Carbon footprint AI accounting method and system based on full-link data

Abstract

This invention discloses a carbon footprint AI calculation method and system based on end-to-end data. The method maps heterogeneous data such as coal consumption and electricity consumption into a production topology map. It extracts spatiotemporal state features of the production process using a graph convolutional network with material flow conservation constraints. Based on the fluctuation characteristics of operating conditions, it dynamically generates multilayer perceptron weights using an adaptive weight generation network, mapping them to a basic carbon emission factor library to output real-time dynamic factors. Subsequently, it cascades and aligns the spatiotemporal state features with the dynamic factors using a dual-flow feature alignment and fusion module to output carbon emission slice features. Finally, it processes the data through a hierarchical node aggregation module to output the end-to-end carbon footprint calculation results. This invention deeply integrates physical constraints into the model, achieving high-precision, high-reliability, and physically consistent real-time dynamic calculation.

Description

Technical Field

[0001] This invention relates to the technical field of carbon footprint accounting, and in particular to an AI-based carbon footprint accounting method and system based on end-to-end data. Background Technology

[0002] With the increasingly severe global climate change problem, greenhouse gas emission reduction has become a general consensus in the international community, and my country has also clearly put forward the strategic goal of "peaking carbon and achieving carbon neutrality." As the core area of ​​energy consumption and carbon emissions, accurate carbon footprint calculation in industrial production is a fundamental prerequisite for achieving energy conservation and emission reduction, optimizing production processes, and participating in carbon trading. Modern industrial production typically involves multiple complex production stages, involving the consumption and conversion of diverse and heterogeneous energy sources such as coal, electricity, and heat, forming a highly coupled end-to-end production system. Therefore, how to scientifically, accurately, and in real-time calculate the carbon footprint of the entire production chain has become a key technical challenge that urgently needs to be solved by industry and academia.

[0003] Traditional carbon footprint accounting primarily relies on life cycle assessment and emission factor methods. These methods typically perform static calculations based on historical statistical data and fixed baseline carbon emission factors. However, in actual industrial production, production conditions are constantly changing due to factors such as fluctuations in raw material quality, equipment aging, and load adjustments. Traditional methods, using static, outdated data and fixed carbon emission factors, cannot reflect the true operating status of the entire production chain at different times, resulting in significant time lags and large errors in the accounting results. This makes it difficult to meet the demands of modern industry for real-time, refined carbon emission management.

[0004] In recent years, with the development of artificial intelligence and big data technologies, some studies have begun to explore the use of deep learning models for carbon emission accounting. However, existing AI-based accounting methods still have significant limitations. First, industrial production is a complex topological network involving material transfer and energy conversion, with close upstream and downstream dependencies between various production stages. Existing data-driven models often treat each production unit as an isolated entity, failing to fully explore and utilize the topological characteristics of the entire production chain. More seriously, purely data-driven black-box models, when extracting spatiotemporal state features, completely deviate from the physical essence of industrial production, ignoring basic physical constraints such as the conservation of material flow. This leads to models that easily output accounting results that violate physical principles when faced with abnormal data or drastic fluctuations in operating conditions, severely reducing the reliability and interpretability of carbon footprint accounting.

[0005] Secondly, existing carbon footprint accounting models still suffer from the constraint of "static mapping" when dealing with carbon emission factors. Although some models incorporate real-time energy consumption data, their underlying carbon emission factor databases are often fixed. In reality, fluctuations in production conditions directly alter the actual carbon emission efficiency at each stage. Static carbon emission factors cannot adapt to the fluctuating characteristics of current operating conditions, leading to significant deviations in accounting results during operating condition switching or non-steady-state operation. Furthermore, when fusing multi-source heterogeneous data, existing methods typically employ simple feature concatenation, failing to achieve effective alignment between the spatiotemporal state characteristics of production stages and carbon emission factors at a deep semantic level. This crude feature fusion approach cannot accurately capture the complex nonlinear relationship between carbon emission factors and specific production states, limiting further improvements in the accuracy of end-to-end carbon footprint accounting.

[0006] In summary, given the complex and ever-changing industrial production environment, the industry urgently needs a novel carbon footprint accounting method. This method should effectively integrate the entire supply chain topology, incorporate physical constraints such as material flow conservation into deep learning models, dynamically generate carbon emission factors based on real-time operating condition fluctuations, and achieve deep alignment and fusion of multi-source features. This will provide industrial enterprises with high-precision, physically sound, and real-time dynamic end-to-end carbon footprint accounting results. Summary of the Invention

[0007] In view of this, the present invention provides a carbon footprint AI accounting method based on full-link data. The purpose is to achieve high-precision, high-reliability, and physically consistent real-time dynamic accounting of the carbon footprint of the entire industrial production chain by deeply integrating physical constraints into graph neural networks and combining dynamic factor generation with working condition adaptation and dual-stream feature alignment mechanism.

[0008] To achieve the above objectives, this invention provides an AI-based carbon footprint calculation method based on end-to-end data, comprising the following steps: B1: The collected coal consumption, electricity consumption and heat output multi-source heterogeneous flow data are mapped into the node features and edge features of the production topology graph. The spatiotemporal state features of the production process are extracted by the material flow constraint graph convolutional state extraction network that introduces material flow conservation constraints. B2: Input the spatiotemporal state characteristics of the production process into the working condition adaptive weight generation network. The working condition adaptive weight generation network dynamically generates the network weights of the multilayer perceptron based on the fluctuation characteristics of the current production working condition. The multilayer perceptron performs nonlinear mapping on the basic carbon emission factor library and outputs real-time dynamic carbon emission factors. B3: Input the spatiotemporal state characteristics of the production process and the real-time dynamic carbon emission factor into the dual-flow feature alignment and fusion module for feature concatenation and alignment. After processing by the feedforward neural network, the carbon emission slice features of each production process are output. B4: Input the carbon emission slice characteristics of each production link into the hierarchical node aggregation module for graph structure node aggregation processing, and output the carbon footprint calculation results of the whole link data through full connection layer mapping.

