A method for predicting carbon emissions from highway service areas based on spatiotemporal graph embedding

Abstract

This invention relates to the fields of energy transportation and low-carbon operation and maintenance technology, and in particular to a method for predicting carbon emissions from highway service areas based on spatiotemporal graph embedding. The method includes acquiring vehicle count streams at gantry and service area entrances / exits; constructing a dynamic topology and spatiotemporal prediction framework based on corridor-level tensors; generating vehicle flow embedding vectors and comprehensive parking intensity curves; mapping service areas as agents and dividing them into scheduling clusters based on the comprehensive parking intensity curves; training and freezing agent policies through three-stage carbon pulse penalties and distributed policy gradients; generating emission matrices and heatmaps based on power plans and emission factor libraries; and comparing predicted and measured emissions to decompose errors. This invention, by generating corridor-level emission matrices and heatmaps, compares predicted emissions with measured data grid by grid, and uses a dual-pointer method to decompose vehicle flow embedding distortion and equipment policy drift errors, ensuring long-term stability of carbon emission prediction accuracy and scheduling effectiveness.

Description

Technical Field

[0001] This invention relates to the fields of energy transportation and low-carbon operation and maintenance technology, and in particular to a method for predicting carbon emissions from highway service areas based on spatiotemporal graph embedding. Background Technology

[0002] As key nodes in transportation energy consumption and carbon emissions, the energy consumption and emission control of highway service areas are of great significance to the achievement of dual carbon goals in the transportation sector. Currently, the industry mainly focuses on two technical paths for energy consumption and emission control of highway service areas: one is based on single-node traffic volume prediction, using time series models such as seasonal difference ARIMA, support vector regression, and LSTM, combined with external variables such as rainfall, temperature and humidity, and holiday markings to extrapolate short-term passenger flow in service areas, thereby assisting in energy consumption prediction; the other is to implement centralized scheduling-based time-sharing power limits on the load side, relying on traditional building energy management systems to divide high-consumption equipment into independent sub-circuits, and execute load control based on manually set time-sharing tables or rule-based start-stop logic through PLC bus.

[0003] However, existing technical solutions still have significant shortcomings and are difficult to meet the needs of accurate prediction and dynamic management of carbon emissions from service areas at the corridor scale. On the one hand, single-node prediction models ignore the dynamic transmission characteristics of traffic flow between the gantry and service area nodes. Differences in data caliber and sampling period lead to inconsistent measurement of the same traffic flow between upstream and downstream, making it impossible to depict the cross-node progression of traffic flow peaks. When faced with long peaks during holidays or sudden changes in traffic flow caused by severe weather, the response is lagging. On the other hand, load-side control schemes mostly use static energy consumption coefficients to estimate equipment power demand, failing to reflect the differentiated energy response caused by different service areas and different vehicle types of passenger flow. Moreover, the centralized power limit framework treats the adjustable load of the service area as a discrete variable, ignoring the impact of nonlinear changes in traffic flow on energy consumption. A large energy consumption safety margin needs to be reserved to avoid risks, resulting in low energy utilization efficiency. At present, a carbon emission prediction method for highway service areas based on spatiotemporal graph embedding is needed. Summary of the Invention

[0004] To address the challenge that existing technologies cannot meet the requirements for accurate prediction and dynamic management of carbon emissions from service areas at the corridor scale, this invention provides a method for predicting carbon emissions from highway service areas based on spatiotemporal graph embedding.

[0005] This invention provides a method for predicting carbon emissions from highway service areas based on spatiotemporal graph embedding, employing the following technical solution:

[0006] A method for predicting carbon emissions from highway service areas based on spatiotemporal graph embedding, comprising:

[0007] S1. Obtain the vehicle count streams at the gantry and entrances / exits of each service area, process the vehicles to generate vehicle equivalent sequences and personnel equivalent sequences, and aggregate them into corridor-level tensors;

[0008] S2. Based on corridor-level tensor, construct a dynamic topology and spatiotemporal prediction framework, including establishing a dynamic directed graph containing gantry and service area, and using a spatiotemporal graph neural network with embedded gating module to generate vehicle prediction equivalents.

[0009] S3. Calculate the dynamic delay based on the vehicle prediction equivalent, separate the upstream transmission amount from the local fluctuation amount, and generate the vehicle flow embedding vector and the comprehensive parking intensity curve.

[0010] S4. Combine the comprehensive parking intensity curve to map the service area into intelligent agents and divide it into scheduling clusters. Calculate the cluster-level power baseline and broadcast it to the intelligent agents. The intelligent agents perform discrete control of equipment and report operating data.

[0011] S5. Construct a cold start rule table based on historical operation data, and control the neighboring lanes by combining the probability of traffic flow impact. Train and freeze the agent policy through three-stage carbon pulse penalty and distributed policy gradient to obtain the power plan and carbon emission curve.

[0012] S6. Generate emission matrix and heat map based on power plan and emission factor library, compare the emission splitting error between prediction and actual emission, and complete the final output.

[0013] Furthermore, step S1 specifically includes the following sub-steps:

[0014] S11. Read the vehicle count stream, and sequentially perform vehicle length differentiation, axle load label supplementation, and license plate attribute mapping. Classify vehicles using a three-dimensional cross-reference table of length, axle load, and license plate, and merge and accumulate each category to generate a vehicle type-specific count sequence. , , ;

[0015] S12, Call the set of conversion factors Mapping the vehicle type count sequence with the corresponding coefficient product yields the vehicle equivalent sequence and the personnel equivalent sequence;

[0016] S13. Aggregate vehicle equivalent sequences and personnel equivalent sequences according to corridor number, lane direction and inner program number, accumulate them through a daily cumulative window to obtain a corridor-level daily aggregation vector, and write it into a unified tensor structure.

[0017] in, , and These represent the conversion factors for passenger cars, buses, and trucks, respectively. , and These are the category labels for passenger cars, buses, and trucks, respectively.

[0018] Furthermore, the establishment of a dynamic directed graph including gantries and service areas includes writing gantries and service areas into a node list incrementally by kilometer markers and adding virtual boundary nodes, recording node attributes, creating directed edges for nodes with adjacent mileage coordinates in node list V based on the actual vehicle travel direction, forming a set of directed edges E, and constructing an adjacency list A by combining node list V and set of directed edges E to obtain the initial directed graph structure. Based on each directed edge in the adjacency list, the mileage difference between points, the average driving speed within a window, and the traffic event level are extracted, and the propagation impedance of each directed edge is calculated using a weighted normalization overlay method. The expression for the propagation impedance is:

[0019] ,

[0020] in, For the edge At any moment Propagation impedance, These are the weights of the three factors. For nodes To node The difference in kilometers, The maximum kilometer difference between adjacent nodes in the same direction. For free flow reference velocity, For the edge At any moment average speed, To ensure the denominator is a very small positive number, and to prevent the denominator from being zero, For the edge At any moment Event level, This is the event amplification factor, which controls the magnitude of the increase in impedance caused by the disaster.

[0021] Furthermore, the generation of vehicle prediction equivalents using a spatiotemporal graph neural network with an embedded gating module includes: extracting the daily aggregated vehicle equivalent vector of each node based on a unified tensor structure; mapping it to the input feature matrix of the spatiotemporal graph neural network according to the order of the node list; constructing a spatiotemporal graph neural network containing graph convolutional layers and temporal convolutional layers; embedding a gating module between the graph convolutional layers and the temporal convolutional layers; inputting the input feature matrix X into the spatiotemporal graph neural network; capturing spatial correlations through graph convolutional layers combined with real-time propagation impedance; capturing temporal dependencies through temporal convolutional layers; and performing hourly inference after modulation by the gating module to output a vehicle equivalent prediction sequence. The expression of the vehicle equivalent prediction sequence is:

[0022] ,

[0023] in, This is the spatiotemporal graph neural network inference function for the gating module. Let A be the feature submatrix corresponding to the vehicle model, and let A be the adjacency list. Let be the propagation impedance matrix at time t. This indicates the status of the gating module.

