal normalized functions, and determine the residual random vector based on the set of orthogonal normalized functions to obtain the alternative covariance matrix; The feature extraction module is configured to: perform iterative feature extraction using a function reduction algorithm based on the alternative covariance matrix to obtain new linear features under unsupervised conditions, which are used to characterize the key information of the spot market decision sample matrix; identify and remove redundant principal components that can be nonlinearly represented by other principal components using component analysis; perform unsupervised selection of original features based on residual information analysis to obtain key original features; and analyze the functional relationships between original features based on feature analysis methods to identify features that can be approximated by combinations of other variables in the multi-source data of the spot market before the market, thereby determining the final set of retained features. The dimensionality reduction module is configured to fuse the features and feature subsets obtained by function reduction algorithm, component analysis method, residual information analysis and feature analysis method with the preset spot day-ahead market decision sample matrix to obtain the dimensionality-reduced feature matrix; The clustering module is configured to perform clustering on the dimensionality-reduced feature matrix.

[0016] Thirdly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the day-ahead market decision sample processing method based on unsupervised feature extraction as described in the first aspect.

[0017] Fourthly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the program to implement the steps of the day-ahead market decision sample processing method based on unsupervised feature extraction described in the first aspect.

[0018] Fifthly, the present invention also provides a computer program product, the computer program product comprising a computer program, which, when executed by a processor, implements the steps of the day-ahead market decision sample processing method based on unsupervised feature extraction as described in the first aspect.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention, within a unified modeling framework for the day-ahead market decision sample matrix, introduces mechanisms for function orthogonalization, residual vectors, and alternative covariance matrices. Without relying on any supervised labels, it first uses GS function reduction (GFR) and GS component analysis (GCA) to recursively remove complex nonlinear redundant relationships formed by factors such as multi-variable loads, uncertain power output and its stochastic characteristics, climate factors, unit maintenance, and grid power construction and decommissioning in the function space. Only the effective structure, which still possesses high variance and high uncertainty in an information theory sense, is retained, achieving unsupervised linear feature extraction from high-dimensional sample matrices. Compared to traditional PCA, kernel PCA, or deep autoencoders, this method maintains the overall linear projection form and lower computational complexity while explicitly using function families and orthogonalization processes to characterize redundant structures and provides theoretical guarantees such as upper bounds on conditional entropy. This makes the dimensionality reduction results more physically interpretable and controllable, suitable for engineering applications of large-scale day-ahead market samples.

[0020] 2. This invention utilizes GS function selection (GFS) and GS feature analysis (GFA) to comprehensively employ the joint criterion of maximum residual variance and function reconstructability at both the residual space and original feature space levels. It automatically selects key original features with high information content and difficult to be approximated by other feature functions from the perspectives of data information content and redundant structure, forming a low-dimensional, high-quality, and low-redundancy feature subset. This feature subset, integrated with extracted linear features, is used as input for preprocessing the nuclear norm clustering of the spot market decision sample matrix and for constructing typical scenarios. This significantly reduces the dimensionality and noise interference of clustering and subsequent prediction optimization models, improving the clustering's ability to identify typical operating scenarios and abnormal conditions (such as extreme weather, concentrated maintenance, and sudden disturbances). Furthermore, it reduces the workload of modelers relying on experience to manually select features and classify scenarios, improving the efficiency and robustness of business processes such as spot market analysis, pricing strategy optimization, and flexibility assessment. Attached Figure Description

[0021] The accompanying drawings, which form part of this embodiment, are used to provide a further understanding of this embodiment. The illustrative embodiments and their descriptions are used to explain this embodiment and do not constitute an improper limitation of this embodiment.

[0022] Figure 1This is a flowchart of the method in Embodiment 1 of the present invention. Detailed Implementation

[0023] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0025] Example 1: In the day-ahead spot market of the power grid, dispatching agencies and market operators need to consider a large number of time-varying factors for clearing and decision-making. Inputs include, but are not limited to, multi-source loads, uncertain power source output and stochastic characteristics, probabilistic correlations between uncertain factors, independent energy storage, virtual power plants, climate, unit maintenance plans, grid power source construction, and power source decommissioning, among other medium- and long-term information. Multi-source loads include total grid load, regional loads, and industry loads; uncertain power source output and stochastic characteristics refer to the predicted output of renewable energy sources such as wind and solar power and their prediction error distribution; probabilistic correlations between uncertain factors include the correlation between wind and solar power, and between temperature and air conditioning load; independent energy storage includes energy storage power / energy constraints and charge / discharge plans; virtual power plants represent the integrated adjustability of aggregated users and distributed power sources; and climate includes temperature, humidity, wind speed, and solar radiation. Outputs include the day-ahead spot market decision results, such as thermal power unit output, wind / solar power output, independent energy storage charge / discharge power, and virtual power plant output.

[0026] Considering the various time-varying factors mentioned above, it is necessary to analyze their impact mechanisms on the output of thermal power, wind power, photovoltaic power, independent energy storage, and virtual power plants. The data characteristics of the sample matrix sequence are studied from the perspectives of data structure, magnitude, dimensions, and dimensionality. Based on statistical indicators, feature selection and dimensionality compression methods are designed to reduce feature redundancy and noise while preserving key information, thereby improving the computational efficiency and stability of subsequent clustering, prediction, and optimization.

[0027] The day-ahead market decision samples are often in a high-dimensional, nonlinearly correlated state. Traditional methods such as Principal Component Analysis (PCA) and linear discriminant analysis can only eliminate linearly correlated redundancy, but are difficult to effectively identify and remove nonlinear redundant features. While nonlinear methods such as Kernel Principal Component Analysis (KPCA) and deep autoencoders have stronger expressive power, they suffer from poor interpretability, hyperparameter sensitivity, and high training costs. Furthermore, if high-value features are not fully extracted and redundant features are not removed under unsupervised conditions before performing kernel norm clustering and typical scenario extraction on the sample matrix, the clustering results will be sensitive to noisy features, leading to the curse of dimensionality in the high-dimensional feature space, deteriorating clustering performance, and significantly increasing model training time and storage costs.

