A method for observing fuel consumption of logistics vehicles based on reversible neural networks

Abstract

This invention discloses a method for observing fuel consumption of logistics vehicles based on a reversible neural network. Specifically, the method involves: first, real-time acquisition and preprocessing of raw time-series data; extraction of vehicle state sequences at each moment; then, designing a reversible neural network consisting of concatenated reversible blocks with shared parameters for both forward and inverse mappings; training the reversible neural network by constructing a loss function that only includes forward fuel consumption prediction errors. Finally, the sequence is input into the trained reversible neural network, and the estimated fuel consumption value for the current moment is obtained through forward mapping, used for real-time monitoring and scheduling decisions. Simultaneously, a target fuel consumption value is determined, and the theoretical vehicle state sequence required to achieve the target fuel consumption is reconstructed through inverse mapping, and transformed into specific driving suggestions or control commands, realizing a closed-loop inverse feedback from fuel consumption to vehicle state. This invention can accurately capture the fuel consumption patterns under dynamic vehicle operating conditions, with high accuracy in forward observation.

Description

Technical Field

[0001] This invention relates to the field of fuel consumption observation for intelligent logistics transportation vehicles, specifically a method for observing fuel consumption of logistics transportation vehicles based on a reversible neural network. Background Technology

[0002] Smart logistics aims to achieve refined, dynamic, and visualized management of all aspects of logistics through intelligent hardware and software, the Internet of Things, and big data technologies. This enhances the intelligent analysis and decision-making capabilities and automated operation execution of the logistics system, leading to a modern logistics model that improves operational efficiency. Among these technologies, vehicle fuel consumption information holds significant value and importance for the monitoring and control of connected vehicles, decision-making and scheduling of transportation, and the assessment and optimization of logistics costs.

[0003] Generally speaking, the development of vehicle fuel consumption models falls into three main categories: physics-based models, machine learning-based models, and statistical models. Physics-based models use detailed mathematical equations to describe the dynamics of vehicle components at each time step, thus formulaically modeling fuel consumption. Machine learning models are data-driven, mapping an input space consisting of vehicle speed, acceleration, etc., to an output space containing real-time fuel consumption. Statistical models are also data-driven, establishing a mapping between the probability distribution of selected predictors and the target outcome based on probabilistic statistical laws.

[0004] Because the fuel consumption mechanism is complex and changes drastically under vehicle dynamic conditions, models based on physical principles usually require a high level of basic knowledge and long-term experience in parameter tuning to achieve the expected observation level. Therefore, data-driven models are usually more favored by practitioners. Currently, a variety of fuel consumption prediction models have been proposed, which collect operating data over a period of time and construct representative predictive factors of different dimensions to predict the corresponding fuel consumption.

[0005] Vehicle fuel consumption, as an economic indicator, is incorporated into decision-making, planning, and control processes, serving as a link between economic efficiency and optimization / control signal adjustment. This not only creates a demand for accurate forward mapping from vehicle state to real-time fuel consumption but also presents challenges to the closed-loop feedback from estimated fuel consumption to vehicle state. Establishing a lossless inverse information mapping while ensuring accurate forward estimation, and improving the interpretability of state estimation, has become a new challenge under the wave of information technology development. Invertible Neural Networks (INNs) use a reversible structure to construct neural networks, using the same encoding and decoding parameters to achieve lightweight bidirectional mapping and lossless information transfer between input and output. Summary of the Invention

[0006] This invention proposes a fuel consumption observation method for logistics vehicles based on a reversible neural network. It addresses the fuel consumption observation problem of transportation vehicles by constructing a verification model based on the forward observation of fuel consumption based on vehicle status and the inverse correspondence between the observation results and vehicle status. This not only provides an effective solution for the intuitive estimation and prediction of fuel consumption factors, but also provides feedback for energy-driven vehicle optimization control.

[0007] The method for observing fuel consumption of logistics vehicles based on reversible neural networks includes the following steps:

[0008] Step 1: Collect raw time-series data of logistics vehicles in real time during their operation, and preprocess the data to remove outliers and missing values.

