A traffic carbon emission deduction method and system based on multi-party longitudinal federation and conditional diffusion model
Patent Information
- Application Number
- CN202610401785.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-30
- Publication Date
- 2026-08-04
Smart Images
Figure CN122511077A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the interdisciplinary fields of intelligent transportation, privacy computing, generative artificial intelligence and environmental pollution measurement, specifically to a method and system for generating microscopic traffic flow patterns and extrapolating high-precision carbon emissions based on a multi-party vertical federated learning architecture and a conditional time series diffusion model. Background Technology
[0002] With the acceleration of industrialization and urbanization, road transportation has become one of the major sources of global greenhouse gas and ambient air pollutant emissions. Accurately measuring and projecting the carbon emission levels of urban road networks is of great strategic significance for formulating scientific energy conservation and emission reduction policies and achieving the dual carbon goals of carbon peaking and carbon neutrality.
[0003] Current traditional methods for calculating carbon emissions from transportation mostly fall under the category of macro-level accounting, generally relying on regional fuel sales statistics or making rough estimates based on annual average vehicle mileage. These methods are extremely coarse-grained, completely ignoring the impact of dynamic congestion at the micro-spatiotemporal scale of urban road networks on the nonlinear surge in carbon emissions. To achieve accurate carbon emission inventory calculations that meet IPCC standards and MRV (Measurable, Reportable, Verifiable) levels, it is necessary to rely on high-frequency, high-fidelity micro-traffic flow data of road networks, as well as dynamic emission dynamics models based on micro-vehicle speeds.
[0004] However, in the complex digital twin ecosystem of actual transportation, acquiring high-precision micro-level traffic flow patterns faces insurmountable technological hurdles. High-quality traffic dynamic features are scattered among different entities such as municipal transportation management departments, mobility service providers, logistics platforms, and telecommunications operators. Under increasingly stringent data security regulations, data from multiple parties cannot be centrally merged and directly shared at the original plaintext level. This results in existing deep learning traffic data generation models being inefficiently generated using single-modal or single-source data, failing to leverage the ecological synergy advantages of multi-modal data from multiple parties. Furthermore, road traffic flow is a highly nonlinear system subject to multiple factors, exhibiting both long-term evolutionary trends and short-term periodic oscillations. Traditional time-series data generation models, in generating long-series multivariate data, are prone to frequency domain distortion in the generated sequence data due to mean square error smoothing optimization. Without explicit constraints, the denoising diffusion probability model will still lose the characteristics of sudden speed changes in the generated traffic sequence data. Moreover, micro-level emission models are extremely sensitive to nonlinear changes in vehicle speed; distortion in the smoothing of input traffic sequence data can lead to huge exponential errors in the final emission calculation results.
[0005] In summary, breaking down data silos among multiple parties, protecting the underlying data privacy of all participants, utilizing generative artificial intelligence technology to reconstruct the spatiotemporal micro-characteristics of transportation with high fidelity, and scientifically coupling them to a micro-environmental emission model to generate an accurate and compliant MRV carbon emission inventory are core technical challenges that urgently need to be addressed in the interdisciplinary field of intelligent transportation and environmental informatics. Summary of the Invention
[0006] To address the aforementioned technical problems, the present invention aims to provide a method and system for extrapolating traffic carbon emissions based on a multi-party vertical federation and conditional diffusion model. The present invention protects the privacy and security of heterogeneous traffic data at the model's underlying layer through a multi-party separation computation graph mechanism, supporting large-scale, star-shaped multi-node computation. It generates high-precision physical traffic flow sequences through a conditional diffusion model with Fourier frequency domain constraints, and finally achieves reliable and traceable environmental carbon emission accounting through seamless coupling of a microscopic nonlinear emission model.