[0009] As a further improvement of the present invention: Optionally, step B1 further includes: Build a production topology diagram ,in Represents a set of nodes, a set of nodes Each node This corresponds to a production stage in the industrial production process. , Representing the production topology diagram Total number of nodes in the system; Let edge set be an edge set. Each edge Indicates the production process With the production process There is a direct material transfer relationship between them; based on the collected data on coal consumption, electricity consumption, and heat output, the production process... The corresponding coal consumption sample value, electricity consumption sample value, and heat output sample value are concatenated to form the initial feature vector of the node. ,in The dimension representing the initial features of a node; the production process Flow to the production process The material flow rate value is denoted as the edge feature. ; Production topology diagram Each intermediate production node Calculate the deviation term of material flow conservation The material flow conservation deviation term The calculation formula is: ; in, Represents a node The set consisting of all incoming predecessor nodes, Indicates from node Inflow node Material flow, Represents a node The set consisting of all outgoing successor nodes. Indicates from node Flow to node The mass flow rate; when the mass flow conservation deviation term Exceeding the preset conservation deviation threshold At that time, with nodes The ratio of the total outflow mass flow to the total inflow mass flow is used as the normalization coefficient. , will node All inflow edge characteristics Multiply by the normalization factor Complete the correction so that the total inflow of materials equals the total outflow of materials. Construct a convolutional state extraction network for the material flow constraint graph, wherein the material flow constraint graph convolutional state extraction network is composed of... It consists of stacked graph convolutional units, each containing an attention weight calculation sublayer and a material flow constraint correction sublayer; the attention weight calculation sublayer adjusts the node... The current layer input feature vector and its neighbor nodes After concatenating the current layer's input feature vectors, the resulting nodes undergo linear transformation, nonlinear activation, and normalization along the dimensions of neighboring nodes. For nodes Attention weight coefficient The material flow constraint correction sublayer is represented by the constraint strength coefficient in the neighborhood weighted aggregation result. Superimposed material flow conservation deviation term Perform constraint correction; The node feature update formula for each graph convolutional unit is: ; in, Indicates the first The nodes output by each graph convolutional unit eigenvectors, Represents a non-linear activation function. Represents a node In production topology The set of all direct neighbor nodes in the set. Indicates the first Nodes in a graph convolutional unit For nodes Attention weight coefficient, Indicates the first Learnable linear transformation matrix of each graph convolutional unit This represents the strength coefficient of the conservation constraint on material flow. Represents a node The deviation term of the material flow conservation, The input feature vector of the first graph convolutional unit The initial feature vector of the node; after After the feature updates layer by layer of each graph convolutional unit, the final output feature vector of each node is... Arranged sequentially by node number, this is a spatiotemporal state feature matrix of the production process. ,in Dimensions representing spatiotemporal state characteristics The Row corresponding node The spatiotemporal state feature vector.

[0010] Optionally, step B2 further includes: The spatiotemporal state feature matrix obtained in step B1 Perform global average pooling along the node dimension. The Middle List all The mean of the features of each node constitutes the global feature vector of the working condition. The One portion, Global feature vector of working conditions Reflection Moment The overall operating status of the entire production chain; in real time global feature vector of working conditions With time global feature vector of working conditions The difference constitutes the characteristic vector of operating condition fluctuation. ; The global feature vector of the working condition With operating condition fluctuation characteristic vector By concatenating along the dimensions, we obtain the comprehensive feature vector of the working conditions. ; Construct a working condition adaptive weight generation network, wherein the working condition adaptive weight generation network is composed of... The system consists of stacked fully connected layers, forming a comprehensive feature vector based on the working conditions. As input, the output dimension of the output layer is set to the sum of the parameters of the weight matrices and bias vectors of all layers in the multilayer perceptron. The output vector of the adaptive weight generation network is decomposed into segments according to the dimensions of the weight matrix and bias vector of each layer of the multilayer perceptron, and then reassembled to form the complete set of network weight parameters of the multilayer perceptron. ; Construct a basic carbon emission factor library, which stores the basic carbon emission factor vectors pre-calibrated for each production stage under standard operating conditions. ,in Dimensions representing carbon emission factors; As the current effective network weights, the basic carbon emission factor vector is processed through a multilayer perceptron. Perform nonlinear mapping to output a real-time dynamic carbon emission factor vector. The mapping formula is: ; in, Indicates time Real-time dynamic carbon emission factor vector, This represents the basic carbon emission factor vector in the basic carbon emission factor library. This indicates that the adaptive weight generation network is at time [time value missing]. The set of weight parameters of the output multilayer perceptron network. This represents the forward computation process of the multilayer perceptron; it represents the real-time dynamic carbon emission factor vector. The Each component Perform out-of-bounds checks. ,in This indicates the lower bound of the reasonable range for carbon emission factors. This represents the upper bound of the reasonable range for carbon emission factors, when... When Set as ,when When Set as .

[0011] Optionally, step B3 further includes: A dual-stream feature alignment and fusion module is constructed, wherein the dual-stream feature alignment and fusion module uses the spatiotemporal state feature matrix obtained in step B1. and the real-time dynamic carbon emission factor vector obtained in step B2 As input; for the spatiotemporal state feature matrix Learnable linear transformation matrix The linear transformation yields the query matrix. ; Real-time dynamic carbon emission factor vector Considered as being by A sequence consisting of scalars, where the is the ? Each component With learnable weight matrix The row vector Perform scaling embedding to obtain the first The key vector of each carbon emission factor ,Depend on Key vectors Arranged in rows to form a key matrix In the same way, with a learnable weight matrix The row vector right Perform scaling embedding, by Value vectors Arranged by rows to form a value matrix , The formula for calculating dual-stream feature alignment is: ; in, This represents the aligned feature matrix. Represents the query matrix. Represents the key matrix, This represents the transpose of the key matrix. Represents a value matrix, The dimension of the key vector. This represents an exponential function normalized along the second dimension. The Behavioral production process State characteristics for all Attention-weighted fusion results for each carbon emission factor dimension; aligning the feature matrix With spatiotemporal state feature matrix Concatenating along the feature dimensions yields a cascaded feature matrix. ,in The Travel The Action and The Row concatenation; cascaded feature matrix The feedforward neural network performs independent nonlinear transformations on the row vectors corresponding to each node, outputting the carbon emission slice feature matrix for each production stage. , The Corresponding production process The carbon emission slice feature vector.

