A multi-scale feature extraction and association analysis method

Abstract

The present invention discloses a multi-scale feature extraction and association analysis method, which constructs a spatial multi-scale association graph system with multiple hierarchical association graphs, each association graph corresponds to a specific spatial scale, and there is a hierarchical convergence relationship between association graphs at different levels; for each node of the bottom-level association graph, its time series indicator data is extracted, and the time series indicator data is decomposed into periodic, trend and burst components; for association graph nodes other than the bottom-level, convergence features are extracted using statistical convergence and pattern convergence; based on the extracted time series indicator data and the extracted convergence features, the attribute correlation and graph structure correlation of node pairs are calculated in each level, and the two are fused to obtain the intra-level correlation index of the corresponding level; the hierarchical convergence correlation index is defined according to the intra-level correlation index to realize inter-level correlation propagation; a hierarchical attention mechanism is introduced to fuse the correlation information of different levels and calculate the spatial multi-scale correlation.

Description

Technical Field

[0001] The present invention relates to a multi-scale feature extraction and association analysis method, and belongs to the field of data analysis and modeling. Background Art

[0002] In the field of data science and complex system analysis, spatial correlation analysis refers to the study of whether there is statistical interdependence or correlation between different spatial units.

[0003] Spatial correlation occurs when the properties or behavior of one spatial unit are influenced by other spatial units. This correlation exists not only in geographic space (such as economic connections between cities) but also in abstract conceptual space (such as the connections between upstream and downstream enterprises in an industrial chain). With the advent of the big data era, accurately identifying and quantifying this multi-scale spatial correlation has become a key challenge in many fields, with important guiding significance for coordinated regional development, optimized industrial chain layout, and efficient resource allocation.

[0004] At present, the research in the field of multi-scale spatial correlation analysis at home and abroad mainly has the following deficiencies:

[0005] First, existing studies often use single-dimensional or limited-dimensional features when analyzing the correlation between entities, ignoring the comprehensive impact of multi-dimensional features on correlation judgment.

[0006] Second, traditional correlation analysis methods typically operate at a single spatial scale, lacking a cross-scale analytical perspective. However, correlations in real systems often exhibit significant scale-dependence—entities that are unrelated or weakly correlated at one spatial scale may exhibit strong correlations at another. For example, two seemingly unrelated micro-individuals may be closely linked at a higher level, such as an industrial cluster or regional economy.

[0007] Third, existing technologies often use a unified feature definition and extraction method when extracting and representing features at different spatial scales, ignoring the significant differences in the emphasis placed on features of entities at different scales. For example, microscopic individuals may focus more on short-term fluctuations, mesoscopic groups may focus more on cyclical changes, and macroscopic systems may prioritize long-term trends.

[0008] Fourth, existing research lacks exploration of the transmission mechanisms of correlations across spatial scales, making it difficult to effectively integrate cross-scale correlations. For example, micro-individual correlations can, through convergence, influence the correlations between meso-groups, which in turn affect the overall correlations of the macro-system. This lack of transmission mechanisms makes comprehensive cross-scale correlation analysis impossible. Summary of the Invention

[0009] The present invention provides a multi-scale feature extraction and association analysis method in order to solve the problems existing in the above-mentioned prior art. First, the present invention designs differentiated multi-dimensional feature definitions and extraction strategies for different spatial scales to ensure that the feature representation matches the entity scale characteristics. Secondly, by establishing a cross-scale analysis framework, the identification and comparison of correlations at different spatial scales are realized, and the scale-dependent characteristics are effectively captured. Third, the present invention constructs a cross-scale transmission mechanism of correlation, revealing how micro-correlations affect the evolutionary path of meso- and macro-correlations through hierarchical convergence. Finally, when defining correlation, the present invention comprehensively considers the direct correlation of the current spatial scale and the transmission correlation from other scales, so as to obtain a more comprehensive and accurate multi-scale spatial correlation evaluation result.

