CN-121234004-B - Deep learning-based building carbon emission data processing method and computer system

CN121234004BCN 121234004 BCN121234004 BCN 121234004BCN-121234004-B

Abstract

The invention provides a deep learning-based building carbon emission data processing method and a computer system, which are characterized in that feature association mining is performed by acquiring a building carbon emission data set to generate an association intensity tensor, wherein the first dimension of the association intensity tensor corresponds to a functional unit, the second dimension corresponds to a time sequence position, the third dimension corresponds to a feature association type, and tensor elements represent the association degree of carbon emission features of the corresponding functional unit at the corresponding time sequence position and carbon emission features of other functional units; the method comprises the steps of constructing a time-varying dependent network based on a correlation strength tensor to obtain time-varying dependent network parameters, tracking a multi-scale evolution track through the time-varying dependent network parameters to generate a multi-scale evolution track set, executing attention mechanism critical path identification through the multi-scale evolution track set, and outputting a carbon emission data processing result. The invention can improve the effectiveness and reliability of the whole carbon emission data processing result.

Inventors

YAO ZIXUAN
XUE SHIWEI
ZUO TINGTING
AN WEI
WANG HAIBIN
SUN HAIXIA
CAO ZHIBIN

Assignees

中建碳科技有限公司
中建生态环境集团有限公司

Dates

Publication Date: 20260505
Application Date: 20250903

Claims (9)