[0012] Optionally, step B4 further includes: Construct a hierarchical node aggregation module, which uses the carbon emission slice feature matrix obtained in step B3. As the initial input, through A series of stacked node aggregation layers produce a topology graph. All node features are aggregated layer by layer; the normalized adjacency matrix From the production topology diagram The adjacency matrix is ​​obtained by adding the identity matrix and then normalizing the longitude; the first... The calculation formula for the aggregation layer of individual nodes is: ; in, Indicates the first The node feature matrix output by the node aggregation layer. The carbon emission slice feature matrix is ​​input to the hierarchical node aggregation module. Indicates the first The learnable weight matrix of the node aggregation layer Represents a non-linear activation function. , This indicates the total number of node aggregation layers in the hierarchical node aggregation module; go through After the aggregation layers of each node are converged, the final output node feature matrix is... Perform global average pooling along the node dimension, that is, ... All The mean of the rows is used to obtain the global graph representation vector. ; Global graph representation vectors are represented through fully connected layers. Perform linear mapping to output the carbon footprint calculation results of the entire data chain. The calculation formula is: ; in, This represents the carbon footprint calculation results of the entire data chain. This represents the weight matrix of the fully connected layer. This represents the bias vector of the fully connected layer. Represents a global graph as a vector; calculates the carbon footprint. Perform nonnegativity check when When Set to 0 to ensure that the carbon footprint accounting results conform to the physical meaning of non-negative constraints.

[0013] This invention also discloses a carbon footprint AI accounting system based on end-to-end data, comprising: Spatiotemporal state feature extraction module: used to map the collected multi-source heterogeneous flow data such as coal consumption, electricity consumption and heat output into the node features and edge features of the production topology graph, and extract the spatiotemporal state features of the production process by introducing a material flow constraint graph convolutional state extraction network with material flow conservation constraints. Dynamic carbon emission factor generation module: It is used to input the spatiotemporal state characteristics of the production process into the working condition adaptive weight generation network. The network dynamically generates the network weights of the multilayer perceptron according to the fluctuation characteristics of the current production working condition, and performs nonlinear mapping on the basic carbon emission factor library through the multilayer perceptron to output the real-time dynamic carbon emission factor. Dual-stream feature alignment and fusion module: used to concatenate and align the spatiotemporal state features of the production process with the real-time dynamic carbon emission factor, and output the carbon emission slice features of each production process after processing by a feedforward neural network. Carbon footprint accounting result output module: It is used to input the carbon emission slice characteristics of each production link into the hierarchical node aggregation module for graph structure node aggregation processing, and finally output the carbon footprint accounting result of the whole link data through full connection layer mapping.

[0014] Compared with the prior art, the present invention has at least the following beneficial effects: This invention does not directly apply conventional graph convolutional networks for feature extraction, but innovatively constructs a material flow-constrained graph convolutional state extraction network. Based on the traditional node attention aggregation mechanism, this network introduces a material flow conservation bias term to constrain and correct node features. This algorithmic improvement ensures that when extracting spatiotemporal state features of the production process, the model is no longer a purely data-driven black box, but is forced to adhere to the material conservation laws of industrial production. This improvement, which deeply integrates prior physical knowledge with deep learning models, effectively avoids the shortcomings of conventional graph algorithms that easily output results that violate physical principles when faced with abnormal data or drastic fluctuations in operating conditions, significantly improving the physical rationality of feature extraction and the robustness of the model.

[0015] This invention overcomes the limitation of fixed weights in traditional multilayer perceptron networks by designing a condition-adaptive weight generation network. This network dynamically generates multilayer perceptron network weights for mapping carbon emission factors by capturing the global state and temporal fluctuation characteristics of the entire production chain. This improvement enables the basic carbon emission factor to undergo nonlinear adaptive adjustment according to fluctuations in real-time production conditions, completely solving the industry problem that conventional static factors cannot match dynamic production processes, and endowing the model with extremely strong condition adaptability and real-time accuracy.

[0016] In the feature fusion stage, this invention abandons the simple feature concatenation method in conventional algorithms and constructs a dual-stream feature alignment fusion module. This module utilizes an attention mechanism to scale and embed dynamically generated carbon emission factors, achieving deep semantic alignment with spatiotemporal state features. This improved fusion method can accurately capture the complex nonlinear mapping relationship between specific production states and various carbon emission factors, realizing deep coupling and cascading of multi-source heterogeneous information.

[0017] The aforementioned algorithmic improvements are not isolated but rather collectively construct a closely collaborative and mutually reinforcing overall architecture. Spatiotemporal state characteristics with physical conservation constraints provide an accurate and reliable data foundation for extracting operational condition fluctuation characteristics; the precisely captured operational condition fluctuation characteristics directly drive the real-time dynamic evolution of carbon emission factors; finally, the dual-stream feature alignment mechanism deeply integrates physically consistent state characteristics with real-time evolving dynamic factors, and derives the global accounting result through hierarchical node aggregation. Each innovative module is interconnected, with smooth data flow, collectively overcoming the blind spots of traditional pure data-driven models in carbon footprint accounting, significantly improving the accuracy, real-time performance, and interpretability of end-to-end carbon footprint accounting, and possessing extremely high practical industrial application value. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating an embodiment of the carbon footprint AI accounting method based on end-to-end data according to the present invention. Figure 2 A schematic diagram illustrating the effect of real-time dynamic carbon emission factor output by the adaptive weight generation network for operating conditions. Detailed Implementation

[0019] The present invention will be further described below with reference to the accompanying drawings, but this is not intended to limit the present invention in any way. Any modifications or substitutions made based on the teachings of the present invention shall fall within the protection scope of the present invention.