[0010] The technical solutions adopted in the present invention are:

[0011] A multi-scale feature extraction and association analysis method includes the following steps:

[0012] S1: Construct a spatial multi-scale correlation graph system with multiple hierarchical correlation graphs, each of which corresponds to a specific spatial scale, and there is a hierarchical convergence relationship between different hierarchical correlation graphs;

[0013] S2: For each node in the bottom-level association graph, extract its time series indicator data and decompose it into periodic, trend, and burst components;

[0014] S3: For nodes in the non-bottom-level association graph, statistical aggregation and pattern aggregation are used to extract aggregation features;

[0015] S4: Based on the time series indicator data extracted in S2 and the aggregated features extracted in S3, the attribute correlation and graph structure correlation of node pairs are calculated in each layer, and the two are combined to obtain the intra-layer correlation index of the corresponding layer;

[0016] S5: Define hierarchical aggregation correlation indicators based on intra-layer correlation indicators to achieve inter-layer correlation propagation;

[0017] S6: Introduce a hierarchical attention mechanism to fuse correlation information at different levels and calculate spatial multi-scale correlation.

[0018] Furthermore, S1 specifically includes:

[0019] (1.1) Construct a set of spatial multi-scale correlation graphs:

[0020] , , ,

[0021] in: represents the i-th layer association graph, corresponding to a specific spatial scale; Represents the node set corresponding to the i-th spatial scale, representing the unit individual of this scale A collection of Represents the set of association relationships between individual units; Represents a set of node attributes and describes the multidimensional characteristics of each node;

[0022] (1.2) Establish hierarchical convergence relationships between association graphs at different levels, so that the node set V of the i-th layer i The nodes in the i-1th layer are composed of the node set V i−1 Several nodes in the network are formed by convergence, reflecting the hierarchical progressive relationship of spatial scales.

[0023] Furthermore, S2 specifically includes:

[0024] (2.1) For the bottom-level association graph Each node , extract its time series indicator data:

[0025] ,

[0026] in, represents the index value of the jth node at time t, and T represents the length of the time series;

[0027] (2.2) Decompose the time series indicator data into three-dimensional features:

[0028] ,

[0029] in: It represents the periodic component and reflects the periodic pattern of the indicator data; Represents trend components and captures the long-term trend of indicator data; It represents the sudden component, characterizing sudden events and abnormal patterns in the indicator data;

[0030] (2.3) Three-dimensional feature decomposition is achieved through the following optimization model:

[0031] ,

[0032] in, is the reconstruction loss function, which is used to measure the difference between the original time series data and the reconstructed data after decomposition;

[0033] is a periodic constraint, used to ensure the periodic characteristics of the periodic component;

[0034] is a smoothness constraint, used to ensure the smoothness of the trend component;

[0035] is a sparsity constraint, used to ensure the sparsity of the burst component;

[0036] 、 、 It is a trade-off parameter used to balance the impact of different constraints on the total loss.

[0037] Furthermore, S3 is specifically:

[0038] for Layer nodes , i≥1, assuming By the node set of the lower layer If the convergence is formed, the statistical convergence characteristics are calculated respectively. , pattern convergence characteristics ;

[0039] Statistical aggregation feature calculation:

[0040] ,

[0041] in, , , , , , represent mean, standard deviation, lower quartile, upper quartile, skewness, and kurtosis, respectively; hour Corresponding respectively , in order to calculate the load statistical aggregation characteristics from three dimensions: periodicity, trend, and burst;

[0042] Mode convergence feature calculation:

[0043] ,

[0044] in, Represents the homogeneity feature and measures the aggregation node The degree of similarity among internal members on key attributes; is the modularity feature, measuring The degree of internal clustering of features or attributes.

[0045] Targeting homogametic characteristics , can comprehensively consider multi-dimensional attributes such as spatial distance attenuation relationship, attribute similarity, and temporal behavior pattern similarity; for modularity representation , we can comprehensively consider attributes such as community structure compactness, degree of functional clustering, resource distribution density and heterogeneity.

[0046] Furthermore, S4 specifically includes:

[0047] (4.1) Any node pair of the layer , calculate attribute correlation :

[0048] ,

[0049] ,

[0050] ,

[0051] in,

[0052] 、 Obtained from the time series indicator data extracted by S2;

[0053] 、 Obtained by statistical aggregation features extracted from S3;

[0054] 、 Obtained by the pattern convergence features extracted in step S3; It is a multivariate correlation measurement function, and the Pearson correlation coefficient or Spearman rank correlation coefficient can be selected.