[0024] Furthermore, the step of calculating the dynamic delay separation of upstream transmission and local fluctuation based on vehicle prediction equivalents includes aggregating service area nodes within the corridor according to the vehicle equivalent prediction sequence in ascending order of kilometer markers, and establishing a local prediction sequence for each target service area node. Translational Overlap Sequence with Upstream Nodes The system collects the associated physical parameters between the target service area node and each upstream node, calculates the dynamic delay for each vehicle model, and then shifts the prediction sequence of the upstream nodes clockwise based on the dynamic delay, comparing it with the local prediction sequence of the target service area. The upstream conduction amount is obtained by taking the minimum value hourly. The dynamic delay calculation formula is as follows:

[0025] ,

[0026] in, upstream node To the target node The road distance, Vehicle category The reference velocity in free flow. Vehicle category At any moment The average speed of the corridor For the edge At any moment The severity level of the accident, The overall weather severity level for the corridor. The weighting coefficients are based on four factors: distance, speed, accidents, and weather. It is a very small positive number, used to avoid the denominator being zero. Vehicle category From node Drive to the node Dynamic latency.

[0027] Furthermore, the generation of the traffic flow embedding vector and the comprehensive dwell intensity curve includes calculating the upstream transmission ratio and the local fluctuation ratio based on the upstream transmission amount, generating the cross-regional traffic flow embedding vector based on the upstream transmission ratio and the local fluctuation ratio, and calling the historical offline dwell probability surface. By combining the time delay correction coefficients corresponding to the vehicle flow embedding vector, future time intervals are generated. The vehicle arrived on the timeline. Search for car models At any moment Dwell probability factor ,Will Multiplying the signal strength component in the traffic flow embedding vector yields the single-vehicle model parking intensity sequence. For three types of vehicles By performing time alignment and intensity overlay, the comprehensive dwelling intensity curve for the next decision window of the target service area is obtained.

[0028] Furthermore, the step of mapping service areas to intelligent agents and dividing them into scheduling clusters by combining the comprehensive parking intensity curve includes constructing a service area index table based on the comprehensive parking intensity curve of each service area and the static attribute data of all service areas in the corridor, converting the service area corresponding to each row in the index table into multiple independent control intelligent agents according to the mapping relationship, clustering the intelligent agents according to the driving direction of the target service area and the adjacency of the kilometer markers, and grouping several intelligent agents with the same direction and consecutively adjacent kilometer markers into the same scheduling cluster.

[0029] Furthermore, the calculation of the cluster-level power baseline and its broadcast to the agent includes calling upon the service area average load factor and corridor total power limit from the historical operation database to calculate the cluster-level power baseline for each scheduled cluster. Construct a power baseline broadcast channel to The scheduling cluster affiliation is used to broadcast to the corresponding intelligent agent and generate a list of discrete device actions. The formula for calculating the cluster-level power baseline is as follows:

[0030] ,

[0031] Where m is the scheduling cluster The number of intelligent agents within, For intelligent agents The rated transformer capacity of the corresponding service area The historical average load factor, M represents the upper limit of the total power of the corridor, and M represents the total number of scheduling clusters within the corridor.

[0032] Furthermore, step S5 specifically includes the following sub-steps:

[0033] S51. Combine historical operational data and establish an index based on the three-dimensional keys of season, time period and service area type to generate a start-stop rule table. Map the start-stop rule table to a combination of power levels of multiple high-power devices and load the corresponding cold start strategy set for each intelligent agent.

[0034] S52. Calculate the probability of traffic flow impact in the next four hours based on the vehicle equivalent prediction sequence and the comprehensive parking intensity curve, set the traffic flow impact probability threshold, and determine the opening and closing status of the agent's neighboring cell capacity transfer channel based on the probability threshold.

[0035] S53. Based on the measured carbon emissions reported by the agent in the previous hour at the start of the current decision cycle, calculate the rolling median value of carbon emissions for the entire corridor. Using the median value as a benchmark, calculate the emission exceedance, spatial synchronization indication, and rate of increase exceeding the threshold. Construct a three-segment carbon pulse penalty function based on the emission exceedance, spatial synchronization indication, and rate of increase exceeding the threshold.

[0036] S54. The agent is trained using a distributed policy gradient algorithm. A reward function is constructed based on carbon pulse penalty. After each iteration, the reward value of each agent is normalized, and the normalized reward is fed back to the value network of the corresponding agent to update the policy parameters.

[0037] Furthermore, step S6 specifically includes the following sub-steps:

[0038] S61. Map the power values ​​of the five types of equipment in the power level plan (lighting, HVAC, catering, charging piles and refueling islands) to the emission factors of the corresponding energy paths on an hourly basis to obtain the hourly carbon emission value of each service area.

[0039] S62. Write the hourly carbon emission values ​​of each service area into the corridor-level emission matrix. The rows of the corridor-level emission matrix correspond to timestamps and the columns correspond to service area node indices. Use color gradient rendering to generate a real-time carbon emission heat map from the emission matrix.

[0040] S63. Collect hourly measured carbon emission data from each service area to form a measured emission matrix. Compare the emission matrix with the measured emission matrix cell by cell. Use the dual-pointer method to decompose the error and obtain the traffic flow embedding distortion error and the equipment strategy drift error.

[0041] S64. Accumulate the traffic flow embedding distortion error and the equipment strategy drift error by service area nodes respectively to obtain the total distortion error and total drift error of each node. Use the emission matrix, carbon emission heat map and error decomposition results as the final output.

[0042] In summary, the present invention has the following beneficial technical effects:

[0043] This invention performs three-dimensional cross-classification on the vehicle counting stream at the gantry and service area entrances and exits, distinguishing vehicle length, supplementing axle load labels, and mapping license plate attributes. It categorizes heterogeneous vehicle data from different nodes into three standard categories: passenger cars, buses, and trucks. Then, it aggregates vehicle and personnel equivalent sequences according to corridor number, lane direction, and inner program number, and writes them into a unified tensor structure. This process unifies the sampling granularity, statistical caliber, and spatial correlation logic of the entire corridor data, avoiding the defects of chaotic data dimensions and spatiotemporal mismatch in different nodes in traditional solutions. It provides a homogeneous and comparable data foundation for subsequent dynamic topology construction and collaborative scheduling.

[0044] 2. This invention constructs a dynamic directed graph containing gantries and service areas based on corridor-level tensors. By collecting real-time data on mileage differences between nodes, driving speeds, and traffic event levels, it dynamically calculates the propagation impedance of directed edges, characterizing the attenuation and lag patterns of traffic flow peaks advancing along nodes. Simultaneously, a gating module is embedded in the spatiotemporal graph neural network, automatically opening remote information channels during peak holiday periods or severe weather, strengthening cross-regional traffic flow information aggregation, and avoiding the problem of delayed response to sudden traffic flows in traditional static models. This design enables vehicle equivalent prediction to synchronously match the spatial transmission characteristics and temporal dependence features of traffic flow, significantly improving the accuracy of prediction results in adapting to actual traffic flow changes.

[0045] 3. This invention calculates dynamic time delay based on vehicle prediction equivalents, shifts the upstream node prediction sequence according to the actual travel rhythm, and compares it hourly with the service area's local prediction sequence to separate upstream transmission volume and local fluctuation volume. Then, it combines historical dwell probability data to generate a comprehensive dwell intensity curve, clearly distinguishing the differentiated energy needs of transit traffic and local traffic. Compared with the traditional solution that estimates energy consumption using static coefficients, this design can accurately characterize the energy consumption characteristics of traffic from different sources, making subsequent equipment power regulation more in line with actual passenger flow needs.

[0046] 4. In step S4, this invention maps each service area as an independent intelligent agent, divides scheduling clusters according to the adjacency of the driving direction and the kilometer marker, calculates the cluster-level power baseline and distributes it uniformly through the broadcast channel to ensure that the power of the entire corridor is dynamically allocated within the safe capacity; step S5 further introduces a three-stage carbon pulse penalty mechanism, combined with a distributed policy gradient algorithm to train the intelligent agent, so that each service area can dynamically balance carbon emission control and service quality assurance, avoiding the global mismatch problem of traditional centralized scheduling. This collaborative mechanism can effectively smooth the corridor-level power peak, reduce carbon emission pulses caused by concentrated start-up and shutdown of equipment, and improve energy utilization efficiency.