[0009] This includes vehicle speed, acceleration, road gradient, etc.

[0010] Step 2: Extract from preprocessed raw data The instantaneous state variables and window statistical characteristics at each moment constitute the real-time vehicle state sequence. ;

[0011] Instantaneous state variables include: instantaneous vehicle speed Instantaneous acceleration Instantaneous road gradient ;

[0012] Window statistical features include the number of parking times Parking time Average driving speed ; and physical characteristics associated with fuel consumption: characteristic acceleration Aerodynamic drag velocity square Change in kinetic energy and the change in potential energy ;

[0013] Vehicle state sequence For a matrix, containing Time and before The feature vector at each time step; that is:

[0014] ;

[0015] The length of the time window; for The feature vector at time step 1 is represented as:

[0016]

[0017] Among them, characteristic acceleration: ;

[0018] Aerodynamic drag on velocity square: ;

[0019] Change in kinetic energy: ;

[0020] Change in potential energy: ;

[0021] The window length is set to a fixed travel distance. For fixed driving distance window The instantaneous speed of the vehicle at the starting point; For fixed driving distance window The instantaneous speed of the vehicle at the finish line; The total mass of the vehicle; It is the acceleration due to gravity;

[0022] Step 3, Design by An invertible neural network consisting of concatenated invertible blocks, containing forward and inverse mappings with shared parameters. .

[0023] First, the vehicle state feature vector at each time step within the window is divided into two sub-vectors of equal dimension according to the channel. After concatenation, each sub-vector is independently input into the invertible neural network. ;

[0024] No. ( Vehicle state feature vector at each time step Divided into , will include all moments within the time window and Concatenate the columns separately to form a matrix. And use it as the input to the 0th layer of the invertible neural network, that is, let , .

[0025] Then, proceed Forward computation of a reversible block;

[0026] No. The forward computation of a reversible block is defined as follows:

[0027]

[0028] in, It is a nonlinear subnetwork. and It is the first The output matrix of each invertible block, with each column corresponding to a moment within a time window.

[0029] Finally, through After reversible blocks, the output features are concatenated. and The two are spliced ​​together through a linear fully connected layer. Get the moment Fuel consumption estimate:

[0030]

[0031] Based on the aforementioned reversible blocks, construct the inverse mapping corresponding to the forward mapping. This structure ensures that fuel consumption values ​​can be obtained. The corresponding vehicle state sequence can be reconstructed in reverse without loss.

[0032] Specifically: First, set the target fuel consumption The initial features are obtained by expanding the model into a tensor of the same dimension as the feature matrix through a learnable projection layer. .

[0033] Then, starting from layer L, calculate backwards layer by layer:

[0034] No. The reverse calculation of a reversible block is as follows:

[0035]

[0036] And so on, until you get and .

[0037] Finally, and By restoring according to the time window, the reconstructed vehicle state sequence is obtained:

[0038]

[0039] Among them, the reconstructed first Feature vector at time n Including the reconstructed instantaneous vehicle speed Instantaneous acceleration Instantaneous slope Information on all features, etc.

[0040] Step 4: Construct a loss function that includes only the positive fuel consumption prediction error, and use historical datasets to analyze the reversible neural network. Conduct training.

[0041] The loss function is defined as:

[0042]

[0043] in, Indicates the first The actual fuel consumption values ​​of each training sample; The total number of training samples; Indicates the first Fuel consumption estimates for each training sample; Iterative updates of the invertible neural network using the Adam optimizer parameters Continue until the loss converges, and a well-trained network model is obtained; Indicates the first A vehicle state sequence matrix of training samples, that is, containing a sequence of vehicle states at a certain historical moment. as the end point The feature vectors at consecutive time points.

[0044] Step 5: Serialize the real-time vehicle status sequence Input the trained invertible neural network The fuel consumption estimate for the current moment is obtained through forward mapping and transmitted to the vehicle monitoring platform for real-time monitoring and scheduling decisions.