[0055] (4.2) Any node pair of the layer , calculate the correlation of graph structure :

[0056] ,

[0057] in, Representation node The neighborhood subgraph of Indicates the graph kernel method, which can be selected from Weisfeiler-Lehman kernel or random walk kernel.

[0058] (4.3) Fusion of the attribute correlation obtained in (4.1) and the graph structure correlation obtained in (4.2) yields the intra-layer correlation index:

[0059] ,

[0060] in and is the weight parameter;

[0061] Furthermore, in S5, the hierarchical aggregation correlation index is:

[0062] ,

[0063] in, To aggregate into nodes The set of lower nodes of It is a correlation aggregation function that is responsible for extracting the key features of the relationship between lower-level nodes. It can use methods such as maximum value, average value, and attention mechanism;

[0064] Furthermore, S6 specifically includes:

[0065] (6.1) Introduce a hierarchical attention mechanism to assign dynamic weights to each layer:

[0066] ,

[0067] in It is a multi-layer perceptron, which is used to extract the features of each level graph and generate weights.

[0068] (6.2) Based on the intra-layer correlation and inter-layer correlation, calculate the spatial multi-scale correlation:

[0069] ,

[0070] in, , They are , In the layer, i.e. The corresponding node.

[0071] The present invention has the following beneficial effects:

[0072] (1) By constructing a hierarchical spatial multi-scale correlation graph system, a unified representation of the correlation relationship of data at different spatial scales is achieved;

[0073] (2) Through the multi-dimensional feature decomposition method, the accurate extraction of data periodicity, trend and burst characteristics is achieved;

[0074] (3) Through statistical aggregation and pattern aggregation, the problem of cross-scale information transmission is solved, and information abstraction and knowledge extraction from micro to macro are achieved;

[0075] (4) Through intra-layer correlation analysis and inter-layer correlation propagation, the quantitative characterization of the correlation relationship between the same scale and different scales is achieved;

[0076] (5) Through the hierarchical attention mechanism, dynamic fusion of multi-scale correlations is achieved, which improves the accuracy and adaptability of correlation analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0078] The present invention will be further described below with reference to the accompanying drawings.

[0079] like Figure 1 This embodiment applies the present invention's multi-scale feature extraction and correlation analysis method to power load analysis. By constructing a four-level graph structure, ranging from individual users to the entire city, this method achieves multidimensional extraction, cross-scale aggregation, and correlation propagation of power load characteristics. By implementing this method, it is possible to obtain intrinsic correlation patterns of power loads at different spatial scales, providing data support and decision-making basis for accurate prediction, efficient scheduling, and scientific planning of power systems.

[0080] (1) Scenario description:

[0081] In power load analysis, data exists at different spatial scales, including individual user electricity consumption data, industrial park or community electricity consumption data, and urban area electricity consumption data. These data exhibit different characteristics and correlations at different spatial scales. For example, the power load of an individual user may be affected by factors such as their lifestyle and the operating cycle of production equipment, exhibiting distinct periodic and sudden characteristics. The power load of an industrial park or community, on the other hand, is influenced by the combined power consumption behaviors of multiple users, exhibiting more complex trends and homogeneity. The power load of an urban area further integrates the power consumption of multiple industrial parks or communities, reflecting the power demand pattern of the entire city.

[0082] (2) Application of the method:

[0083] Step 1: Construct a multi-scale correlation graph system for power load space.

[0084] In this embodiment, a four-layer spatial multi-scale correlation graph system is constructed. 、 、 、 , corresponding to the following spatial scales:

[0085] : Individual electricity user level, including industrial users, commercial users and residential users.

[0086] : Industrial park or community level, composed of multiple similar users.

[0087] : Urban area level, consisting of multiple industrial parks or communities.

[0088] : The overall urban level, consisting of multiple urban areas.