[0047] 5. This invention generates a corridor-level emission matrix and heat map, compares the predicted emissions with the measured data grid by grid, uses a dual-pointer method to separate traffic flow embedding distortion and equipment strategy drift error, and corrects the model parameters of the previous steps in reverse according to the error results. It can continuously adapt to changes in traffic flow and fluctuations in equipment operating conditions, and ensure the long-term stability of carbon emission prediction accuracy and scheduling effect. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the overall process of a method for predicting carbon emissions from highway service areas based on spatiotemporal graph embedding, according to an embodiment of the present invention.

[0049] Figure 2 This is a comparison diagram of the embodiments of the present invention and the solutions of conventional technologies.

[0050] Figure 3 This is a flowchart of step S3 in an embodiment of the present invention.

[0051] Figure 4 This is a flowchart of step S4 in an embodiment of the present invention.

[0052] Figure 5 This is a flowchart of step S5 in an embodiment of the present invention.

[0053] Figure 6 This is a flowchart of step S6 in an embodiment of the present invention. Detailed Implementation

[0054] The present invention will be further described in detail below with reference to the accompanying drawings.

[0055] Example 1

[0056] Reference Figure 1 This embodiment of a method for predicting carbon emissions from highway service areas based on spatiotemporal graph embedding includes:

[0057] S1. Obtain the vehicle count streams at the gantry and entrances / exits of each service area, process the vehicles to generate vehicle equivalent sequences and personnel equivalent sequences, and aggregate them into corridor-level tensors;

[0058] S2. Based on corridor-level tensor, construct a dynamic topology and spatiotemporal prediction framework, including establishing a dynamic directed graph containing gantry and service area, and using a spatiotemporal graph neural network with embedded gating module to generate vehicle prediction equivalents.

[0059] S3. Calculate the dynamic delay based on the vehicle prediction equivalent, separate the upstream transmission amount from the local fluctuation amount, and generate the vehicle flow embedding vector and the comprehensive parking intensity curve.

[0060] S4. Combine the comprehensive parking intensity curve to map the service area into intelligent agents and divide it into scheduling clusters. Calculate the cluster-level power baseline and broadcast it to the intelligent agents. The intelligent agents perform discrete control of equipment and report operating data.

[0061] S5. Construct a cold start rule table based on historical operation data, and control the neighboring lanes by combining the probability of traffic flow impact. Train and freeze the agent policy through three-stage carbon pulse penalty and distributed policy gradient to obtain the power plan and carbon emission curve.

[0062] S6. Generate emission matrix and heat map based on power plan and emission factor library, compare the emission splitting error between prediction and actual emission, and complete the final output.

[0063] Specifically, a method for predicting carbon emissions from highway service areas based on spatiotemporal graph embedding includes the following steps:

[0064] S1. Obtain the vehicle count streams at the gantry and entrances / exits of each service area, process the vehicles to generate vehicle equivalent sequences and personnel equivalent sequences, and aggregate them into corridor-level tensors;

[0065] like Figure 1 , Figure 2 As shown, S11 is the total vehicle counting stream where the gantry and the entrances / exits of each service area are on the same corridor. Read in chronological order, and perform three mapping operations on each vehicle record sequentially:

[0066] First, based on the vehicle length threshold This embodiment differentiates between long and short vehicles, defining passenger cars as ≤6m, buses as 6-12m, and trucks as >12m. By comparing the length data of each vehicle passage record, the vehicle category range can be preliminarily determined;

[0067] Second, load labels are added to the classification results based on the axle load range. For vehicles that have been initially classified, load labels are added according to the axle load range: passenger car axle load ≤ 10t, bus axle load 10-20t, and truck axle load > 20t. If a vehicle is 5m long (initially classified as a passenger car) but has an axle load of 12t (exceeding the passenger car axle load threshold), it is corrected to be classified as a bus candidate, thus realizing cross-verification of length and axle load.

[0068] Third, based on the license plate classification code, supplement the usage attributes of the vehicles with labels, map the license plate attributes of the vehicles with supplemented axle load labels, and complete the final vehicle attribute confirmation.

[0069] After all three mappings are completed, the mapping results are used to locate the category number using a three-dimensional cross-reference table based on length, axle load, and license plate number. Each record is then uniquely assigned to one of the three major categories: passenger cars, buses, or trucks. Subsequently, the counts of vehicles of the same category within the same node and time interval are merged and accumulated to generate a vehicle-type counting sequence that maintains the original sampling interval, including a passenger car counting sequence. Bus counting sequence , and truck counting sequence Where t is the time interval identifier, and the sampling granularity is locked during the generation of this sequence. It does not change with node differences.

[0070] S12. Based on the past 12 months of operational data for this highway corridor (including energy consumption test data by vehicle type and passenger volume statistics for typical vehicles), a set of preset conversion factors is established. ,in, , and These represent the conversion factors for passenger cars, buses, and trucks, respectively. Specifically, taking passenger cars as the benchmark, their vehicle equivalent contribution is set to 1.0, i.e. Because large buses have a larger passenger capacity and longer dwell time, their energy consumption requirement is 3.5 times that of small buses. This embodiment sets... Because trucks have high load capacities, require maintenance, and refuel frequently, this embodiment sets... Similarly, based on historical data statistics, the equivalent conversion factors for the three types of personnel are determined.

[0071] The three counting sequences obtained in sub-step S11 are respectively called by the fixed vehicle and personnel conversion factor sets. ;Will , , At the same time By matching the corresponding coefficients and performing a product mapping, the vehicle equivalent sequence can be directly obtained. Human equivalent sequence The vehicle equivalent sequence The calculation formula is:

[0072] ,

[0073] in, , and These are the category markers for passenger cars, buses, and trucks, respectively. The physical dimensions of the vehicle are used for subsequent calculations of the electrical and energy modules; Based on typical vehicle passenger-energy consumption test values, a dual-dimensional scale is established to ensure that both vehicle and passenger perspectives are available at the same time.

[0074] S13. Convert the vehicle equivalent sequence generated by each node in sub-step S12. Human equivalent sequence The sequence is aggregated and sorted using a three-element ordered key-value pair: corridor number, lane direction, and inner program number; a daily cumulative window is then applied to the sorted sequence. Perform the accumulation to obtain the corridor-level daily aggregation vector. and ;in, For the cumulative number of passenger cars on a single day, For the cumulative number of large passenger buses on a single day, This refers to the cumulative vehicle equivalent of trucks on a single day. This refers to the cumulative passenger volume of passenger vehicles on a single day. This refers to the cumulative passenger equivalent of a single day's large passenger bus. The daily cumulative number of personnel equivalent for trucks will be... and The corresponding kilometer marker index and direction identifier are also written into the unified tensor structure. Tensors are directly used as the node inputs of the spatiotemporal graph neural network and as the common data benchmark for subsequent collaborative energy consumption decisions. Through aggregation, upstream and downstream service areas reference vehicle and personnel equivalent information of the same source, same dimension, and same direction at any subsequent time point.

[0075] S2. Based on corridor-level tensor, construct a dynamic topology and spatiotemporal prediction framework, including establishing a dynamic directed graph containing gantry and service area, and using a spatiotemporal graph neural network with embedded gating module to generate vehicle prediction equivalents.

[0076] First, read the corridor-level tensor. The coordinates and node types of all gantry and service area kilometer markers recorded in the system are written into node list V in ascending order of kilometer marker values. Simultaneously, one virtual boundary node is added to the beginning and end of the list to close the actual vehicle travel path. For each node in node list V, three immutable attributes are recorded simultaneously: type identifier. (Value taken from gantry service area or virtual), traffic flow direction (Value taken as up or down), mileage coordinates (Unit: kilometers) Next, based on the actual direction of vehicle travel (incrementing by kilometer markers for the uphill direction and decreasing by kilometer markers for the downhill direction), directed edges are created for nodes with adjacent kilometer markers in node list V. For example, a directed edge is created between gantry K100 (node ​​i) and gantry K105 (node ​​j) in the uphill direction. A directed edge is created between the K105 gantry (node ​​j) and service area A (node ​​k). This forms a set of directed edges E. Combining the node list V and the set of directed edges E, an adjacency list A is constructed using the adjacency list format. Each row in adjacency list A corresponds to one node, recording the neighboring nodes, directed edge identifiers, and basic attributes of that node, thus obtaining the initial directed graph structure. .