[0028] To address at least one of the aforementioned problems, this embodiment provides a method for processing day-ahead market decision samples based on unsupervised feature extraction. Combining a historical data sample matrix of day-ahead spot market decisions, unsupervised feature extraction and selection are performed based on Gram-Schmidt (GS) function orthogonalization. This can serve as a preprocessing step for the day-ahead spot market decision sample matrix before it enters the nuclear norm clustering, typical scenario construction, and prediction optimization model. Figure 1 As shown, the method includes: S101. Modeling of Spot Day-ahead Market Decision Sample Vectors and Random Vectors: Historical decision-making data from the day-ahead spot market is collected and vectorized for modeling. First, the time granularity and statistical period of the day-ahead spot market decision-making samples are determined. Taking the day-ahead market as an example, the granularity can be based on days, treating each day-ahead clearing day as a sample; alternatively, it can be based on time periods, treating each rolling optimization window as a sample. The historical time range of the samples can be selected from the past 1 to 3 years, ensuring sufficient sample size while covering various climatic conditions, load levels, and unit maintenance statuses.

[0029] S1011. Construct a high-dimensional sample vector for each previous market operation day (or each rolling optimization period), assuming: ; in, It represents an abstract high-dimensional random vector used to uniformly characterize all input and output quantities of a day-ahead market operation scenario (such as a day-ahead clearing day); Indicates the first Each random variable corresponds to a specific feature, such as the total grid load at a certain time period, the predicted output of a certain wind farm, the predicted output of a certain photovoltaic station, the maintenance status of a certain unit (0 / 1), the commissioning status of a certain line, the output of a certain virtual power plant, the charging power of a certain independent energy storage, etc. It represents the total dimension of features, including the number of features such as multi-source loads, uncertain power output, probabilistic correlation indicators, independent energy storage status, virtual power plant parameters, climate variables, unit maintenance and grid power construction, decommissioning status, and the total output of thermal power, wind power, photovoltaic, energy storage and / or virtual power plants.

[0030] S1012. Organize the historical data to obtain a sample set: ; in, Indicates the first The feature vector of each historical sample is a random vector. One observation; Indicates the first The sample at the th The specific numerical value of the feature; This indicates the number of historical samples, such as the number of days before the clearing date in history or the number of historical rolling windows.

[0031] In this embodiment, in order to construct the expectation and covariance in subsequent steps, all forms are... In practical calculations, the mathematical expectation is approximated using the sample mean: ; in, Represents random variables Mathematical expectation of the sample distribution; Indicates the first random variables in a sample The value of ; This represents the number of historical samples.

[0032] In step S101, the complex historical data of the spot market before the next trading day is uniformly abstracted into a random vector. The sample observation set provides a unified data carrier for subsequent Gram-Schmidt orthogonalization and the construction of alternative covariance matrices in the function space, and also lays the foundation for comprehensive analysis of multi-source information such as load, wind and solar, energy storage, virtual power plants, maintenance, and decommissioning.

[0033] S102, Construction of Function Families and Definition of Function Space Inner Product: Optionally, around the random vector constructed in step S101 We establish a family of functions to characterize the nonlinear relationships between features, and define the inner product of the function space based on this, providing a mathematical basis for the subsequent Gram-Schmidt orthogonalization.

[0034] S1021. Define a formal variable vector: ; in, Represents a set of formal variables used to construct function expressions at the symbolic level; Indicates the first These are formal variables that will be replaced by their corresponding random variables in the function in the future. replace; This represents the total dimension of the features.

[0035] S1022. Define a family of linearly independent functions: ; in, Represents a set of functions, containing One function; Indicates the first A function, which can be a single-variable function (such as...) ), multivariate linear functions, multivariate polynomials, and product functions (such as...) Examples of nonlinear relationships between load and temperature, wind speed and wind power output, etc., are used to describe nonlinear relationships between load and temperature, wind speed and wind power output, etc. The number of functions can be selected by the user or system designer as needed, which helps to balance expressive power and computational complexity.

[0036] S1022. Define an arbitrary set of indices as follows: ; in, Represented by formal variables The middle belongs to the index set The subvector composed of the components; Representing a family of functions China only depends on A subset of functions, i.e., all functions that satisfy The function.

[0037] S1024, When using random vectors Replacement of formal variables When this happens, we obtain a family of random functions: ; in, Represents a random vector The middle belongs to the index set The subvector composed of the components; Indicates random variables Substitute into function The random variable obtained afterwards; Let represent the set of these random functions, where each random function has a finite second moment, i.e., belongs to . .

[0038] S1025, in random vector distribution In the corresponding function space, define the inner product to form the Hilbert function space: ; in, This represents two random functions, which can come from... Or a linear combination thereof; Representation function AND function exist Inner product in space; Represents random variables and The mathematical expectation of the product of .

[0039] In step S102, the original features are mapped to the function space, and the correlation between functions is characterized by the inner product structure. This provides a basis for the subsequent Gram-Schmidt orthogonalization to eliminate redundancy caused by the nonlinear relationship between multiple sources such as load, wind and solar power, energy storage, and climate. It also provides a unified mathematical framework for unsupervised feature extraction and feature selection.

[0040] S103. Function orthogonalization and residual vector construction based on Gram–Schmidt: For a family of random functions in the function space Perform Gram-Schmidt orthogonalization to generate a set of orthonormal functions. And calculate the residual random vector based on this set. Thus, an alternative covariance matrix is ​​constructed. In this step, the function family defined in step S102 is... ,exist Perform Gram-Schmidt orthogonalization in the space to construct a set of orthogonal normalized functions.

