A method and system for tracking carbon emission responsibility of electric vehicle life cycle by coupling dynamic interval prediction, graph neural network and input-output table

Abstract

This invention discloses a method and system for cross-domain tracking of carbon emission responsibility throughout the lifecycle of electric vehicles, coupled with dynamic interval prediction, graph neural networks, and input-output tables. It relates to the interdisciplinary fields of carbon emission accounting and artificial intelligence. The method outputs emission intervals for the three stages of production, use, and scrapping through a Bayesian quantile regression neural network, and updates the model online based on physical thresholds. It constructs a heterogeneous graph including a forward supply chain and a reverse recycling network, and uses a relational graph convolutional network to learn multi-hop emission transfer coefficients. The interval prediction and transfer amount are coupled to generate the total lifecycle responsibility interval and the cross-provincial responsibility matrix. This invention achieves dynamic adaptation, high-precision attribution, and verifiable traceability.

Description

Technical Field

[0001] This invention relates to the field of carbon emission accounting and artificial intelligence, and more specifically to a method and system for cross-domain tracking of carbon emission responsibility throughout the life cycle of electric vehicles that couples dynamic interval prediction, graph neural networks and input-output tables. Background Technology

[0002] The lifecycle carbon emissions of electric vehicles encompass multiple stages, including upstream material production, vehicle manufacturing, charging during use, and end-of-life dismantling and recycling. Each stage involves material flow and emission transfer between different provinces. Accurately calculating and allocating cross-regional carbon emission responsibilities is crucial for formulating fair carbon reduction policies, optimizing supply chain layout, and promoting the development of a circular economy. Current research attempts to utilize artificial intelligence to capture data, construct lifecycle databases, and couple artificial neural networks with emission accounting models to establish a fast CO2 solver. Simultaneously, it analyzes the spatial distribution of vehicle lifecycles based on multi-regional input-output tables (MRIO). However, these methods still suffer from the following fundamental shortcomings: The output is a point estimate or a static interval, lacking dynamic adaptive capability: Existing models (including the aforementioned fast solvers) output a single numerical value or a fixed quantile, failing to express uncertainty. More importantly, the model parameters are frozen once training is complete, unable to perceive data distribution drift caused by changes in energy structure (such as a sudden drop in the proportion of thermal power in a province) or improvements in battery recycling technology. When drift occurs, the model needs to be retrained offline, with response lags ranging from weeks to months, leading to a decrease in the probability of interval coverage (PICP).

[0003] Linear input-output models fail to capture multi-hop and reverse paths: Traditional MRIO uses the Leontief inverse matrix, assuming a constant consumption coefficient, and can only calculate the linear proportion of direct or total consumption. Real-world supply chains contain numerous multi-hop indirect paths (e.g., steel in province B → parts processing in province C → vehicle manufacturing in province A), and the reverse recycling network at the end-of-life stage (waste batteries flowing from the province of use to the province of processing) is completely unmodeled. Existing methods ignore the emission contributions of intermediate provinces, resulting in significant attribution errors.

[0004] The contradiction of "low-use emissions, high material emissions" is not expressed in a range: emissions from electric vehicles are lower during use due to the cleanliness of charging, but emissions from upstream material production may be higher, while the end-of-life stage may bring net emission reductions (e.g., emissions from recycling materials are much lower than those from mining, processing, smelting, and manufacturing processes). Existing methods simply add up point values, failing to output a joint range.

[0005] In summary, there is an urgent need for a lifecycle tracking method and system that integrates physical deterministic models and intelligent enhancement technologies to achieve dynamic adaptive interval output and nonlinear multi-hop path capture. Summary of the Invention

[0006] In view of this, the present invention provides a method and system for cross-domain tracking of carbon emission responsibility in the life cycle of electric vehicles by coupling dynamic interval prediction, graph neural network and input-output table, so as to solve the problems in the background art.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: On one hand, this invention discloses a cross-domain tracing method for the lifecycle carbon emission responsibility of electric vehicles that couples dynamic interval prediction, graph neural networks, and input-output tables, including: Acquire multi-source time-series data on electric vehicles during the production, use, and scrapping stages, and construct a heterogeneous graph including nodes for provinces, industries, products, and scrapped products. The emission ranges for the production, use, and disposal stages are output using a Bayesian quantile regression neural network, and the network parameters for each stage are updated online based on data distribution drift. Message passing is performed on the heterogeneous graph through a relational graph convolutional network, where the set of edge types includes input-output edges and inter-provincial trade edges in the forward supply chain, as well as end-of-life recycling edges and recycled material edges in the reverse recycling network, in order to learn the forward multi-hop emission transfer coefficient and the reverse multi-hop emission transfer coefficient in parallel. The cross-regional transfer amount of implicit emissions is calculated based on the forward and reverse multi-hop emission transfer coefficients, and the three-stage emission intervals are coupled with the transfer amount according to the interval arithmetic rules to generate the life cycle total carbon emission responsibility interval and cross-regional responsibility allocation results.