[0077] Then, for each directed edge in adjacency list A Three physical parameters are collected in real time: the distance difference between node i and node j. (Unit: km) and the average speed within the sliding window over the past 10 minutes. (Unit: km·h) Third, the traffic event level of the directed edge at time t. Values ​​1–5 correspond to normal to severe congestion, with 1 representing normal traffic flow, 2 representing minor congestion, 3 representing moderate congestion, 4 representing severe congestion, and 5 representing severe congestion (reported in real-time by the traffic incident monitoring system). Finally, each directed edge is calculated using a weighted normalization overlay method. Real-time propagation impedance at time t The propagation impedance expression is:

[0078] ,

[0079] in, For the edge At any moment Propagation impedance, These are the weights of the three factors. For nodes To node The difference in kilometers, The maximum kilometer difference between adjacent nodes in the same direction. For free flow reference velocity, For the edge At any moment average speed, To ensure the denominator is a very small positive number, and to prevent the denominator from being zero, For the edge At any moment Event level, This is the event amplification factor, controlling the magnitude of the impedance increase caused by the disaster. The propagation impedance of each directed edge is calculated in real time to ensure the initial directed graph structure... Transformed into a dynamically changing directed graph over time. ,in, Let be the propagation impedance matrix at time t, according to a two-level threshold. and right Perform dynamic segmentation if Then Marking a graph as weakly connected causes graph convolutions to retain only first-order information and block further downstream propagation; if Immediately restore the strong connection, allowing information to be transmitted without attenuation.

[0080] Through the aforementioned continuously evolving impedance weights, the framework maintains lightweight calculations for nearest neighbors under normal conditions, while automatically weakening long-range diffusion paths during peak holiday periods or severe weather, ensuring that downstream service areas obtain traffic flow prediction benchmarks that better match the physical rhythm of travel in advance.

[0081] The corridor-level tensor generated in step S1 Extract the daily aggregated vehicle equivalent vector for each node (including gantries and service areas). The daily aggregated vehicle equivalent data (including passenger cars, buses, and trucks) is mapped to the input feature matrix X of the spatiotemporal graph neural network according to the order of the node list V. The dimension of the input feature matrix X is the number of nodes × time step × feature dimension. The "number of nodes" is the total number of nodes in the node list V. In this embodiment, the corridor node list V contains 143 nodes, so the number of nodes = 143. The time step is the number of time intervals for historical data sampling (taking 10-minute data from the past 7 days, 144 time intervals per day, so the time step = 7 × 144 = 1008). The feature dimension is 3, corresponding to the vehicle equivalent data of the three types of vehicles: passenger cars, buses, and trucks.

[0082] Next, a spatiotemporal graph neural network is constructed, comprising graph convolutional layers and temporal convolutional layers: the graph convolutional layers use the ChebNet graph convolution operator, with the input being the input feature matrix X and the adjacency list A, combined with the propagation impedance matrix at time t. The system captures spatial correlation features between nodes and adjusts the information transmission weights between nodes by dynamically changing the propagation impedance. For example, directed edges with high propagation impedance correspond to reduced information transmission strength between nodes, simulating the delayed transmission of traffic flow information under congestion conditions. The temporal convolutional layer uses a 1D convolution operator with a kernel size of 3, and the input is the output features of the graph convolutional layer, used to capture the temporal dependencies of vehicle equivalent data, such as the traffic flow time sequence patterns during morning and evening peak hours. A holiday-weather gating module is embedded between the graph convolutional layer and the temporal convolutional layer. The gating module The control logic is driven by real-time holiday information and weather data: when the current period is detected to be a statutory holiday or the weather level is ≥2 (1 for sunny, 2 for cloudy / light rain, 3 for heavy rain / strong wind, 4 for heavy rain / snowstorm), the gating module... Output control signal "1" to open remote jump connection channels across 3 or more nodes, allowing traffic flow information from upstream nodes to be directly transmitted to remote downstream nodes, enhancing the rapid aggregation of cross-regional traffic flow information in abnormal scenarios; when it is a normal time period and the weather level is <2, the gating module... Output control signal "0" to close the remote jump connection channel, retaining only the information transmission between adjacent nodes to reduce unnecessary computational redundancy. Then, input feature matrix X is input to the above-mentioned embedded gating module. The spatiotemporal graph neural network, through graph convolutional layers and propagation impedance matrix The synergistic effect accurately captures the spatial correlation features of traffic flow between different nodes; the temporal convolutional layer captures the dependence of vehicle equivalent in the time dimension; and the holiday-weather gating module... The information transmission paths are dynamically modulated under different scenarios to ensure the network's responsiveness to abnormal traffic flow. After multiple rounds of inter-layer iterative inference (10 iterations, with the Adam optimizer updating network parameters in each iteration and a learning rate of 0.001), the spatiotemporal graph neural network outputs a node-level vehicle equivalent prediction sequence for the next 24 hours, categorized by vehicle type. Finally, the network output is formatted to generate a passenger car vehicle equivalent prediction sequence. Bus vehicle equivalent prediction sequence Truck vehicle equivalent prediction sequence Where t is a 10-minute time interval identifier for the next 24 hours (t=1 corresponds to the first 10 minutes of the prediction start time, and t=144 corresponds to the last 10 minutes of the prediction end time), the above prediction sequence is generated according to the formula:

[0083] ,

[0084] in, They represent passenger cars, buses, and trucks, respectively. This is the spatiotemporal graph neural network inference function for the gating module. Let A be the feature submatrix corresponding to the vehicle model, and let A be the adjacency list. Let be the propagation impedance matrix at time t. As the state of the gating module, this prediction sequence provides accurate prior traffic data for the subsequent steps S3, including traffic source segmentation and parking intensity calculation.

[0085] S3. Calculate the dynamic delay based on the vehicle prediction equivalent, separate the upstream transmission amount from the local fluctuation amount, and generate the vehicle flow embedding vector and the comprehensive parking intensity curve.

[0086] like Figure 3 As shown, S31 is the 24-hour prediction tensor output from step S2, which has been sorted by vehicle type by node. Aggregation is performed in ascending order of kilometer markers; for each service area node in the aggregation... Two buffer queues are established simultaneously to record the predicted sequence for this region, denoted as . ;

[0087] Another summary is the sequence obtained by translating and superimposing all upstream nodes, denoted as... .

[0088] To maintain a consistent inference pace, the scrolling window length is set uniformly for both queues. And lock step size Subsequently, vehicle type indexes were established for the three categories of vehicles: passenger cars, buses, and trucks. , , To ensure that subsequent disassembly actions are performed under the same vehicle model criteria, and finally, to explicitly mark the time intervals in the predicted sequence that fall within the holiday or official accident duration ranges. This provides peak position clues for the decomposition operator and reserves high-risk warning positions for subsequent energy-consuming agents.

[0089] S32. Decomposition of execution latency-source for upstream-target node pairs:

[0090] 1. Combine all upstream nodes within the aggregation. Target service area By vehicle type Extract the 24-hour prediction sequence one by one Then, the driving distance was used. Real-time average speed in the corridor Free flow velocity Accident level Weather rating Five-item generation dynamic delay The distance and speed terms provide the base values ​​for the time scale, while the accident and weather terms are multiplicatively increased under abnormal scenarios.

[0091] II. Based on each upstream sequence After performing a clockwise shift, sum the results for the same vehicle model, and then combine this sum with the predicted sequence for this region. By taking the minimum value hourly, the conduction quantity explained by the upstream is obtained. The upstream transmission ratio can be obtained by summing the values ​​over 24 hours. Local fluctuation ratio ,in .

[0092] The two-stage action described above, through a progressive path of physical time delay correction and minimum value truncation, explicitly separates explainable external shocks from unexplainable local fluctuations, providing clear leading signals for subsequent agents:

[0093] ,

[0094] ,

[0095] in, upstream node To the target node Road distance, in km; Vehicle category Reference velocity in free flow, in km· ; Vehicle category At any moment The average speed of the corridor; For the edge At any moment Accident severity level (1–5); The overall weather severity level for the corridor is (0–3). The four weighting coefficients for distance, speed, accident, and weather are preset and meet industry safety standards; It is a very small positive number, used to avoid the denominator being zero; Vehicle category From node Drive to the node Dynamic latency, in hours (h); , Vehicle category At the node or The The estimated equivalent for the next hour; To predict the time step, it is fixed at 1 hour; Vehicle category At the node The The conduction equivalent that can be explained upstream within an hour; Vehicle category Upstream transmission ratio; Vehicle category Local fluctuation ratio, always satisfying .