[0007] To achieve the above objectives, the first technical solution adopted by this invention is to provide a method for extrapolating traffic carbon emissions based on a multi-party vertical federation and conditional diffusion model, comprising the following steps:
[0008] Step S1: In a multi-party vertical federation scenario, acquire the first traffic observation data of the first participant (as the master node) and the second traffic observation data of at least one second participant (as multiple guest nodes).
[0009] Step S2: At each of the second participating parties, the second traffic observation feature sequence is extracted using the pre-constructed second low-level encoder to obtain the second local latent features, and all the second local latent features are sent to the first participating party through a secure communication channel.
[0010] Step S3: At the first participant's end, the first local latent feature is extracted from the first traffic observation data using the pre-constructed first bottom-level encoder; the first local latent feature is concatenated and nonlinearly fused with all the received second local latent features using the fusion layer network to generate a global conditional feature vector.
[0011] Step S4: Input the global conditional feature vector into the conditional time series diffusion model; generate a multivariate traffic flow situation tensor through the inverse denoising sampling process of the conditional time series diffusion model, wherein the traffic flow situation tensor includes lane flow and lane density of each road segment in the target road network;
[0012] Step S5: Calculate the average driving speed of each road segment based on the lane flow rate and lane density, combined with fluid dynamics formulas;
[0013] Step S6: Input the total traffic flow and average driving speed of each road segment into the preset micro-vehicle emission calculation engine, and combine it with the preset mixed vehicle fleet composition ratio distribution to calculate the emissions of various air pollutants and carbon emissions of each road segment in the target road network, and finally generate a carbon footprint checklist that meets international standards.
[0014] Preferably, the method includes a multi-party joint training phase of the model based on a multi-party longitudinal federated learning architecture, specifically comprising:
[0015] In the feature fusion calculation graph of the first participant, each of the received second local latent features is truncated, and its gradient tracking attributes are retained before being input into the fusion layer network for joint multi-party mapping.
[0016] At the first participant's end, the traffic flow situation tensor generated by the conditional time series diffusion model and the real traffic label data are substituted into the loss function to calculate the loss value, and the gradients of the fusion layer network, the first bottom layer encoder, and the conditional time series diffusion model are calculated through the error backpropagation algorithm to update the local network parameters of the first participant.
[0017] At the first participant's end, the partial derivatives of the loss function with respect to each truncated second local latent feature are extracted as the corresponding return gradients, and each return gradient is routed and sent down to the corresponding second participant.
[0018] Each of the second participants receives its own return gradient and independently calculates and updates the network parameters of its underlying encoder based on the chain rule.
[0019] Preferably, the underlying encoders deployed by each participant adopt a hybrid network architecture, which includes: a temporal one-dimensional convolutional module, used to extract the dynamic evolution features of the time series through one-dimensional convolutional layers and nonlinear activation functions in the time dimension; and a spatial graph convolutional network module, which uses the adjacency matrix of the target road network to perform graph convolutional aggregation on the extracted temporal features in order to capture the spatial dependencies between topologically connected road segments in the road network.
[0020] Preferably, the conditional time series diffusion model includes a denoising network module, which is a Transformer architecture based on conditional feature injection and includes an encoder and a decoder.
[0021] The global conditional features, after being mapped by linear projection, are superimposed on the feature embedding representation of the input sequence as an external guiding conditional injection network; simultaneously, they are fused with the output features of the encoder through a cross-attention mechanism.
[0022] The decoder internally includes a trend building module and a seasonal building module in parallel: the trend building module extracts the long-term trend component of the traffic time series based on multinomial regression; the seasonal building module extracts the maximum amplitude frequency of high frequencies based on fast Fourier transform and reconstructs the periodic seasonal components of the traffic sequence data through inverse discrete Fourier transform.
[0023] Preferably, the loss function used in the joint training phase of the model includes a weighted sum of temporal reconstruction loss and frequency-domain Fourier loss;
[0024] The frequency domain Fourier loss is calculated as follows: the output sequence of the conditional time series diffusion model and the real traffic label sequence are respectively subjected to fast Fourier forward transform, the absolute error between the real part features and the absolute error between the imaginary part features under the frequency domain features are calculated, and the real part absolute error and the imaginary part absolute error are added together.