[0020] Example 1: A carbon footprint AI accounting method based on end-to-end data, such as Figure 1 As shown, it includes the following steps: B1: The collected multi-source heterogeneous flow data of coal consumption, electricity consumption, and heat output are mapped to the node and edge features of the production topology. A material flow constraint graph convolutional state extraction network, incorporating material flow conservation constraints, is used to extract the spatiotemporal state features of the production process, including: Build a production topology diagram ,in Represents a set of nodes, a set of nodes Each node This corresponds to a production stage in the industrial production process. , Representing the production topology diagram Total number of nodes in the system; Let edge set be an edge set. Each edge Indicates the production process With the production process There is a direct material transfer relationship between them; based on the collected data on coal consumption, electricity consumption, and heat output, the production process... The corresponding coal consumption sample value, electricity consumption sample value, and heat output sample value are concatenated to form the initial feature vector of the node. ,in The dimension representing the initial features of a node; the production process Flow to the production process The material flow rate value is denoted as the edge feature. In this embodiment, the entire production chain of a coal chemical enterprise is taken as the object, and the production topology diagram is shown. Total of Each production node corresponds to 12 main production units, including raw material pretreatment, gasification, conversion, purification, synthesis, and distillation; edge set It contains 18 directed edges, representing material transfer paths between 12 production stages; the sampling interval for coal consumption, electricity consumption, and heat output data is set to 1 hour; the initial feature vector of each node... The node's initial feature dimension is formed by concatenating the above three components. ; Production topology diagram Each intermediate production node Calculate the deviation term of material flow conservation The material flow conservation deviation term The calculation formula is: ; in, Represents a node The set consisting of all incoming predecessor nodes, Indicates from node Inflow node Material flow, Represents a node The set consisting of all outgoing successor nodes. Indicates from node Flow to node The mass flow rate; when the mass flow conservation deviation term Exceeding the preset conservation deviation threshold At that time, with nodes The ratio of the total outflow mass flow to the total inflow mass flow is used as the normalization coefficient. , will node All inflow edge characteristics Multiply by the normalization factor The correction is completed so that the total inflow mass flow rate equals the total outflow mass flow rate; in this embodiment, the conservation deviation threshold is... Set to 0.05, that is, when the node Normalization correction is triggered when the absolute difference between the inflow and outflow mass flows exceeds 5% of the total inflow mass flow; Production topology diagram There are 10 intermediate production nodes involved in the conservation deviation calculation. The raw material pretreatment unit, which is the first node, and the distillation unit, which is the last node, do not participate in the conservation deviation verification. Construct a convolutional state extraction network for the material flow constraint graph, wherein the material flow constraint graph convolutional state extraction network is composed of... It consists of stacked graph convolutional units, each containing an attention weight calculation sublayer and a material flow constraint correction sublayer; the attention weight calculation sublayer adjusts the node... The current layer input feature vector and its neighbor nodes After concatenating the current layer's input feature vectors, the resulting nodes undergo linear transformation, nonlinear activation, and normalization along the dimensions of neighboring nodes. For nodes Attention weight coefficient In this embodiment, the attention weight coefficient The specific calculation process is as follows: [The node is...] In the Input feature vector of each graph convolutional unit with neighboring nodes In the Input feature vector of each graph convolutional unit By concatenating along the feature dimensions, we obtain the concatenated feature vector. , dimension ,in Indicates the first Each graph convolutional unit takes the dimension of the input feature vector as its input; the concatenated feature vector is then compared with the dimension... Learnable attention parameter vector After performing a dot product, the nodes are processed by the LeakyReLU activation function with a negative slope coefficient of 0.2 to obtain the nodes. For nodes raw attention score ; for nodes The original attention scores of all neighboring nodes are Softmax normalized along the dimension of the neighboring nodes to obtain the attention weight coefficients. The material flow constraint correction sublayer is represented by the constraint strength coefficient in the neighborhood weighted aggregation result. Superimposed material flow conservation deviation term Perform constraint correction; The node feature update formula for each graph convolutional unit is: ; in, Indicates the first The nodes output by each graph convolutional unit eigenvectors, Represents a non-linear activation function. Represents a node In production topology The set of all direct neighbor nodes in the set. Indicates the first Nodes in a graph convolutional unit For nodes Attention weight coefficient, Indicates the first Learnable linear transformation matrix of each graph convolutional unit This represents the strength coefficient of the conservation constraint on material flow. Represents a node The deviation term of the material flow conservation, The input feature vector of the first graph convolutional unit The initial feature vector of the node; after After the feature updates layer by layer of each graph convolutional unit, the final output feature vector of each node is... Arranged sequentially by node number, this is a spatiotemporal state feature matrix of the production process. ,in Dimensions representing spatiotemporal state characteristics The Row corresponding node The spatiotemporal state feature vector; in this embodiment, Nonlinear activation function Take the ReLU function, constraint strength coefficient The learnable linear transformation matrix of three graph convolutional units , and The dimensions are as follows: , and Dimensions of spatiotemporal state features Spatiotemporal state feature matrix The dimension is .

[0021] Optionally, in the material flow constraint graph convolutional state extraction network, edge features are... Explicit learnable embedding is performed, and material flow information is integrated into the attention weight calculation sublayer using additive bias modulation. This ensures that the attention weights between nodes are simultaneously constrained by both node feature similarity and actual material flow intensity, enhancing the graph convolution feature propagation's ability to physically perceive the production topology and logistics structure. In this embodiment, for the first... Production process in a graph convolutional unit neighboring nodes The corresponding directed edge With learnable scalar weights and learnable bias scalar opposite edge features Perform a linear affine transformation to obtain the first... Directed edges in each graph convolution unit Flow modulation term The calculation formula is: ; in, Indicates the first Directed edges in each graph convolution unit Flow modulation term, Indicates the first The learnable edge feature weight scalar corresponding to each graph convolutional unit Indicates the production process To the production stage Material flow edge characteristics, Indicates the first The learnable edge feature bias scalar corresponding to each graph convolutional unit; the flow modulation term The original attention score of the sub-layer is calculated by incorporating attention weights in an additive manner. The attention score after flow modulation is obtained. The calculation formula is: ; in, Indicates the first Nodes in a graph convolutional unit For nodes Attention score after flow modulation This represents the original attention score output after the attention weight calculation sublayer is processed by the LeakyReLU activation function. This represents the corresponding flow modulation term; the attention score is calculated based on the flow modulation. Replace the original attention score For nodes The attention scores of all neighboring nodes after traffic modulation are softmax normalized along the neighboring node dimension to obtain the traffic modulation attention weight coefficients. The flow modulation attention weight coefficient Replace the original attention weight coefficient Participating in the Node feature update of each graph convolutional unit.