[0008] This technical solution is the first to unify the three stages of production, use, and disposal within a single intelligent framework, solving fundamental problems such as the fragmentation of existing technologies, output point estimation, and inability to handle reverse logistics. By organically integrating dynamic interval prediction with a graph neural network input-output model, the system can simultaneously output statistically guaranteed emission intervals and automatically track multi-hop implicit emission paths across provinces, achieving a leap from "post-event offline accounting" to "real-time adaptive traceability."

[0009] Preferably, the data distribution drift includes physical threshold judgments of the provincial power structure change rate and the battery recycling technology efficiency improvement rate; the online update adopts Bayesian quantile learning with conjugate normal-inverse gamma prior, which only applies to the weights of the output layer of the Bayesian quantile regression neural network at each stage, assuming that the output layer is a linear combination of the input feature embeddings; for univariate input features The posterior parameter update formula is:

[0010] For multidimensional input features, a diagonal approximation is used, and the above update is applied independently to each feature dimension; where... and These are the prior mean and variance; This refers to the newly arrived small batch of data; and Input characteristics and actual emission values; To observe the noise variance; and The updated posterior mean and variance; the retraining trigger condition for the model at each stage is that the interval coverage probability (PICP) is below 90% for three consecutive periods.

[0011] Compared to existing technologies that rely solely on static models trained offline in batches or general data-driven drift detection, this solution employs physically interpretable threshold judgments and combines them with an analytical update formula based on conjugate normal inverse gamma priors. This enables the model to rapidly adapt to scenarios such as sudden changes in energy structure and advancements in recycling technology, whereas traditional methods require weeks to months. Furthermore, using a 90% interval coverage probability as the industry reliability standard to trigger retraining ensures the credibility of the prediction results in practical policy applications. This significantly improves the model's robustness and practicality in dynamic environments, overcoming the technical bias that "static models cannot adapt to non-stationary data flows."

[0012] Preferably, the propagation formula of the relational graph convolutional network is:

[0013] The set of edge types This includes positive input-output sides, inter-provincial trade sides, emission factor sides, as well as reverse end-of-life product recycling sides and recycled material sides; For the first Layer nodes The hidden state vector; It is a non-linear activation function; To pass through relationships With nodes A set of connected neighbor nodes; This is the normalization constant; For relationship In the The trainable weight matrix of the layer; It is a self-loop transformation matrix; The multi-hop emission transfer coefficient is calculated using path integral:

[0014] in, For the node ( ) to intermediate node The transfer coefficient; To start from the intermediate node To the target node ( The transfer coefficient of ) intermediate node Attention weights; This is the attention parameter vector; and Transformation matrix for query and key; This is a vector concatenation operation.

[0015] Compared to existing linear input-output models (which can only handle direct consumption coefficients) or general graph neural networks (which lack multi-hop path integral design for carbon emissions), this solution, by explicitly defining a two-hop path integral formula and utilizing a graph attention mechanism to weight intermediate nodes, can automatically capture indirect supply chain paths such as "steel in province B → processing in province C → complete vehicle in province A," as well as closed-loop circular paths such as "waste batteries in province A → recycling in province D → recycled materials → province E." This is a significant improvement over traditional input-output analysis, providing a learnable graph neural network solution for nonlinear, multi-hop material flows and emissions responsibility tracking.

[0016] Preferably, the amount of tacit cross-regional transfer of the implicit emissions includes: Forward transition period in the production stage:

[0017] Reverse transfer interval during the scrapping stage:

[0018] in, This refers to the amount of emissions implicitly transferred from province B to province A during the production phase. For the quality of upstream materials; This is the positive transition coefficient vector; The emission factor for producing this material in Province B; This refers to the amount of emissions transferred from province A to province B during the decommissioning phase. For the quality of used batteries; This is the reverse transfer coefficient vector; This represents the net emission factor for waste battery treatment in Province B (emissions during treatment minus emissions from virgin materials replaced by recycled materials; this can be negative); the superscripts low and high represent the lower and upper limits of the interval, respectively.

[0019] The aforementioned scheme organically combines the transfer coefficients learned by the graph neural network with physical mass flow and emission factors. This ensures that the emission responsibility transfer amount not only reflects the multi-hop relationships in the network topology but also retains the physical meaning of mass conservation and emission factors. In particular, defining the end-of-life transfer amount as a transfer from the point of use to the point of disposal, and adopting a mathematical form symmetrical to the forward transfer, unifies the implicit emission accounting of the forward supply chain and the reverse recycling network within an input-output framework for the first time. This allows responsibility allocation to possess both the self-learning capability of graph neural networks and physical interpretability, avoiding the shortcomings of pure black-box models that cannot trace the source of responsibility.

