A Road System Energy Consumption Estimation Method Based on Data Fusion and Generative Adversarial Networks

Abstract

This invention relates to a road system energy consumption estimation method based on data fusion and generative adversarial networks (GANs). The method includes: collecting macroscopic traffic data, including average vehicle speed and vehicle spacing distribution across cross sections, using cross-section detectors; constructing a GAN containing a fuzzy logic module to generate individual vehicle speed sequences from the macroscopic traffic data; obtaining individual vehicle acceleration sequences through differential operations on the individual speed sequences; establishing a differentiable fuzzy logic module to identify driving style categories based on speed and acceleration features; constructing a style-aware fusion module to fuse kinematic features and driving style features to estimate individual vehicle fuel consumption; and aggregating individual vehicle fuel consumption to obtain a road segment-level fuel consumption estimate. This invention enables high-precision estimation of road segment-level fuel consumption under low-cost observation conditions such as cross-section detectors. By explicitly modeling driving styles using differentiable fuzzy logic, it effectively improves the accuracy and interpretability of fuel consumption estimation.

Description

Technical Field

[0001] This invention relates to the fields of intelligent transportation and deep learning technology, specifically to a method for estimating the energy consumption of road systems based on data fusion and generative adversarial networks. Background Technology

[0002] Road segment-level fuel consumption estimation is an important foundation for traffic management and environmental protection, and can be used for applications such as traffic signal optimization, flow control strategy formulation, carbon emission assessment, and energy efficiency management of intelligent connected vehicles.

[0003] Traditional methods for obtaining road segment fuel consumption data mainly rely on devices such as GPS equipment, on-board OBD terminals, or high-precision radar to collect speed and acceleration data of individual vehicles, and then calculate fuel consumption using fuel consumption models. However, these methods generally suffer from problems such as high equipment costs, difficulty in data acquisition, and limited coverage, making large-scale deployment across the road network difficult. For example, the penetration rate of GPS floating car data is typically only 5%-15%, which is insufficient to represent the entire traffic flow; on-board OBD devices need to be installed on a vehicle-by-vehicle basis, resulting in high deployment costs and difficult maintenance.

[0004] In contrast, cross-sectional detectors such as loop coils and microwave radars have been widely deployed on highways and urban roads, and can reliably acquire macroscopic data such as cross-sectional traffic volume, average vehicle speed, and headway, but cannot directly provide information on the speed distribution and behavioral characteristics of individual vehicles, and therefore cannot directly support fuel consumption estimation.

[0005] In existing research, some scholars have tried to use deep learning methods to infer micro-states from macro data, but there are the following shortcomings: (1) Traditional generative adversarial networks have pattern collapse problems, which leads to unstable generation results; (2) Most methods do not consider the impact of subjective behavioral factors such as driving style on fuel consumption, while the impact of driving style (cautious / normal / aggressive) on fuel consumption can reach 15%-30%; (3) Some methods use attention mechanisms to implicitly model driving behavior, which lacks interpretability and has high training costs. Summary of the Invention

[0006] The purpose of this invention is to overcome the problems existing in the prior art and provide a road system energy consumption estimation method based on data fusion and generative adversarial networks. It aims to achieve high-precision estimation of road segment-level fuel consumption based solely on macroscopic data collected by low-cost observation equipment such as cross-section detectors, thereby providing data support for energy conservation and emission reduction in transportation.

[0007] To achieve the above-mentioned technical objectives and effects, the present invention is implemented through the following technical solution: A method for estimating energy consumption of road systems based on data fusion and generative adversarial networks. The method includes the following steps: Step S1: Collect macroscopic traffic data using a cross-section detector to obtain the average vehicle speed at time t. Vehicle spacing distribution ; Step S2: Construct a generative adversarial network (GAN). The generator G generates individual vehicle speed sequences based on the average vehicle speed, spacing distribution, and random noise vector. ; Step S3: Obtain individual vehicle acceleration sequences through differential operations. ; Step S4: Establish a differentiable fuzzy logic module, and identify driving styles based on speed and acceleration characteristics through fuzzy reasoning, classifying drivers into three categories: cautious, normal, and aggressive. Step S5: Construct a style perception fusion module that integrates kinematic features and driving style features, and estimate the individual vehicle fuel consumption e through a fuel consumption regression network. t,n ; Step S6: Aggregate individual vehicle fuel consumption to obtain road segment-level fuel consumption estimates. .

