Expressway service area carbon emission hotspot prediction method
By using lightweight data packet aggregation and multi-task spatiotemporal graph neural networks on edge computing nodes, combined with supply-side constraints and asynchronous table lookup, the computational power and physical mechanism problems in carbon emission prediction of highway service areas are solved, and efficient and accurate carbon emission hotspot prediction is achieved under low computing power.
Patent Information
- Application Number
- CN202610679132.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-18
- Publication Date
- 2026-06-16
AI Technical Summary
Existing technologies for predicting carbon emissions in highway service areas suffer from problems such as excessive reliance on computing power, lack of mapping of the physical mechanisms of carbon emissions, neglect of supply-side constraints in charging demand prediction, and slow response in real-time solution of microgrid equations, resulting in inaccurate prediction results and insufficient response capabilities.
By aggregating lightweight data packets at edge computing nodes, a dynamic spatiotemporal map of the road network is constructed. A multi-task spatiotemporal graph neural network is used to predict carbon emissions. By combining a saturation-based mapping function and a supply-side constraint function, direct and indirect carbon emissions are calculated. Carbon emission coefficients are obtained through asynchronous table lookup, enabling accurate prediction of carbon emission hotspots.
It achieves accurate prediction of service area-level carbon emission hotspots with low computing power overhead, reduces system response latency, improves prediction accuracy and response capability, and can truly reflect the micro-level congestion status of service areas and the nonlinear impact of charging facility supply constraints.
Smart Images
Figure CN122222216A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of transportation carbon emission monitoring and intelligent prediction technology, and in particular to a method for predicting carbon emission hotspots in highway service areas based on cloud-edge collaboration and multi-task graph networks. Background Technology
[0002] With the popularization of new energy vehicles and the advancement of carbon peaking and carbon neutrality goals in the transportation sector, the carbon emission management of highway service areas, as convergence points of energy consumption in the road network, is receiving increasing attention. Carbon emissions from service areas exhibit typical "multi-source heterogeneous" characteristics: they include both direct carbon emissions from idling gasoline vehicles entering the station and indirect carbon emissions from the power grid caused by electric vehicle charging. These two types of carbon sources show highly dynamic fluctuations in time and space, especially during peak holiday periods, when carbon emission intensity may surge several times higher than normal levels.
[0003] Currently, several technical solutions exist for predicting carbon emissions from transportation. In the field of traffic flow prediction, graph neural network technologies such as spatiotemporal graph convolutional networks have been widely used to model the spatial topological relationships and temporal evolution patterns between nodes in a road network, achieving good prediction accuracy. In the field of service area energy management, there are existing technical solutions that utilize microgrid planning methods to optimize the capacity configuration of photovoltaic, energy storage, and charging loads in service areas. However, the cross-integration of these two technologies to achieve accurate carbon emission prediction at the service area level still faces the following prominent challenges in engineering practice: First, the reliance on computing power is too high, making it extremely prone to failure during holidays due to congestion. Existing graph network traffic prediction solutions typically rely on accurate tracking of individual vehicle trajectories and real-time extraction of section speeds to construct the edge weights of the graph. During peak passenger flow periods like holidays, the concurrent transmission of massive amounts of individual vehicle data can cause edge computing power to collapse or communication latency to increase dramatically, resulting in severe lag in the update of graph network edge weights and a significant decrease in the timeliness of prediction results.
[0004] Second, the physical mechanisms underlying carbon emission sources are lacking. Most existing technologies only predict vehicle arrivals or service area dwell times and establish a simple linear relationship between these and carbon emissions. However, in actual operation, carbon emissions from gasoline-powered vehicles are highly correlated with the micro-level congestion at service areas—vehicles only generate significant carbon emissions during idling while searching for parking spaces and waiting in line. Ignoring the nonlinear impact of service area saturation on idling time and directly equating dwell time with carbon emissions leads to predictions that deviate significantly from actual carbon emission levels.
[0005] Third, the supply-side constraints of charging infrastructure are not considered. Existing electric vehicle charging load forecasting schemes typically only focus on the number of electric vehicles entering the service area, ignoring the inhibitory effect of the real-time occupancy status of charging piles on users' willingness to charge. When charging piles are near full capacity, many arriving electric vehicle users will abandon charging and choose to leave due to excessive waiting time, resulting in actual charging demand that is far lower than the forecast value. This "blind box effect" leads to generally inflated forecasts of indirect carbon emissions.
[0006] Fourth, real-time solution of microgrid energy equations leads to slow system response. To calculate the indirect carbon emissions corresponding to charging electricity consumption in the service area, existing technologies typically solve the energy dispatch equations of microgrids such as "photovoltaic-storage-DC-flexible" in real time during the predictive inference process to obtain the grid carbon emission factor. This solution process has huge computational overhead and is difficult to meet the millisecond-level response requirements in real-time carbon emission hotspot early warning scenarios for large-scale road networks.