[0096] S33, Output the code from sub-step S32 , and holiday markers Accident markers The four types of dominant factors are concatenated in the order of transmission intensity, fluctuation intensity, and special state to generate the cross-regional traffic flow embedding vector. .

[0097] Replicating sequentially across three vehicle model dimensions And add a vehicle model index to form This vector set provides interpretable components while maintaining directionality, enabling subsequent energy agents to instantly identify the proportion of external peaks and the degree of local fluctuations during reading.

[0098] The meanings and origins of the symbols in S33 are as follows: This represents the transmission ratio sequence that can be explained by the upstream. This represents the local fluctuation ratio sequence; For hourly holiday markers, For accident status labeling, these four types of factors are all directly obtained from the time delay-source decomposition of S32. Specifically, for service areas... With car model S32 based on dynamic delay Shift each upstream prediction sequence and compare it with the prediction sequence for this region. The upstream conduction quantity is obtained by taking the minimum value hourly. and cause local volatility ,then , and ; Provided by the holiday calendar, Based on the traffic event grading flow at the hourly granularity (normalized values ​​can be used), the cross-regional traffic flow embedding vector is defined as follows: And add a vehicle index form ,gather This is the result described in S33.

[0099] S34, will Corresponding time delay correction coefficient Combine them, and call the linear time-shift mapping for each vehicle type; first, based on Generate future The arrival timeline of a moment, and then the probability surface of offline historical dwell time. Searching for corresponding probability factors Finally Multiply by the conduction strength component to output the dwell strength. sequence The three types of vehicles By performing time alignment and intensity overlay, the service area is obtained. Overall dwell curve of the next decision window and complete embedding pairs .

[0100] S4. Combine the comprehensive parking intensity curve to map the service area into intelligent agents and divide it into scheduling clusters. Calculate the cluster-level power baseline and broadcast it to the intelligent agents. The intelligent agents perform discrete control of equipment and report operating data.

[0101] like Figure 4 As shown, an index table is constructed by combining the comprehensive occupancy intensity curve and the static attributes of service areas, and then mapped to the agent: Multiple pairs of comprehensive occupancy intensity curves for each service area are extracted from the output of step S3, while simultaneously retrieving the static attribute data of all service areas within the corridor, including service area type identifiers. Rated transformer capacity (Unit: kVA) Main building functions The comprehensive parking intensity curve is generated according to the service area number, driving direction, and kilometer marker coordinates. Integrate with the aforementioned static attribute data to construct a service area index table. Each row in the index table corresponds to one service area, including the service area number, direction, kilometer marker, and stopping intensity curve. ,type Rated transformer capacity Main building functions Then, according to the one-to-one mapping relationship, the index table is... Each row of service areas is transformed into an independent control agent, forming a set of agents. Each of the intelligent agents , Each agent uniquely corresponds to one service area and possesses core functions such as receiving power commands, executing equipment control, and reporting operational data. Furthermore, the agent's ID and directional attributes are consistent with the information of the corresponding service area in the index table.

[0102] Next, scheduling clusters are divided based on the adjacency of driving direction and kilometer markers: For the generated agent set A, it is first initially classified according to driving direction, completely separating agents in the up direction from those in the down direction, ensuring that agents in the same scheduling cluster correspond to the same traffic flow direction in their respective service areas, avoiding interference from differences in the spatiotemporal transmission characteristics of cross-directional traffic flow with the power coordination scheduling logic; then, agents in the same direction are sorted in ascending order of kilometer marker coordinates of their associated service areas, ensuring that the sorting of agents is consistent with the spatial order of the actual travel path of vehicles in the corridor, laying a spatial foundation for subsequent power coordination based on traffic flow transmission; finally, according to the clustering principle of same direction and continuous, uninterrupted kilometer markers, the sorted agents are divided into several scheduling clusters. Each scheduling cluster contains several consecutive adjacent agents, and there are no breaks in the kilometer marker intervals of the service areas corresponding to all agents within the cluster. Record the number of agents m in the cluster and the rated transformer capacity corresponding to each agent.

[0103] Subsequently, cluster-level power baselines were calculated and broadcast channels were constructed for targeted distribution. Service area average load factors were extracted based on historical operational databases. This factor is defined as the average ratio of the actual operating power of the service area to the rated transformer capacity during the same historical period. It reflects the load utilization level of the service area under normal operating conditions, with a value range of 0-1. It is calculated from the load data of the same period in the past 6-12 months, providing historical operating basis for the power baseline and avoiding the baseline setting from deviating from the actual operating scenario. At the same time, it retrieves the upper limit of the total power of the corridor. (Determined by the power supply contract signed between the regional substation and the highway operator, a global constraint upper limit is set for the total power consumption of all service areas in the entire corridor to prevent the risk of power outages caused by exceeding the power supply capacity); Based on the above parameters, a calculation method with dual constraints of intra-cluster equipment capacity constraint + corridor total power constraint is adopted, and the cluster-level power baseline calculation formula is as follows:

[0104] ,

[0105] Where m is the scheduling cluster The number of intelligent agents within, For intelligent agents The rated transformer capacity of the corresponding service area The historical average load factor, M represents the upper limit of the total power of the corridor, and M represents the total number of scheduling clusters in the corridor. The total power is allocated according to the proportion of the number of intelligent agents in the cluster to ensure that the power demand of each cluster matches the total capacity and avoid some clusters having excessive power while others have insufficient power.

[0106] Create a broadcast channel at the top of the corridor This allows the scheduling layer to forward power adjustment commands to all agents at once; whenever traffic flow embeds... When the wave crest is advancing upstream and the downstream clusters have not yet been impacted, the scheduling layer first tightens control on the upstream clusters where the wave crest has not yet arrived. Then, the reserved capacity is released to the downstream clusters that are about to be impacted, and finally the increase or decrease is calculated. Broadcast to This enables cross-node power rolling, allowing all service areas to share a consistent total power constraint at the same time, thus suppressing serial overflow from the source.

[0107] Finally, the agents perform discrete control of the devices and report operational data: each agent receives the cluster-level power baseline of the corresponding scheduling cluster. With additional instructions, the system maps the five types of high-consumption equipment it controls—lighting, HVAC, catering, charging stations, and refueling islands—to a set of discrete power levels preset in this document. Within each intelligent agent, five types of high-consumption equipment—lighting, HVAC, catering, charging stations, and refueling islands—are mapped to a fixed power level set. Each gear level and its duration Combining into discrete actions and write it to the action list. .for Generate device-level channels so that any action in the list can trigger the closing or opening of field branch switches during the corresponding time period, thus preventing the strategy from remaining only at the digital level.

[0108] For each intelligent agent Assemble the observation vector And record the sampling timestamp for all information fragments. The vectors are concatenated in the following order: outdoor temperature and humidity of the area, local electricity price, service area type fragment, and traffic flow embedding generated in step S3. The latest broadcast power baseline at the scheduling layer enables... Simultaneously presenting cross-regional traffic impact, station operating costs, and real-time power constraints, it provides multi-dimensional decision-making basis for the strategy network, thereby enabling the agent to balance carbon peak suppression and customer comfort in actual execution.

[0109] In the broadcast channel Add feedback pathway At the end of each decision cycle, the agent... Two indicators were proactively reported for actual power usage. Based on the proportion of unmet demands, after the scheduling layer has collected all reports, a linear penalty function is applied. For the cluster-level baseline of the next cycle Make corrections; through The new baseline will be redistributed.

[0110] The penalty function is data-driven and used to perform monotonic and stable closed-loop correction of the cluster-level baseline in the next cycle within the total power constraint of the corridor.

[0111] Specifically, it involves using samples from each decision-making cycle. (in The slope is obtained by performing a least-squares fit on the proportion of unmet demand. In conjunction with capacity clamping and smoothing to avoid oscillation, a newer type is formed.

[0112] ,

[0113] in, , This represents the upper limit of capacity within the cluster and the corridor. This is the smoothing coefficient. The linear form originates from the empirical monotonic marginal relationship and convexity requirement, facilitating rapid online iteration; By using grid search and cross-validation selection on the historical validation set, the non-satisfaction rate is reduced and the baseline does not overshoot. Then, the S4 process is performed. It is used only for baseline correction and is independent of the three-segment carbon pulse penalty in S5 training; the latter uses the rolling median as a reference to segmentally weight overamplitude, spatial synchronization and rate overthreshold, and is used for policy learning, but does not directly participate in real-time baseline correction.