[0022] B2: The spatiotemporal state characteristics of the production process are input into the adaptive weight generation network. The adaptive weight generation network dynamically generates the network weights of the multilayer perceptron based on the fluctuation characteristics of the current production conditions. The multilayer perceptron performs a nonlinear mapping on the basic carbon emission factor library and outputs real-time dynamic carbon emission factors, including: The spatiotemporal state feature matrix obtained in step B1 Perform global average pooling along the node dimension. The Middle List all The mean of the features of each node constitutes the global feature vector of the working condition. The One portion, Global feature vector of working conditions Reflection Moment The overall operating status of the entire production chain; in real time global feature vector of working conditions With time global feature vector of working conditions The difference constitutes the characteristic vector of operating condition fluctuation. ; The global feature vector of the working condition With operating condition fluctuation characteristic vector By concatenating along the dimensions, we obtain the comprehensive feature vector of the working conditions. In this embodiment, the global feature vector of the working condition The dimension is 64, and the characteristic vector of operating condition fluctuation is... The dimension is 64, and the comprehensive feature vector of the working condition is 64. The dimension is 128; Construct a working condition adaptive weight generation network, wherein the working condition adaptive weight generation network is composed of... The system consists of stacked fully connected layers, forming a comprehensive feature vector based on the working conditions. As input, the output dimension of the output layer is set to the sum of the parameters of the weight matrices and bias vectors of all layers in the multilayer perceptron. The output vector of the adaptive weight generation network is decomposed into segments according to the dimensions of the weight matrix and bias vector of each layer of the multilayer perceptron, and then reassembled to form the complete set of network weight parameters of the multilayer perceptron. In this embodiment, The number of output neurons in the three fully connected layers of the adaptive weight generation network are 512, 1024, and 1868, respectively. The first two fully connected layers use the ReLU activation function, while the third fully connected layer does not. The multilayer perceptron uses a three-layer fully connected structure, with the first layer having an input dimension of... The first layer has an output dimension of 32 and uses the ReLU activation function; the second layer has an input dimension of 32 and an output dimension of 32, also using the ReLU activation function; the third layer has an input dimension of 32 and an output dimension of... No activation function is used; the dimension of the weight matrix of the first layer of the multilayer perceptron is... The bias vector has a dimension of 32, and the weight matrix of the second layer has a dimension of [missing value]. The bias vector has a dimension of 32, and the weight matrix of the third layer has a dimension of [missing value]. The bias vector has a dimension of 12; Construct a basic carbon emission factor library, which stores the basic carbon emission factor vectors pre-calibrated for each production stage under standard operating conditions. ,in Dimensions representing carbon emission factors; As the current effective network weights, the basic carbon emission factor vector is processed through a multilayer perceptron. Perform nonlinear mapping to output a real-time dynamic carbon emission factor vector. The mapping formula is: ; in, Indicates time Real-time dynamic carbon emission factor vector, This represents the basic carbon emission factor vector in the basic carbon emission factor library. This indicates that the adaptive weight generation network is at time [time value missing]. The set of weight parameters of the output multilayer perceptron network. This represents the forward computation process of the multilayer perceptron; in this embodiment, the basic carbon emission factor vector... The 12 components were initially calibrated for each of the 12 production stages according to the industry benchmark carbon emission factor specifications. The dimensions of each component are as follows: Equivalent / standard unit; for real-time dynamic carbon emission factor vector The Each component Perform out-of-bounds checks. ,in This indicates the lower bound of the reasonable range for carbon emission factors. This represents the upper bound of the reasonable range for carbon emission factors, when... When Set as ,when When Set as In this embodiment, Equivalent / standard unit Equivalent / standard unit, out-of-bounds check performed on each forward inference step. Each component is executed one by one; such as Figure 2 As shown, the dynamic carbon emission factors of the three core production processes of gasification, purification and synthesis have been adaptively varied with operating conditions over a continuous 24-hour period.

[0023] Optionally, during the construction of the comprehensive feature vector of operating conditions, in the global feature vector of operating conditions... With operating condition fluctuation characteristic vector Building upon this foundation, a time-series sliding window historical state enhancement mechanism is further introduced, which enhances the historical state of continuous sliding windows. Lightweight self-attention convergence is performed on the global feature vectors of the operating conditions at each historical moment to obtain the historical trend vector of the operating conditions. This trend vector is then concatenated with the global feature vector and the fluctuation feature vector of the operating conditions to form an enhanced comprehensive feature vector. This vector is then input into the adaptive weight generation network for the operating conditions. This allows the adaptive weight generation network to simultaneously perceive the current state, short-term fluctuations, and the evolution trend of the operating conditions over a longer timescale when generating weights for the multilayer perceptron, thus improving its ability to identify and distinguish gradual changes in the operating conditions and transient disturbances. In this embodiment, the time-series sliding window length... Set to 8; set the time... At that time common The global feature vectors of the operating conditions at each historical moment are arranged in chronological order to form a historical operating condition sequence matrix. ,in The OK, Corresponding time The global feature vector of the working condition at that location. ; Historical operating condition sequence matrix Projection matrix queried through learnable history Learnable historical key projection matrix and learnable history projection matrix Perform a linear transformation to obtain the historical self-attention query matrix. Historical self-attention key matrix and historical self-attention value matrix The historical self-attention output matrix The calculation formula is: ; in, This represents the historical self-attention output matrix. Represents the historical self-attention query matrix. Representing the historical self-attention key matrix transpose, Represents the historical self-attention value matrix. The projection dimension representing historical self-attention. Represents the normalized exponential function; takes the historical self-attention output matrix. The Row vectors as historical trend vectors of operating conditions The historical trend vector of the operating conditions Near-terminal signals were encoded using a self-attention approach. The temporal correlation information between the global feature vectors of the operating condition at each historical moment; the global feature vectors of the operating condition Operating condition fluctuation characteristic vector Historical trend vector of working conditions By concatenating the features along the dimensional order, we obtain the enhanced working condition comprehensive feature vector. To enhance the comprehensive feature vector of operating conditions Replace the original working condition comprehensive feature vector An adaptive weight generation network is used to generate weights based on input conditions; in this embodiment, the projection dimension of historical self-attention is... Enhance the comprehensive feature vector of working conditions The dimension is The input dimension of the adaptive weight generation network is adjusted from 128 to 160, while the structure and output dimension of the remaining network layers remain unchanged.