[0020] Preferably, the coupling interval operation specifically includes: Total lifecycle emissions liability range for target province A:

[0021] Interval addition and subtraction follow the standard interval arithmetic rules. , ; The ultimate responsibility area for supply / processing province B consists of three parts: production responsibility, processing responsibility, and optional extended producer responsibility. This refers to the direct emission range during the local production phase in Province A. This refers to the direct emission range during the local use phase in Province A. This refers to the direct emission range during the local scrapping phase in Province A. This refers to the amount of emissions implicitly transferred from province B to province A during the production phase. This refers to the amount of emissions transferred from province A to province B during the decommissioning phase.

[0022] in, This is an indicator function. It takes the value 1 when Province B is the main supplier of key materials for electric vehicles in Province A, and 0 otherwise. The main supplier is determined by a threshold based on the supply ratio. For example, it takes the value 1 when the value of materials supplied by Province B accounts for more than 20% of the total value of all key materials in Province A, and 0 otherwise. The value of materials supplied to Province B is a proportion of the total value of all materials, and is used only to allocate emissions during the production phase; emissions during the usage phase are separately accounted for by Province A, the place of use, and the province supplying electricity to it, based on the proportion of electricity inflow; emissions during the disposal phase are borne by Province B, the place of disposal.

[0023] Unlike existing technologies that simply add up the values ​​at each stage, the above scheme uses interval addition to preserve the uncertainty of each stage and achieves dynamic allocation of cross-provincial responsibility through transfer adjustments. Specifically, emissions during the usage stage are allocated to upstream provinces according to the proportion of material value or treatment volume, thus fairly reflecting the contradiction that "clean electricity use in the usage area leads to emission reductions, while the material production area bears high emissions." The above calculation formula is a systematic integration of interval uncertainty theory and input-output responsibility allocation rules, providing a quantifiable and comparable interval-based responsibility matrix for cross-provincial carbon assessment, supporting policymakers in making probabilistic decisions.

[0024] It is necessary to clearly define the areas of responsibility for the aforementioned provinces. and This is a policy-driven responsibility-sharing indicator, designed to strengthen the coordinated emission reduction efforts across the supply chain through the Extended Producer Responsibility (EPR) mechanism. This responsibility matrix is ​​used for inter-provincial comparisons, green procurement evaluations, and carbon offset negotiations. Adding an EPR item to Province B's responsibility framework links the carbon balance sheet of the upstream core supplier (Province B) to the technological level of the downstream vehicle manufacturer (Province A). This responsibility-linked mechanism is completely impossible to achieve with existing linear input-output models. Double-counting is a necessary technical means, not a technical flaw, for achieving two-way carbon reduction traction across the supply chain. This approach incentivizes Province B to assist Province A in improving its vehicle manufacturing processes, thereby achieving carbon reduction across the entire supply chain at its source.

[0025] Although the Graph Neural Network (R-GCN) of this invention strictly adheres to the law of conservation of matter at the physical layer (i.e., the transfer coefficient satisfies the condition that inflow equals outflow), the interval operations at the responsibility allocation layer allow for the superposition of responsibilities. This superposition is not used to directly generate a national carbon emission physical inventory. If the total amount of physical conservation is required, the responsibilities of each province need to be normalized and deduplicated. This normalization step is a derivative application of this invention but does not constitute a necessary technical feature of the main accounting process.

[0026] Preferably, the heterogeneous graph of the scrapping stage includes scrapped product nodes, whose attributes include battery type, retirement year, remaining capacity, and hazardous substance content. The weight of the reverse recycling edge is determined by actual reverse logistics statistics and the spatial distribution of recycling and processing enterprises.

[0027] Preferably, the Bayesian quantile regression neural network shares the underlying feature extraction layer at each stage, but the output layer is independent, in order to retain stage specificity and reduce the number of parameters.

[0028] The above scheme not only takes advantage of the commonalities in vehicle physical parameters (such as mass and material composition) during the production, use, and scrapping stages, but also reduces the number of parameters, prevents overfitting, and accelerates training by sharing the underlying layer; it also retains the unique emission patterns of each stage through an independent output layer (for example, the use stage is sensitive to the charging structure, and the scrapping stage is sensitive to the recycling technology).

[0029] On the other hand, the present invention discloses a cross-domain tracking system for the lifecycle carbon emission responsibility of electric vehicles that implements the method, comprising: The full-cycle data acquisition module has a built-in physical drift detector for acquiring multi-source data during the production, use, and disposal phases. A three-stage Bayesian quantile regression neural network engine supports online posterior updates and outputs emission ranges for production, use, and scrapping, respectively. The forward and reverse heterogeneous graph construction and R-GCN inference module embeds a graph attention mechanism to simultaneously learn the multi-hop emission transfer coefficients of the forward supply chain and the reverse recycling network; The full-cycle interval coupling responsibility allocation calculator is used to perform interval coupling calculations between the three-stage emission intervals and the forward and reverse transfer amounts; A visual front-end is used to display three-stage intervals, forward and reverse Sankey diagrams, and a cross-provincial responsibility matrix for the life cycle.