[0008] Furthermore, in step S1, the cross-section detector includes a loop detector, a microwave radar detector, or a video detector; macroscopic traffic data is collected at fixed time intervals ∆t, wherein the average vehicle speed at the cross-section is... To detect the average speed of all vehicles passing through the cross section, the vehicle spacing distribution... The data was obtained by calculating the time difference and speed of adjacent vehicles passing each other; during data preprocessing, min-max normalization was used to normalize the data to the range [0,1]. , in, and These represent the minimum and maximum velocity values ​​among all training samples, respectively.

[0009] Furthermore, in step S2, the specific structure and generation process of generator G are as follows: Step S201: The generator G input includes: the average vehicle speed corresponding to time step t. Vehicle spacing distribution and random noise vector z t ; Step S202: Generator G generates individual vehicle speed distributions through nonlinear mapping: , Where, θ G These are the generator parameters; Step S203: To preserve the sparsity of the generated results, output the generator G. Element-wise multiplication with the input sparsity mask M: , Where ⊙ represents element-wise multiplication; the mask M takes the value of M[n]=1 when vehicle n exists, otherwise M[n]=0; Step S204: The loss function of generator G is: , Where, λ loss This represents the weights of the loss function.

[0010] Furthermore, in step S3, the specific method for the difference operation is as follows: Step S301: Calculate the acceleration of vehicle n at time t using the first-order finite difference method: , Where ∆t is the sampling time interval, v t,n and v t+1,n Let be the speeds of vehicle n at times t and t+1, respectively; Step S302: Obtain the effective vehicle acceleration through masking calculation: , Among them, A t M is the acceleration matrix for all vehicles. t The vehicle at time t has a mask.

[0011] Furthermore, in step S4, the specific method for modeling driving style using the differentiable fuzzy logic module is as follows: Step S401: Define trapezoidal membership functions for velocity and acceleration, mapping the input variables to the interval [0,1]. The trapezoidal membership function is defined as follows: , Where a1, a2, a3, and a4 are the left base, left peak, right peak, and right base parameters of the trapezoidal function, respectively; Step S402: The fuzzy set of velocity variables includes low speed. medium speed ,high speed Three categories of fuzzy sets of acceleration variables include small... ,middle ,big Three categories; Step S403: Use the T-norm operator to process the fuzzy AND relationship for inferring aggressive driving style: ; Step S404: Use the S-conorm operator to process the fuzzy "OR" relationship to infer cautious driving style: ; Step S405: Employ the Reichenbach implication function and introduce the sigmoid activation function to enhance differentiability: , Where σ(·) is the sigmoid function and k is the scaling parameter, used to control the smoothness of the inference boundary; Step S406: Define the driving style reasoning rule set: Rule 1: If the speed is high and the acceleration is large, the driving style is aggressive. Rule 2: If the speed is low or the acceleration is small, the driving style is cautious. Rule 3: In all other cases, the driving style is normal; Step S407: Obtain the final driving style inference result through the product aggregator: .

[0012] Furthermore, in step S5, the style-aware fusion module estimates individual fuel consumption using the following specific method: Step S501: Construct the velocity-acceleration feature projection module H v Mapping kinematic features to a high-dimensional space: , Among them, [v t,n ,a t,n ] represents the concatenated vector of velocity and acceleration; Step S502: Construct the style feature projection module H s Encode driving style categories into feature vectors: , Among them, st n For vehicle n, define the driving style category; Step S503: Construct the style modulation feature vector through element-wise multiplication: ; Step S504: Input the modulation features into the fuel consumption regression network f SAF Output the estimated fuel consumption of individual vehicles: .

[0013] Furthermore, in step S6, the specific method for road segment-level fuel consumption aggregation is as follows: Step S601: For time t, the time-level fuel consumption is obtained by averaging the fuel consumption estimates of all vehicles: , Where, N t This represents the number of valid vehicles in the road segment at time t. Step S602: For the time period [t1, t2], obtain the total fuel consumption of the road segment through time integration: , Where ∆t is the sampling time interval.