[0007] Therefore, there is an urgent need for a method for predicting carbon emission hotspots in highway service areas that can accurately integrate traffic flow forecasting and carbon emission physical mechanisms with low computing power while taking into account supply-side constraints. Summary of the Invention
[0008] The purpose of this invention is to provide a method for predicting carbon emission hotspots in highway service areas, in order to solve the problems of excessive computing power in graph network traffic prediction, lack of mapping of carbon emission physical mechanisms, neglect of supply-side constraints in charging demand prediction, and slow response in real-time solution of microgrid equations in the existing technology.
[0009] To achieve the above-mentioned objectives, the technical solution provided by this invention includes: Methods for predicting carbon emission hotspots at highway service areas include: Receive lightweight data packets reported from edge computing nodes deployed on highways. These lightweight data packets are generated by the edge computing nodes after classifying, counting, and aggregating vehicles collected by traffic detection equipment at a preset time window. A dynamic spatiotemporal graph of the road network is constructed based on the lightweight data package. Each service area is used as a node, and the ratio of the total number of vehicles currently in the service area to the maximum parking capacity is used as the node dynamic saturation. The edge weights are dynamically updated based on the interval flow difference between adjacent nodes. The dynamic spatiotemporal graph is input into the spatiotemporal graph neural network, and the predicted number of fuel vehicles and electric vehicles in each service area within a future preset time period is output through at least two prediction branches. Based on the dynamic saturation, the idling time of a single vehicle in each service area is calculated through a mapping function, and the direct carbon emissions are calculated by combining the predicted number of fuel vehicles, idling fuel consumption rate and carbon emission conversion factor. The supply status parameters of charging facilities in each service area are obtained. The predicted number of electric vehicles is corrected by the supply-side constraint function to obtain the effective charging demand. The carbon emission coefficients corresponding to each service area and the future preset time period are obtained. The indirect carbon emissions are calculated in combination with the effective charging demand. The direct carbon emissions are superimposed with the indirect carbon emissions to obtain a comprehensive carbon emission prediction value, and hotspot marking and early warning are performed on the road network topology map.
[0010] Preferably, the mapping function is a nonlinear piecewise function based on a saturation threshold: When the dynamic saturation is not greater than the first threshold, the single vehicle idling time is taken as the reference idling time; When the dynamic saturation is greater than the first threshold and not greater than the second threshold, the single vehicle idling time is equal to the sum of the baseline idling time and the linear congestion growth coefficient multiplied by the portion of the saturation exceeding the first threshold. When the dynamic saturation is greater than the second threshold, the vehicle idling time is superimposed with an exponential deterioration term on the sum of the baseline idling time and the linear congestion growth coefficient multiplied by the portion of saturation exceeding the first threshold. The exponential deterioration term is the product of the exponential deterioration parameter and an exponential function of the portion of saturation exceeding the second threshold.
[0011] Preferably, the supply status parameter is the real-time occupancy rate of charging piles, the supply-side constraint function is a variant of the Sigmoid function, the effective charging demand is equal to the difference between the predicted number of electric vehicles multiplied by 1 and the demand loss penalty value, the demand loss penalty value is equal to the sum of the maximum loss rate upper limit divided by 1 plus the sum of the product of the negative sensitivity parameter and the real-time occupancy rate minus the queuing congestion threshold, where the queuing congestion threshold ranges from 85% to 95%.
[0012] Preferably, the step of dynamically updating the edge weights based on the interval flow difference between adjacent nodes includes: The difference between the upstream inflow and the downstream outflow of adjacent nodes is calculated as the interval flow difference value; When the interval flow difference is less than the preset normal fluctuation threshold, the edge weight takes the base time value; When the interval flow difference value is greater than the normal fluctuation threshold, the edge weight is dynamically increased based on the interval flow difference value through an incrementing function to reflect the spatiotemporal transmission delay.
[0013] Preferably, the carbon emission coefficient is obtained through a preset multidimensional carbon emission coefficient lookup table. The lookup table is indexed by at least two dimensions, namely season, weather conditions and time period. It is calculated in advance based on the weighted average of the green electricity consumption ratio and the thermal power supply ratio, which is obtained by statistically analyzing the historical operation logs or simulation model data of the microgrid energy management system and combining them with the main grid carbon emission factor. It is updated and maintained asynchronously outside of the real-time inference process.
[0014] Preferably, the classification counting aggregation is implemented using a dual-channel method: The first channel uses license plate recognition equipment to distinguish between new energy vehicles and fuel vehicles based on the color of the license plate. The second channel verifies the vehicle type registration feature code by reading the vehicle's electronic tag unit through the electronic non-stop toll collection gantry.
[0015] Preferably, the spatiotemporal graph neural network is trained offline using an adaptive joint loss function. The adaptive joint loss function is the sum of the product of the first weight parameter and the prediction error of fuel vehicles, the product of the second weight parameter and the prediction error of electric vehicles, and the product of the third weight parameter and the prediction error of carbon emissions. The first weight parameter and the second weight parameter are gradient-based learnable weights.