1. The method for processing the building carbon emission data based on deep learning is characterized by comprising the following steps of: Acquiring a building carbon emission data set, wherein the building carbon emission data set comprises a carbon emission monitoring sequence of a building functional system cluster in a continuous time sequence and an associated influence sequence, the building functional system cluster is a functional unit combination for generating carbon emission in a building operation process, and the associated influence sequence is an internal and external environment parameter sequence for influencing the change of the carbon emission characteristics of the functional units along with the time sequence; Performing feature association mining on the building carbon emission data set to generate an association strength tensor, wherein the first dimension of the association strength tensor corresponds to a functional unit, the second dimension corresponds to a time sequence position, the third dimension corresponds to a feature association type, and tensor elements represent association degrees of carbon emission features of the corresponding functional unit at the corresponding time sequence position and carbon emission features of other functional units; Constructing a time-varying dependent network based on the correlation strength tensor to obtain a time-varying dependent network parameter; Tracking a multi-scale evolution track through the time-varying dependent network parameters to generate a multi-scale evolution track set; Executing attention mechanism key path identification through the multi-scale evolution track set, and outputting a carbon emission data processing result; performing feature association mining on the building carbon emission data set to generate an association strength tensor, including: Separating a carbon emission monitoring subsequence and an associated influence subsequence of each functional unit from the building carbon emission data set to obtain a functional unit data sequence set, wherein each element in the functional unit data sequence set is a combined sequence of the carbon emission monitoring subsequence and the associated influence subsequence corresponding to one functional unit; Performing feature coupling processing on the combined sequences in the functional unit data sequence set to generate a coupling feature sequence, wherein the feature coupling processing is used for enabling feature vectors of each time sequence position to simultaneously comprise carbon emission monitoring features and associated influence features by carrying out feature dimension interleaving on the carbon emission monitoring sub-sequence and the associated influence sub-sequence according to time sequence positions; Constructing a correlation mining network, wherein the correlation mining network comprises a sequence coding layer, a cross correlation layer and a tensor output layer, the sequence coding layer is used for extracting time sequence characteristics of the coupling characteristic sequences, the cross correlation layer is used for calculating correlations among the coupling characteristic sequences of different functional units, and the tensor output layer is used for integrating correlation results to generate tensors; inputting the functional unit data sequence set into a sequence coding layer of an associated mining network, and coding the coupling characteristic sequences of all the functional units through stacked gating circulating units to obtain functional unit coding characteristic sequences, wherein the functional unit coding characteristic sequences comprise time dependent characteristics of the coupling characteristic sequences; Inputting the functional unit coding feature sequences to a cross-correlation layer of a correlation mining network, calculating the correlation score of any two functional unit coding feature sequences at each time sequence position through a multi-head cross attention mechanism, wherein the correlation score is obtained through the weighted combination of an attention weight matrix and a coding feature vector; And arranging the association scores according to the dimension of the functional unit, the dimension of the time sequence position and the dimension of the attention head, inputting the association scores to a tensor output layer of an association mining network, and carrying out feature integration through a three-dimensional convolution layer to obtain the association strength tensor.
2. The method of claim 1, wherein performing feature coupling processing on the combined sequences in the set of functional unit data sequences to generate a coupled feature sequence comprises: Acquiring a carbon emission monitoring subsequence and an associated influence subsequence in the combined sequence, and determining the time sequence length and characteristic dimension of the two subsequences, wherein the time sequence length is the number of positions of a continuous time sequence, and the characteristic dimension is the number of characteristic parameters contained in each time sequence position; performing time sequence alignment verification on the carbon emission monitoring subsequence and the associated influencing subsequence, so that the time sequence lengths of the two subsequences are consistent, and the time sequence positions are in one-to-one correspondence; Performing dimension interleaving on the characteristic vector of each time sequence position of the carbon emission monitoring subsequence and the characteristic vector of the time sequence position corresponding to the associated influencing subsequence, wherein the dimension interleaving generates an interleaving characteristic vector by alternately arranging two characteristic vectors according to the characteristic parameter sequence; performing feature enhancement on the interleaving feature vector, and inserting a difference feature of adjacent time sequence positions into the interleaving feature vector, wherein the difference feature is an interleaving feature vector difference value between a current time sequence position and a previous time sequence position; and arranging the enhanced interweaved feature vectors of all the time sequence positions in a time sequence order to obtain the coupling feature sequence, wherein the time sequence length of the coupling feature sequence is consistent with that of the original combined sequence, and the feature dimension is a preset multiple of the original feature dimension.
3. The method of claim 1, wherein said inputting the functional unit encoded feature sequences into a cross-correlation layer of an association mining network calculates an association score for any two functional unit encoded feature sequences at each time-series location via a multi-headed cross-attention mechanism, comprising: Selecting the coding feature sequences of any two functional units from the coding feature sequences of the functional units as target coding pairs to obtain a first coding sequence and a second coding sequence; The first coding sequence and the second coding sequence are respectively input into a query vector generation module and a key value vector generation module of a multi-head cross attention mechanism to generate a query vector sequence, a key vector sequence and a value vector sequence, wherein the query vector sequence is obtained by linear transformation of the first coding sequence, and the key vector sequence and the value vector sequence are obtained by different linear transformations of the second coding sequence; Carrying out dimension splitting on the query vector sequence, the key vector sequence and the value vector sequence according to the attention head quantity to obtain a plurality of head components, wherein each head component comprises a corresponding query sub-vector sequence, key sub-vector sequence and value sub-vector sequence; Performing dot product operation on the query sub-vector sequence and the key sub-vector sequence of each head component to obtain an original correlation matrix, wherein the elements of the original correlation matrix are dot product results of the query sub-vector and the key sub-vector; performing scale scaling on the original incidence matrix, dividing each element by the square root of the key sub-vector dimension to obtain a scaled incidence matrix; Activating the scaling incidence matrix to obtain an attention weight matrix, wherein elements of the attention weight matrix represent incidence weights of a query sub-vector sequence and a key sub-vector sequence; carrying out weighted summation on the attention weight matrix and the value sub-vector sequence to obtain a head association characteristic sequence of each head component; And performing dimension splicing on the head association characteristic sequences of all the head components, and integrating through a linear transformation layer to obtain association scores of target coding pairs at each time sequence position.