[0045] The estimated fuel consumption is calculated as follows:

[0046]

[0047] Step 6: Determine the target fuel consumption value based on the energy consumption optimization objectives of logistics and transportation. And input the trained reversible neural network. By reconstructing the theoretical vehicle state sequence required to achieve the target fuel consumption through inverse mapping, and transforming it into specific driving suggestions or control commands, a closed-loop inverse feedback from fuel consumption to vehicle state is realized.

[0048] The theoretical vehicle state sequence is:

[0049]

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] (1) High-precision forward observation: By combining the key features extracted through physical mechanism analysis with a reversible neural network, the fuel consumption pattern under dynamic vehicle conditions can be accurately captured, and the forward observation accuracy is high.

[0052] (2) Lossless inverse feedback: By utilizing the reversible structure of the network, the inverse mapping from fuel consumption to vehicle state can be achieved without additional training, providing a theoretical state reference for energy consumption-oriented optimization control and enhancing the interpretability of the model.

[0053] (3) Lightweight deployment: The reversible neural network adopts bidirectional mapping with shared parameters, with fewer parameters and higher computational efficiency, making it suitable for real-time operation of vehicle-mounted embedded systems.

[0054] (4) Physical consistency: Input features are derived based on dynamics to ensure that the model output conforms to physical laws and avoid the risk of overfitting of pure data-driven models. Attached Figure Description

[0055] Figure 1 This is a flowchart of a method for observing fuel consumption of logistics vehicles based on a reversible neural network, according to the present invention.

[0056] Figure 2 This is a graph showing the positive prediction result of fuel consumption using the reversible neural network employed in this invention.

[0057] Figure 3 This is a graph showing the change in loss value with the number of iterations during the training process of the reversible neural network of the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0059] This invention proposes a fuel consumption observation method for logistics vehicles based on a reversible neural network. It achieves high-precision forward fuel consumption observation and lossless reverse state feedback. The model has few parameters and high computational efficiency, meeting the lightweight deployment requirements of in-vehicle embedded systems. Furthermore, the input features are derived based on dynamics, ensuring that the output conforms to physical laws. This method is suitable for real-time fuel consumption observation and energy consumption optimization control scenarios for intelligent logistics vehicles. Its principles include the following:

[0060] The first step involves establishing a driving resistance model based on vehicle longitudinal dynamics, deriving the correlation between engine power and fuel consumption rate, extracting instantaneous state variables and window statistical features reflecting the vehicle's operating state, and determining the input feature set of the reversible neural network. The second step involves dividing the vehicle state feature vector into two sub-vectors as independent input branches, designing a reversible neural network composed of multiple cascaded reversible blocks, where each reversible block achieves forward mapping and inverse reconstruction through a dedicated reversible transformation, preserving the integrity of the input information. The output features of the reversible blocks are then concatenated and passed through a linear fully connected layer to obtain the fuel consumption estimate output. The third step involves constructing a loss function that only includes the positive fuel consumption prediction error. The first step involves using the Adam optimizer to iteratively train the reversible neural network using actual vehicle operating data until the loss value converges. No explicit reconstruction loss term is needed; the training process only optimizes the accuracy of the forward prediction. The second step involves inputting the real-time collected vehicle state sequence into the trained reversible neural network to complete the forward observation of the current fuel consumption estimate, providing data support for vehicle monitoring and scheduling decisions. The third step involves, given a target fuel consumption value, reconstructing the theoretical vehicle state sequence required to achieve that fuel consumption through the inverse mapping function of the reversible neural network. The reconstruction result is then converted into driving suggestions or control commands, realizing the inverse feedback from fuel consumption to vehicle state.

[0061] like Figure 1 As shown, the specific steps include:

[0062] Step 1: Collect raw time-series data of logistics vehicles in real time during their operation, and preprocess the data to remove outliers and missing values.

[0063] This includes vehicle speed, acceleration, road gradient, etc.