[0041] S1031, Let the first Before this iteration, several random variables have been selected: ; in, Indicates the first A constructed random variable may be a component of the original feature or it may be a random vector. A linear combination in a certain direction; This indicates the current iteration step.

[0042] S103.2, Based on these variables, by modifying the function family... By performing Gram-Schmidt orthogonalization, we can obtain a set of orthonormal functions: ; in, This represents the pre-constructed set of orthogonal normalized functions; Represents a family of functions The set of functions obtained after Gram-Schmidt orthogonalization and normalization; Indicates the first A set of orthogonal normal functions that satisfy and (when hour); This indicates the number of orthogonal functions.

[0043] For a new function In the current set of independent variables The following random function is constructed: ; Then from the already obtained set of orthogonal functions Remove projection from the middle: ; in, Represents a random function With random functions The inner product; Represents a random function In the current orthogonal subspace Orthogonal projection on.

[0044] S1033, then normalization is performed: ; in, Represents random variables The second moment; This represents the root mean square of the random variable, used to represent the random function. Normalized to the unit norm.

[0045] Normalized random function Add to the set of orthogonal normal functions In this process, a new set of orthogonal functions is formed.

[0046] S1034, Obtaining the set of orthogonal functions Then, define the residual random vector: ; in, Indicates the first The residual vector in the step state; Represents a random vector With random functions The covariance vector, its th Each component is ; In function space China Regression in the least squares sense, i.e. In the Projection in Zhang Cheng's function subspace.

[0047] S1035. Based on the residual vector, define the alternative covariance matrix: ; in, Indicates the first The alternative covariance matrix during step iteration; The outer product of the residual vectors is represented by the first term. Yuanwei ; It represents the term-by-term mathematical expectation of the product matrix.

[0048] In step S103, Gram-Schmidt orthogonalization is used to remove the portion of the function space that can be explained by the nonlinear relationships between factors such as multivariate loads, uncertain power sources, climate, and maintenance status. Only the residual structure, which still has considerable uncertainty, is retained, and the covariance matrix is ​​replaced. Quantify this information that has not yet been explained.

[0049] S104, GS function reduction (GFR) for unsupervised feature extraction: In this step, the alternative covariance matrix obtained in step S103 is used... The GS function reduction (GFR) algorithm is used for iterative feature extraction to obtain new linear features under unsupervised conditions, which are used to compactly represent the key information of the spot market decision sample matrix.

[0050] S1041. First, define the initial covariance matrix: ; in, Represents the original random vector The covariance matrix; Represents a random vector The outer product matrix with itself, its th Yuanwei ; This represents the matrix obtained by taking the mathematical expectation of each term in the product matrix.

[0051] S1042, in the In the next iteration, the following maximum eigenvalue variable is solved: ; in Represents a candidate direction vector; express The Euclidean norm is 1, that is... ; Represents a random vector In direction variance on (in the first) (in the sense of replacing the covariance matrix) Indicates the first The optimal direction vector obtained in the second iteration corresponds to The direction of the largest eigenvalue.

[0052] Set variance threshold ,like If no direction with variance greater than the threshold exists in the current alternative covariance matrix, the feature extraction process is stopped. Indicates along direction The residual variance; The square of the variance threshold controls the amount of remaining unexplained information.

[0053] If the above conditions are not met, then define a new linear characteristic: ; in Indicates the first The extracted new features; Represents a random vector In direction The projection on, i.e. .

[0054] New linear features Add to the variable set, update the orthogonal function set and residual vector. Alternative covariance matrix Then return to the search step in the direction of the largest feature and continue iterating.

[0055] In the case of discrete characteristics, if the random vector Since each component has a finite number of values, the feature vector output by GFR is: ; The following conditional upper bound for entropy is satisfied: ; in, Indicates the extraction of features given Under the condition, the original random vector Conditional entropy; Indicates the original feature dimension; Indicates threshold The sign of the order of magnitude that is linearly proportional indicates that when When the conditional entropy is small enough, it is on the same order of magnitude as... For terms of the same order, the value approaches zero.

[0056] Through step S104, GFR continuously seeks new high-variance directions as it gradually removes nonlinear redundancy. This results in a small number of linearly combined features that are rich in information. These features, in an information theory sense, can almost determine the original spot market decision sample before the future, thus achieving efficient unsupervised feature extraction.

[0057] S105, GS component analysis (GCA) is based on redundancy elimination of principal components: In this step, for the principal component set obtained by traditional principal component analysis (PCA), the GS component analysis (GCA) method is used to identify and remove redundant principal components that can be nonlinearly represented by other principal components, so as to avoid retaining a large number of principal components with the same function after dimensionality reduction.

[0058] S1051, Regarding the initial covariance matrix Perform eigenvalue decomposition to obtain eigenvectors: ; in, Indicates the first The principal direction corresponds to the first principal direction of the covariance matrix. Large eigenvalues; eigenvalues ​​are sorted from largest to smallest such that the variances of the corresponding principal components satisfy... .

[0059] S1052. Definition of principal components in the traditional sense: ; in, Indicates the first Each principal component is a linear combination of the original variables; Represents a random vector In the eigenvector Projection in the direction.

[0060] To characterize the redundancy relationships between principal components, a redundancy function is defined: ; If a constant exists , making Then it is said that it exists in the principal component space. - Redundancy .in, This indicates that the redundant function is in the principal component vector. The value of ; This represents the second moment of the function on the data; The upper bound representing the degree of redundancy, when When the value is small, it indicates that the data points are highly concentrated. The described relationship is near.