[0030] This system forms a closed loop from data source to final visualization. The various modules work collaboratively to achieve core functions such as dynamic adaptive interval prediction, forward and reverse multi-hop transfer learning, and interval coupling allocation. Compared with existing decentralized accounting tools, this system is the first to integrate uncertainty quantification and graph neural networks on a single platform. It can support large-scale real-time accounting from single vehicle models to different models in different provinces, and can output visualized reports for use by governments, enterprises, and third-party auditors. It has strong engineering practicality and promotional value.

[0031] Preferably, the three-stage Bayesian quantile regression neural network engine is deployed on a heterogeneous computing platform, wherein the stage engine is deployed on an FPGA accelerator card to achieve millisecond-level online inference, and the production and scrap stage engines are deployed on a CPU cluster due to their lower update frequency; the R-GCN inference module uses a dedicated graph neural network processor to handle large-scale message passing of forward and inverse mixed graphs.

[0032] This heterogeneous computing architecture fully considers the different real-time requirements of different stages. The usage stage needs to respond to real-time charging data from tens of thousands of electric vehicles and is sensitive to latency; while the production and scrapping stages have a lower update frequency and higher throughput requirements.

[0033] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method and system for cross-domain tracking of carbon emission responsibility of electric vehicles by coupling dynamic interval prediction, graph neural network and input-output table. By organically integrating dynamic adaptive interval prediction, graph neural network-enhanced input-output model and life cycle interval coupling operation, it realizes cross-domain dynamic tracking of carbon emission responsibility in the three stages of electric vehicle production, use and scrapping: it uses physical threshold-driven online Bayesian quantile learning to automatically respond to changes in energy structure and recycling technology, and always maintains a high reliability of emission interval coverage; with the help of relation graph convolutional network and attention-weighted multi-hop path integral, it captures the implicit emission transfer of the positive supply chain and the reverse recycling network in the input-output framework for the first time, which significantly reduces cross-provincial attribution error; it outputs the total life cycle responsibility interval and cross-provincial responsibility matrix through interval coupling operation, supporting probabilistic decision-making; and it meets the industrial-grade real-time accounting requirements by combining heterogeneous computing architecture, thus achieving substantial progress in dynamic adaptability, attribution accuracy, uncertainty expression and engineering performance. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0035] Figure 1 A flowchart of the method provided by the present invention; Figure 2 This is a system block diagram provided for the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] like Figure 1 As shown, the present invention provides a method for cross-domain tracing of electric vehicle lifecycle carbon emission responsibility by coupling dynamic interval prediction, graph neural network, and input-output table, comprising the following steps: Data Acquisition and Heterogeneous Graph Construction: Acquire multi-source time-series data on electric vehicles during the production, use, and scrapping stages, and construct a heterogeneous graph including nodes for provinces, industries, products, and scrapped products.

[0038] Dynamic range prediction: The emission ranges for the production, use and disposal stages are output by Bayesian quantile regression neural network (BQRNN), and the network parameters for each stage are updated online according to the data distribution drift (only the output layer is updated).

[0039] Multi-hop transfer coefficient learning: Message passing is performed on the heterogeneous graph through a relational graph convolutional network (R-GCN), while learning the multi-hop emission transfer coefficients of the forward supply chain and the reverse recycling network, and an attention mechanism is used to weight the inter-link paths.

[0040] Inter-regional transfer volume calculation: The uncertainty of transfer coefficients, material quality, and emission factors is expressed in an inter-regional manner to calculate the inter-regional transfer volume of implicit emissions at each stage.

[0041] Coupled interval calculation and responsibility allocation: Couple interval calculation is performed between the three-stage emission interval and the transfer interval to generate the life cycle total carbon emission responsibility interval and cross-regional responsibility allocation results.

[0042] This embodiment takes the 100,000 new electric private cars added in Province A in 2025 as the object, and tracks their life cycle carbon emissions from upstream raw material acquisition, vehicle manufacturing, use and charging to final scrapping and recycling, and quantifies the emission responsibility range that Provinces B, C, D and other provinces should bear.

[0043] Acquire multi-source time-series data on electric vehicles during the production, use, and scrapping stages, and construct a heterogeneous graph including nodes for provinces, industries, products, and scrapped products; configure a distributed asynchronous crawler to retrieve data from the websites and databases listed in Table 1 for 2023-2025.