[0014] Furthermore, in step S2, the objective function for optimizing the discriminator D of the generative adversarial network is: , in, This represents the actual vehicle speed distribution; To predict vehicle speed distribution; λ is the gradient penalty weight, with a value range of 0.1≤λ≤10, used to ensure that the discriminator D satisfies the 1-Lipschitz condition; During training, interpolated samples are introduced to calculate the gradient penalty term: , The joint objective function is: .

[0015] The beneficial effects of this invention are: 1. In environments with limited observation conditions, this invention utilizes adversarial network technology to provide a road segment-level fuel consumption estimation method based on macroscopic cross-sectional data. Fuel consumption estimation can be achieved using only macroscopic traffic data provided by cross-sectional detectors, eliminating the need to deploy high-cost trajectory acquisition equipment such as GPS and OBD, significantly reducing data acquisition costs, and possessing broad practical engineering application value.

[0016] 2. Compared with the existing technology, the present invention introduces a differentiable fuzzy logic module to explicitly model driving style, classifying drivers into three categories: cautious, normal, and aggressive. This has better interpretability compared with implicit modeling methods. At the same time, fuzzy inference rules are used to handle the uncertainty and heterogeneity of driving behavior, avoiding dependence on large-scale parameter learning and significantly improving training efficiency.

[0017] 3. Compared with the prior art, the present invention embeds behavioral semantics into the fuel consumption estimation process through a style-aware fusion module, ensuring semantic consistency between the generated behavior and the fuel consumption output, and effectively improving the estimation accuracy. Attached Figure Description

[0018] Figure 1 This is a general framework diagram of the present invention; Figure 2 This is a flowchart illustrating the road section fuel consumption estimation method of the present invention. Detailed Implementation

[0019] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] It should be noted that this embodiment is based on the following assumptions: (1) Observation conditions are limited, and only macroscopic traffic data can be obtained through cross-section detectors; (2) Delays that may exist in information transmission, data processing and calculation and model reasoning are ignored; (3) The vehicle type is uniformly small passenger car, and the impact of vehicle type heterogeneity on fuel consumption is not considered for the time being.

[0021] like Figure 1 and Figure 2 As shown, a method for estimating the energy consumption of a road system based on data fusion and generative adversarial networks is presented. This method includes the following steps: Step S1: Collect macroscopic traffic data using a cross-section detector to obtain the average vehicle speed at time t. Vehicle spacing distribution ; Step S2: Construct a generative adversarial network (GAN). The generator G generates individual vehicle speed sequences based on the average vehicle speed, spacing distribution, and random noise vector. ; Step S3: Obtain individual vehicle acceleration sequences through differential operations. ; Step S4: Establish a differentiable fuzzy logic module, and identify driving styles based on speed and acceleration characteristics through fuzzy reasoning, classifying drivers into three categories: cautious, normal, and aggressive. Step S5: Construct a style perception fusion module that integrates kinematic features and driving style features, and estimate the individual vehicle fuel consumption e through a fuel consumption regression network. t,n ; Step S6: Aggregate individual vehicle fuel consumption to obtain road segment-level fuel consumption estimates. .

[0022] In step S1, the cross-section detector includes a loop detector, a microwave radar detector, or a video detector; macroscopic traffic data is collected at a fixed time interval ∆t. In this embodiment, ∆t is set to 0.1 seconds. The collected data includes: (1) Average speed of the cross section at time t The average speed of all vehicles passing through the detection section is calculated using the following formula: , Where, N t Let t be the number of vehicles passing through the detection section. Let n be the measured speed of vehicle n; (2) Distribution of vehicle spacing : Calculated by the time difference and speed of adjacent vehicles passing each other: , Where, ∆t n,n-1 Let n be the time difference between vehicle n and vehicle n-1 passing through the detection section; (3) Number of vehicles N in the scene: obtained by counting the number of unique vehicle identifiers at each time point; During data preprocessing, min-max normalization is used to normalize the data to the range [0,1]. , in, and These represent the minimum and maximum velocity values ​​among all training samples, respectively; in this embodiment, =0m / s, =40m / s.