[0064] Step 2: Extract from preprocessed raw data The instantaneous state variables and window statistical characteristics at each moment constitute the real-time vehicle state sequence. ;

[0065] Instantaneous state variables include: instantaneous vehicle speed Instantaneous acceleration Instantaneous road gradient ;

[0066] Window statistical features include the number of parking times Parking time Average driving speed ; and physical characteristics associated with fuel consumption: characteristic acceleration Aerodynamic drag velocity square Change in kinetic energy and the change in potential energy ;

[0067] Vehicle state sequence For a matrix, containing Time and before The feature vector at each time step; that is:

[0068] ;

[0069] The length of the time window; for The feature vector at time step 1 is represented as:

[0070]

[0071] Among them, based on the derivation of vehicle longitudinal dynamics and fuel consumption physical mechanisms, we obtain:

[0072] Characteristic acceleration: ;

[0073] Aerodynamic drag on velocity square: ;

[0074] Change in kinetic energy: ;

[0075] Change in potential energy: ;

[0076] The window length is set to a fixed travel distance. For fixed driving distance window The instantaneous speed of the vehicle at the starting point; For fixed driving distance window The instantaneous speed of the vehicle at the finish line; The total mass of the vehicle (including the vehicle's own weight and load). Let be the acceleration due to gravity, taken as 9.8. ; for Road gradient angle at any given time; It represents the sine value of the slope angle, used to calculate the component of gravity along the slope direction.

[0077] Step 3, Design by An invertible neural network consisting of concatenated invertible blocks, containing forward and inverse mappings with shared parameters. .

[0078] First, the vehicle state feature vector at each time step within the window is divided into two sub-vectors of equal dimension according to the channel. After concatenation, each sub-vector is independently input into the invertible neural network. ;

[0079] No. ( Vehicle state feature vector at each time step (dimension is) Divide into two sub-vectors of equal dimension according to the channel. Each subvector has a dimension of ,like If the number is odd, the dimension of the subvector can be made to be zero by padding with zeros. This ensures the feasibility of network computation. The feature vector at each time step k... Both contain two types of information: instantaneous values ​​and historical cumulative values. They are technical features that use historical statistical characteristics to capture the cumulative effect of driving behavior.

[0080] All moments within the time window and Concatenate the columns separately to form a matrix. And use it as the input to the 0th layer of the invertible neural network, that is, let , .

[0081] in , .

[0082] matrix The matrix contains The column vector at each time step.

[0083] Then, proceed Forward computation of a reversible block;

[0084] No. The forward computation of a reversible block is defined as follows:

[0085]

[0086] in, It is a nonlinear subnetwork. and It is the first The output matrix of each invertible block, with each column corresponding to a moment within a time window.

[0087] Finally, through After reversible blocks, the output features are concatenated. and The two are spliced ​​together through a linear fully connected layer. Get the moment scalar fuel consumption estimate :

[0088]

[0089] Based on the inverse reversible blocks, the inverse mapping corresponding to the forward mapping is constructed layer by layer. This structure ensures that fuel consumption values ​​can be obtained. Lossless reverse reconstruction of the corresponding vehicle state sequence .

[0090] Specifically: First, set the target fuel consumption The initial features are obtained by expanding the model into a tensor of the same dimension as the feature matrix through a learnable projection layer. .

[0091] Then, starting from layer L, calculate backwards layer by layer:

[0092] No. The reverse calculation of a reversible block is as follows: ;

[0093] And so on, until you get and .

[0094] Finally, and By restoring according to the time window, the reconstructed vehicle state sequence is obtained:

[0095]

[0096] Among them, the reconstructed first Feature vector at time n Including the reconstructed instantaneous vehicle speed Instantaneous acceleration Instantaneous slope Information on all features, etc.

[0097] Step 4: Construct a loss function that includes only the positive fuel consumption prediction error, and use historical datasets to analyze the reversible neural network. Conduct training.