[0016] Preferably, the spatiotemporal graph neural network is a multi-task spatiotemporal graph convolutional network, including an input layer that receives historical multi-time step node features and dynamic edge weight matrices, a shared representation layer that extracts spatial topological interaction features and temporal evolution features, and a task branch output layer with separate branches for fuel vehicle prediction and electric vehicle prediction.
[0017] Preferably, the lightweight data packet includes a timestamp, node identifier, traffic cross-section throughput, total number of fuel vehicles entering the station, total number of electric vehicles entering the station, and total number of vehicles exiting the station.
[0018] Beneficial effects This invention receives lightweight data packets from edge computing nodes, constructs a dynamic spatiotemporal map of the road network with service areas as nodes in the cloud, dynamically updates edge weights using traffic difference values, and performs parallel prediction by vehicle type using a multi-task spatiotemporal graph neural network. Then, it calculates direct and indirect carbon emissions using a saturation-based mapping function and a supply-side constraint function, respectively. This achieves accurate prediction of service area-level carbon emission hotspots without tracking individual vehicle trajectories or measuring speeds in real time.
[0019] Since the edge side only performs classification, counting, and aggregation operations, the cloud replaces vehicle speed measurement with traffic differential and equation solving with table lookup, significantly reducing the overall computing power consumption of the system and maintaining stable predictive response capabilities when dealing with high concurrency traffic during holidays.
[0020] Furthermore, by employing a nonlinear piecewise mapping function based on a saturation threshold to calculate the idling time of fuel vehicles, and by using a supply-side constraint function with a Sigmoid variant to correct the charging demand of electric vehicles, the carbon emission prediction results can truly reflect the nonlinear impact of service area micro-congestion and charging facility supply constraints on carbon emission intensity, thus avoiding the systematic bias caused by directly linearly mapping dwell time to carbon emissions.
[0021] Furthermore, by maintaining a multidimensional carbon emission coefficient lookup table asynchronously outside the real-time inference process, the acquisition of indirect carbon emission coefficients is transformed from the real-time solution of complex microgrid energy equations into a direct table lookup operation with constant complexity. This improves the inference speed by orders of magnitude without sacrificing the spatiotemporal accuracy of the carbon emission coefficients. Attached Figure Description
[0022] Figure 1 This is a flowchart illustrating a preferred embodiment of the present invention for predicting carbon emission hotspots in highway service areas. Detailed Implementation
[0023] The technical solution of the present invention will be described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the scope of protection of the present invention. Unless otherwise specified, the features in the following embodiments can be combined with each other.
[0024] The carbon emission hotspot prediction method for highway service areas described in this invention can be implemented in a cloud-edge collaborative computing architecture. This architecture includes edge computing nodes deployed at highway gantries and service area computer rooms, and a cloud server deployed at the highway network dispatch center. The edge computing nodes are equipped with embedded processors, communication interfaces, and local caches, and connect to existing ETC gantry systems and video streams from checkpoint cameras. The cloud server is equipped with a GPU computing cluster, large-capacity storage, and network communication modules, and is responsible for graph network maintenance, model inference, and carbon emission mapping calculations. Data interaction between the edge computing nodes and the cloud server occurs via a dedicated highway communication network.
[0025] As shown in Figure 1, the method of this invention uses a cloud server as the execution entity. It receives aggregated data from the edge side, constructs a dynamic road network map, performs multi-task prediction and physical mechanism mapping, and finally outputs the comprehensive carbon emission prediction value for each service area. The technical implementation of each step is described in detail below.
[0026] Step S1: Receive lightweight data packets reported from edge computing nodes deployed on highways. The lightweight data packets are generated by the edge computing nodes after classifying, counting and aggregating vehicles collected by traffic detection equipment at a preset time window period.
[0027] The cloud server continuously receives lightweight data packets reported from various edge computing nodes via a communication interface. These edge computing nodes are connected to ETC antennas deployed on highway gantries and checkpoint cameras at service area entrances and exits, within a preset time window. A data aggregation operation is performed periodically. Within each time window, edge computing nodes classify and count vehicles detected by traffic detection equipment, encapsulate the statistical results into a lightweight data packet, and report it to the cloud. (Preset time window) The time window can be set from 3 minutes to 10 minutes, with 5 minutes being the preferred setting. The reason for choosing 5 minutes is that a time window that is too short will result in too high a frequency of data packets and an increased communication load; while a time window that is too long will reduce the time resolution of dynamic saturation, which is not conducive to capturing the rapid changes in traffic during peak holiday periods.
[0028] Existing traffic map network prediction schemes typically require tracking individual vehicle trajectories and extracting high-precision features such as interval vehicle speeds. This can easily lead to edge computing power collapse and communication congestion during holidays when a massive number of vehicles are running concurrently.
[0029] In some preferred embodiments, the classification and counting aggregation is implemented through a dual-channel approach: the first channel uses a license plate recognition device to distinguish between new energy vehicles and fuel vehicles based on license plate color; the second channel uses an electronic toll collection (ETC) gantry to read the vehicle type registration feature code from the on-board unit (OBU) for verification. Specifically, the first channel uses the checkpoint camera at the service area entrance / exit to visually classify the license plate color of passing vehicles—green license plates are identified as new energy vehicles, and blue or yellow license plates as fuel vehicles. The second channel uses the ETC gantry to conduct short-range communication with the on-board unit (OBU) when a vehicle passes through, reading the pre-registered vehicle type feature code in the OBU and cross-validating the visual classification result of the first channel. When the results of the two channels are inconsistent, the OBU registration information prevails.