[0025] Preferably, step S5, the step of calculating the average driving speed, specifically includes:
[0026] For the lane flow and the lane density Perform a non-negative truncation operation;
[0027] According to the formula Calculate average driving speed ,in The preset minimum density threshold for preventing division by zero anomalies is used, and the upper limit of the average driving speed is cut off to not exceed the free flow speed of the road segment;
[0028] Preferably, in step S6, the preset microscopic vehicle emission model is an emission coefficient equation that is nonlinearly fitted for different pollutant categories and different fleet types, where specific categories of vehicles at speeds of [missing information]. Single vehicle emission factor The calculation formula is:
[0029] in, The average driving speed is limited to a preset effective speed physical range; , , , , For specific empirical coefficients determined in a pre-defined mapping dictionary by vehicle type, fuel type, European emission standards, and pollutant category;
[0030] The total traffic flow, time step interval, and physical length of each road segment are multiplied to obtain the total vehicle mileage. Based on the mixed vehicle fleet composition distribution, the emission factors for each type of vehicle at the corresponding average driving speed are then calculated. The total emissions are obtained by performing a weighted summation and then multiplying it by the total mileage of the vehicles.
[0031] Preferably, the emissions of the multiple pollutants include at least carbon monoxide, nitrogen oxides, and fine particulate matter; the carbon emissions are carbon dioxide emissions.
[0032] The method also includes: aggregating the calculated emission data for each time window and road segment according to the spatial corridor network and time dimension to generate a structured carbon footprint inventory report that meets the international MRV (Measurable, Reportable, Verifiable) criteria. The report includes the total vehicle mileage of the corridor, the total greenhouse gas emissions, the total emissions of various pollutants, and the carbon emission intensity index per unit mileage.
[0033] Furthermore, the conditional time series diffusion model performs multiple independent random denoising generation samplings under the guidance of the same global conditional features, calculates the mean and standard deviation of multiple parallel carbon emission total projection result sets, and obtains the coefficient of variation as the percentage uncertainty of carbon emission projection, which is written into the inventory list report to provide a verification confidence evaluation.
[0034] A second aspect of the present invention provides a transportation carbon emission extrapolation system based on a multi-party longitudinal federation and conditional diffusion model, comprising:
[0035] The first server includes the first underlying encoder, the feature fusion layer network, the conditional time series diffusion model, and the traffic flow situation and micro-emission analysis module.
[0036] At least one second server contains its own corresponding second underlying encoder;
[0037] Each of the second servers is used to encode local underlying multivariate traffic observation data into a second local latent feature and independently encrypt and send it to the first server; the first server is used to encode local core traffic observation data to obtain a first local latent feature, and to nonlinearly concatenate and fuse the local latent features of the first server with all received local latent features of the second server through a feature fusion layer network to obtain a global conditional feature vector. The global conditional feature vector is sent into the conditional time series diffusion model to generate a traffic flow situation tensor through inverse denoising, and the traffic flow situation and micro emission analysis module is used to calculate the dynamic average vehicle speed of the road segment and verify the micro dynamic carbon emissions of the road network.
[0038] A third aspect of the present invention provides an electronic device comprising: a processor; a memory for storing one or more computer programs; wherein, when the one or more computer programs are executed by the processor, the processor causes the processor to perform the method as described in any of the preceding claims.
[0039] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described in any of the preceding claims.