[0114] Power limit across the entire corridor The internal dynamic allocation of capacity allows each service area to flexibly acquire or relinquish capacity based on real-time traffic flow, thus suppressing serial overflow and carbon emission spikes caused by information silos from the source.

[0115] S5. Construct a cold start rule table based on historical operation data, and control the neighboring lanes by combining the probability of traffic flow impact. Train and freeze the agent policy through three-stage carbon pulse penalty and distributed policy gradient to obtain the power plan and carbon emission curve.

[0116] like Figure 5 As shown, firstly, a cold start rule table is constructed by combining historical operational data, and a cold start strategy set is loaded: The historical operational database defined in this document is retrieved, and three types of core historical data are extracted: historical passenger flow logs, branch power curves (including hourly power data for five types of high-consumption equipment: lighting, HVAC, catering, charging piles, and refueling islands, reflecting the correlation between equipment power and passenger flow), and emission ledgers (including hourly measured carbon emission values, used to correlate power consumption and carbon emission patterns). The above data is indexed using a three-dimensional key based on season, time period, and service area type. "Season" is divided into four categories: spring, summer, autumn, and winter; time period is divided into five categories: early morning, morning peak, off-peak, evening peak, and nighttime; and service area type is consistent with step S4. The historical data is categorized through the three-dimensional index, generating a start / stop rule table. ;For the start / stop rule table Each index record is mapped to a power level combination of five types of high-power devices based on historical power distribution curves. This ensures that each rule corresponds to the optimal equipment level scheme under a specific season, specific time period, and specific service area type. Subsequently, according to the service area type, each agent generated in step S4 is loaded with a dedicated cold start strategy set. That is, the comprehensive agent is loaded with the rule subset corresponding to the comprehensive service area, and the refueling-side agent is loaded with the rule subset corresponding to the refueling-side service area. This enables each agent to have a basic control strategy that conforms to its own functional attributes in the initial stage of training, avoiding equipment operation disorder or service quality degradation caused by random strategies in the early stage of training.

[0117] Based on the dynamic control of the agent's neighboring area capacity transfer channel based on traffic flow impact probability: The system calls upon the vehicle equivalent prediction sequence for the next 24 hours by vehicle type output in step S2 and the comprehensive parking intensity curve generated in step S3, using the next four hours as the prediction window, to calculate the traffic flow impact probability for each agent's corresponding service area. The calculation logic for the probability of traffic flow impact is to first determine the 90th percentile of the historical comprehensive parking intensity of the service area corresponding to the agent. As a benchmark for judging traffic flow impact, the predicted stopping intensity per hour for the next four hours is then calculated. The formula for calculating the probability of traffic flow impact is:

[0118] ,

[0119] in, This is an indicator function; it takes the value 1 if the condition within the parentheses is met, and 0 otherwise. A traffic flow impact probability threshold is set based on the 90th percentile of the historical comprehensive parking intensity of the service area corresponding to the intelligent agent. ,when When it is determined that the service area corresponding to the intelligent agent will face traffic congestion in the next four hours, it automatically opens the capacity transfer channel of its neighboring area. This allows the agent to report backup capacity requests to neighboring agents within the same scheduling cluster, while also allowing neighboring agents to report their remaining capacity, thus achieving complementary capacity replenishment within the cluster. If the traffic flow is determined to be at a normal level, the capacity transfer lane in the adjacent area will be closed.

[0120] Subsequently, a three-stage carbon pulse penalty function was constructed and the penalty calculation logic was determined: based on the measured carbon emission values ​​reported by all agents in the corridor over the past hour, a rolling median algorithm was used to calculate the dynamic baseline. For each agent in the current hour h, calculate three core derived indicators: emission exceedance, etc. Spatial synchronization indication When three consecutive service areas in the same direction simultaneously meet the requirements Set to 1 if the rate of ascent exceeds the threshold, otherwise set to 0; ,in, To calculate the slope of emissions in this area over one hour, first calculate the current hourly carbon emission increase rate of the agent. Then, compared with the preset maximum allowable rate of ascent. For comparison, the formula is:

[0121] ,

[0122] Based on the above three indicators, a three-segment carbon pulse penalty function is constructed. :

[0123] ,

[0124] in, , The threshold for exceeding emission limits, The base penalty coefficient satisfies , This is the spatial synchronization magnification factor. It is a rate over-threshold amplification factor (used to amplify the penalty for a sharp increase in carbon emissions). Through segmented penalties, it achieves precise control of mild over-amplitude suppression and severe over-amplitude suppression, while taking into account spatial coordination and rate control, and avoiding the omission of carbon emission pulses caused by single-dimensional penalties.

[0125] Among them, such as Figure 1As shown, error backfeeding refers to the automatic triggering of model correction and weight updates during continuous system operation when the deviation between the predicted result and the actual monitored value exceeds a set threshold (when the average error or relative error exceeds a set range (generally 5% to 20% of the maximum power or 5% to 15% of the average error)). The "penalty coefficient enhancement" refers to the system applying higher penalty weights to periods with larger errors during carbon emission prediction or scheduling optimization to encourage faster model convergence.

[0126] The penalty coefficient is divided into three levels:

[0127] The first level is the safe range, corresponding to normal deviation, and uses the basic weight;

[0128] The second level is the alarm range. When the error exceeds the set first threshold (usually 2% to 5% of the maximum power), the system gradually increases the penalty coefficient.

[0129] The third level is the over-limit range. When the error exceeds the second threshold (usually 5% to 10% of the maximum power), the penalty coefficient will rise rapidly to the upper limit range (generally 50 to 200 times the base value) to force the algorithm to converge and reduce subsequent deviations.

[0130] Finally, the distributed policy gradient algorithm is used to train the agents and freeze the policy: the distributed policy gradient algorithm (PG) defined in this document is used to train the agent set synchronously, and the reward function is constructed with the goal of maximizing the long-term comprehensive benefit during the training process. The reward function consists of three parts: the actual power reported in step S4, and the reward function itself. unmet demand ratio Three-stage carbon pulse punishment The specific expression is:

[0131] ,

[0132] in, , and Each agent has a different weighting coefficient. After each iteration, the corridor scheduling layer collects the reward values ​​of all agents. The min-max normalization method is used to map the reward value to the [0,1] interval. The reward value of each agent is normalized to avoid training bias caused by differences in the magnitude of reward values ​​between different agents. The reward term comes from the formal modeling of the corridor-level operational objective. Power constraint compliance, quality of service assurance, and carbon pulse suppression are mapped to three observable and dimensionless quantities, resulting in... ,in, The effect of power bias (the deviation between actual power and cluster-level baseline, normalized by min–max). The proportion of unmet demand. A three-segment carbon pulse penalty based on a rolling median benchmark. Obtained through offline calibration, representing the weights of different coefficients: in the past Seeking answers in the historical log. ,in, Let q be the expected reward value of the agent, constrained by the upper limit of the total power of the corridor and the service quality threshold. Bayesian optimization is used and cross-validation is performed to select stable weights; all quantities are uniformly calculated. Interval normalization is used to eliminate dimensional differences; the required data are derived from the actual power reported by the agent. On-site service completion rate And the amplitude exceeding the limit, spatial synchronization indication and slope exceeding the threshold constructed from the emission measurement sequence, combined as .

[0133] The normalized reward is fed back into the value network of the corresponding agent. The agent's policy parameters are updated using a policy gradient descent algorithm, enabling the agent to gradually learn the optimal control policy that balances power constraints, service quality, and carbon emission suppression; a policy convergence threshold is set. The change in policy parameters of each agent during three consecutive iterations. Upon successful convergence of the policy training, the policy network parameters of the agent are frozen to prevent parameter drift during subsequent operation. The frozen policy network then performs rolling inference on traffic demand for the next 24 hours, outputting power level plans for five types of high-consumption equipment. Simultaneously, it calls the emission factor library defined in this document (containing carbon emission factors for three energy pathways: fuel, electricity, and self-generated power) to generate the power level plans. The formula for mapping hourly particle size to carbon emissions is as follows:

[0134] ,

[0135] in, Let the planned power of device i in hour h be denoted as . Hourly carbon emission curves are generated based on the carbon emission factors corresponding to the energy pathways.