[0030] Compared with solutions that rely solely on visual license plate recognition, this invention effectively solves the problem of misclassification caused by license plate obstruction in rainy or foggy weather through a dual-channel cross-verification mechanism, thereby improving the accuracy of classifying fuel vehicles and electric vehicles.
[0031] In some preferred embodiments, the lightweight data packet includes a timestamp, node identifier, traffic cross-section throughput, total number of fuel-powered vehicles entering the station, total number of electric vehicles entering the station, and total number of vehicles exiting the station. The node identifier is a unique code for the edge computing node within the road network topology. This data packet has a compact structure, with a single packet size not exceeding 100 bytes, reducing communication bandwidth usage by at least two orders of magnitude compared to transmitting single-vehicle trajectory data.
[0032] It should be understood that, in addition to the dual-channel approach mentioned above, vehicle type classification can also be achieved by using deep learning classification models based on vehicle appearance features, visual solutions based on charging port location detection, or relying solely on OBU single-channel recognition.
[0033] Step S2: Construct a dynamic spatiotemporal graph of the road network based on the lightweight data packet. With each service area as a node, the ratio of the total number of vehicles currently in the service area to the maximum parking capacity is used as the node dynamic saturation. The total number of vehicles currently in the service area is determined based on the difference between the historical cumulative number of vehicles entering the station and the cumulative number of vehicles leaving the station. The edge weights are dynamically updated based on the interval flow difference scores between adjacent nodes.
[0034] After receiving lightweight data packets reported by each edge computing node, the cloud server constructs and maintains a dynamic spatiotemporal graph G=(V,E) covering the target highway segment. In the graph, each service area is used as the node set V, and the road segments between adjacent service areas or adjacent gantry nodes are used as the edge set E.
[0035] Furthermore, the dynamic saturation, as a node feature, is calculated using the following methods: For the i-th service area node, the cloud continuously accumulates the total number of vehicles entering and leaving the service area since the start time, and uses the difference between the two as the total number of vehicles currently in the area, which is then divided by the maximum parking capacity of the service area. The dynamic saturation of the nodes is obtained. This value reflects the level of congestion within the service area in real time: when When the number of vehicles approaches or exceeds 100%, it indicates that the service area is nearing or has exceeded its full capacity, and vehicles entering the area will face significant idling wait times.
[0036] In existing traffic map networks, node features typically only include traffic state parameters such as traffic flow and speed, lacking a depiction of the micro-level congestion within service areas. Introducing saturation into node features enables the graph network to perceive the real-time load pressure of service areas, providing a physical basis for subsequent idling carbon emission mapping.
[0037] The edge weights can be dynamically updated based on traffic difference. Specifically, in the existing traffic map network, the update of edge weights usually depends on the real-time acquisition of the average vehicle speed in the section. This requires the deployment of speed measurement radar or reliance on floating car GPS data, which is difficult to achieve in road sections with poor infrastructure.
[0038] In some preferred embodiments, the step of dynamically updating edge weights based on the interval flow difference between adjacent nodes includes: calculating the difference between the upstream inflow and the downstream outflow of adjacent nodes as the interval flow difference; when the interval flow difference is less than a preset normal fluctuation threshold, the edge weight is taken as a base time value; when the interval flow difference is greater than the normal fluctuation threshold, the edge weight is dynamically increased based on the interval flow difference using an incrementing function to reflect the spatiotemporal transmission delay.
[0039] Specifically, for adjacent gantry nodes A and B, the inflow volume at point A within the same time window is extracted. and the outflow at point B Calculate the interval flow difference value .when When the fluctuation is less than the normal fluctuation threshold (which can be set to twice the standard deviation of the daily average flow difference), the interval operation is considered smooth, and the edge weights are adjusted accordingly. Take the baseline free-flow travel time. When the value is significantly greater than the threshold, it indicates that vehicles are congested within the interval. The system then uses an incrementing function (such as...) ,in The growth coefficient, For the growth index, the optimal value is selected. Dynamically increasing edge weights reflects the spatiotemporal transmission delay caused by traffic congestion between sections. This refers to the aforementioned normal fluctuation threshold.
[0040] Compared with edge weight update methods based on real-time speed measurement, this invention uses a flow differential method to infer the congestion status of a section using only existing ETC gantry counting data, without the need to deploy additional speed measurement equipment, which significantly reduces the hardware cost and computing power threshold for obtaining spatial features of graph networks.
[0041] It should be understood that the increasing function is not limited to the power function form mentioned above. Monotonically increasing functions such as linear functions, piecewise linear functions, or logarithmic functions can also be used to establish the mapping relationship between the interval flow difference value and the edge weight increment.