[0098] The loss function is defined as:

[0099]

[0100] in, Indicates the first The actual fuel consumption values ​​of each training sample; The total number of training samples; Indicates the first Fuel consumption estimates for each training sample; Iterative updates of the invertible neural network using the Adam optimizer parameters The process continues until the loss converges, resulting in a well-trained network model. Because the network structure is strictly invertible, no reconstruction loss term needs to be added during training. Indicates the first The vehicle state sequence matrix of each training sample has the same structure as described in step two. Same, that is, including a certain historical moment as the end point The feature vectors at consecutive time points.

[0101] Step 5: Serialize the real-time vehicle status sequence Input the trained invertible neural network The fuel consumption estimate for the current moment is obtained through forward mapping and transmitted to the vehicle monitoring platform for real-time monitoring and scheduling decisions.

[0102] The estimated fuel consumption is calculated as follows:

[0103]

[0104] Step 6: Determine the target fuel consumption value based on the energy consumption optimization objectives of logistics and transportation. And input the trained reversible neural network. By reconstructing the theoretical vehicle state sequence required to achieve the target fuel consumption through inverse mapping, and transforming it into specific driving suggestions or control commands, a closed-loop inverse feedback from fuel consumption to vehicle state is realized.

[0105] Example:

[0106] Step 1: Data Acquisition and Physical Mechanism Analysis;

[0107] 1.1 Vehicle longitudinal dynamics modeling;

[0108] According to vehicle dynamics theory, a vehicle must overcome rolling resistance, air resistance, acceleration resistance, and gradient resistance during operation. The total driving force acting on the drive wheels is... It can be represented as: ;

[0109] The resistance components are as follows:

[0110] Rolling resistance: On a gentle slope, it is approximately equal to ;

[0111] Air resistance: ;

[0112] Acceleration resistance: ,in It is instantaneous acceleration;

[0113] Slope resistance: ;

[0114] in, Rolling resistance coefficient; Air density ( ); Air drag coefficient; Vehicle frontal area ( ); Instantaneous vehicle speed ( ); Instantaneous acceleration ( ); Road slope ( );

[0115] Therefore, the expression for the total driving force is:

[0116]

[0117] Mechanical power on drive wheels The product of driving force and vehicle speed:

[0118]

[0119] Considering transmission system efficiency The power required by the engine for:

[0120]

[0121] 1.2 Modeling of Engine Efficiency and Fuel Consumption Rate

[0122] Engine instantaneous fuel consumption With engine power and effective fuel consumption rate The relationship is:

[0123]

[0124] If fuel consumption is measured by volume Therefore, fuel density needs to be introduced. ( ):

[0125]

[0126] Effective fuel consumption rate Engine speed and torque The function. Based on the universal characteristics of the engine, It can be expressed in the following polynomial form:

[0127]

[0128] In the formula These are the fitting coefficients. The fitting order (usually taken as...) ).

[0129] Engine speed With vehicle speed Related by transmission ratio:

[0130]

[0131] in: : Gearbox ratio; : Main reduction ratio; Wheel rolling radius The constant 0.377 is the unit conversion factor, derived from... Export;

[0132] Engine torque With driving force The relationship is:

[0133]

[0134] Substituting equation (2) into equation (9):

[0135]

[0136] Substituting equations (4) and (7) into equation (6), we obtain the complete expression for instantaneous fuel consumption:

[0137]

[0138] in and The equations (8) and (10) are given respectively, and are both... The function.

[0139] 1.3 Feature Aggregation Based on Distance Domain

[0140] For a fixed driving distance window Average fuel consumption within the window Defined as:

[0141]

[0142] According to vehicle dynamics, the average fuel consumption within the window can be expressed as a nonlinear function of the following statistical characteristics:

[0143] Number of parking times Parking time Average driving speed Characteristic acceleration Aerodynamic drag on the square of velocity ; change in kinetic energy ; change in potential energy ;

[0144] 1.4 Selection Logic of Input Features in Neural Networks

[0145] Based on the above analysis of physical mechanisms, this invention determines that the input features of a neural network include two types of variables:

[0146] (1) Instantaneous state variables (reflecting the dynamic state at the current moment):