[0024] B3: The spatiotemporal characteristics of the production process and real-time dynamic carbon emission factors are input into the dual-stream feature alignment and fusion module for feature concatenation and alignment. After processing by a feedforward neural network, the carbon emission slice features of each production process are output, including: A dual-stream feature alignment and fusion module is constructed, wherein the dual-stream feature alignment and fusion module uses the spatiotemporal state feature matrix obtained in step B1. and the real-time dynamic carbon emission factor vector obtained in step B2 As input; for the spatiotemporal state feature matrix Learnable linear transformation matrix The linear transformation yields the query matrix. ; Real-time dynamic carbon emission factor vector Considered as being by A sequence consisting of scalars, where the is the ? Each component With learnable weight matrix The row vector Perform scaling embedding to obtain the first The key vector of each carbon emission factor ,Depend on Key vectors Arranged in rows to form a key matrix In the same way, with a learnable weight matrix The row vector right Perform scaling embedding, by Value vectors Arranged by rows to form a value matrix , In this embodiment, the linear transformation matrix can be learned. The dimension is Learnable weight matrix The dimension is Learnable weight matrix The dimension is Query matrix The dimension is Key matrix The dimension is Value matrix The dimension is The formula for calculating dual-stream feature alignment is: ; in, This represents the aligned feature matrix. Represents the query matrix. Represents the key matrix, This represents the transpose of the key matrix. Represents a value matrix, The dimension of the key vector. This represents an exponential function normalized along the second dimension. The Behavioral production process State characteristics for all Attention-weighted fusion results for each carbon emission factor dimension; aligning the feature matrix With spatiotemporal state feature matrix Concatenating along the feature dimensions yields a cascaded feature matrix. ,in The Travel The Action and The Row concatenation; cascaded feature matrix The feedforward neural network performs independent nonlinear transformations on the row vectors corresponding to each node, outputting the carbon emission slice feature matrix for each production stage. , The Corresponding production process The carbon emission slice feature vector; in this embodiment, the feedforward neural network consists of two fully connected layers. The first fully connected layer maps the input dimension of 96 to 128, and the activation function is the ReLU function; the second fully connected layer maps the 128 dimensions to 64 dimensions, and does not use an activation function; the feedforward neural network... The 12 rows of vectors are processed independently, and the nodes share the same feedforward neural network parameters.

[0025] B4: Input the carbon emission slice characteristics of each production stage into the hierarchical node aggregation module for graph structure node aggregation processing, and output the carbon footprint calculation results of the entire link data through fully connected layer mapping, including: Construct a hierarchical node aggregation module, which uses the carbon emission slice feature matrix obtained in step B3. As the initial input, through A series of stacked node aggregation layers produce a topology graph. All node features are aggregated layer by layer; the normalized adjacency matrix From the production topology diagram The adjacency matrix is ​​obtained by adding the identity matrix and then normalizing the longitude; in this embodiment, the normalized adjacency matrix is... The specific calculation steps are as follows: Construct the production topology map Corresponding adjacency matrix If the production process With the production process There are directed edges between them. but ,otherwise In the adjacency matrix Superimposed 12th order identity matrices Obtain the self-loop adjacency matrix ; Calculate the self-loop adjacency matrix degree matrix , Let be a diagonal matrix, and its first... diagonal elements The normalized adjacency matrix is ​​obtained through symmetric normalization. ;No. The calculation formula for the aggregation layer of individual nodes is: ; in, Indicates the first The node feature matrix output by the node aggregation layer. The carbon emission slice feature matrix is ​​input to the hierarchical node aggregation module. Indicates the first The learnable weight matrix of the node aggregation layer Represents a non-linear activation function. , This indicates the total number of node aggregation layers in the hierarchical node aggregation module; in this embodiment, Nonlinear activation function Using the ReLU function, the learnable weight matrix of the three-node aggregation layer , and All dimensions are ; go through After the aggregation layers of each node are converged, the final output node feature matrix is... Perform global average pooling along the node dimension, that is, ... All The mean of the rows is used to obtain the global graph representation vector. ; Global graph representation vectors are represented through fully connected layers. Perform linear mapping to output the carbon footprint calculation results of the entire data chain. The calculation formula is: ; in, This represents the carbon footprint calculation results of the entire data chain. This represents the weight matrix of the fully connected layer. This represents the bias vector of the fully connected layer. The global graph representation vector is used; in this embodiment, the weight matrix of the fully connected layer is used. The dimension is Bias vector As a scalar, the carbon footprint accounting results It is a scalar with dimensions of . Equivalent; carbon footprint accounting results Perform nonnegativity check when When Set to 0 to ensure that the carbon footprint accounting results conform to the physical meaning of non-negative constraints.

[0026] The material flow constraint graph convolutional state extraction network, the working condition adaptive weight generation network, the multilayer perceptron, the feedforward neural network in the dual-flow feature alignment and fusion module, and all learnable parameters in the hierarchical node aggregation module involved in steps B1 to B4 above are optimized using an end-to-end joint training method. In this embodiment, the training dataset consists of hourly production records from a coal chemical enterprise for 12 consecutive months, containing a total of 8640 training samples. Each training sample corresponds to the full-link operation data within one hour of production and the carbon footprint label value certified by a third-party verification agency. The training set and validation set were split in an 8:2 ratio, with 6912 samples in the training set and 1728 samples in the validation set. Mean squared error was used as the main loss function for carbon footprint regression. The Adam optimizer was used to update all learnable parameters, with an initial learning rate of 0.001. The first-order momentum decay coefficient of the Adam optimizer was... Second-order momentum decay coefficient Numerical stability term The training consisted of 300 rounds, using a cosine annealing strategy to adjust the learning rate. The minimum learning rate for cosine annealing was set to... During the backpropagation process of each training batch, the gradient norm of all learnable parameters of the adaptive weight generation network is clipped, with the gradient clipping threshold set to 1.0 to prevent gradient explosion during the generation of multilayer perceptron weights. The model training uses the root mean square error on the validation set as the evaluation metric, and the model parameters corresponding to the minimum value of the root mean square error on the validation set are saved as the final model weights.