[0042] Step S3: Input the dynamic spatiotemporal graph into the spatiotemporal graph neural network, and output the predicted number of fuel vehicles and electric vehicles in each service area within a preset future time period through at least two prediction branches.
[0043] The cloud server inputs the constructed weighted dynamic spatiotemporal graph into the pre-trained spatiotemporal graph neural network model, performs forward inference, and outputs the predicted number of fuel vehicles in each service area within the preset time period H. and the predicted number of electric vehicles The preset time period H can be set from 15 minutes to 2 hours, preferably from 30 minutes to 1 hour. The reason for using parallel prediction by vehicle type instead of just predicting the total flow is that the carbon emission paths of fuel vehicles and electric vehicles are completely different. The former emits carbon directly through idling, while the latter indirectly generates carbon emissions from the power grid through charging. Therefore, it is necessary to obtain the predicted number of each type of vehicle separately in order to perform subsequent differentiated carbon emission mapping calculations.
[0044] Existing traffic prediction models typically only output a single traffic flow prediction or vehicle speed prediction, lacking the ability to output values in parallel for different vehicle types.
[0045] In some preferred embodiments, the spatiotemporal graph neural network is a multi-task spatiotemporal graph convolutional network, comprising an input layer that receives node features and dynamic edge weight matrices from historical multi-timesteps, a shared representation layer that extracts spatial topological interaction features and temporal evolution features, and task branch output layers with separate branches for fuel-powered vehicle prediction and electric vehicle prediction. Specifically, the input layer receives node feature matrices and dynamic edge weight adjacency matrices from the past T timesteps (e.g., T=12, corresponding to 1 hour of historical data). The shared representation layer consists of several spatiotemporal convolutional modules connected in series. Each module first aggregates neighbor node information along the spatial dimension through graph convolution operations, and then extracts sequence evolution patterns along the temporal dimension through temporal convolution operations. The task branch output layer is divided into a fuel-powered vehicle prediction branch and an electric vehicle prediction branch. The two branches share the underlying representation but each has an independent fully connected output layer, which respectively outputs the predicted number of vehicle types for the next H time period.
[0046] Compared to single-task prediction models, the multi-task architecture allows fuel vehicles and electric vehicles to share the learning of road network spatial features (both types of vehicles are affected by the same road network topology), while maintaining their respective prediction independence at the output layer. This reduces the number of model parameters and improves the prediction accuracy for small sample vehicle types (such as some road sections with a low proportion of EVs).
[0047] In some preferred embodiments, the spatiotemporal graph neural network is trained offline using an adaptive joint loss function: the adaptive joint loss function is the sum of the product of a first weight parameter and the prediction error of a gasoline vehicle, the product of a second weight parameter and the prediction error of an electric vehicle, and the product of a third weight parameter and the prediction error of carbon emissions, wherein the first weight parameter and the second weight parameter are gradient-based learnable weights. In conventional multi-task learning, the loss weights of each task usually need to be manually adjusted and kept fixed. When the loss magnitudes of different tasks differ significantly, it is easy for one task to dominate gradient updates while other tasks do not converge sufficiently. This invention addresses this issue by using a first weight parameter... Second weight parameter Designed as learnable parameters that are adaptively updated based on gradient information, this allows the loss weight allocation to be dynamically adjusted during training based on the real-time convergence state of each task, avoiding the tediousness of manual parameter tuning and the risk of the model getting trapped in local optima. In one implementation, the third weight parameter... It can be set to a fixed value of 1.0 to prioritize the strength of the monitoring signal for carbon emission prediction, while and In the early stages of training, uniform initialization is used, followed by adaptive adjustment through gradient backpropagation. Training data can be extracted from the full volume of anonymized traffic logs of the highway network over the past 1 to 2 years as input samples. The true labels for intermediate tasks are taken from historical vehicle passage statistics, and the true labels for carbon emissions are obtained by offline inverse calculation using historical microgrid EMS discharge logs and a high-precision fuel consumption model.
[0048] It should be understood that, in addition to spatiotemporal graph convolutional networks, the spatiotemporal graph neural network can also adopt graph attention networks (GAT), graph Transformers, or other graph neural network architectures based on message passing mechanisms, as long as they can simultaneously extract spatial topological features and time series features and support multi-task output.
[0049] Step S4: Based on the dynamic saturation, calculate the idling time of a single vehicle in each service area using a mapping function, and calculate the direct carbon emissions by combining the predicted number of fuel vehicles, idling fuel consumption rate, and carbon emission conversion factor.
[0050] The standard idle fuel consumption rate can be obtained from the vehicle-specific idle condition data in national emission standards or emission models such as MOVES, and the carbon emission conversion factor can be determined by referring to the fossil fuel carbon emission coefficient published in the IPCC National Greenhouse Gas Inventory Guidelines.
[0051] After obtaining the predicted number of gasoline-powered vehicles, it is necessary to convert this into actual carbon emissions. The core of this conversion lies in determining the idling time of vehicles within service areas. Vehicles only generate significant carbon emissions during idling while searching for parking spaces and waiting in queues. Clearly, idling time is highly correlated with the congestion level of service areas: when there is little traffic, vehicles can quickly enter and exit, resulting in extremely short idling times; while during congestion, vehicles wait in queues for extended periods, leading to a sharp increase in idling time.