[0023] In step S2, the specific structure and generation process of generator G are as follows: Step S201: The generator G input includes: the average vehicle speed corresponding to time step t. Vehicle spacing distribution and random noise vector z t (Dimensions are 100); Step S202: Generator G generates individual vehicle speed distributions through nonlinear mapping: , Where, θ G For the generator parameters, the generator G adopts a fully connected neural network structure, containing 4 hidden layers, each with 256 neurons, and the activation function is LeakyReLU; Step S203: To preserve the sparsity of the generated results, output the generator G. Element-wise multiplication with the input sparsity mask M: , Where ⊙ represents element-wise multiplication; the mask M takes the value of M[n]=1 when vehicle n exists, otherwise M[n]=0; Step S204: The loss function of generator G is: , Where, λ loss This represents the weight of the loss function; in this embodiment, λ is taken. loss =0.5.

[0024] In step S2, the goal of the discriminator D is to distinguish the generated vehicle speed distributions. and the actual vehicle speed distribution The objective function for optimizing the discriminator D in the generative adversarial network is: , in, This represents the actual vehicle speed distribution; To predict vehicle speed distribution; λ is the gradient penalty weight, with a value range of 0.1≤λ≤10, used to ensure that the discriminator D satisfies the 1-Lipschitz condition, and in this embodiment, λ=10; During training, interpolated samples are introduced to calculate the gradient penalty term: , The joint objective function is: ; Training parameter settings: The noise dimension z is set to 100; the learning rate of the generator and discriminator is initialized to 0.0002, and the Adam optimizer is used (β1=0.5, β2=0.999); the batch size is set to 32, and the training is conducted for 10 epochs; the Wasserstein loss is used in the diffusion model, and the gradient penalty weight λ is set to 10 to satisfy the 1-Lipschitz condition and ensure training stability.

[0025] In step S3, the specific method of difference operation is as follows: Step S301: Calculate the acceleration of vehicle n at time t using the first-order finite difference method: , Where ∆t = 0.1 seconds is the sampling time interval, v t,n and v t+1,n Let be the speeds of vehicle n at times t and t+1, respectively; Step S302: Obtain the effective vehicle acceleration through masking calculation: , Among them, A t M is the acceleration matrix for all vehicles. t For vehicles at time t, there exists a mask. In this embodiment, the acceleration range is limited to [-4,4] m / s². 2 Values ​​outside the specified range are truncated to ensure physical validity.

[0026] In step S4, the specific method for modeling driving style using a differentiable fuzzy logic module is as follows: Step S401: Fuzzification: Define trapezoidal membership functions for velocity and acceleration to map the input variables to the [0,1] interval. The trapezoidal membership function is defined as follows: , Where a1, a2, a3, and a4 are the left base, left peak, right peak, and right base parameters of the trapezoidal function, respectively; Step S402: The fuzzy set of velocity variables includes low speed. medium speed ,high speed Three categories of fuzzy sets of acceleration variables include small... ,middle ,big Three categories; in this embodiment, the fuzzy set of velocity variables and parameter settings are as follows: low speed : a1=0m / s, a2=0m / s, a3=10m / s, a4=15m / s, medium speed : a1=10m / s, a2=15m / s, a3=25m / s, a4=30m / s, high speed : a1=25m / s, a2=30m / s, a3=40m / s, a4=40m / s; The fuzzy set of acceleration variables and parameter settings are as follows: Small : a1=-4m / s², a2=-4m / s², a3=-1m / s², a4=0m / s², middle : a1=-1m / s², a2=0m / s², a3=0m / s², a4=1m / s², big : a1=0m / s², a2=1m / s², a3=4m / s², a4=4m / s²; Step S403: Fuzzy Inference: The T-norm operator is used to process fuzzy "AND" relationships for inferring aggressive driving styles. ; Step S404: Use the S-conorm operator to process the fuzzy "OR" relationship to infer cautious driving style: ; Step S405: Employ the Reichenbach implication function and introduce the sigmoid activation function to enhance differentiability: , Where σ(·) is the sigmoid function, and k is the scaling parameter. In this embodiment, k=10 is used to control the smoothness of the inference boundary. Step S406: Define the driving style reasoning rule set: Rule 1: If the speed is high And the acceleration is large The driving style is then considered aggressive. ; Rule 2: If the speed is low Or the acceleration is small If so, the driving style is cautious. ; Rule 3: In all other cases, the driving style is normal. ; Step S407: Defuzzification: Obtain the final driving style inference result through the product aggregator. Meanwhile, the membership values ​​of each category are retained as soft labels for subsequent feature fusion.