4. The method of claim 1, wherein said constructing a time-dependent network based on said correlation-strength tensor results in time-dependent network parameters comprising: taking a functional unit corresponding to the first dimension of the association strength tensor as an initial node set of the time-varying dependent network, wherein each node in the initial node set corresponds to one functional unit; extracting an associated intensity matrix of each time sequence position from the associated intensity tensor, wherein the associated intensity matrix is a two-dimensional matrix formed by a first dimension and a third dimension when the associated intensity tensor is fixed in the second dimension; Determining a connection relation between nodes of each time sequence position based on the correlation intensity matrix, and if the corresponding element value in the correlation intensity matrix is larger than a preset correlation threshold value for any two nodes, establishing temporary edge connection between the nodes in the time sequence position, wherein the edge weight is the corresponding element value; integrating the temporary edge connection and the edge weight of all the time sequence positions according to the time sequence order to obtain a time-varying edge set, wherein the time-varying edge set comprises edge connection states and corresponding edge weight values of nodes at different time sequence positions; Extracting a feature vector corresponding to each node at each time sequence position, wherein the node feature vector is obtained by carrying out element-level multiplication on a first dimension vector of the node in an associated intensity tensor and a feature vector of the node at a time sequence position corresponding to a coupling feature sequence; Arranging the node feature vectors in a time sequence order to obtain a node feature vector time sequence; calculating the change rate of the edge weight of each edge in the time-varying edge set at the adjacent time sequence position to obtain an edge weight evolution gradient, wherein the evolution gradient is the difference value between the edge weight of the current time sequence position and the edge weight of the previous time sequence position; and integrating the node feature vector time sequence and the edge weight evolution gradient to obtain the time-varying dependent network parameters.
5. The method of claim 4, wherein the determining the inter-node connection relationship for each time series location based on the correlation strength matrix comprises: performing matrix standardization processing on the correlation intensity matrix, and adjusting matrix element values to a preset numerical value interval to enable the correlation intensity matrix at different time sequence positions to be comparable; Processing the standardized association intensity matrix by adopting a non-maximum suppression algorithm, reserving a preset number of elements with the maximum value in each row, and setting other elements to be zero values to obtain a sparse association matrix; carrying out symmetry processing on the sparse incidence matrix, and if the ith row and the jth column elements in the matrix are nonzero and the ith row and the jth column elements in the matrix are zero, assigning the ith row and the ith column elements as the ith row and the jth column element values; Traversing non-zero elements in the symmetric sparse incidence matrix, and recording row indexes and column indexes corresponding to each non-zero element, wherein the row indexes and the column indexes respectively correspond to identifications of two nodes; And taking the node identification pair corresponding to each non-zero element as the connection relation among the nodes, and taking the non-zero element value as the edge weight of the corresponding connection relation to obtain the connection relation among the nodes of the time sequence position.
6. The method of claim 4, wherein extracting the feature vector for each node at each time-series location comprises: extracting a first dimension vector corresponding to a target node from the association strength tensor to obtain a node association feature vector, wherein the dimension of the node association feature vector is consistent with a third dimension of the association strength tensor, and the element value is the association degree of the target node and other nodes; Extracting a coupling feature sub-vector of the target node at a corresponding time sequence position from the coupling feature sequence, wherein the coupling feature sub-vector comprises a carbon emission monitoring feature and an associated influence feature of the target node; Performing dimension adaptation processing on the node association feature vector and the coupling feature sub-vector, and adjusting the dimensions of the two vectors to be the same dimension through linear transformation to obtain an adaptation association vector and an adaptation coupling vector; Performing element-level product operation on the adaptive correlation vector and the adaptive coupling vector to obtain a preliminary node feature vector; and carrying out feature smoothing treatment on the preliminary node feature vector, and taking the smoothed preliminary node feature vector as a node feature vector of the target node at the time sequence position.
7. The method of claim 1, wherein the tracking the multi-scale evolution trace through the time-varying dependent network parameters to generate the set of multi-scale evolution traces comprises: determining a multi-scale time window set, wherein the multi-scale time window set comprises a short-term time window, a medium-term time window and a long-term time window, and the window length of each time window is different proportions of the number of continuous time sequence positions; Based on the node feature vector time sequence in the time-varying dependent network parameters, window division is carried out on the node feature vector time sequence according to a multi-scale time window set to obtain a short-term feature window sequence, a medium-term feature window sequence and a long-term feature window sequence of each node, wherein the feature window sequence consists of node feature vectors at a plurality of continuous time sequence positions; Performing track fitting on the short-term characteristic window sequences of each node, and obtaining a short-term characteristic evolution track through a polynomial curve fitting algorithm, wherein the short-term characteristic evolution track represents the change trend of the node characteristic vector in a short-term time window; performing track fitting on the middle-term characteristic window sequence and the long-term characteristic window sequence by adopting the same method to obtain a middle-term characteristic evolution track and a long-term characteristic evolution track; extracting gradient values of edge weights among nodes at each time sequence position from the edge weight evolution gradient in the time-varying dependent network parameters, and dividing the gradient values according to a multi-scale time window set to obtain a short-term gradient window sequence, a medium-term gradient window sequence and a long-term gradient window sequence; Carrying out gradient accumulation processing on the short-term gradient window sequences of the edges between each node, and calculating the accumulation sum of gradient values in the windows to obtain a short-term gradient change track, wherein the short-term gradient change track represents the accumulated change quantity of the edge weights in a short-term time window; Performing gradient accumulation treatment on the middle gradient window sequence and the long gradient window sequence by adopting the same method to obtain a middle gradient change track and a long gradient change track; And integrating the short-term characteristic evolution track, the medium-term characteristic evolution track, the long-term characteristic evolution track, the short-term gradient change track, the medium-term gradient change track and the long-term gradient change track of all sides of all nodes to obtain the multi-scale evolution track set.
8. The method of claim 7, wherein the windowing the time series of node feature vectors based on the time series of node feature vectors in the time-varying dependent network parameters according to a set of multi-scale time windows results in a short-term feature window sequence, a mid-term feature window sequence, and a long-term feature window sequence for each node, comprising: Acquiring the total time length of the node feature vector time sequence, wherein the total time length is the total number of time sequence positions; Determining the sliding step length of the short-term time window according to the total time length and the window length of the short-term time window, wherein the sliding step length is the number of time sequence positions of window movement, so that a preset proportion of overlapping areas exist between adjacent short-term time windows; starting from the initial position of the node characteristic vector time sequence, intercepting the subsequence according to the short-term time window length and the sliding step length to obtain a plurality of short-term characteristic windows, and arranging the windows according to the intercepting order to obtain a short-term characteristic window sequence; Determining a middle sliding step length based on the window length of a middle time window and a preset overlapping proportion, and obtaining a middle characteristic window sequence by adopting the same interception method; Determining a long-term sliding step length based on the window length of a long-term time window and a preset overlapping proportion, and obtaining a long-term characteristic window sequence by adopting the same interception method; and carrying out boundary processing on windows in each characteristic window sequence, and if the length of the last window is smaller than the length of the corresponding time window, filling the window length by copying the characteristic vector of the last node so that the lengths of all the windows are consistent.
9. A computer system comprising a memory and a processor, the memory storing a computer program executable on the processor, wherein the processor implements the steps of the method of any one of claims 1 to 8 when the program is executed.