[0044] Table 1

[0045] Constructing heterogeneous graphs ,in: Node set These represent 31 provinces, 42 industries, 26 key materials (steel, aluminum, copper, etc.), and end-of-life products (used batteries).

[0046] edge type Including: positive input-output sides Inter-provincial trade border Emission factor and reverse recycling of scrapped products , .

[0047] The node attributes of scrapped products include battery type (ternary lithium), retirement year (2025), remaining capacity (70% of nominal capacity), and content of hazardous substances (cobalt, nickel, etc.). The weight of the reverse recycling side is determined by the actual reverse logistics statistics (e.g., 60% flows to province E and 40% flows to province F) and the spatial distribution of recycling and processing enterprises.

[0048] Bayesian Quantile Regression Neural Network (BQRNN) Dynamic Interval Prediction: BQRNNs are constructed for the three stages of production, use, and disposal.

[0049] Network input features include: vehicle type code (private car), province code. Network structure: input layer. Fully connected layers (128 neurons per layer, ReLU activation) The output layer consists of two neurons, outputting the 0.05th and 0.95th quantiles respectively. The loss function is quantile loss. ,in .

[0050] Prior settings and online updates: Output layer weights Follows conjugate normal-inverse gamma prior , Online updates only apply to the output layer, which is assumed to be a linear combination of the input feature embeddings. When physical drift is detected (e.g., provincial power structure change rate, battery recycling technology efficiency improvement rate exceeding a 5% threshold), newly arrived mini-batch data is utilized. Update the posterior parameters. For a one-dimensional feature, the update formula is:

[0051] For multidimensional input features, a diagonal approximation is used, and the above update is applied independently to each feature dimension. Noise variance. The inverse gamma hyperparameter is updated to:

[0052] Retraining trigger condition: The probability of interval coverage (PICP) is below 90% for three consecutive periods (each period is one week). PICP is defined as the proportion of true values ​​falling within the predicted interval on the validation set. .

[0053] Training and prediction used data from eight provinces and cities from 2020 to 2024, with 70% used as the training set and 30% as the validation set. A deterministic physical model was used to calculate the precise emission values ​​for each sample as a supervision signal, and physical constraint regularization terms were added. This makes the median of the output interval approximate the value calculated by the physical model.

[0054] In June 2025, the system detected a decrease in the proportion of thermal power in Province B from 75% to 63% (a change rate of 12% > the 5% threshold), triggering an online update. The system automatically extracted 2000 new data points from the most recent 7 days (including new emission factors for Province B) and updated the output layer posterior parameters according to the above formula, completing the deployment in 2.8 hours. Prediction result: For the number of newly added electric private vehicles in Province A in 2025, the production stage emission range output by BQRNN is... The vehicle's usage period is as follows: Vehicle (total lifespan 10 years), the scrapping stage range is as follows On the validation set, the PICP (Personalized Interest Rate) was 94.2%, and the average interval width (MPIW) was 0.32. In contrast, the static model (without online updates) showed a decrease in PICP after a sudden change in the proportion of thermal power plants. This invention consistently maintains a success rate of over 93%.

[0055] Heterogeneous graphs and R-GCN multi-hop transition coefficient learning: Baseline MRIO Construction: Based on the input-output tables of 42 departments in 31 provinces from the National Bureau of Statistics, cross-provincial and sectoral material flows are estimated using a gravity model.

[0056] in, Indicates from the region department Input to area The flow of matter, For the region department Total output, For the region department Total input, Let be the distance between the two regions. Construct the direct consumption coefficient matrix. and Leontief inverse matrix .

[0057] R-GCN propagation: in heterogeneous graphs The propagation formula for the above relational graph convolutional network is:

[0058] in, To pass through relationships With nodes The set of connected neighbor nodes, The normalization constant (usually taken as ) ), For relationship In the The trainable weight matrix of the layer.

[0059] Multi-hop transfer coefficient: The formula for calculating the two-hop path integral is:

[0060] The weight of the intermediate node Automatic learning through graph attention mechanisms:

[0061] Case study demonstrating specific calculations (using steel as an example): Direct route: Steel plant in Province B → Vehicle factory in Province A .

[0062] Indirect path: Steel in Province B → Steel processing in Province D → Vehicle assembly plant in Province A. R-GCN learns the attention weights of the intermediate node in Province D. Two-hop coefficient .

[0063] Total transfer factor (take) (Preferring direct path)

[0064] The aluminum transfer coefficient from province C to province A is calculated similarly to 0.25 (0.22 directly + 0.03 indirectly).

[0065] Reverse transfer during the end-of-life stage: Waste batteries from Province A flow to Provinces E (60%) and F (40%), R-GCN learns the reverse transfer coefficient. , .

[0066] Interval-based transfer calculation: The transfer coefficient, material quality, and emission factor learned by the graph neural network are all expressed as intervals (the upper and lower limits are based on parameter uncertainty and model prediction intervals) to convey uncertainty.