[0040] Compared with the prior art, the present invention has at least the following beneficial effects:
[0041] (1) This invention breaks down data silos among multiple parties and ensures data privacy compliance throughout the entire chain. This invention adopts a star-shaped vertical federated learning architecture with a split computation graph supporting multiple nodes. The leading first participant and at least one second participant only interact with abstract latent feature vectors and digital gradient ciphertexts that have undergone nonlinear dimensionality reduction. The original sensitive data of any party will never leave the local data center. This breaks down data silos among multiple parties from the physical and algorithmic levels, fully leverages the synergistic advantages of multi-source heterogeneous modal features in the generation of complex traffic flow situation data, meets data security and compliance requirements, and has strong commercial application and business expansion capabilities.
[0042] (2) This invention greatly improves the physical fidelity of generating complex spatiotemporal traffic sequence data. It abandons the traditional deterministic deep learning time series model and applies a generative conditional diffusion model to traffic sequence data generation. By embedding a multinomial trend decomposition block and a periodic reconstruction layer based on inverse discrete Fourier transform in the diffusion decoder, and introducing a composite frequency domain loss function containing Fourier transform, it ensures that the traffic flow and density generated by the model denoising are not only accurate in macroscopic trends, but also retain the microscopic high-frequency speed change and start-stop cycle jitter characteristics.
[0043] (3) This invention pioneers an end-to-end scientific closed loop for the evolution of microscopic traffic flow patterns into macroscopic carbon emission MRV inventories. This invention delves into the core of traffic physics mechanisms, using fluid dynamics formulas to accurately extrapolate the generated flow rate and density into constrained dynamic vehicle speeds, solving the disconnect between previous intelligent transportation systems that only measured flow rate and not emissions. The system is tightly coupled with a standard-compliant environmental emission calculation method, precisely covering the absolute emissions of carbon dioxide, carbon monoxide, nitrogen oxides, and fine particulate matter from mixed vehicle fleets at the microscopic level. Furthermore, by solving for uncertainties through multiple probabilistic diffusion sampling, it automatically generates MRV greenhouse gas inventory reports that meet the standards of the Intergovernmental Panel on Climate Change (IPCC), providing the most rigorous and traceable quantitative foundation for building digital, intelligent, and green cities. Attached Figure Description
[0044] The invention will now be further described with reference to the accompanying drawings.
[0045] Figure 1 This is a flowchart of a traffic carbon emission extrapolation method and system based on a multi-party vertical federation and conditional diffusion model, according to the present invention. Detailed Implementation
[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0047] This invention provides an end-to-end integrated hardware and software system, which includes a multi-party longitudinal federated training module for multi-source heterogeneous traffic data, a time-series conditional diffusion generation module, a micro-level traffic flow situation inference and analysis branch, and a micro-level carbon emission measurement and verification branch. The execution logic is described in detail below.
[0048] For urban trunk road networks with the same time dimension and business scenario space, the system sets up one primary participant and multiple secondary participants. The primary participant is responsible for acquiring first-type traffic observation data, such as vehicle counting sensors distributed across road cross sections; each secondary participant independently extracts second-type traffic observation data, such as average travel time, travel distance, and dwell status for each road segment, using its proprietary floating car massive trajectory data and signaling data.
[0049] During model training and inference, a uniform sliding time window length is set. Spatial graph convolutional low-level encoders are independently deployed and configured on the local servers of both the first participant and each of the second participants. When extracting features, each low-level encoder first loads the adjacency matrix corresponding to the road network and performs graph convolution to aggregate the spatial traffic states of topologically connected road segments in the road network. Subsequently, the data feature map is transposed and a one-dimensional convolutional layer is used to capture the dynamic evolution relationship in the time dimension. Finally, the high-dimensional traffic spatiotemporal matrix is compressed into a local latent feature tensor of a specified dimension through the GELU activation function and a fully connected linear mapping layer.
[0050] To ensure the privacy of the original plaintext features of each participant while integrating data from multiple parties, this embodiment deploys a multi-party scalable, separate computation graph routing mechanism between the master node and multiple guest nodes.
[0051] Each guest node computes its local latent feature sequence in parallel and transmits it independently across domains to the master node through its own encrypted communication channel. After receiving the latent features from multiple guest nodes, the master node uniformly executes an operation to block the reverse computation graph chain, ensuring that the computation graphs of all parties are physically disconnected.