[0061] S1053. If there are redundant relationships such that a certain principal component can be represented by almost all other principal component functions, for example, if there is an index... with subset , so that: ; Furthermore, this function can be used by a family of functions. A good approximation is one in which a projection exists: ; satisfy: ; in, Indicates from arrive The function; Indicates a set of indices The sub-vectors formed by concatenating the principal components in the vector; Indicates a family of functions in variables The set of functions after orthogonalization; The variance represents the projection error; This indicates the tolerance threshold.

[0062] Under the above conditions, GCA can be considered as the principal component. For the selected principal component set It is redundant, so it is skipped during the principal component screening process, and only those that cannot pass the function family are retained. Principal components reconstructed from other principal components.

[0063] Through step S105, the principal components obtained by traditional PCA, after being processed by GCA, will have a large number of redundant components that can be reconstructed by nonlinear combination of other principal components. This makes the remaining principal component set retain both directions with large variance and reduced hidden nonlinear redundancy, making it more suitable for describing independent patterns in the spot market decision sample matrix and providing a clearer subspace structure for subsequent clustering and feature analysis.

[0064] S106, GS function selection (GFS) is based on the selection of original features according to residual variance: In this step, alternative variance vectors are constructed using the variances of each dimension of the residual random vector in step S103. Unsupervised selection of original features is performed based on the residual information content, thereby obtaining a set of key original features that still have high information content after the nonlinear structure is stripped away.

[0065] S1061. Define the squares of the components of the residual vector (Hadamard power): ; in, Indicates the first The residual random vector of the step iteration; The residual vector is represented at the th... Components on the dimension; This represents the vector obtained by squaring each dimension of the residual vector. This represents the Hadamard power based on the square of the elements.

[0066] S1062. Define the alternative variance vector: ; in, Indicates the first The alternative variance vector during step iteration; Indicates the first The residual vector at the first step The variance of the dimension; The residual vector is represented at the th... The components of a dimension.

[0067] Initially, the first step of replacing the variance vector is: ; in, It is the vector obtained by squaring each dimension of the original feature vector.

[0068] In the In each iteration, the feature index with the largest residual variance is selected: ; in, The substitution variance vector represents the first... Dimensional components; Represents the set of all feature indices; Indicates the first The feature index selected in the next iteration corresponds to the feature with the largest residual variance.

[0069] S1063. Define the infinite norm: ; in, This represents the maximum component value of the vector.

[0070] like If the variance of all features in the current residual space does not exceed the threshold, the feature selection process ends.

[0071] If the above conditions are not met, then select the index. The corresponding original features: ; in, Represents the original vector The Each component corresponds to a specific characteristic, such as the power output of a wind farm, the power output of an important unit, or the power flow of a critical line. The newly selected random variable will be included in the variable set for Gram-Schmidt orthogonalization update.

[0072] Repeat the above process until the stopping condition is met to obtain the final set of selected feature indices: ; in, Represents the set of original feature indices that were selected; This indicates the number of iterations at the end of the feature selection process, i.e., the number of features selected.

[0073] In the case of discrete features with finite values, the feature set output by GFS is... Satisfies the upper bound of conditional entropy: ; in, This indicates that only the index belongs to... eigenvectors; This represents the conditional entropy of the original feature vector given a selected subset of features.

[0074] In step S106, GFS automatically selects the original features with the largest residual variance from the residual space, which has already partially removed nonlinear redundancy, to obtain a set of key variables with high information content that are difficult to be nonlinearly reconstructed by other features. These variables often correspond to key power sources, key network channels, and typical load indicators in the spot market scenario, which helps to significantly reduce the dimensionality of the sample matrix while retaining the main information, providing a concise and effective feature set for subsequent clustering and decision optimization.

[0075] S107, GS Feature Analysis (GFA) Redundancy Structure Analysis in the Original Feature Space: In this step, the GS Feature Analysis (GFA) method is used to directly analyze redundant structures in the original feature space, that is, to determine whether certain features can be approximated by functions of other features, thereby further guiding feature selection.

[0076] S1071, Order: ; The meaning of each component is the same as in step S101.

[0077] Suppose there exists a measurable function: ; and parameters ,like Then it is assumed that there exists in the original feature space. - Redundancy .in, Representation function In random vectors The value of ; This represents the second moment of the redundant function on the data; This represents the upper bound of the degree of redundancy.

[0078] S1072, If an indicator exists with subset and functions , so that: ; And a family of functions exists. Projection: ; satisfy: ; in, Indicates index by feature Subvectors formed by various units, such as sub-feature sets composed of the output of several generating units and the output of several wind power units; Indicates from arrive A function can be derived from a family of functions. Approximate representation; Indicates in variable The set of functions after orthogonalization; This indicates the use of a family of functions to analyze features. The mean square error after approximation; This indicates the allowable error threshold.

[0079] When the above conditions are met, the feature can be considered... In the mean square sense, it can be determined by the characteristics The function representation of these features is redundant. During feature analysis and selection, such features can be marked as redundant to avoid duplication with the key features selected by GFS.

[0080] Through step S107, GFA analyzes the functional relationships between the original features from the perspective of redundant structure, and identifies those features that can be approximated by combinations of other variables in the multi-source data of the spot market before the spot date. This makes the final retained feature set not only highly informative but also less redundant, providing a cleaner and more interpretable feature subspace for the subsequent clustering and optimization of the sample matrix of the spot market before the spot date.

[0081] S108. Integration with the spot day-ahead market decision sample matrix and clustering preprocessing workflow: In this step, the features and feature subsets obtained from GFR, GCA, GFS, and GFA are fused with the spot market decision sample matrix structure to provide high-quality input for subsequent clustering preprocessing based on nuclear norm.