[0067] Forward transfer range during production phase (Province B → Province A, Steel): Steel mass per vehicle (consider Fluctuation ), the transfer coefficient is taken Steel emission factors in Province B .but:

[0068] Aluminum contribution range in Province C .

[0069] Other materials total .

[0070] Total transfer volume range during the production phase:

[0071] Reverse transfer interval during the scrapping stage (Province A → Province E): Waste battery quality (Pick ), reverse transfer coefficient E Province recovers emission factors [-0.46, -0.44] : ; Upper limit (closest to zero): 245 × 0.57 × (-0.44) = -61.446; Lower limit (most negative): 255 × 0.59 × (-0.46) = -69.207;

[0072] Transfer volume range in Province F The coefficients are [0.37, 0.39]: Lower limit (most negative): 255 × 0.39 × (-0.42) = -41.769; Upper limit (closest to zero): 245 × 0.37 × (-0.40) = -36.260;

[0073] Total amount transferred during the scrapping phase: = The negative transfer amount during the aforementioned end-of-life phase means that provinces E and F, through advanced recycling processes, avoided emissions exceeding the emissions from the dismantling of waste batteries and the recycling of materials compared to virgin material mining, thus generating net emission reduction benefits. This negative value will be credited as "emission reduction credits" to the accounts of the provinces handling the waste batteries in subsequent responsibility allocation.

[0074] Coupled interval operations and responsibility allocation: Total emission liability range for target province A (per vehicle):

[0075]

[0076]

[0077] Interval addition and subtraction follow the standard interval arithmetic rules. , .

[0078] The final responsibility area for supplying / processing province B: According to the responsibility allocation rules of this invention, emissions during the production stage are borne by the material supplying provinces, emissions during the usage stage are borne by the usage location and the power supply province (not apportioned according to material source), and emissions during the disposal stage are borne by the disposal province (negative values ​​indicate emission reduction benefits). Furthermore, the extended producer responsibility principle is adopted, allowing emissions during the production stage to be further apportioned to upstream supplying provinces in proportion to the material value.

[0079] Taking Province B as an example (steel supply share) ,and ): Resp(Province B) = -0 + 1 × 0.4 × [6.9, 8.3] = [0.240, 0.256] + [2.76, 3.32] = [3.00, 3.576] / car; Province C (30% of aluminum supply): Resp(Province C) = [0.170, 0.195] + 0.3 × [6.9, 8.3] = [0.170, 0.195] + [2.07, 2.49] = [2.240, 2.685] / car; Province E (not responsible for production, only for liability): Resp(E Province) = 0 - =[-0.0692, -0.0614] / car Similar to Province F; Resp(Province F) = [-0.0418, -0.0363] / car; Province A, the place of use, bears the emissions during the use phase, direct emissions during the scrapping phase, emissions during the remaining production phase that are not shared by upstream provinces (1-40%-30%=30%), and transfer emissions of "other materials" that are not separately allocated to Provinces B and C [0.320,0.380] t; =[10.2,11.8]+[0.8,1.5]+0.3×[6.9,8.3]+[0.320,0.380]=[10.2,11.8]+[0.8,1.5]+[2.07,2.49]+[0.320,0.380]=[13.390,16.170] / car Total responsibility of each province: Resp(B) + Resp(C) + Resp(A) + Resp(E) + Resp(F) =[3.000,3.576]+[2.240,2.685]+[13.390,16.170]+[-0.0692,-0.0614]+[-0.0418,-0.0363]; =[3.000+2.240+13.390-0.0692-0.0418,3.576+2.685+16.170-0.0614-0.0363]=[18.5190,22.3333] tCO2 / vehicle=[18.5190,22.3333] / car; The result is completely consistent with the total responsibility range of target province A [18.5190, 22.3333], proving that the responsibility allocation mathematically achieves a precise balance of the overall accounts, and there is no double counting or omission.

[0080] The final cross-provincial responsibility matrix is ​​as follows: Table 2

[0081] Note: Negative values ​​indicate that the province should actually receive emission reduction credits because the recycling and treatment reduced the responsibility of province A. In the example, province D only appears as an "intermediate node" in the multi-hop emission transfer path, and its emission contribution has been merged into the final responsibility of province B. Therefore, it is not included in the final matrix as an independent responsible entity.

[0082] Furthermore, it also includes visual displays: Front-end interface (image omitted): Three-stage emission interval box diagram (production, use, and disposal).

[0083] Forward and reverse Sankey diagrams: On the left, arrows from provinces B and C flow to province A; on the right, waste batteries from province A flow to provinces E and F.

[0084] Cross-provincial responsibility heat map: hovering the mouse displays the range value.