[0052] The master node uses its internal feature fusion network to concatenate its extracted latent features with the latent features of all the truncated guest nodes in the feature dimension, and outputs a unified global conditional feature vector.
[0053] After completing global loss calculation and backpropagation of its own network error, the master node extracts the mathematical gradient matrix corresponding to each guest node from the truncated set of latent feature nodes. The master node then transmits these gradient matrices, which contain no plaintext business information, as ciphertext to the corresponding guest nodes on a one-to-one basis. Each guest node independently performs backpropagation using its received dedicated backpropagation gradients, following the chain rule, to update the network weights of its local underlying encoder. This process enables joint optimization of a multi-party federated network under conditions of no plaintext data exchange.
[0054] The master node uses the high-dimensional global conditional features obtained through fusion to guide the core conditional time series diffusion model to generate traffic flow situation data.
[0055] The core of the model is a conditional Transformer network. During the inverse denoising generation process, the input global conditional features are embedded into the hidden dimension of the diffusion model through a linear projection layer and injected into the decoder through a cross-attention mechanism, which forces the evolution direction of the generated sequence to be constrained in the form of stride conditions.
[0056] In particular, since traffic flow has strong macro-long-term trends and micro-level seasonal characteristics such as traffic light cycles, the model incorporates a trend building module and a Fourier seasonality building module within the decoder to extract and fit the baseline long-term trend and hidden micro-level periodic fluctuations of traffic sequence data.
[0057] In federated training backpropagation, in addition to L1 regression loss constraints on time-series numerical data, this invention specifically incorporates frequency-domain Fourier loss. Specifically, a Fast Fourier Transform is performed on the sequence data output by the diffusion model and the actual traffic label sequence, calculating the real and imaginary absolute losses of their complex results respectively. The introduction of Fourier loss completely avoids high-frequency smoothing distortion in the microscopic time series of the diffusion-generated sequence, preserving crucial microscopic jitter details of vehicle acceleration and deceleration, and providing accurate pre-input data for microscopic dynamic emission estimation.
[0058] Traffic features generated by rapid denoising sampling are normalized data. The system first restores them to the real physical space through the inverse normalization module to obtain the single-lane flow and single-lane density of each road segment at each time step.
[0059] To accurately deduce dynamic micro-vehicle speeds to support subsequent high-precision environmental carbon emission calculations, the system incorporates a traffic fluid dynamics derivation engine: In this process, to prevent singular value divergence in vehicle speed calculation caused by density approaching zero, a safety lower limit threshold for preventing loss to zero is established.
[0060] Subsequently, through numerical boundary pruning, it is ensured that the output physical vehicle speed is forcibly truncated and constrained within the upper limit of the physical free-flow speed of the real road network.
[0061] The key innovation of this invention lies in its close integration of cutting-edge multi-party federated AI generation algorithms with professional environmental engineering computational theory. It utilizes the derived microscopic dynamic vehicle speed... Total traffic flow of road sections It is fed into the pre-installed microscopic vehicle emission calculation engine.
[0062] This engine is based on the latest authoritative COPERT-IV technical guidelines from the European Environment Agency and incorporates the mixed fleet composition of various cities. For example, a certain region may have 60% Euro 5 gasoline passenger cars, 25% Euro 6 diesel passenger cars, 10% diesel light trucks and 5% Euro 6 heavy trucks.
[0063] For greenhouse gas carbon dioxide and core hazardous air pollutants such as carbon monoxide, nitrogen oxides, and fine particulate matter, the system calls upon a large matrix of high-order polynomial empirical coefficients in its internal mapping dictionary to perform conversions based on velocity-dependent nonlinear microdynamic emission formulas.