[0052] The dynamic saturation of the nodes calculated in the cloud based on step S2 Through mapping function Calculate the average idling time per vehicle in each service area under the current conditions. Then multiply the average idling time per vehicle by the predicted number of fuel-powered vehicles. By calculating the standard idle fuel consumption rate (in liters per hour) and the carbon emission conversion factor (in kilograms of CO2 per liter of fuel), the direct carbon emissions of the service area can be obtained. The standard idle fuel consumption rate and carbon emission conversion factor can be determined through... Existing carbon emission estimation methods typically multiply the number of vehicles entering the station by a fixed carbon emission coefficient, ignoring the nonlinear impact of service area congestion on idling time, resulting in predictions that are significantly underestimated under congested conditions.
[0053] In some preferred embodiments, the mapping function is a nonlinear piecewise function based on a saturation threshold. Let the reference idling time be... (That is, the fixed time for vehicles to enter and exit the service area normally when it is idle, preferably 2 to 5 minutes), then: when the dynamic saturation is not greater than the first threshold (preferably set to 80%), the service area is in a smooth state, vehicles do not need to queue, and the idling time of a single vehicle is taken as the baseline idling time. When the dynamic saturation is greater than the first threshold but not greater than the second threshold (preferably set to 100%), the service area enters a slow-moving state, vehicles begin to form a slight queue, and the idling time of a single vehicle is equal to the baseline idling time. Add the linear congestion growth factor Multiply by the product of the portion of saturation exceeding the first threshold (i.e.) ),in The preferred value range is 0.5 to 2.0 minutes per percentage point; when the dynamic saturation exceeds the second threshold, the service area enters a state of extreme congestion and queuing, and the idling time is superimposed with an exponential deterioration term on the basis of the aforementioned linear growth, wherein the exponential deterioration term is the exponential deterioration parameter. The exponential function of saturation exceeding the second threshold The accumulation of, among which and To optimize the values of parameters fitted using historical micro-trajectory data. =0.1~0.5, =3~8. The reason for using a three-segment piecewise function is that it accurately reproduces the gradual process of a service area from idle to slow traffic to severe congestion. When idle, the idling time is constant; when moderately congested, it increases linearly; and when extremely congested, it deteriorates exponentially. This is highly consistent with the micro-queuing behavior of service areas observed in traffic engineering.
[0054] Compared with methods that directly and linearly map dwell time to carbon emissions, this invention introduces a piecewise mapping function related to saturation, enabling carbon emission prediction results to reflect the true physical characteristics of idling carbon emissions under different congestion conditions in the service area, thus avoiding the systematic bias of severely underestimating carbon emissions in congestion scenarios.
[0055] It should be understood that the mapping function is not limited to the above three-segment form. It can also be a sigmoid curve, a high-order polynomial fitting, or a lookup table mapping based on historical data training, as long as it can reflect the nonlinear positive correlation between saturation and idling time.
[0056] Step S5: Obtain the supply status parameters of charging facilities in each service area, correct the predicted number of electric vehicles through the supply-side constraint function to obtain the effective charging demand, and obtain the carbon emission coefficient corresponding to each service area and the future preset time period, and calculate the indirect carbon emissions in combination with the effective charging demand.
[0057] After obtaining the predicted number of electric vehicles, two practical constraints need to be considered: first, the supply-side constraints of charging facilities. When charging piles are close to full capacity, some electric vehicle users who arrive at the station will choose to give up charging due to the long waiting time; second, the grid carbon emission intensity varies in different service areas at different times, which depends on the current photovoltaic output of the microgrid and the power supply ratio of the main grid.
[0058] It should be understood that supply-side constraints need to be corrected. The supply status parameters of charging facilities in each service area are obtained from the cloud, and the demand loss penalty value is calculated through the supply-side constraint function. The effective charging demand is obtained by multiplying the predicted number of electric vehicles by the loss correction coefficient.
[0059] Existing charging load forecasting schemes typically focus only on the number of electric vehicles entering service areas, including them all in the charging demand, while ignoring the inhibitory effect of charging station occupancy status on users' willingness to charge. This "blind box" forecasting leads to generally inflated indirect carbon emission predictions.
[0060] In some preferred embodiments, the supply status parameter is the real-time occupancy rate of the charging pile. The supply-side constraint function is a variant of the Sigmoid function. The effective charging demand equals the predicted number of electric vehicles multiplied by... Among them, the penalty value for demand churn In the formula The maximum loss rate is set at 0.6 to 0.8 (preferably, this means that under extreme full load conditions, up to 60% to 80% of EVs will abandon charging). The queuing congestion threshold (ranging from 85% to 95%, preferably 90%, indicating that users begin to show a significant willingness to give up when the occupancy rate exceeds this threshold). This is a sensitivity parameter (preferably 8-15, to control the steepness of the churn rate increase). The technical principle of using the Sigmoid variant function is that users' charging abandonment behavior does not increase linearly with the occupancy rate, but presents an S-shaped curve of "almost no abandonment when the occupancy rate is low, a sharp abandonment when approaching the threshold, and saturation after full load". The Sigmoid function precisely describes this psychological threshold effect.