[0027] In step S5, the style-aware fusion module estimates individual fuel consumption using the following specific method: Step S501: Velocity-Acceleration Feature Projection: Constructing the velocity-acceleration feature projection module H v A two-layer fully connected network is used to map kinematic features to a 64-dimensional high-dimensional space:

[0028] = , Among them, [v t,n ,a t,n ] represents the concatenated vector of velocity and acceleration. , The weight matrix is ​​a learnable weight matrix; Step S502: Style Feature Projection: Construct the style feature projection module H s The soft labels for driving style categories are encoded as 64-dimensional feature vectors: ; Step S503: Feature Fusion: Construct a style modulation feature vector through element-wise multiplication to achieve the interaction between kinematic features and style features. This fusion method allows the impact of different driving styles on fuel consumption estimation to be explicitly modeled; Step S504: Fuel Consumption Regression: Input the modulation features into the fuel consumption regression network f SAF Output the estimated fuel consumption of individual vehicles: The fuel consumption regression network adopts a three-layer fully connected network structure with hidden layer dimensions of 64, 32, and 1 respectively. The output layer does not use an activation function to ensure the continuity of fuel consumption values. Supervision signal: The training loss function uses mean squared error (MSE). , in, This is the true value of fuel consumption calculated based on actual speed and acceleration.

[0029] In step S6, the specific method for road segment-level fuel consumption aggregation is as follows: Step S601: For time t, the time-level fuel consumption is obtained by averaging the fuel consumption estimates of all vehicles: , Where, N t This represents the number of valid vehicles in the road segment at time t. Step S602: For the time period [t1, t2], obtain the total fuel consumption of the road segment through time integration: , Where ∆t is the sampling time interval.

[0030] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for estimating energy consumption of road systems based on data fusion and generative adversarial networks, characterized in that, The method includes the following steps: Step S1: Collect macroscopic traffic data using a cross-section detector to obtain the average vehicle speed at time t. Vehicle spacing distribution ; Step S2: Construct a generative adversarial network (GAN). The generator G generates individual vehicle speed sequences based on the average vehicle speed, spacing distribution, and random noise vector. ; Step S3: Obtain individual vehicle acceleration sequences through differential operations. ; Step S4: Establish a differentiable fuzzy logic module, and identify driving styles based on speed and acceleration characteristics through fuzzy reasoning, classifying drivers into three categories: cautious, normal, and aggressive. Step S5: Construct a style perception fusion module that integrates kinematic features and driving style features, and estimate the individual vehicle fuel consumption e through a fuel consumption regression network. t,n ; Step S6: Aggregate individual vehicle fuel consumption to obtain road segment-level fuel consumption estimates. .

2. The road system energy consumption estimation method based on data fusion and generative adversarial networks according to claim 1, characterized in that, In step S1, the cross-section detector includes a loop detector, a microwave radar detector, or a video detector; macroscopic traffic data is collected at fixed time intervals ∆t, where the average vehicle speed at the cross-section is... To detect the average speed of all vehicles passing through the cross section, the vehicle spacing distribution... The data was obtained by calculating the time difference and speed of adjacent vehicles passing each other; during data preprocessing, min-max normalization was used to normalize the data to the range [0,1]. ， in, and These represent the minimum and maximum velocity values ​​among all training samples, respectively.