[0089] In specific implementation, take the electricity users in a certain city as an example:

[0090] exist Level, select 1000 typical electricity users in the city as nodes , including 500 industrial users, 300 commercial users and 200 residential users.

[0091] Create edge sets based on the similarity of user geographic locations and electricity usage types ,When the geographical distance between two users is less than 500 meters and the similarity of ,electricity usage types is greater than 0.7, a connection is established between them.

[0092] exist level, will The users in the cluster are aggregated according to the park or community they belong to, forming 50 nodes. , representing 50 industrial parks or communities.

[0093] exist level, will The parks or communities in the city are gathered according to their respective urban areas to form 10 nodes. , representing 10 urban areas.

[0094] exist level, will The urban areas in the , indicating the entire city.

[0095] Step 2: Extraction and characterization of multi-dimensional features of power load.

[0096] for Each power user node in the hierarchy , collect its 15-minute granularity power load data for one year to form time series data .

[0097] Decompose time series indicator data into three-dimensional features:

[0098] ,

[0099] The three-dimensional feature decomposition is achieved through the following optimization model:

[0100] ,

[0101] Among them, the loss function uses the L2 norm, =0.1, =0.05, =0.2.

[0102] Periodicity Constraints Defined as:

[0103] ,

[0104] Among them, T1=96 (representing daily cycle) and T2=672 (representing weekly cycle).

[0105] Smoothness constraints Defined as:

[0106] ,

[0107] in represents the second-order difference operator.

[0108] Sparsity constraints Defined as:

[0109] ,

[0110] The L1 norm is used to promote sparsity.

[0111] Take an industrial user For example, the power load is decomposed into:

[0112] Periodic components Capture load differences between working days and rest days, as well as cyclical changes in production shifts.

[0113] Trend components Capture long-term trends caused by capacity expansion or seasonal changes.

[0114] Sudden component Capture abnormal events such as equipment maintenance and sudden failures.

[0115] Step 3: Aggregation of cross-scale characteristics of power load.

[0116] for Hierarchical industrial park nodes , assuming that it is 20 user nodes at the level The features are aggregated in the following ways:

[0117] Statistical aggregation feature calculation:

[0118] ,

[0119] ,

[0120] ,

[0121] in, 、 、 They are The set of periodic, trend and burst components of all nodes in .

[0122] Mode convergence feature calculation:

[0123] ,

[0124] in, Represents the homogeneity feature and measures the aggregation node The degree of similarity among internal members on key attributes; is the modularity feature, measuring the degree of internal clustering of functions or attributes;

[0125] Homogeneity feature r: Calculates the homogeneity of the power consumption type distribution and voltage level distribution of users in the park:

[0126] ,

[0127] in, is the information entropy of the electricity usage type, is the information entropy of the voltage level, N is the number of electricity types, and M is the number of voltage level types.

[0128] Modularity characteristic q: Calculates the tightness of the connection between nodes within the park:

[0129] - ,

[0130] A high-tech industrial park As an example, the mean value of periodic load characteristics is obtained through statistical aggregation. =5.2MW, standard deviation =1.8MW, indicating that the overall load of the park has obvious working / non-working period characteristics; through pattern aggregation, its homogeneity characteristic r=0.78 and modularity characteristic q=0.65 are obtained, indicating that the power users within the park have high homogeneity and close functional associations.

[0131] Step 4: Correlation analysis within the power load layer.

[0132] right Any two industrial park nodes at the same level and , perform intra-layer correlation analysis according to the following steps:

[0133] Calculate attribute correlations:

[0134] ,

[0135] in, =0.6, =0.4, represents the Pearson correlation coefficient, Represents cosine similarity.

[0136] Compute graph structure relevance:

[0137] ,

[0138] Among them, WL represents the Weisfeiler-Lehman kernel, express The neighborhood subgraph of Indicates the number of iterations.

[0139] Fusion of attribute and structural dependencies:

[0140] ,

[0141] Two industrial parks (High-tech Park) and (Traditional Manufacturing Park) Example: Attribute Correlation =0.42, indicating that there are certain differences in the load characteristics of the two parks; =0.35, indicating that the network structures of the two parks are quite different; the intra-layer correlation after fusion =0.4, indicating that the overall correlation between the power loads of the two parks is low.