[0061] Compared to the approach of directly including all EVs in charging demand without considering supply-side constraints, this invention effectively filters out "pseudo-charging demand" by introducing a Sigmoid loss penalty mechanism based on real-time occupancy rate, making the predicted indirect carbon emissions closer to the actual values.
[0062] After obtaining the effective charging demand, it is also necessary to obtain the carbon emission coefficients corresponding to each service area and the future preset time period in order to convert the charging power into indirect carbon emissions.
[0063] In existing charging carbon emission accounting schemes, the “photovoltaic-storage-DC-flexible” energy dispatch equations of microgrids are usually solved in real time within the inference flow to obtain the grid carbon emission factor for the current period, which has huge computational overhead.
[0064] In some preferred embodiments, the carbon emission coefficient is obtained through a pre-defined multidimensional carbon emission coefficient lookup table. This lookup table is indexed by at least two dimensions: season, weather conditions, and time period. It is pre-calculated based on the weighted average of the green electricity consumption ratio and the thermal power supply ratio, calculated using historical operation logs or simulation model data from the microgrid energy management system, combined with the main grid carbon emission factor. This table is updated and maintained asynchronously outside of the real-time inference process. Specifically, a three-dimensional grid is used: "season (spring / summer / autumn / winter) × weather (sunny / cloudy / rainy) × time period (24-hour clock, one grid per hour)". Each grid cell stores the equivalent carbon emission power factor (unit: kg CO2 / kWh) under the corresponding conditions. This lookup table is recalculated and updated periodically (e.g., weekly or monthly) by an independent offline module based on the latest EMS logs. During online inference, only one lookup operation with constant-time complexity is required.
[0065] Compared with the scheme of solving the microgrid energy equation in real time, the present invention transforms the acquisition of indirect carbon emission coefficients from solving complex equations to O(1) binary direct table lookup through asynchronous table lookup mapping, which improves the inference speed by orders of magnitude. At the same time, the asynchronous update mechanism of the table lookup ensures that the carbon emission coefficients can track seasonal and meteorological changes without sacrificing online inference performance.
[0066] Finally, the indirect carbon emissions of each service area are obtained by multiplying the effective charging demand by the average charging kWh per vehicle (which can be obtained from historical charging pile data) and the equivalent carbon emission electricity factor. .
[0067] It should be understood that, in addition to the multidimensional lookup table method mentioned above, simplified single-dimensional lookup table (indexed only by time period), carbon emission factor prediction based on regression model, or obtaining real-time factors through direct interface with microgrid EMS can also be used to determine the carbon emission coefficient.
[0068] Step S6: The direct carbon emissions are superimposed with the indirect carbon emissions to obtain a comprehensive carbon emission prediction value, and hotspot marking and early warning are performed on the road network topology map.
[0069] The cloud will calculate the direct carbon emissions of each service area obtained in step S4. The indirect carbon emissions calculated in step S5 The values are then superimposed to obtain the overall carbon emission forecast for each service area. Subsequently, Mapped onto the road network topology, each service area node is marked with a hotspot level according to a preset carbon emission intensity classification standard. Optionally, when the predicted comprehensive carbon emission value of a service area exceeds a preset warning threshold, the system automatically triggers a warning message to be pushed to the road network dispatch center and relevant microgrid management platforms, so as to conduct early intervention for traffic guidance and energy storage dispatch.
[0070] In summary, the carbon emission hotspot prediction method for highway service areas provided by this invention solves the computational bottleneck of high concurrency during holidays through lightweight edge-side data aggregation; reduces the hardware cost of graph network construction through speed-weightless edge updates based on traffic flow differences; provides a foundation for differentiated carbon emission mapping through multi-task parallel prediction based on vehicle type; restores the true physical process of idling carbon emissions through a piecewise mapping function based on saturation; eliminates the "blind box effect" by filtering out pseudo-charging demand through Sigmoid supply-side constraints; and transforms indirect carbon emission accounting from real-time solution of complex equations to constant-level table lookup operations through asynchronous multidimensional carbon emission coefficient lookup. The organic integration of technologies in each stage enables the system to maintain stable real-time early warning capabilities even under high-concurrency scenarios, while the carbon emission prediction results possess strong physical realism and engineering deployability.
[0071] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.