[0064] The denominator of the formula is passed through the safety constant. Lower bound truncation to prevent overflow can effectively prevent model numerical divergence under conditions of extreme congestion and idling, while also controlling the input... Dynamic constraints are within the legal domain of the theory.
[0065] After calculating the emission factor per unit mileage for various vehicle types, the engine then combines this with the total mileage traveled by vehicles within a unit time step on that road segment. (in By weighting the data according to the market penetration rate of the fleet (the physical length of the road segment), the absolute quality of pollution and carbon emissions for each road segment within a micro time series can be obtained.
[0066] The system's built-in inventory checker integrates and sums the emission result matrices across all time and spatial dimensions of the entire network. Since the deep diffusion model falls under the category of probabilistic generation, to scientifically quantify the scientific confidence of the projection results, the system controls the conditional diffusion model by injecting different random Gaussian noise under the same global guidance features for multiple parallel Monte Carlo sampling, thereby obtaining a set of projection probability distribution results for total carbon emissions. The module calculates the mean and standard deviation of this distribution set, deriving the coefficient of variation as the percentage uncertainty of the carbon emission projection.
[0067] Finally, the system loads the spatiotemporal statistical totals and uncertainty results together and outputs them as a highly structured electronic MRV report file, which contains a globally unique identifier, a derivation generation timestamp, the methodology version used, the total emissions of various pollutants in the corridor, the carbon emission intensity per unit mileage, and the overall model uncertainty assessment, forming a high-value multi-party collaborative MRV traffic carbon emission verification list that fully complies with international IPCC standards.
[0068] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.
Claims
1. A method for extrapolating traffic carbon emissions based on a multi-party vertical federation and conditional diffusion model, characterized in that, Includes the following steps: Step S1: Under the multi-party vertical federated learning framework, the first participant uses the local first bottom-level encoder to extract features from the traffic observation data to be processed, and obtains the first local latent features. At least one second participant uses its local second-level encoder to extract features from its respective traffic observation data to be processed, thereby obtaining at least one corresponding second local latent feature. In step S2, each of the second participants securely transmits the extracted second local latent features to the first participant. The first participant then concatenates and nonlinearly maps the first local latent features with all the received second local latent features to generate a global conditional feature vector. Step S3: The global conditional feature vector is injected into the pre-trained conditional time series diffusion model as a guiding condition. Through a multi-step reverse denoising sampling process, a multivariate traffic flow situation sequence of the target road network during the data observation period is generated. The multivariate traffic flow situation sequence includes at least the traffic flow and traffic density of each road segment under each time window. Step S4: Input the multivariate traffic flow situation sequence into the micro-environmental pollution emission calculation engine, calculate the average vehicle speed and vehicle mileage of each road segment according to the principle of fluid dynamics, and combine the preset vehicle formation composition ratio distribution to calculate the various environmental pollution emissions of each road segment at each time step using the micro-emission factor model. Step S5: Input the calculated environmental pollution emissions of each road segment into the MRV accounting module, aggregate them in time and space within the specified accounting period, and generate a carbon emission MRV inventory report.
2. The method for extrapolating traffic carbon emissions based on a multi-party vertical federation and conditional diffusion model according to claim 1, characterized in that, The conditional time series diffusion model, the first low-level encoder, and all the second low-level encoders are trained collaboratively through the following split learning mechanism: Forward propagation phase: Each of the second participants separates the second local latent features obtained from the forward calculation from its own computation graph, retains the gradient tracking attribute, and sends it to the first participant; the first participant performs multi-party feature fusion and inputs it into the conditional time series diffusion model, and calculates the global composite loss function in combination with real traffic flow label data; Backpropagation phase: The first participant calculates the gradient based on the global composite loss function, updates the network parameters of the local conditional time series diffusion model, feature fusion layer and first bottom encoder, and obtains the partial derivative of the global composite loss function with respect to each of the second local latent features as the corresponding backpropagation gradient vector, and sends each of the backpropagation gradient vectors to the corresponding second participant. Each of the second participants receives the corresponding backpropagation gradient vector and independently performs backpropagation of its local network using the chain rule to update the network parameters of its local second underlying encoder.