[0142] Step 5: Correlation propagation between power load layers.

[0143] for Hierarchical industrial park nodes , and its hierarchical convergence correlation is calculated as follows:

[0144] Calculate the average correlation between lower-level nodes:

[0145] ,

[0146] in, express Any two nodes in the hierarchy and The correlation coefficient between , Representing a collection The number of nodes in the set, so N represents the number of all possible different node pairs in the set. This formula is used to characterize and its lower-level nodes (i.e. The correlation of power loads of nodes in the network.

[0147] Step 6: Fusion of multi-scale correlation of power loads.

[0148] Finally, the correlation information at different levels is integrated to calculate the spatial multi-scale correlation:

[0149] The attention weights of each layer are calculated by multi-layer perceptron:

[0150] ,

[0151] The input of MLP is the feature vector of each level graph, including the number of nodes, edge density, average degree, etc.

[0152] Assume that the obtained attention weight is α=[0.15, 0.35, 0.3, 0.2], which indicates the importance attached to each level from G0 to G3.

[0153] Compute the spatial multi-scale correlation between any two nodes:

[0154] ,

[0155] in, , They are , In the layer, i.e. The corresponding node.

[0156] To evaluate high-tech parks Compared with traditional manufacturing parks As an example, the multi-scale correlation between levels, their correlation is =0.4; Level, the correlation of their corresponding urban area nodes is =0.55; Level, they correspond to the same city node, the correlation is =1.0; their respective hierarchical convergence correlations are , ; The final multi-scale correlation is calculated as:

[0157]

[0158] ,

[0159] The results show that although the two parks The direct correlations of the hierarchies were low, but their combined correlations reached a moderate level after accounting for higher-level associations and the internal structure of the hierarchies.

[0160] Through the above-mentioned specific implementation steps, the method of the present invention can comprehensively capture the characteristics and correlations of power loads at different spatial scales, providing a basis for power load forecasting, power system scheduling, and power system planning. Compared with traditional methods, the present invention can more comprehensively and accurately capture the multi-scale characteristics of power loads, improve the operating efficiency and reliability of the power system, and reduce operating costs.

[0161] The above description is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements can be made without departing from the principles of the present invention. These improvements should also be regarded as the scope of protection of the present invention.

Claims

1. A multi-scale feature extraction and association analysis method, characterized by: The method is applied to the power load analysis scenario and includes the following steps: S1: Constructing a multi-scale correlation graph system G of the power load space with four hierarchical correlation graphs o , G1, G2, G3, each association graph corresponds to a specific spatial scale, where: G o : Single power user level; G1: Industrial park or community level; G2: Urban area level; G3: Overall city level, and there is a hierarchical convergence relationship between the association diagrams of different levels. S2: For each node in the bottom-level association graph, extract its time series indicator data and decompose it into periodic, trend, and sudden components. The time series indicator data is the 15-minute granularity power load data collected for each node for one year. Using the optimization model, it is decomposed into three-dimensional features, where the periodic component reflects the load difference between weekdays and weekends, as well as the cyclical changes in production shifts; the trend component captures long-term trends caused by capacity expansion or seasonal changes; and the sudden component captures equipment maintenance and sudden failures. S3: For nodes in the non-bottom-level association graph, statistical aggregation and pattern aggregation are used to extract aggregation features; S4: Based on the time series indicator data extracted in S2 and the aggregated features extracted in S3, the attribute correlation and graph structure correlation of node pairs are calculated in each layer, and the two are combined to obtain the intra-layer correlation index of the corresponding layer; S5: Define hierarchical aggregation correlation indicators based on intra-layer correlation indicators to achieve inter-layer correlation propagation; S6: Introducing a hierarchical attention mechanism to fuse correlation information at different levels and calculate spatial multi-scale correlations; S2 specifically includes: (2.1) For the bottom-level association graph Each power user node , extract its time series indicator data: ， in, represents the index value of the jth node at time t, and T represents the length of the time series; (2.2) Decompose the time series indicator data into three-dimensional features: ， in: It represents the periodic component, reflecting the periodic regularity of the indicator data, the load difference between working days and rest days, and the periodic changes of production shifts; Represents the trend component, capturing the long-term trend of indicator data, that is, capturing the long-term trend caused by capacity expansion or seasonal changes; It represents the sudden component, characterizing sudden events and abnormal patterns in the indicator data, i.e., capturing equipment maintenance and sudden failures; (2.3) Three-dimensional feature decomposition is achieved through the following optimization model: ， in, is the reconstruction loss function; 、 、 are periodic constraints, smoothness constraints, and sparsity constraints, respectively. 、 、 is a trade-off parameter.