[0027] Optionally, after completing the full-chain carbon footprint accounting results After outputting the results, the carbon footprint accounting results are decomposed according to the contribution of each production stage, and the carbon footprint contribution value of each production stage is output. This supports targeted traceability of carbon emission reduction production processes; in this embodiment, the node... carbon footprint contribution value The node feature matrix is ​​finally output by the hierarchical node aggregation module. The The row feature vector is substituted into the fully connected layer for mapping and the bias vector is evenly distributed to obtain the result. The calculation formula is as follows: ; in, Indicates the production process carbon footprint contribution value, This represents the node feature matrix that is ultimately output by the hierarchical node aggregation module. The row feature vector, Representing the production topology diagram The total number of nodes; in this embodiment, The carbon footprint traceability results include the carbon footprint contribution value and contribution ratio of each of the 12 production stages.

[0028] Example 2: This invention also discloses a carbon footprint AI accounting system based on end-to-end data, comprising the following modules: Spatiotemporal state feature extraction module: used to map the collected multi-source heterogeneous flow data such as coal consumption, electricity consumption and heat output into the node features and edge features of the production topology graph, and extract the spatiotemporal state features of the production process by introducing a material flow constraint graph convolutional state extraction network with material flow conservation constraints. Dynamic carbon emission factor generation module: It is used to input the spatiotemporal state characteristics of the production process into the working condition adaptive weight generation network. The network dynamically generates the network weights of the multilayer perceptron according to the fluctuation characteristics of the current production working condition, and performs nonlinear mapping on the basic carbon emission factor library through the multilayer perceptron to output the real-time dynamic carbon emission factor. Dual-stream feature alignment and fusion module: used to concatenate and align the spatiotemporal state features of the production process with the real-time dynamic carbon emission factor, and output the carbon emission slice features of each production process after processing by a feedforward neural network. Carbon footprint accounting result output module: It is used to input the carbon emission slice characteristics of each production link into the hierarchical node aggregation module for graph structure node aggregation processing, and finally output the carbon footprint accounting result of the whole link data through full connection layer mapping.

[0029] It should be noted that the sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0030] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0031] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A carbon footprint AI accounting method based on end-to-end data, characterized in that, Includes the following steps: B1: The collected coal consumption, electricity consumption and heat output multi-source heterogeneous flow data are mapped into the node features and edge features of the production topology graph. The spatiotemporal state features of the production process are extracted by the material flow constraint graph convolutional state extraction network that introduces material flow conservation constraints. B2: Input the spatiotemporal state characteristics of the production process into the working condition adaptive weight generation network. The working condition adaptive weight generation network dynamically generates the network weights of the multilayer perceptron based on the fluctuation characteristics of the current production working condition. The multilayer perceptron performs nonlinear mapping on the basic carbon emission factor library and outputs real-time dynamic carbon emission factors. B3: Input the spatiotemporal state characteristics of the production process and the real-time dynamic carbon emission factor into the dual-flow feature alignment and fusion module for feature concatenation and alignment. After processing by the feedforward neural network, the carbon emission slice features of each production process are output. B4: Input the carbon emission slice characteristics of each production link into the hierarchical node aggregation module for graph structure node aggregation processing, and output the carbon footprint calculation results of the whole link data through full connection layer mapping.

2. The carbon footprint AI accounting method based on end-to-end data as described in claim 1, characterized in that, Step B1 includes: Build a production topology diagram ,in Represents a set of nodes, a set of nodes Each node This corresponds to a production stage in the industrial production process. , Representing the production topology diagram Total number of nodes in the system; Let edge set be an edge set. Each edge Indicates the production process With the production process There is a direct material transfer relationship between them; based on the collected data on coal consumption, electricity consumption, and heat output, the production process... The corresponding coal consumption sample value, electricity consumption sample value, and heat output sample value are concatenated to form the initial feature vector of the node. ,in The dimension representing the initial features of a node; the production process Flow to the production process The material flow rate value is denoted as the edge feature. .

3. The carbon footprint AI accounting method based on end-to-end data as described in claim 2, characterized in that, Step B1 further includes: Production topology diagram Each intermediate production node Calculate the deviation term of material flow conservation The material flow conservation deviation term The calculation formula is: ； in, Represents a node The set consisting of all incoming predecessor nodes, Indicates from node Inflow node Material flow, Represents a node The set consisting of all outgoing successor nodes. Indicates from node Flow to node The mass flow rate; when the mass flow conservation deviation term Exceeding the preset conservation deviation threshold At that time, with nodes The ratio of the total outflow mass flow to the total inflow mass flow is used as the normalization coefficient. , will node All inflow edge characteristics Multiply by the normalization factor Complete the correction so that the total inflow of materials equals the total outflow of materials. Construct a convolutional state extraction network for the material flow constraint graph, wherein the material flow constraint graph convolutional state extraction network is composed of... It consists of stacked graph convolutional units, each containing an attention weight calculation sublayer and a material flow constraint correction sublayer; the attention weight calculation sublayer adjusts the node... The current layer input feature vector and its neighbor nodes After concatenating the current layer's input feature vectors, the resulting nodes undergo linear transformation, nonlinear activation, and normalization along the dimensions of neighboring nodes. For nodes Attention weight coefficient The material flow constraint correction sublayer is represented by the constraint strength coefficient in the neighborhood weighted aggregation result. Superimposed material flow conservation deviation term Perform constraint correction; The node feature update formula for each graph convolutional unit is: ； in, Indicates the first The nodes output by each graph convolutional unit eigenvectors, Represents a non-linear activation function. Represents a node In production topology The set of all direct neighbor nodes in the set. Indicates the first Nodes in a graph convolutional unit For nodes Attention weight coefficient, Indicates the first Learnable linear transformation matrix of each graph convolutional unit This represents the strength coefficient of the conservation constraint on material flow. Represents a node The deviation term of the material flow conservation, The input feature vector of the first graph convolutional unit The initial feature vector of the node; after After the feature updates layer by layer of each graph convolutional unit, the final output feature vector of each node is... Arranged sequentially by node number, this is a spatiotemporal state feature matrix of the production process. ,in Dimensions representing spatiotemporal state characteristics The Row corresponding node The spatiotemporal state feature vector.