[0147] Instantaneous speed Instantaneous acceleration Instantaneous road gradient ;

[0148] (2) Window statistical characteristics (reflecting the cumulative effect over a period of time in the past):

[0149] Number of parking times Parking time Average driving speed Characteristic acceleration Aerodynamic drag on the square of velocity ; change in kinetic energy ; change in potential energy ;

[0150] The above features can all be directly calculated from the time series of vehicle speed, acceleration, and gradient, without the need for additional sensors, meeting the requirements for low-cost deployment. The input-output relationship can be uniformly expressed as:

[0151]

[0152] This mapping implicitly contains all the physical relationships described by equations (1) to (11).

[0153] Step 2: Design of Reversible Neural Network Structure

[0154] 2.1 State Vector Segmentation

[0155] For each moment eigenvectors Divide it into two sub-vectors according to the channel. All within the time window T and Concatenate them into a matrix , which serve as two independent input branches of the network.

[0156] 2.2 Reversible Block Design

[0157] The network consists of L reversible blocks connected in series. The i-th reversible block... Forward computation is defined as:

[0158]

[0159] in It is a nonlinear subnetwork consisting of fully connected layers and ReLU activation functions, with both input and output dimensions being [missing information]. , .

[0160] This structure guarantees strict reversibility, that is, given... The output of the previous layer can be uniquely solved:

[0161]

[0162] 2.3 Output Layer

[0163] go through After obtaining a number of reversible blocks, the final feature is obtained. and The two are then concatenated and connected via a linear fully connected layer. Get the moment Fuel consumption estimate:

[0164]

[0165] 2.4 Reverse Mapping

[0166] Given target fuel consumption value First, the features are expanded into vectors of the same dimension as the features through a learnable projection layer to obtain the initial features. Then proceed in sequence. The reverse operation of each reversible block (according to equation (15)) yields Finally, it is restored to a state sequence:

[0167]

[0168] Step 3: Loss Function and Training

[0169] First, we construct a loss function that includes only the positive fuel consumption prediction error. The expression for the loss function is as follows:

[0170]

[0171] Next, the actual vehicle operation data is divided into training set, validation set and test set according to a preset ratio. The training set is used for iterative updates of network parameters, the validation set is used to monitor overfitting during the training process, and the test set is used for final model performance evaluation. All data are normalized to eliminate the impact of differences in the dimensions of different features on model training.

[0172] Then, the reversible neural network is trained using the Adam optimizer, taking the vehicle state feature set in the training set as input and the actual fuel consumption value as the label, and iteratively updating the network parameters. After each iteration, the loss values ​​of the training set and the validation set are calculated. If the loss value of the validation set does not decrease for several consecutive rounds, training is stopped until the loss value of the training set converges to a stable range, thus obtaining the trained reversible neural network model. The model parameters are saved for subsequent fuel consumption observation.

[0173] Step 4: Forward Fuel Consumption Observation

[0174] First, a real-time data acquisition system for logistics and transportation vehicles is built to collect raw time-series data such as vehicle speed, acceleration, and road slope during the vehicle's journey. The acquisition frequency is matched with the time window during model training.

[0175] Next, the raw data collected in real time is preprocessed to remove outliers and missing values. Following the method in step (1), instantaneous state variables and window statistical features are extracted from the raw data to form a real-time vehicle state sequence. The features are processed using the same normalization method as during the training process.

[0176] Then, the preprocessed real-time vehicle state sequence St is input into the trained reversible neural network, and the fuel consumption estimate for the current moment is obtained through the network's forward mapping calculation.

[0177]

[0178] The fuel consumption estimate is transmitted to the vehicle monitoring platform in real time, providing accurate energy consumption data support for logistics companies' decision-making in real-time vehicle monitoring, transportation route scheduling, and driving behavior assessment.