[0082] Let the original sample matrix be: ; in, Indicates the first The feature vector of each sample, and the feature vector in step S101 Consistent; Let the matrix consist of all historical samples, and let its i-th... Behavior No. The nth sample, the th Listed as the number One characteristic.

[0083] The linear extraction features obtained in step S104 This corresponds to a new feature matrix at the sample level: ; in, Indicates the first The values ​​of each sample in the extracted features; This indicates the number of features extracted.

[0084] The original feature index set obtained through steps S106 and S107: ; It can be used to construct the original eigenvalue submatrix: ; in, This means retaining only items belonging to the set on the column index. The column, that is, only retains the original key features selected.

[0085] Depending on the specific application requirements, the following two integration methods can be adopted: Use only As input for subsequent nuclear norm clustering, the expressive power of abstract factors is emphasized; And, will and The features are then concatenated to form an enhanced feature matrix, which maintains interpretability while also taking into account the ability to express abstract concepts.

[0086] The resulting dimensionality-reduced feature matrix can be directly input into the nuclear norm clustering preprocessing process of the spot day-ahead market decision sample matrix, realizing the overall technical route of "first using the GS method for unsupervised feature extraction and selection, and then using nuclear norm high-dimensional clustering for sample matrix clustering preprocessing". This significantly improves the efficiency, robustness and interpretability of spot day-ahead market data processing in engineering practice.

[0087] This embodiment performs unsupervised feature extraction and selection on the spot market decision sample matrix without relying on label information, effectively identifying and eliminating nonlinear redundancy. While retaining key statistical information and typical operating modes, it significantly reduces the dimensionality of the sample matrix, providing high-quality low-dimensional data representation for subsequent nuclear norm clustering, typical scenario construction, and prediction model training. It provides entropy constraints and redundancy elimination theoretical guarantees in the sense of information theory and probability statistics, making the feature extraction / selection process interpretable and quantifiable.

[0088] This embodiment constructs a series of alternative covariance matrices based on Gram–Schmidt orthogonalization of the function space, iteratively extracts new high-variance directions, achieves unsupervised feature extraction, and provides conditional entropy. Boundedness is guaranteed; based on the principal components, a family of functions is used to eliminate nonlinear redundancy between principal components, and the principal component subsets after "redundancy removal" are output; using alternative variance vectors, unsupervised feature selection is performed directly on the original features, selecting features with high variance that are difficult to be nonlinearly represented by the selected features; based on the GCA approach, redundancy is directly eliminated on the original features, realizing feature selection from the perspective of redundant structure.

[0089] Example 2: In this embodiment, the spot-day market operation data of a certain regional power grid for the past three years are selected as the research object to illustrate and verify the method of the present invention. The renewable energy penetration rate of this regional power grid is relatively high, with wind power and photovoltaic installed capacity accounting for more than 35% of the total installed capacity, and independent energy storage and virtual power plants have participated in the spot-day market clearing.

[0090] S201. Time range and sample size: The historical data time range can be selected from January 1, 2022 to December 31, 2024; using days as the sample granularity, after removing days with severely missing data, the number of samples before the valid date is obtained: .in, This represents the total number of samples, i.e., the number of days before the effective date.

[0091] S202, Feature Dimension Composition: For each day-ahead sample, the input and output quantities are uniformly modeled as random vectors. in, For the first One feature; The overall feature dimensions are specifically composed of: Multi-dimensional load features: including time-sharing total grid load, regional load, and daily peak-valley values, totaling 60 dimensions; Uncertain power source features: wind power / solar power predicted output, prediction error statistics (mean, variance, confidence interval, etc.), totaling 80 dimensions; Meteorological and climate features: temperature, humidity, wind speed, irradiance, and extreme weather indicators, totaling 30 dimensions; Unit and grid status features: unit maintenance status, line commissioning status, power source decommissioning indicators, etc., totaling 20 dimensions; Output decision features: time-sharing output or aggregated indicators of thermal power, wind power, solar power, independent energy storage, and virtual power plants, totaling 110 dimensions.

[0092] In summary, the various feature dimensions are superimposed: That is, each daily sample is a 300-dimensional feature vector.

[0093] S203, Sample matrix representation: Stack all samples row-wise to obtain the sample matrix: .in, This is a sample matrix of day-ahead market decisions; Indicates the number of samples; Indicates the number of feature dimensions; the first Behavior No. The feature vector of the sample from the previous day, the th Listed as the number One characteristic.

[0094] S204. To verify the effectiveness of the method of the present invention, the scheme of the present invention is compared with a variety of common dimensionality reduction or feature selection methods, and the following four schemes are set: Option A: Baseline without dimensionality reduction: Directly use the original 300-dimensional feature vector Subsequent nuclear norm clustering and modeling are performed; no feature extraction or feature selection is performed; this serves as a benchmark for performance and computational complexity.

[0095] Option B: Traditional Principal Component Analysis (PCA) Dimensionality Reduction Solution: For the sample matrix Standardization is performed; eigenvalue decomposition is performed on the covariance matrix to obtain principal components; the first... One principal component, resulting in a cumulative variance contribution rate of 95%; in this example, we obtain Subsequent clustering was performed in a 40-dimensional principal component space.

[0096] Option C: Autoencoder (AE) nonlinear dimensionality reduction scheme: A three-layer symmetric autoencoder is constructed: a 300-dimensional input layer, a 64-dimensional hidden layer, and a 32-dimensional bottleneck layer, followed by symmetric decoding. Unsupervised training is performed using 1,050 samples, with the goal of minimizing the reconstruction error. The 32-dimensional features of the bottleneck layer are taken as the dimensionality-reduced representation for subsequent clustering.