[0136] The policy network is obtained by training and converging a multi-agent PG network, with each agent observing... (Considering factors such as dwell intensity, cluster-level power baseline, weather conditions, and electricity prices) Input to both ends of the network: Strategy With value The network uses a single temporal convolutional layer for extraction. The time characteristics are represented by two fully connected layers, outputting discrete power level distributions; parameters are initialized using Xavier, the optimizer is Adam, and the policy follows the time characteristics. Updated, value is trained using TD error. Training is performed in distributed synchronous rounds. At the end of each round, the reward for each agent is normalized to min–max and backpropagated; when three consecutive rounds meet the condition... Convergence is determined and parameters are frozen. The frozen strategy network is used to generate power tier plans for the next 24 hours on a rolling basis, and mapped with the emission factor library to obtain hourly carbon emission curves.

[0137] S6. Generate emission matrix and heat map based on power plan and emission factor library, compare the emission splitting error between prediction and actual emission, and complete the final output.

[0138] like Figure 6 As shown, S61, the power level plan for each service area output in step S5 is... With emission factor library A step-by-step mapping was performed. During the mapping process, factor groups were selected for each of the three energy pathways: oil-fired boilers, purchased electricity, and self-generated electricity. The carbon emission values ​​for five types of equipment were calculated at the hourly level. Aggregate all nodes in the same direction of travel according to the kilometer marker order and write them into the corridor emission matrix. The rows of the matrix represent timestamps. The column represents the node index. .right Color gradient rendering is used to generate a real-time emission heat map, which is then projected onto a large scheduling screen.

[0139] S62, the predicted emission matrix formed in sub-step S61 With real-time meter reading matrix Grid-by-grid comparison is performed, and the source and load dual-pointer method is used to split the error. The first pointer is anchored to the source of traffic flow, which is the stopping intensity of the same node in step S3. On-site passenger flow count Difference attribution to vehicle flow embedding distortion The second pointer anchors the equipment load by comparing the planned power at the same node. Power collected on site Difference attribution to device policy drift The two types of errors are accumulated separately as follows: and When the cumulative value exceeds the threshold or At that time, for the node Mark the source error or load error to clarify the direction of subsequent backtracking.

[0140] S63. Perform difference analysis on the marked nodes and calculate the upstream energy saving downgrade rate using the proportional method. With downstream advance reserve capacity .Will and Write the coordinated start-up and shutdown schedule for the next 24 hours in ascending order of kilometer markers. At the same time, Feedback is sent to step S2, and the false weather gating threshold is adjusted proportionally. To enhance the model's sensitivity to abnormal traffic flow; Feedback is sent to step S5 to linearly increase the carbon pulse penalty amplification factor. and This strengthens the strategy's ability to suppress device drift. Finally, a new cluster-level power baseline is established. With start and stop schedule via broadcast channel The data is uniformly distributed to all intelligent agents to achieve corridor-level closed-loop calibration and initiate the next round of collaborative operation.

[0141] Example 2

[0142] A specific application of a method for predicting carbon emissions from highway service areas based on spatiotemporal graph embedding is provided:

[0143] During the data preparation phase, 112 million original vehicle passage records were collected from 107 sets of highway gantry gantries, 28 entrance weighing stations, and 62 sets of ETC gantries. All records were first mapped to vehicle categories using a three-dimensional mapping table from length to axle load and from axle load to license plate number. Vehicle and passenger equivalent dual sequences were then generated using fixed conversion factors. Subsequently, edge weight functions were constructed using mileage difference, average window speed, and traffic event level to generate a dynamic directed graph covering 34 service area pairs and 296 directional edges. The impedance matrix was refreshed every five minutes and written to the time-series graph database. Based on this graph structure, a spatiotemporal graph neural network underwent 14-day sliding window training on an A100 GPU cluster. The holiday-weather gating module focused on the return peak from the sixth to the eighth day of the Lunar New Year, amplifying the information channel across a 120-kilometer section through gating weights.

[0144] During the commissioning phase, the system deployed cluster-level control on the EMS master station side, dividing the 34 service areas into 6 load clusters, each cluster integrating a power baseline broadcast channel. The local PLCs in the service areas were loaded with multi-agent strategy execution modules, and these modules were linked in real-time with five fixed-level execution channels (lighting, HVAC, catering, fast charging, and refueling) via Modbus-TCP. From 17:00 on February 17th to 02:00 on February 19th, 2024 (the peak return travel period), the system continuously output a 24×12 pane vehicle equivalent prediction matrix, and then obtained the upstream transmission and local fluctuation results through time delay and source dual-level operators. The multi-agent strategy used this as input to calculate power increases and decreases and distribute them to each cluster. Distribution SCADA records show that the 10 service areas in the upstream Shanfen section reduced their HVAC and lighting power by an average of 17.4% between 18:00 and 22:00. Meanwhile, the 12 service areas in the downstream Jingzhu section simultaneously obtained 3.7MW of transferable capacity. The peak load dropped from 41.3MW in the same period of history to 34.6MW, and the peak duration was reduced from 110 minutes to 48 minutes.

[0145] Table 1 shows the specific data.

[0146]

[0147] To verify the effectiveness, the core indicators of the technical method in this application were compared with those of three existing commonly used solutions in the same time period. The results are summarized as follows:

[0148] The data shows that the technology in this application has achieved better performance in terms of prediction accuracy, peak load control and carbon emission reduction, while maintaining minute-level response. Although centralized planning has a lower prediction error, it takes a long time to solve and cannot support five-minute rolling prediction. The single-node model and manual scheduling lack perception of cross-node peak transmission during peak periods, resulting in limited peak shaving and emission reduction effects.

[0149] like Figure 2 As shown, existing solutions mostly employ single-agent regression prediction models, such as time-series regression (e.g., ARIMA) or ensemble learning algorithms (e.g., XGBoost). These models can only rely on historical power or carbon emission univariate data for short-term predictions and cannot effectively capture the spatial correlation between service areas and the transmission relationship between upstream and downstream areas. Their output is usually a static prediction curve, requiring manual correction and periodic parameter adjustment. The prediction accuracy of such models drops significantly during holiday peaks or sudden traffic events, and they cannot adaptively optimize according to changes in the external environment.

[0150] The algorithm proposed in this invention introduces a dynamic spatiotemporal graph neural network framework in its structure. By embedding spatial relationship nodes between the gantry and the service area, and combining dynamic data from multiple sources such as traffic flow, energy consumption, and emissions, a three-dimensional information coupling of nodes, edges, and time is formed.

[0151] Based on a real-time closed-loop system for load and emissions, the proposed method resulted in an 11.6% decrease in overall corridor electricity demand despite an 8.2% increase in total traffic volume. The emissions heatmap and power migration logs automatically generated by the operations center were recognized by a third-party auditing agency as a complete and traceable accounting chain, providing a standardized interface for subsequent replication to port transport corridors or intercity railway complexes. Therefore, this embodiment verifies the practicality and potential for widespread application of the proposed technology in large-scale, multi-node, and highly volatile transportation energy systems.

[0152] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting carbon emissions of a highway service area based on spatio-temporal graph embedding, characterized in that, Comprise: S1, obtain gantry and each service area entrance and exit vehicle counting flow, generate vehicle equivalent sequence and personnel equivalent sequence through vehicle classification processing, and aggregate into corridor level tensor; S2, based on the corridor level tensor, construct a dynamic topology and a space-time prediction framework, including establishing a dynamic directed graph containing the gantry and the service area, and generating vehicle predicted equivalent using a space-time graph neural network with an embedded gating module; S3, according to the vehicle predicted equivalent, calculate the dynamic time delay to split the upstream conduction quantity and the local fluctuation quantity, generate the traffic flow embedding vector and the comprehensive parking intensity curve; S4, combine the comprehensive parking intensity curve to map the service area to an agent and divide the dispatch cluster, calculate the cluster level power baseline and broadcast it to the agent, and the agent executes device discrete control and reports operation data; S5, according to the historical operation data, build a cold start rule table, and control the adjacent area channel combined with the traffic impact probability, train and freeze the agent strategy through three-stage carbon pulse punishment and distributed policy gradient, obtain the power plan and carbon emission curve; S6, based on the power plan and the emission factor library, generate the emission matrix and the heat map, compare the prediction and the measured emission split error, and complete the final output.