4. The carbon footprint AI accounting method based on end-to-end data as described in claim 1, characterized in that, Step B2 includes: The spatiotemporal state feature matrix obtained in step B1 Perform global average pooling along the node dimension. The Middle List all The mean of the features of each node constitutes the global feature vector of the working condition. The One portion, Global feature vector of working conditions Reflection Moment The overall operating status of the entire production chain; in real time global feature vector of working conditions With time global feature vector of working conditions The difference constitutes the characteristic vector of operating condition fluctuation. ; The global feature vector of the working condition With operating condition fluctuation characteristic vector By concatenating along the dimensions, we obtain the comprehensive feature vector of the working conditions. ; Construct a working condition adaptive weight generation network, wherein the working condition adaptive weight generation network is composed of... The system consists of stacked fully connected layers, forming a comprehensive feature vector based on the working conditions. As input, the output dimension of the output layer is set to the sum of the parameters of the weight matrices and bias vectors of all layers in the multilayer perceptron. The output vector of the adaptive weight generation network is decomposed into segments according to the dimensions of the weight matrix and bias vector of each layer of the multilayer perceptron, and then reassembled to form the complete set of network weight parameters of the multilayer perceptron. .

5. The carbon footprint AI accounting method based on end-to-end data according to claim 4, characterized in that, Step B2 further includes: Construct a basic carbon emission factor library, which stores the basic carbon emission factor vectors pre-calibrated for each production stage under standard operating conditions. ,in Dimensions representing carbon emission factors; As the current effective network weights, the basic carbon emission factor vector is processed through a multilayer perceptron. Perform nonlinear mapping to output a real-time dynamic carbon emission factor vector. The mapping formula is: ； in, Indicates time Real-time dynamic carbon emission factor vector, This represents the basic carbon emission factor vector in the basic carbon emission factor library. This indicates that the adaptive weight generation network is at time [time value missing]. The set of weight parameters of the output multilayer perceptron network. This represents the forward computation process of the multilayer perceptron; it represents the real-time dynamic carbon emission factor vector. The Each component Perform out-of-bounds checks. ,in This indicates the lower bound of the reasonable range for carbon emission factors. This represents the upper bound of the reasonable range for carbon emission factors, when... When Set as ,when When Set as .

6. The carbon footprint AI accounting method based on end-to-end data as described in claim 1, characterized in that, Step B3 includes: A dual-stream feature alignment and fusion module is constructed, wherein the dual-stream feature alignment and fusion module uses the spatiotemporal state feature matrix obtained in step B1. and the real-time dynamic carbon emission factor vector obtained in step B2 As input; for the spatiotemporal state feature matrix Learnable linear transformation matrix The linear transformation yields the query matrix. ; Real-time dynamic carbon emission factor vector Considered as being by A sequence consisting of scalars, where the is the ? Each component With learnable weight matrix The row vector Perform scaling embedding to obtain the first The key vector of each carbon emission factor ,Depend on Key vectors Arranged in rows to form a key matrix In the same way, with a learnable weight matrix The row vector right Perform scaling embedding, by Value vectors Arranged by rows to form a value matrix , The formula for calculating dual-stream feature alignment is: ； in, This represents the aligned feature matrix. Represents the query matrix. Represents the key matrix. This represents the transpose of the key matrix. Represents a value matrix, This represents the dimension of the key vector. This represents an exponential function normalized along the second dimension. The Behavioral production process State characteristics for all Attention-weighted fusion results for each carbon emission factor dimension; aligning the feature matrix With spatiotemporal state feature matrix Concatenating along the feature dimensions yields a cascaded feature matrix. ,in The Travel The Action and The Row concatenation; cascaded feature matrix The feedforward neural network performs independent nonlinear transformations on the row vectors corresponding to each node, outputting the carbon emission slice feature matrix for each production stage. , The Corresponding production process The carbon emission slice feature vector.

7. The carbon footprint AI accounting method based on end-to-end data as described in claim 1, characterized in that, Step B4 includes: Construct a hierarchical node aggregation module, which uses the carbon emission slice feature matrix obtained in step B3. As the initial input, through A series of stacked node aggregation layers produce a topology graph. All node features are aggregated layer by layer; the adjacency matrix is ​​normalized. From the production topology diagram The adjacency matrix is ​​obtained by adding the identity matrix and then normalizing the longitude; the first... The calculation formula for the aggregation layer of individual nodes is: ； in, Indicates the first The node feature matrix output by the node aggregation layer. The carbon emission slice feature matrix is ​​input to the hierarchical node aggregation module. Indicates the first The learnable weight matrix of the node aggregation layer Represents a non-linear activation function. , This indicates the total number of node aggregation layers in the hierarchical node aggregation module; go through After the aggregation layers of each node are converged, the final output node feature matrix is... Perform global average pooling along the node dimension, that is, ... All The mean of the rows is used to obtain the global graph representation vector. ; Global graph representation vectors are represented through fully connected layers. Perform linear mapping to output the carbon footprint calculation results of the entire data chain. The calculation formula is: ； in, This represents the carbon footprint calculation results of the entire data chain. This represents the weight matrix of the fully connected layer. This represents the bias vector of the fully connected layer. Represents a global graph as a vector; calculates the carbon footprint. Perform nonnegativity check when When Set to 0 to ensure that the carbon footprint accounting results conform to the physical meaning of non-negative constraints.

8. A carbon footprint AI accounting system based on end-to-end data, characterized in that, include: Spatiotemporal state feature extraction module: used to map the collected multi-source heterogeneous flow data such as coal consumption, electricity consumption and heat output into the node features and edge features of the production topology graph, and extract the spatiotemporal state features of the production process by introducing a material flow constraint graph convolutional state extraction network with material flow conservation constraints. Dynamic carbon emission factor generation module: It is used to input the spatiotemporal state characteristics of the production process into the working condition adaptive weight generation network. The network dynamically generates the network weights of the multilayer perceptron according to the fluctuation characteristics of the current production working condition, and performs nonlinear mapping on the basic carbon emission factor library through the multilayer perceptron to output the real-time dynamic carbon emission factor. Dual-stream feature alignment and fusion module: used to concatenate and align the spatiotemporal state features of the production process with the real-time dynamic carbon emission factor, and output the carbon emission slice features of each production process after processing by a feedforward neural network. Carbon footprint accounting result output module: It is used to input the carbon emission slice characteristics of each production link into the hierarchical node aggregation module for graph structure node aggregation processing, and finally output the carbon footprint accounting result of the whole link data through full connection layer mapping. To achieve the carbon footprint AI accounting method based on full-link data as described in any one of claims 1-7.

Images