3. The road system energy consumption estimation method based on data fusion and generative adversarial networks according to claim 1, characterized in that, In step S2, the specific structure and generation process of generator G are as follows: Step S201: The generator G input includes: the average vehicle speed corresponding to time step t. Vehicle spacing distribution and random noise vector z t ; Step S202: Generator G generates individual vehicle speed distributions through nonlinear mapping: ， in, θ G These are the generator parameters; Step S203: To preserve the sparsity of the generated results, output the generator G. Element-wise multiplication with the input sparsity mask M: ， Where ⊙ represents element-wise multiplication; the mask M takes the value of M[n]=1 when vehicle n exists, otherwise M[n]=0; Step S204: The loss function of generator G is: ， Where, λ loss This represents the weights of the loss function.

4. The road system energy consumption estimation method based on data fusion and generative adversarial networks according to claim 3, characterized in that, In step S3, the specific method of difference operation is as follows: Step S301: Calculate the acceleration of vehicle n at time t using the first-order finite difference method: ， Where ∆t is the sampling time interval, v t,n and v t+1,n Let be the speeds of vehicle n at times t and t+1, respectively; Step S302: Obtain the effective vehicle acceleration through masking calculation: ， Among them, A t M is the acceleration matrix for all vehicles. t The vehicle at time t has a mask.

5. The road system energy consumption estimation method based on data fusion and generative adversarial networks according to claim 4, characterized in that, In step S4, the specific method for modeling driving style using a differentiable fuzzy logic module is as follows: Step S401: Define trapezoidal membership functions for velocity and acceleration, mapping the input variables to the interval [0,1]. The trapezoidal membership function is defined as follows: ， Where a1, a2, a3, and a4 are the left base, left peak, right peak, and right base parameters of the trapezoidal function, respectively; Step S402: The fuzzy set of velocity variables includes low speed. medium speed ,high speed Three categories of fuzzy sets of acceleration variables include small... ,middle ,big Three categories; Step S403: Use the T-norm operator to process the fuzzy AND relationship for inferring aggressive driving style: ； Step S404: Use the S-conorm operator to process the fuzzy "OR" relationship to infer cautious driving style: ； Step S405: Employ the Reichenbach implication function and introduce the sigmoid activation function to enhance differentiability: ， Where σ(·) is the sigmoid function and k is the scaling parameter, used to control the smoothness of the inference boundary; Step S406: Define the driving style reasoning rule set: Rule 1: If the speed is high and the acceleration is large, the driving style is aggressive. Rule 2: If the speed is low or the acceleration is small, the driving style is cautious. Rule 3: In all other cases, the driving style is normal; Step S407: Obtain the final driving style inference result through the product aggregator: 。 6. The road system energy consumption estimation method based on data fusion and generative adversarial networks according to claim 5, characterized in that, In step S5, the style-aware fusion module estimates individual fuel consumption using the following specific method: Step S501: Construct the velocity-acceleration feature projection module H v Mapping kinematic features to a high-dimensional space: ， Among them, [v t,n ,a t,n ] represents the concatenated vector of velocity and acceleration; Step S502: Construct the style feature projection module H s Encode driving style categories into feature vectors: ， Among them, st n For vehicle n, define the driving style category; Step S503: Construct the style modulation feature vector through element-wise multiplication: ； Step S504: Input the modulation features into the fuel consumption regression network f SAF Output the estimated fuel consumption of individual vehicles: 。 7. The road system energy consumption estimation method based on data fusion and generative adversarial networks according to claim 6, characterized in that, In step S6, the specific method for road segment-level fuel consumption aggregation is as follows: Step S601: For time t, the time-level fuel consumption is obtained by averaging the fuel consumption estimates of all vehicles: ， Where, N t This represents the number of valid vehicles in the road segment at time t. Step S602: For the time period [t1, t2], obtain the total fuel consumption of the road segment through time integration: ， Where ∆t is the sampling time interval.

8. The method for estimating the energy consumption of a road system based on data fusion and generative adversarial networks according to claim 3 or 7, characterized in that, In step S2, the objective function for optimizing the discriminator D of the generative adversarial network is: ， in, This represents the actual vehicle speed distribution; To predict vehicle speed distribution; λ is the gradient penalty weight, with a value range of 0.1≤λ≤10, used to ensure that the discriminator D satisfies the 1-Lipschitz condition; During training, interpolated samples are introduced to calculate the gradient penalty term: ， The joint objective function is: .

Images