[0097] Scheme D: The present invention's unsupervised feature extraction and feature selection scheme based on GS: Construct a family of constructors based on the sample matrix. This includes univariate functions and second-order product functions; several linear features are extracted using GS function reduction (GFR) until the residual variance threshold is reached. The example results show the extraction of linear feature numbers. In the residual space, the Gross-Free Search (GFS) function is used to select the original features, and a threshold is applied. : Obtain the number of key original features The 25-dimensional linear feature is concatenated with the original 35-dimensional feature to form a dimension of: The comprehensive feature vector is used as the input for subsequent nuclear norm clustering.

[0098] S205. Evaluation Indicators and Numerical Norm Clustering Settings: Clustering Algorithm Configuration: Four schemes are compared within the same high-dimensional clustering framework with the same nuclear norm. Let the number of clusters be: It is proposed to divide the historical days into eight typical operating scenarios (such as summer high load days, spring high wind days, autumn high photovoltaic days, low load maintenance days, etc.).

[0099] Clustering evaluation metrics: Sum of Squared Errors within clusters (SSE, Sum of Squared Errors): ; in, For the first A sample set of each cluster; For the first The vector of each sample in the reduced-dimensional feature space; For clusters The center vector; It is the Euclidean norm; It is the sum of the squared errors within all clusters; the smaller the sum, the better the cluster compactness.

[0100] Silhouette Coefficient: Defined for each sample: ; in, For the sample Average distance from other samples in the same cluster; For the sample The average distance to the nearest other cluster samples; For the sample The contour coefficient, range of values The closer the value is to 1, the better the clustering effect.

[0101] The global contour coefficient is defined as: ; in, The average silhouette coefficient is used to evaluate the overall clustering effect.

[0102] Predictive performance metrics based on clustering results: To examine the impact of clustering on subsequent models, each cluster was used as a scene label, and daily load prediction models were trained separately under each scheme (e.g., training linear regression or lightweight machine learning models by cluster), and the root mean square error of the prediction was examined. ; in, For the first The actual daily total load of each sample; For the first Each sample corresponds to the predicted daily total load value of the prediction model; This represents the root mean square error of daily load forecast; the smaller the value, the better the forecast.

[0103] S206. Comparison of example results and data: S2061. Comparison of dimensionality reduction effect and computational cost: Table 1 shows a comparison of the four schemes in terms of feature dimension and clustering computation time (time is the average running time under the same hardware environment): Table 1. Comparison of the four schemes in terms of feature dimension and clustering computation time.

[0104] As can be seen from Table 1, the dimension of the present invention scheme D is reduced from 300 to 60, with a dimension reduction rate of 80%, and the clustering running time is about 38% of that of the benchmark scheme A; although the dimensions of PCA and autoencoder are even lower.

[0105] S2062, Clustering Quality Comparison: Under the same nuclear norm clustering parameters, the SSE and average silhouette coefficient obtained by each scheme are as follows. As shown in Table 2: Table 2 SSE and Mean Profile Coefficient

[0106] For ease of comparison, SSE is normalized to 1.00 for Scheme A, with the remaining schemes given proportionally. It can be seen that Scheme D reduces SSE by approximately 28% compared to the baseline Scheme A, approximately 17% compared to the PCA scheme, and approximately 13% compared to the autoencoder scheme. Regarding the profile coefficient, Scheme D has an average profile coefficient of 0.58, significantly higher than other schemes, indicating more compact clusters, greater separation between clusters, and a clearer clustering structure. From a business perspective, the eight clusters obtained using Scheme D exhibit typical daily load curves and wind / solar output patterns that better align with dispatchers' experience: several high-wind, high-solar-solar combination scenarios appear separately; high-load, extreme-temperature scenarios are clustered into a separate category; and days with concentrated maintenance and tight available unit capacity also form clusters with distinct physical characteristics.

[0107] S2063. Impact on subsequent prediction models: Based on the clustering results of the four methods, cluster labels were used as scene categories. Simple daily total load prediction models (e.g., linear regression or lightweight gradient boosting trees) were trained within each cluster, and the overall RMSE was calculated. The results are shown in Table 3. Table 3 Overall RMSE Statistics

[0108] As can be seen, compared with the baseline scheme A without dimensionality reduction, the prediction RMSE of the present invention scheme D is reduced from 310MW to 252MW, a reduction of about 18.7%, indicating that the feature space obtained by GS feature extraction and selection can better support the construction of intra-cluster prediction models. Compared with PCA and autoencoder schemes, the present invention scheme provides better prediction performance while maintaining a lower dimensionality, which is consistent with the design of the present invention that can explicitly remove nonlinear redundancy and select physically interpretable key features.

[0109] S2064, Abnormal scene recognition capability: Based on nuclear norm clustering, this example also identifies and statistically analyzes anomalous scenarios (such as extreme cold waves, heat waves, and large-scale maintenance days). After marking samples whose weighted distance from the nearest cluster center exceeds a threshold and whose membership degrees are all low as anomalous samples, the number of anomalous scenarios identified by each scheme is shown in Table 4. Table 4 Number of Abnormal Scenarios

[0110] As can be seen, the number of abnormal samples identified by the present invention's scheme D is closer to the actual abnormal scenarios confirmed by humans, with the lowest false alarm rate; the accuracy rate reaches 81.3%, which is about 8 and 6.6 percentage points higher than the traditional PCA and autoencoder schemes, respectively, which is conducive to operation and maintenance personnel focusing on extreme and risky scenarios.