[0085] Example 2: Batch accounting and comparison experiment for multiple vehicle models and multiple provinces, as shown in Table 3; To verify the universality and performance improvement of this invention, life cycle carbon emission accounting was performed on eight vehicle types (taxi, bus, private car, medium and large passenger vehicle, and micro / light / medium / heavy truck) in 31 provinces across China in 2025, and compared with existing technologies (static point estimation + linear MRIO).

[0086] Data set: 744 records in total, covering 31 provinces, 8 vehicle models, and 3 years (2023-2025). Each record contains complete parameters for the production, use, and scrapping stages.

[0087] Table 3

[0088] Conclusion: This invention significantly outperforms existing technologies in terms of uncertainty expression, adaptability, attribution accuracy, response speed, and verifiability.

[0089] Example 3: Sensitivity Analysis of Reverse Logistics in the End-of-Life Stage To verify the importance of handling the positive and negative intervals during the disposal phase, a control group was set up: ignoring reverse transfer, i.e., assuming all used batteries are processed locally in Province A (no cross-provincial responsibility transfer occurs). In this case, the total responsibility interval of Province A increases, while the responsibility of processing provinces such as Province E and Province F is 0.

[0090] Compared with the actual survey data (E Province actually processed 60% of the batteries, and the net emissions during the processing were low), ignoring the reverse transfer led to an overestimation of A Province's responsibility by about 12%, and the emission reduction contribution of the processing province was ignored.

[0091] This invention accurately captures the transfer through the reverse edge of R-GCN, making the allocation of responsibility more equitable.

[0092] Example 4: Multi-objective decision support for logistics fleet optimization A logistics company plans to purchase 100 electric light trucks, needing to consider both carbon emissions and operating costs. This invention's system outputs the carbon emission range [350, 520 g] for each vehicle. The cost range is [1.2, 2.1] yuan / km. Based on the Pareto frontier (NSGA-III), the company selected the "medium carbon emission" option (450 g). After six months of actual operation, the fleet's total carbon emissions decreased by 34% compared to the original fuel vehicles, while costs increased by only 9%, which is better than the single-point prediction scheme (without interval trade-offs).

[0093] Example 5: A cross-domain tracking system for the lifecycle carbon emission responsibility of electric vehicles, such as Figure 2 As shown, it includes: The full-cycle data acquisition module has a built-in physical drift detector for acquiring multi-source data during the production, use, and disposal phases. A three-stage Bayesian quantile regression neural network engine supports online posterior updates and outputs emission ranges for production, use, and scrapping, respectively. The forward and reverse heterogeneous graph construction and R-GCN inference module embeds a graph attention mechanism to simultaneously learn the multi-hop emission transfer coefficients of the forward supply chain and the reverse recycling network; The full-cycle interval coupling responsibility allocation calculator is used to perform interval coupling calculations between the three-stage emission intervals and the forward and reverse transfer amounts; A visual front-end is used to display three-stage intervals, forward and reverse Sankey diagrams, and a cross-provincial responsibility matrix for the life cycle.

[0094] To further optimize the above technical solution, the three-stage Bayesian quantile regression neural network engine is deployed on a heterogeneous computing platform. The stage engine is deployed on an FPGA accelerator card to achieve millisecond-level online inference, while the production and scrap stage engines are deployed on a CPU cluster due to their lower update frequency. The R-GCN inference module uses a dedicated graph neural network processor to handle large-scale message passing of forward and inverse mixed graphs.

[0095] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0096] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for cross-domain tracing of carbon emission liability across the lifecycle of electric vehicles, coupled with dynamic interval prediction, graph neural networks, and input-output tables, characterized in that, include: Acquire multi-source time-series data on electric vehicles during the production, use, and scrapping stages, and construct a heterogeneous graph including nodes for provinces, industries, products, and scrapped products. A Bayesian quantile regression neural network is used to output emission intervals for the production, use, and disposal stages, respectively. These emission intervals are prediction intervals with a preset confidence level. The network parameters for each stage are updated online based on the data distribution drift. Message passing is performed on the heterogeneous graph using a relational graph convolutional network, where the set of edge types includes input-output edges and inter-provincial trade edges in the forward supply chain, as well as waste product recycling edges and recycled material edges in the reverse recycling network, to learn the forward multi-hop emission transfer coefficient and the reverse multi-hop emission transfer coefficient in parallel; the loss function of the relational graph convolutional network includes a physical constraint regularization term to ensure that the transfer coefficient satisfies the mass conservation principle; The cross-regional transfer amount of implicit emissions is calculated based on the forward and reverse multi-hop emission transfer coefficients, and the three-stage emission intervals are coupled with the transfer amount according to the interval arithmetic rules to generate the life cycle total carbon emission responsibility interval and cross-regional responsibility allocation results.