[0179] Step 5: Reverse State Feedback

[0180] First, based on the energy consumption optimization goals of logistics and transportation, the target fuel consumption value is determined. The target fuel consumption value can be set by comprehensively considering factors such as vehicle model, transportation route, and cargo weight. Next, the target fuel consumption value... The trained reversible neural network is input, and through the inverse mapping structure designed in step (2), the projection layer expansion, reversible block inverse operation, and feature splitting recovery are completed in sequence to reconstruct the theoretical vehicle state sequence required to achieve the target fuel consumption:

[0181]

[0182] This sequence contains information such as theoretical instantaneous vehicle speed, instantaneous acceleration, road gradient, and window statistical characteristics.

[0183] Then, the reconstructed theoretical vehicle state sequence is analyzed, and combined with actual road conditions and vehicle performance limitations, the theoretical state information is transformed into specific and executable driving suggestions or control commands.

[0184] Driving suggestions include suggested speed ranges, acceleration limits, and gear shifting timings, and control commands can be directly interfaced with the execution module of the autonomous driving system.

[0185] Driving suggestions or control commands are pushed to the driver or the autonomous driving system to achieve energy-driven vehicle optimization control and complete the closed-loop reverse feedback from fuel consumption to vehicle status.

[0186] like Figure 2 The figure shows the results of the reversible neural network proposed in this invention on the task of forward fuel consumption prediction. The horizontal axis of the figure represents the time window number, with each window corresponding to a prediction time, and the vertical axis represents the estimated fuel consumption value (unit: L / 100km). The solid blue line represents the actual fuel consumption of the vehicle, the solid red line represents the model's prediction value on the training set, and the dashed yellow line represents the model's prediction value on the test set.

[0187] from Figure 2As can be seen, the predicted curve (red solid line) of this method highly overlaps with the actual fuel consumption curve (blue solid line) on the training set; on the test set (yellow dashed line), the predicted curve still tracks the changing trend of actual fuel consumption well, without significant deviation or divergence. Quantitative calculations show that the mean absolute error (MAE) of the method of this invention on the test set is 1.88 L / 100km, and the coefficient of determination is... The result reached 0.91. These results verify that the method of this invention has high-precision forward fuel consumption observation capability and good generalization performance.

[0188] like Figure 3 The figure shows the curve of the loss value changing with the number of iterations during the training process of the reversible neural network of this invention. The horizontal axis represents the number of iterations, and the vertical axis represents the loss value, using mean squared error (MSE) as the loss function. As the number of iterations increases, the loss value on the vertical axis shows the following trend: In the early stages of training (the first 50 iterations), the loss value decreases rapidly, then the rate of decrease gradually slows down. Around the 300th iteration, the loss value converges to below 0.002 and no longer fluctuates significantly. No overfitting or oscillations occurred throughout the training process, verifying the effectiveness and stability of the reversible structure and simplified loss function used in this invention.

[0189] The above examples are for illustrative purposes only and are not intended to limit the invention. Those skilled in the art should understand that any modifications, variations, or equivalent substitutions to the invention without departing from its spirit and scope should be included in the claims of the invention to make the objectives, technical solutions, and advantages of the invention clearer. The embodiments of the invention will be further described in detail below with reference to the accompanying drawings.