[0111] Based on the above examples, the application of the Gram-Schmidt-based unsupervised feature extraction and selection method in the actual spot market decision sample matrix demonstrates that: With the dimension reduced from 300 to 60, this invention not only significantly reduces the computation time of nuclear norm clustering, but also outperforms the original features, PCA, and autoencoders in clustering metrics such as intra-cluster sum of squares and silhouette coefficient. In daily load forecasting tasks using clusters as scene labels, the proposed solution achieves a significantly lower RMSE, substantially improving the subsequent prediction model. Regarding anomaly day identification, this solution maintains a moderate number of identifications while improving the identification rate and accuracy of truly abnormal scenarios, better meeting the safety and economic requirements of power system operation. Therefore, the results of the examples demonstrate that this invention possesses comprehensive advantages in unsupervised feature extraction and selection of the spot market decision sample matrix before the spot market date, including dimensionality reduction efficiency, clustering quality, improved prediction performance, and enhanced anomaly identification capabilities, making it highly valuable for engineering applications.

[0112] Example 3: This embodiment provides a day-ahead market decision sample processing system based on unsupervised feature extraction, including: The data acquisition module is configured to: acquire current day market decision samples and construct random vectors based on the current day market decision samples; The random function family building module is configured to: obtain a random function family based on a random vector and a preset linearly independent function family; The alternative covariance matrix establishment module is configured to: orthogonalize the family of random functions in the function space, generate a set of orthogonal normalized functions, and determine the residual random vector based on the set of orthogonal normalized functions to obtain the alternative covariance matrix; The feature extraction module is configured to: perform iterative feature extraction using a function reduction algorithm based on the alternative covariance matrix to obtain new linear features under unsupervised conditions, which are used to characterize the key information of the spot market decision sample matrix; identify and remove redundant principal components that can be nonlinearly represented by other principal components using component analysis; perform unsupervised selection of original features based on residual information analysis to obtain key original features; and analyze the functional relationships between original features based on feature analysis methods to identify features that can be approximated by combinations of other variables in the multi-source data of the spot market before the market, thereby determining the final set of retained features. The dimensionality reduction module is configured to fuse the features and feature subsets obtained by function reduction algorithm, component analysis method, residual information analysis and feature analysis method with the preset spot day-ahead market decision sample matrix to obtain the dimensionality-reduced feature matrix; The clustering module is configured to perform clustering on the dimensionality-reduced feature matrix.

[0113] The working method of the system is the same as that of the day-ahead market decision sample processing method based on unsupervised feature extraction in Example 1, and will not be repeated here.

[0114] Example 4: This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the day-ahead market decision sample processing method based on unsupervised feature extraction described in Embodiment 1.

[0115] Example 5: This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. When the processor executes the program, it implements the steps of the day-ahead market decision sample processing method based on unsupervised feature extraction described in Embodiment 1.

[0116] Example 6: This embodiment provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the steps of the day-ahead market decision sample processing method based on unsupervised feature extraction described in Embodiment 1.

[0117] The above description is merely a preferred embodiment of this practice and is not intended to limit the scope of this practice. Various modifications and variations can be made to this practice by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this practice should be included within the protection scope of this practice.

Claims

1. A day-ahead market decision sample processing method based on unsupervised feature extraction, characterized in that, The method comprises the following steps: obtaining a day-ahead market decision sample, and constructing a random vector according to the day-ahead market decision sample; obtaining a random function family according to the random vector and a preset linearly independent function family; orthogonalizing the random function family in a function space to generate an orthogonal normalized function set, and determining a residual random vector based on the orthogonal normalized function set to obtain a substitute covariance matrix; performing iterative feature extraction by using a function reduction algorithm according to the substitute covariance matrix, and obtaining new linear features under an unsupervised condition, which are used to represent key information of the spot day-ahead market decision sample matrix; identifying and removing redundant principal components that can be nonlinearly represented by other principal components by using a component analysis method; performing unsupervised selection of original features based on residual information quantity analysis to obtain key original features; analyzing function relationships among the original features based on a feature analysis method, identifying features that can be approximately represented by combinations of other variables in the spot day-ahead market multi-source data, and determining a final retained feature set; fusing the features and feature subsets obtained by the function reduction algorithm, the component analysis method, the residual information quantity analysis and the feature analysis method with a preset spot day-ahead market decision sample matrix to obtain a dimension-reduced feature matrix; performing clustering processing on the dimension-reduced feature matrix.

2. The unsupervised feature extraction based day-ahead market decision sample processing method of claim 1, wherein, The construction of the sample vector and the random vector comprises the following steps: constructing each day-ahead clearing day as a sample, and uniformly representing multivariate load, uncertain power output and random characteristics, probability correlation among uncertain factors, independent energy storage, virtual power plant, climate, unit maintenance plan, power grid power supply construction and power supply retirement input, and thermal power, wind power, photovoltaic power, independent energy storage and virtual power plant output as a random vector.

3. The unsupervised feature extraction based day-ahead market decision sample processing method of claim 1, wherein, The random function family is obtained according to the random vector and the preset linearly independent function family, which comprises the following steps: ； wherein, denotes a set of formal variables; denotes the thformal variable, which is replaced in the function by the corresponding random variable x; denotes the total dimension of the feature; defining a form variable vector: ； wherein represents a set of functions; represents the function; represents the number of functions; defining a set of linearly independent functions: ； wherein, denotes a subvector consisting of the components of that belong to the index set ; denotes the subset of the function family that only depends on , i.e. all functions that satisfy ; When the random vector is replaced by the form variable a family of random functions is obtained: ； wherein, denotes a random vector belonging to the index set consisting of components of denotes the random variable obtained after substituting the random variable into the function denotes the set of these random functions.