2. The method according to claim 1, characterized in that, The data distribution drift includes physical threshold judgments of the provincial power structure change rate and the battery recycling technology efficiency improvement rate; the online update adopts Bayesian quantile learning with conjugate normal-inverse gamma prior, and its retraining triggering conditions include: the interval coverage probability PICP is less than 90% for three consecutive periods; the Bayesian quantile learning only applies to the weights of the output layer of the neural network at each stage.

3. The method according to claim 1, characterized in that, The propagation formula of the relational graph convolutional network is: The set of edge types This includes positive input-output sides, inter-provincial trade sides, emission factor sides, as well as reverse end-of-life product recycling sides and recycled material sides; For the first Layer nodes The hidden state vector; It is a non-linear activation function; To pass through relationships With nodes A set of connected neighbor nodes; This is the normalization constant; For relationship In the The trainable weight matrix of the layer; The self-loop transformation matrix is ​​used; the multi-hop emission transfer coefficient is calculated using path integrals. in, For the node To the intermediate node The transfer coefficient; To start from the intermediate node To the target node The transfer coefficient; intermediate node Attention weights; This is the attention parameter vector; and Transformation matrix for query and key; This is a vector concatenation operation.

4. The method according to claim 1, characterized in that, The inter-regional transfer range of the implicit emissions is calculated by expressing the transfer coefficients learned from the graph neural network and the uncertainties of mass and emission factors in a range: Forward transition period in the production stage: Reverse transfer interval during the scrapping stage: in, This refers to the amount of emissions implicitly transferred from province B to province A during the production phase. For the quality of upstream materials; This is the positive transition coefficient vector; The emission factor for producing this material in Province B; This refers to the amount of emissions transferred from province A to province B during the decommissioning phase. For the quality of used batteries; This is the reverse transfer coefficient vector; This represents the net emission factor for waste battery treatment in Province B; the superscripts low and high indicate the lower and upper limits of the interval, respectively.

5. The method according to claim 1, characterized in that, The coupling interval calculation specifically refers to: the total lifecycle emission responsibility interval for target province A: Interval addition and subtraction follow the standard interval arithmetic rules. , ; This refers to the direct emission range during the local production phase in Province A. This refers to the direct emission range during the local use phase in Province A. This refers to the direct emission range during the local scrapping phase in Province A. This refers to the amount of emissions implicitly transferred from province B to province A during the production phase. This refers to the amount of emissions transferred from province A to province B during the decommissioning phase. The ultimate responsibility area for supply / processing province B consists of three parts: production responsibility, processing responsibility, and optional extended producer responsibility. in, This is an indicator function; it takes the value 1 when Province B is the main supplier of key materials for electric vehicles in Province A, and 0 otherwise. The value of materials supplied to Province B is a proportion of the total value of all materials, and is used only to allocate emissions during the production phase; emissions during the usage phase are separately accounted for by Province A, the place of use, and the province supplying electricity to it, based on the proportion of electricity inflow; emissions during the disposal phase are borne by Province B, the place of disposal.

6. The method according to claim 1, characterized in that, The heterogeneous graph of the scrapping stage includes scrapped product nodes, whose attributes include battery type, retirement year, remaining capacity, and hazardous substance content. The weight of the reverse recycling edge is determined by actual reverse logistics statistics and the spatial distribution of recycling and processing enterprises.

7. The method according to claim 1, characterized in that, The Bayesian quantile regression neural network shares the underlying feature extraction layer across all stages, but the output layer is independent to preserve stage specificity and reduce the number of parameters.

8. A cross-domain tracking system for the lifecycle carbon emission responsibility of electric vehicles implementing the method of any one of claims 1-7, characterized in that, include: The full-cycle data acquisition module has a built-in physical drift detector for acquiring multi-source data during the production, use, and disposal phases. A three-stage Bayesian quantile regression neural network engine supports online posterior updates and outputs emission ranges for production, use, and scrapping, respectively. The forward and reverse heterogeneous graph construction and R-GCN inference module embeds a graph attention mechanism to simultaneously learn the multi-hop emission transfer coefficients of the forward supply chain and the reverse recycling network; The full-cycle interval coupling responsibility allocation calculator is used to perform interval coupling calculations between the three-stage emission intervals and the forward and reverse transfer amounts; A visual front-end is used to display three-stage intervals, forward and reverse Sankey diagrams, and a cross-provincial responsibility matrix for the life cycle.

9. The system according to claim 8, characterized in that, The three-stage Bayesian quantile regression neural network engine is deployed on a heterogeneous computing platform. The stage engine is deployed on an FPGA accelerator card to achieve millisecond-level online inference, while the production and scrap stage engines are deployed on a CPU cluster due to their lower update frequency. The R-GCN inference module uses a dedicated graph neural network processor to handle large-scale message passing of forward and inverse mixed graphs.

Images