Claims
1. A method for predicting carbon emission hotspots in highway service areas, characterized in that, include: Receive lightweight data packets reported from edge computing nodes deployed on highways. These lightweight data packets are generated by the edge computing nodes after classifying, counting, and aggregating vehicles collected by traffic detection equipment at a preset time window. A dynamic spatiotemporal graph of the road network is constructed based on the lightweight data package. Each service area is used as a node, and the ratio of the total number of vehicles currently in the service area to the maximum parking capacity is used as the node dynamic saturation. The edge weights are dynamically updated based on the interval flow difference between adjacent nodes. The dynamic spatiotemporal graph is input into the spatiotemporal graph neural network, and the predicted number of fuel vehicles and electric vehicles in each service area within a future preset time period is output through at least two prediction branches. Based on the dynamic saturation, the idling time of a single vehicle in each service area is calculated through a mapping function, and the direct carbon emissions are calculated by combining the predicted number of fuel vehicles, idling fuel consumption rate and carbon emission conversion factor. The supply status parameters of charging facilities in each service area are obtained. The predicted number of electric vehicles is corrected by the supply-side constraint function to obtain the effective charging demand. The carbon emission coefficients corresponding to each service area and the future preset time period are obtained. The indirect carbon emissions are calculated in combination with the effective charging demand. The direct carbon emissions are superimposed with the indirect carbon emissions to obtain a comprehensive carbon emission prediction value, and hotspot marking and early warning are performed on the road network topology map.
2. The method for predicting carbon emission hotspots in highway service areas according to claim 1, characterized in that, The mapping function is a non-linear piecewise function based on a saturation threshold: When the dynamic saturation is not greater than the first threshold, the single vehicle idling time is taken as the reference idling time; When the dynamic saturation is greater than the first threshold and not greater than the second threshold, the single vehicle idling time is equal to the sum of the baseline idling time and the linear congestion growth coefficient multiplied by the portion of the saturation exceeding the first threshold. When the dynamic saturation is greater than the second threshold, the vehicle idling time is superimposed with an exponential deterioration term on the sum of the baseline idling time and the linear congestion growth coefficient multiplied by the portion of saturation exceeding the first threshold. The exponential deterioration term is the product of the exponential deterioration parameter and an exponential function of the portion of saturation exceeding the second threshold.
3. The method for predicting carbon emission hotspots in highway service areas according to claim 1, characterized in that, The supply status parameter is the real-time occupancy rate of charging piles, the supply-side constraint function is a variant of the Sigmoid function, the effective charging demand is equal to the difference between the predicted number of electric vehicles multiplied by 1 and the demand loss penalty value, the demand loss penalty value is equal to the sum of the maximum loss rate upper limit divided by 1 plus the sum of the product of the negative sensitivity parameter and the real-time occupancy rate minus the queuing congestion threshold, where the queuing congestion threshold ranges from 85% to 95%.
4. The method for predicting carbon emission hotspots in highway service areas according to claim 1, characterized in that, The step of dynamically updating edge weights based on the interval flow difference between adjacent nodes includes: The difference between the upstream inflow and the downstream outflow of adjacent nodes is calculated as the interval flow difference value; When the interval flow difference is less than the preset normal fluctuation threshold, the edge weight takes the base time value; When the interval flow difference value is greater than the normal fluctuation threshold, the edge weight is dynamically increased based on the interval flow difference value through an incrementing function to reflect the spatiotemporal transmission delay.
5. The method for predicting carbon emission hotspots in highway service areas according to claim 1, characterized in that, The carbon emission coefficient is obtained through a pre-set multidimensional carbon emission coefficient lookup table. The lookup table is indexed by at least two dimensions, namely season, weather conditions and time period. It is calculated in advance based on the weighted average of the green electricity consumption ratio and the thermal power supply ratio, which is based on the historical operation logs or simulation model data of the microgrid energy management system and combined with the main grid carbon emission factor. It is updated and maintained asynchronously outside the real-time inference process.
6. The method for predicting carbon emission hotspots in highway service areas according to claim 1, characterized in that, The classification and counting aggregation is implemented using a dual-channel approach: The first channel uses license plate recognition equipment to distinguish between new energy vehicles and fuel vehicles based on the color of the license plate. The second channel verifies the vehicle type registration feature code by reading the vehicle's electronic tag unit through the electronic non-stop toll collection gantry.
7. The method for predicting carbon emission hotspots in highway service areas according to claim 1, characterized in that, The spatiotemporal graph neural network is trained offline using an adaptive joint loss function. The adaptive joint loss function is the sum of the product of the first weight parameter and the prediction error of fuel vehicles, the product of the second weight parameter and the prediction error of electric vehicles, and the product of the third weight parameter and the prediction error of carbon emissions. The first weight parameter and the second weight parameter are gradient-based learnable weights.
8. The method for predicting carbon emission hotspots in highway service areas according to claim 1, characterized in that, The spatiotemporal graph neural network is a multi-task spatiotemporal graph convolutional network, including an input layer that receives historical multi-time step node features and dynamic edge weight matrices, a shared representation layer that extracts spatial topological interaction features and temporal evolution features, and a task branch output layer with separate branches for fuel vehicle prediction and electric vehicle prediction.
9. The method for predicting carbon emission hotspots in highway service areas according to claim 1, characterized in that, The lightweight data packet includes timestamps, node identifiers, traffic cross-section throughput, total number of fuel vehicles entering the station, total number of electric vehicles entering the station, and total number of vehicles exiting the station.