4. The method for day-ahead market decision sample processing based on unsupervised feature extraction of claim 3, wherein, In the random vector distribution In the corresponding function space, define the inner product: ； wherein, denotes two random functions; denotes the function denotes the function denotes the inner product in space; denotes the random variable denotes the mathematical expectation of the product of and 5. The unsupervised feature extraction based day-ahead market decision sample processing method of claim 1, wherein, defining an arbitrary index set as:

6. The unsupervised feature extraction based day-ahead market decision sample processing method of claim 1, wherein, The substitute covariance matrix is determined, which comprises the following steps: selecting a plurality of random variables; obtaining a set of orthogonal normalized functions by performing Gram-Schmidt orthogonalization on the function family based on the plurality of random variables; constructing a random function under the current independent variable set for the new function; then removing the projection from the obtained orthogonal function set; then normalizing; and adding the normalized random function to the orthogonal normalized function set to form a new orthogonal function set. ； ； where denotes the residual vector at step ; denotes a random vector with covariance vector ; denotes the regression of in the least squares sense in the function space ; denotes the surrogate covariance matrix at iteration step ; denotes the outer product of the residual vector ; denotes the entry-wise mathematical expectation of the outer product matrix.

7. The unsupervised feature extraction based day-ahead market decision sample processing method of claim 1, wherein, Iterative feature extraction using function reduction algorithms includes: defining the initial covariance matrix; in the... In the next iteration, the largest eigenvalue is determined; ； wherein, denotes a candidate direction vector; denotes the Euclidean norm of denotes a random vector the variance of in direction denotes the optimal direction vector obtained at the th iteration; Setting a variance threshold If , it is considered that there is no longer a direction in the current substitute covariance matrix with variance greater than the threshold, and the feature extraction process is stopped; wherein, represents the residual variance along the direction ; is the square of the variance threshold, which controls the size of the remaining unexplained information; otherwise, a new linear feature is defined; the new linear feature is added to the variable set, the orthogonal function set and the residual vector, the substitute covariance matrix are updated, and the maximum feature direction search step is returned to continue iteration.

8. The unsupervised feature extraction based day-ahead market decision sample processing method of claim 1, wherein, The substitute covariance matrix is:

9. The unsupervised feature extraction based day-ahead market decision sample processing method of claim 1, wherein, The component analysis method is used to identify and remove redundant principal components that can be nonlinearly represented by other principal components, which comprises the following steps: performing eigenvalue decomposition on the initial covariance matrix to obtain an eigenvector; defining traditional principal components and redundant functions; if there is a redundant relationship that enables a principal component to be represented by other principal components, the principal component is considered to be redundant to the selected principal component set, so that the principal component is skipped in the principal component screening process, and only the principal component that cannot be reconstructed on other principal components by the function family is retained. The unsupervised selection of the original features based on the residual information quantity analysis to obtain the key original features comprises the following steps: defining a component square sum residual vector and a substitute variance vector; and defining an infinite norm: ； wherein, represents the maximum component value of a vector; represents the first represents the alternative variance vector at the step iteration; represents the component of the residual vector; whether all features in the current residual space do not exceed the threshold value is determined by the infinite norm, if yes, the feature selection process ends, otherwise, the original feature corresponding to the selected index is selected, and the unsupervised selection is repeated.

10. The unsupervised feature extraction based day-ahead market decision sample processing method of claim 1, wherein, The function relationship between the original features is analyzed based on a feature analysis method, including: ； wherein, represents an abstract high-dimensional random vector; represents the th random variable; represents the total dimension of the features; let there exist a measurable function: ； and parameters , if , then it is considered that there exists -redundant relationships , where denotes the function evaluated on the random vector ; denotes the second moment of the redundant function on the data; denotes an upper bound on the degree of redundancy; if there exist indices and a subset , and a function , such that: ； And there exists a family of functions whose projections: ； satisfying: ； Then it is considered a feature In the mean square sense, it can be determined by the characteristics The functional representation of is a redundant feature; among which, Indicates index by feature Subvectors formed; Indicates from arrive The function; Indicates in variable The set of functions after orthogonalization; This indicates the use of a family of functions to analyze features. The mean square error after approximation; This indicates the allowable error threshold.

11. A day-ahead market decision sample processing system based on unsupervised feature extraction, characterized by, including: The data acquisition module is configured to: acquire a day-ahead market decision sample, and construct a random vector according to the day-ahead market decision sample; The random function family establishment module is configured to: obtain a random function family according to the random vector and a preset linearly independent function family; The substitute covariance matrix establishment module is configured to: orthogonalize the random function family in a function space to generate an orthogonal normalized function set, determine a residual random vector based on the orthogonal normalized function set, and obtain a substitute covariance matrix; The feature extraction module is configured to: perform iterative feature extraction by using a function reduction algorithm according to the substitute covariance matrix, and obtain new linear features under an unsupervised condition, which are used to represent key information of the spot day-ahead market decision sample matrix; Redundant principal components that can be nonlinearly represented by other principal components are identified and removed by using a component analysis method; key original features are obtained through unsupervised selection of the original features based on residual information quantity analysis; The function relationship between the original features is analyzed based on a feature analysis method, identifying features in the spot day-ahead market multi-source data that can be approximately represented by combinations of other variables, thereby determining a final retained feature set; The dimension reduction module is configured to: fuse the features and feature subsets obtained by the function reduction algorithm, the component analysis method, the residual information quantity analysis, and the feature analysis method with a preset spot day-ahead market decision sample matrix to obtain a dimension reduction feature matrix; The clustering processing module is configured to: perform clustering processing on the dimension reduction feature matrix.

12. A computer readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the steps of the day-ahead market decision sample processing method based on unsupervised feature extraction according to any one of claims 1-10.

13. An electronic device comprising a memory, a processor, and a computer program stored on the memory and capable of running on the processor, characterized in that, The processor executes the program to implement the steps of the day-ahead market decision sample processing method based on unsupervised feature extraction according to any one of claims 1-10.

14. A computer program product, characterised in that, The computer program product includes a computer program that, when executed by a processor, implements the steps of the day-ahead market decision sample processing method based on unsupervised feature extraction according to any one of claims 1-10.

Images