3. The method for extrapolating traffic carbon emissions based on a multi-party vertical federation and conditional diffusion model according to claim 1, characterized in that, The first traffic observation data and the second traffic observation data are multi-source heterogeneous data distributed on different business entities, including but not limited to one or more of the following: traffic node cross-sectional flow data, floating vehicle trajectory data, vehicle travel time data, or location-based mobile terminal positioning data.
4. The method for extrapolating traffic carbon emissions based on a multi-party vertical federation and conditional diffusion model according to claim 1, characterized in that, The process of calculating various environmental pollutant emissions using a micro-emission factor model in step S4 specifically includes: Step S41, based on the formula Calculate the average speed of each road segment Where Q is the traffic flow rate per lane and K is the traffic density per lane. To prevent the extremely low density threshold from being divided by zero; and the average vehicle speed The effective physical velocity is limited to the closed interval preset by the micro-emission factor model. Step S42: Based on the given vehicle category, fuel type, and emission standards, use the emission factor formula per unit mileage. Calculate the emission factor per unit mileage ,in , , , , For pre-calibrated empirical coefficients of the polynomial, This is the lower limit of the safe minimum value of the denominator; Step S43: Multiply the traffic flow by the time interval step and the road segment length to calculate the total mileage traveled by vehicles on each road segment per unit time. Step S44, calculate the emission factor per unit mileage for each vehicle model. The market penetration rate corresponding to the preset vehicle platoon composition ratio distribution is multiplied by the total mileage of the vehicles and summed to obtain the various environmental pollution emissions of the corresponding road segment within the target time step.
5. The method for extrapolating traffic carbon emissions based on a multi-party vertical federation and conditional diffusion model according to claim 4, characterized in that, The various environmental pollution emissions include, but are not limited to, one or more of the absolute emission masses of carbon dioxide, carbon monoxide, nitrogen oxides, and fine particulate matter.
6. The method for extrapolating traffic carbon emissions based on a multi-party vertical federation and conditional diffusion model according to claim 1, characterized in that, When generating the carbon emission MRV inventory report in step S5, an assessment step for the uncertainty of deep model extrapolation is also included: Under the guidance of the same global conditional feature vector, the conditional time series diffusion model performs multiple independent random denoising generation samplings based on different random noise distributions, and obtains multiple parallel sets of total carbon emission extrapolation results through the micro-environmental pollution emission calculation engine. Calculate the mean and standard deviation of multiple parallel total carbon emission projection results, obtain the coefficient of variation as the percentage of uncertainty in the carbon emission projection, and write it into the carbon emission MRV inventory report to provide an assessment of confidence in international verification.
7. A transportation carbon emission projection system based on a multi-party vertical federation and conditional diffusion model, characterized in that, include: The first server includes the first underlying encoder, the feature fusion layer network, the conditional time series diffusion model, and the traffic flow situation and micro-emission analysis module. At least one second server, each second server containing its own second underlying encoder; Each of the second servers is used to independently encode local heterogeneous traffic observation data into second local latent features and send them to the first server through a secure communication channel; the first server is used to encode local first traffic observation data to obtain first local latent features, and to nonlinearly concatenate and fuse the first local latent features with all received second local latent features through a feature fusion layer network to obtain a global conditional feature vector; the global conditional feature vector is sent into the conditional time series diffusion model, and a traffic flow situation sequence is generated through inverse denoising, and the traffic flow situation and micro-emission analysis module is used to calculate the dynamic average vehicle speed of the road segment and verify the micro-dynamic carbon emissions of the road network.
8. An electronic device, characterized in that, include: processor; And a memory for storing one or more computer programs; when the one or more computer programs are executed by the processor, the processor causes the processor to implement the method as described in any one of claims 1 to 6.