2. The multi-scale feature extraction and association analysis method according to claim 1, wherein: S1 specifically includes: (1.1) Construct a set of spatial multi-scale correlation graphs: G o =(V o ,E o ,P o ),G1=(V1,E1,P1),G2=(V2,E2,P2),G3=(V3,E3,P3), Among them: G0 represents the zero-level association map, G1 represents the first-level association map, G2 represents the second-level association map, and G3 represents the third-level association map, which correspond to specific spatial scales, specifically: G o : Individual electricity user level, including industrial users, commercial users and residential users; G1: industrial park or community level, consisting of multiple similar users; G2: urban area level, consisting of multiple industrial parks or communities; G3: overall urban level, consisting of multiple urban areas; Represents the node set corresponding to the i-th spatial scale, representing the unit individual of this scale A collection of Represents the set of association relationships between individual units; Represents a set of node attributes and describes the multidimensional characteristics of each node; In G o At the hierarchical level, electricity users are selected as nodes; edge sets are established based on the similarity of user geographical locations and electricity usage types; At the G1 level, G o The users in the group are aggregated according to the parks or communities they belong to; At the G2 level, the parks or communities in G1 are clustered according to the urban areas to which they belong; At the G3 level, the urban areas in G2 are aggregated into one node, representing the entire city; (1.2) Establish hierarchical convergence relationships between association graphs at different levels, so that the node set V of the i-th layer i The nodes in the i-1th layer are composed of the node set V i−1 Several nodes in the network are formed by convergence, reflecting the hierarchical progressive relationship of spatial scales.

3. The multi-scale feature extraction and association analysis method according to claim 1, wherein: S3 specifically: for Layer nodes , i≥1, assuming By the node set of the lower layer Convergence forms, m represents belonging layers, converged into The number of nodes is calculated separately. , pattern convergence characteristics ; Statistical aggregation feature calculation: ， in, , , , , , represent mean, standard deviation, lower quartile, upper quartile, skewness, and kurtosis, respectively; hour Corresponding respectively , in order to calculate the load statistical aggregation characteristics from three dimensions: periodicity, trend, and burst; Mode convergence feature calculation: ， in, Represents the homogeneity feature and measures the aggregation node The degree of similarity among internal members on key attributes; is the modularity feature, measuring The degree of internal clustering of features or attributes.

4. The multi-scale feature extraction and association analysis method according to claim 3, wherein: S4 specifically includes: (4.1) Any node pair of the layer , calculate attribute correlation : ， ， ， in, is the multivariate correlation measurement function; (4.2) Any node pair of the layer , computing graph structure relevance : ， in, Representation node The neighborhood subgraph of Representation graph kernel method; (4.3) Fusion of the attribute correlation obtained in (4.1) and the graph structure correlation obtained in (4.2) yields the intra-layer correlation index: ， in and is the weight parameter.

5. The multi-scale feature extraction and association analysis method according to claim 4, wherein: In S5, the hierarchical aggregation correlation index is: ， in, To aggregate into nodes The set of lower nodes of is the correlation aggregation function.

6. The multi-scale feature extraction and association analysis method according to claim 5, wherein: S6 specifically includes: (6.1) Introduce a hierarchical attention mechanism to assign dynamic weights to each layer: ， in is a multi-layer perceptron; (6.2) Based on the intra-layer correlation and inter-layer correlation, calculate the spatial multi-scale correlation: ， in, , They are , In the layer, i.e. The corresponding node.

Images