Description

Deep learning-based building carbon emission data processing method and computer system Technical Field The invention relates to the technical field of deep learning and data processing, in particular to a building carbon emission data processing method and a computer system based on deep learning. Background Along with the increase of carbon emission reduction demands in the building industry, research on building energy conservation and carbon management is focused, and a key path affecting carbon emission characteristics is identified by analyzing carbon emission monitoring data and associated influence data of a building functional system, so that data support is provided for optimizing low-carbon operation of a building. At present, building carbon emission data processing adopts a two-dimensional matrix to analyze association relations among systems, tracks evolution tracks of carbon emission characteristics by constructing dependency relations among static network description functional units, and screens key influence paths depending on preset thresholds or manual rules. However, the related characteristics of different types cannot be distinguished in the prior art, the connection relation between the nodes is fixed by the static network, the change of the related strength along with time is difficult to capture, and the effectiveness and the reliability of the carbon emission data processing result are affected. Disclosure of Invention In view of the above, the present invention provides a deep learning-based building carbon emission data processing method and a computer system. The technical scheme of the embodiment of the invention is realized as follows: On one hand, the embodiment of the invention provides a deep learning-based building carbon emission data processing method, which comprises the steps of obtaining a building carbon emission data set, wherein the building carbon emission data set comprises a carbon emission monitoring sequence and an associated influence sequence of a building functional system cluster in a continuous time sequence, the building functional system cluster is a functional unit combination for generating carbon emission in a building operation process, the associated influence sequence is an internal and external environment parameter sequence for influencing the change of the carbon emission characteristics of the functional unit along the time sequence, performing feature association mining on the building carbon emission data set to generate an associated intensity tensor, the first dimension of the associated intensity tensor corresponds to a functional unit, the second dimension corresponds to a time sequence position, the third dimension corresponds to a feature association type, tensor elements represent the association degree of the carbon emission characteristics of the corresponding functional unit at the corresponding time sequence position and the carbon emission characteristics of other functional units, constructing a time-varying dependent network based on the associated intensity tensor to obtain a time-varying network parameter, tracking a multi-scale evolution track by the time-varying dependent network parameter, and performing attention critical path recognition by the multi-scale evolution track set, and outputting a carbon emission data processing result. In another aspect, embodiments of the present invention provide a computer system comprising a memory and a processor, the memory storing a computer program executable on the processor, the processor implementing the steps of the method described above when the program is executed. According to the deep learning-based building carbon emission data processing method, a building carbon emission data set is obtained, feature association mining is performed to generate an association strength tensor, a time-varying dependence network is constructed to obtain time-varying dependence network parameters, a multi-scale evolution track is tracked to generate a multi-scale evolution track set, and a key path is identified through an attention mechanism to output a processing result. According to the invention, association analysis is expanded from two dimensions to three dimensions through association intensity tensors, characterization of multiple types of association features is realized, dependency modeling is evolved along with time by combining a time-varying dependent network, short-term, medium-term and long-term evolution rules are captured in a layered manner by multi-scale evolution track tracking, a key path identification driven by an attention mechanism can be used for adaptively screening a path sequence with contribution degree dominant, accuracy of association feature characterization, dependency modeling, comprehensiveness of evolution rule capture and accuracy of key path identification in a building carbon emission data processing process are effectively improved, and effectiveness