2. The highway service area carbon emission prediction method based on space-time graph embedding according to claim 1, characterized in that Step S1, specifically comprising the following substeps: S11, read the vehicle counting flow, sequentially execute vehicle length distinction, axle load label supplement and license plate attribute mapping, classify vehicles through three-dimensional cross-retrieval table of length, axle load and license plate, and generate sub-vehicle type counting sequence by merging and accumulating each class 、 、 ; S12, calling a set of conversion coefficients The vehicle equivalent sequence and the personnel equivalent sequence are obtained by mapping the vehicle type count sequence and the corresponding coefficient product. S13, aggregate the vehicle equivalent sequence and the personnel equivalent sequence according to the corridor number, the lane direction and the inner procedure number, accumulate the corridor level daily aggregated vector through the single day accumulation window, and write it into a unified tensor structure; wherein, , and represent the conversion factors for small passenger cars, large passenger cars, and trucks, respectively, , and represent the class markers for small passenger cars, large passenger cars, and trucks, respectively.

3. The method of claim 1, wherein, The dynamic directed graph containing the gantry and the service area includes writing the gantry and the service area into the node list in increasing order of kilometer post and appending virtual boundary nodes, recording node attributes, creating directed edges between nodes with adjacent mile coordinates in the node list V according to the actual travel direction of vehicles, forming a directed edge set E, constructing an adjacency list A combining the node list V and the directed edge set E, obtaining an initial directed graph structure, extracting the inter-node mile difference, the window average speed and the traffic event level from each directed edge of the adjacency list, and calculating the propagation impedance of each directed edge by weight normalization superposition. The propagation impedance expression is: ， in, For the edge At any moment Propagation impedance, These are the weights of the three factors. For nodes To node The difference in kilometers, The maximum kilometer difference between adjacent nodes in the same direction. For free flow reference velocity, For the edge At any moment average speed, To ensure the denominator is a very small positive number, and to prevent the denominator from being zero, For the edge At any moment Event level, This is the event amplification factor, which controls the magnitude of the increase in impedance caused by the disaster.

4. The method of claim 1, wherein, The space-time graph neural network with an embedded gating module generates vehicle predicted equivalent, including extracting the daily aggregated vehicle equivalent vector of each node based on the unified tensor structure, mapping the input feature matrix of the space-time graph neural network according to the order of the node list, constructing a space-time graph neural network containing a graph convolution layer and a time convolution layer, embedding a gating module between the graph convolution layer and the time convolution layer, inputting the input feature matrix X into the space-time graph neural network, capturing spatial correlation through the graph convolution layer combined with real-time propagation impedance, capturing time sequence dependence through the time convolution layer, executing time-by-time reasoning after modulation by the gating module, and outputting the vehicle equivalent prediction sequence. The vehicle equivalent prediction sequence expression is: ， wherein, is a spatio-temporal graph neural network inference function of a gating module, is a corresponding vehicle model feature sub-matrix, A is an adjacency list, is a propagation impedance matrix at time t, is a gating module state.

5. The method of claim 1, wherein, The dynamic time delay is calculated according to the vehicle prediction equivalent, the upstream conduction quantity is split from the local fluctuation quantity, which comprises aggregating the service area nodes in the corridor according to the vehicle equivalent prediction sequence in the order of kilometer post increment, and establishing the local prediction sequence for each target service area node The upstream node translation superposition sequence The associated physical parameters of the target service area node and each upstream node are collected, the dynamic time delay is calculated for each vehicle type one by one, and the prediction sequence of the upstream node is translated clockwise after the dynamic time delay, and then superposed with the local prediction sequence of the target service area The minimum value is taken every hour to obtain the upstream conduction quantity every hour, and the dynamic time delay calculation formula is: ， wherein, is the road distance from the upstream node to the target node , is the vehicle category reference speed in free flow state, is the vehicle category corridor average speed at time , is the accident severity level at time for the edge , is the corridor overall weather severity level, are the distance, speed, accident, and weather weight coefficients, is a very small positive number to avoid zero denominator, is the dynamic delay for the vehicle category from the node to the node .

6. The method of claim 1, wherein, The process of generating the traffic flow embedding vector and the comprehensive dwell intensity curve includes calculating the upstream transmission ratio and the local fluctuation ratio based on the upstream transmission volume, generating the cross-regional traffic flow embedding vector based on the upstream transmission ratio and the local fluctuation ratio, and calling the historical offline dwell probability surface. By combining the time delay correction coefficients corresponding to the vehicle flow embedding vector, future time intervals are generated. The vehicle arrived on the timeline. Search for car models At any moment Dwell probability factor ,Will Multiplying the signal strength component in the traffic flow embedding vector yields the single-vehicle model parking intensity sequence. For three types of vehicles By performing time alignment and intensity overlay, the comprehensive dwelling intensity curve for the next decision window of the target service area is obtained.

7. The method of claim 1, wherein, The combined comprehensive parking intensity curve maps the service area as an agent and divides the scheduling cluster, including constructing a service area index table based on the comprehensive parking intensity curve of each service area and the static attribute data of all service areas in the corridor, converting the corresponding service area of each row in the index table into multiple independent control agents according to the mapping relationship, clustering and dividing the agents according to the driving direction of the target service area and the adjacent nature of the kilometer post, and grouping the agents with the same direction and kilometer post into the same scheduling cluster.

8. The method of claim 1, wherein, The cluster-level power baseline is broadcast to the intelligent agent, including calling the average load factor of the service area and the upper limit of the total power of the corridor in the historical operation database, and calculating the cluster-level power baseline of each dispatch cluster , a power baseline broadcast channel is constructed, and is broadcast to the corresponding intelligent agent according to the dispatch cluster affiliation relationship and a device discrete action list is generated, and the cluster-level power baseline calculation formula is: ， wherein m is the number of agents in the cluster , is the rated transformation capacity of the corresponding service area of the agent , is the historical average load factor is the corridor total power cap, and M is the total number of clusters in the corridor.

9. The method of claim 1, wherein, Step S5 specifically includes the following sub-steps: S51, combine historical operation data and establish an index according to three dimensions of season, time period and service area type to generate a start-stop rule table, map the start-stop rule table to a plurality of high-consumption device power gear combinations, and load the corresponding cold start strategy set for each agent; S52, calculate the future four-hour traffic impact probability according to the vehicle equivalent prediction sequence and the comprehensive parking intensity curve, set a traffic impact probability threshold, and determine the switch state of the neighbor area capacity transfer channel of the agent according to the probability threshold; S53, take the carbon emission measured value reported by the agent in the last hour at the start time of the current decision cycle as the basis data, calculate the rolling median value of the carbon emission of the entire corridor, and take the median value as the reference to calculate the emission amplitude, spatial synchronization indication and rising rate threshold, and construct a three-section carbon pulse penalty function according to the emission amplitude, spatial synchronization indication and rising rate threshold; S54, train the agent using a distributed policy gradient algorithm, construct a reward function based on the carbon pulse penalty, perform normalization processing on the reward value of each agent after each iteration, and feed the normalized reward back to the value network of the corresponding agent to update the policy parameters.

10. The method of claim 1, wherein, Step S6 specifically includes the following sub-steps: S61, map and calculate the power values of the five types of devices in the power gear plan, i.e. lighting, heating, catering, charging piles and fueling islands, by hour granularity and corresponding emission factors of energy paths to obtain the hourly carbon emission value of each service area; S62, write the hourly carbon emission value of each service area into a corridor-level emission matrix, which corresponds to the time stamp in the row and the service area node index in the column, and generate a real-time carbon emission heat map by rendering the emission matrix using a color gradient; S63, collect hourly measured carbon emission data of each service area to form a measured emission matrix, compare the emission matrix with the measured emission matrix by grid, split the error using the double-pointer method, and obtain the traffic embedding distortion error and device strategy drift error; S64, accumulate the traffic embedding distortion error and the device strategy drift error according to the service area nodes respectively to obtain the total distortion error and the total drift error of each node, and take the emission matrix, the carbon emission heat map and the error splitting result as the final output.

Images