Claims

1. A method for observing fuel consumption of logistics vehicles based on a reversible neural network, characterized in that, Includes the following steps: Step 1: Collect raw time-series data of logistics vehicles in real time during their operation and perform preprocessing. Step 2: Extract from preprocessed raw data The instantaneous state variables and window statistical characteristics at each moment constitute the real-time vehicle state sequence. ; Vehicle state sequence For a matrix, containing Time and before The feature vector at each time step; that is: ； The length of the time window; for The feature vector at time step; Step 3, Design by An invertible neural network consisting of concatenated invertible blocks, containing forward and inverse mappings with shared parameters. ; First, the vehicle state feature vector at each time step within the window is divided into two sub-vectors of equal dimension according to the channel. After concatenation, each sub-vector is independently input into the invertible neural network. ; The first in the window Vehicle state feature vector at each time step Divided into , will include all moments within the time window and Concatenate the columns separately to form a matrix. And use it as the input to the 0th layer of the invertible neural network, that is, let , ; Then, proceed Forward computation of a reversible block; No. The forward computation of a reversible block is defined as follows: in, It is a nonlinear subnetwork. and It is the first The output matrix of each invertible block, with each column corresponding to a moment within a time window; Finally, through After reversible blocks, the output features are concatenated. and The two are spliced ​​together through a linear fully connected layer. Get the moment Fuel consumption estimate: Based on reversible blocks, construct the inverse mapping corresponding to the forward mapping. This structure ensures that fuel consumption values ​​can be achieved. The corresponding vehicle state sequence can be reconstructed in reverse without loss. Step 4: Construct a loss function that includes only the positive fuel consumption prediction error, and use historical datasets to analyze the reversible neural network. Conduct training; Step 5: Serialize the real-time vehicle status sequence Input the trained invertible neural network The fuel consumption estimate for the current moment is obtained through forward mapping and transmitted to the vehicle monitoring platform for real-time monitoring and scheduling decisions. The estimated fuel consumption is calculated as follows: Step 6: Determine the target fuel consumption value based on the energy consumption optimization objectives of logistics and transportation. And input the trained reversible neural network. By reconstructing the theoretical vehicle state sequence required to achieve the target fuel consumption through inverse mapping, and transforming it into specific driving suggestions or control commands, a closed-loop inverse feedback from fuel consumption to vehicle state is realized. The theoretical vehicle state sequence is: .

2. The method as described in claim 1, characterized in that, The original time series data in step one includes vehicle speed, acceleration, and road slope; preprocessing includes removing outliers and missing values.

3. The method as described in claim 1, characterized in that, In step two, the instantaneous state variables include: instantaneous vehicle speed. Instantaneous acceleration Instantaneous road gradient ; Window statistical features include the number of parking times Parking time Average driving speed ; and physical characteristics associated with fuel consumption: characteristic acceleration Aerodynamic drag velocity square Change in kinetic energy and the change in potential energy ; in, Characteristic acceleration at time: ; aerodynamic drag at time t: ; Change in kinetic energy at time t: ; Change in potential energy at time: ; The window length is set to a fixed travel distance. For fixed driving distance window The instantaneous speed of the vehicle at the starting point; For fixed driving distance window The instantaneous speed of the vehicle at the finish line; The total mass of the vehicle; This is the acceleration due to gravity.

4. The method as described in claim 1 or 3, characterized in that, In step two Represented as: for Instantaneous acceleration at a given moment; for Instantaneous speed at any given moment; for Instantaneous road gradient at any given moment; for Number of stops at any given time; for Parking time at any given moment; for Average speed at any given time; for Characteristic acceleration at any given moment; for The aerodynamic drag at time t is the square of the velocity. for The change in kinetic energy at any given moment; for The change in potential energy at any given moment.

5. The method as described in claim 4, characterized in that, In step three, the reverse mapping specifically involves: First, set the target fuel consumption The initial features are obtained by expanding the model into a tensor of the same dimension as the feature matrix through a learnable projection layer. ; Then, starting from layer L, calculate backwards layer by layer: No. The reverse calculation of a reversible block is as follows: And so on, until you get and ; Finally, and By restoring according to the time window, the reconstructed vehicle state sequence is obtained: Among them, the reconstructed first Feature vector at time n Including the reconstructed instantaneous vehicle speed Instantaneous acceleration Instantaneous slope Information on all features, etc.

6. The method as described in claim 1, characterized in that, In step four, the loss function is defined as: in, Indicates the first The actual fuel consumption values ​​of each training sample; The total number of training samples; Indicates the first Fuel consumption estimates for each training sample; Iterative updates of the invertible neural network using the Adam optimizer parameters Continue until the loss converges, and a well-trained network model is obtained; Indicates the first A vehicle state sequence matrix of training samples, that is, containing a sequence of vehicle states at a certain historical moment. as the end point The feature vectors at consecutive time points.

Images