A slope-adaptive predictive energy-saving control method for heavy trucks

Abstract

The present invention discloses a gradient-adaptive predictive energy-saving control method for heavy-duty trucks, comprising the following steps: Step 1: establishing a longitudinal dynamics model of the heavy-duty truck; Step 2: establishing an engine fuel consumption characteristic model; Step 3: establishing a NMPC controller for the heavy-duty truck power system and constructing a multi-objective cost function, wherein the multi-objective cost function includes a speed traceability weight w1, a fuel economy weight w2, and a torque variation weight w3; Step 4: constructing a neural network that captures the mapping relationship between the gradient sequence and the fuel economy weight w2, wherein the neural network outputs the optimal fuel economy weight w2 to the NMPC controller; Step 5: constructing a feature data set for neural network training; Step 6: establishing an automated simulation scheduling system; Step 7: the automated simulation scheduling system generates labels; and Step 8: training the neural network using a data set consisting of the labels and the long-scale gradient sequence. The present invention improves fuel efficiency under variable gradients.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent control of power systems, and in particular to a gradient adaptive predictive energy-saving control method for heavy trucks. Background Art

[0002] In road transportation, heavy trucks account for a significant portion of fuel consumption, and their fuel efficiency is largely affected by road conditions, especially changes in road slope. To address this issue, it is crucial to obtain the road slope in advance and use this information to adjust the controller so that the vehicle's power system can respond quickly and effectively to slope changes during cruising. With the development of digital maps, the integration of slope information has increased the possibility of achieving energy-saving driving based on slope data. However, truck cruise control faces new challenges, such as balancing multiple objectives (vehicle dynamics, fuel economy, and ride comfort) and maintaining control performance under different slope conditions.

[0003] To address these issues, experts and scholars have conducted in-depth research, and various control schemes have been proposed, which can be broadly categorized as rule-based, optimization-based, and learning-based. Rule-based methods primarily abstract mathematically described laws from driving experience and use these laws to design control strategies to achieve energy conservation. However, these rule-based strategies rely heavily on human experience, which limits the full potential of vehicle energy conservation. Optimization-based control methods originate from the field of optimal control, with the mainstream model predictive control (MPC) used to find the optimal solution for energy conservation. However, the high computational complexity of MPC algorithms limits the length of their rolling prediction horizon. This makes it difficult to obtain global information during long-term optimization, thus limiting overall performance. In recent years, learning-based control methods have gained increasing popularity in the field of predictive energy-saving control. Artificial intelligence technologies such as deep learning and reinforcement learning have rapidly developed, demonstrating powerful data feature mining capabilities and providing new avenues for discovering and utilizing slope information for energy conservation. However, these learning-based methods rely on large amounts of high-quality training data, and data acquisition is time-consuming and costly. Furthermore, a lack of interpretability is another factor hindering their practical application.

[0004] In summary, MPC methods offer good robustness and interpretability but struggle to capture long-scale features. Deep learning methods excel at extracting feature information (including long-scale information) from a variety of data types, but suffer from poor interpretability and data acquisition challenges. Therefore, this study aims to combine these two approaches, leveraging their respective strengths and achieving complementary results. Furthermore, to address the challenge of training data acquisition, an automated simulation scheduling system based on a large language model (LLM) agent is introduced to replace manual simulation and data collection, demonstrating the potential of LLMs for controller optimization. Summary of the Invention

[0005] The present invention addresses the technical deficiencies in existing predictive energy-saving control technologies by providing a method for adaptive slope predictive energy-saving control of heavy trucks. The controller, based on a neural network, achieves long-scale slope adaptation and dynamically adjusts controller parameters to ultimately achieve fuel savings. Furthermore, an automated simulation scheduling system based on a large language model (LLM) agent replaces manual simulation and data collection, addressing the difficulty and high cost of data collection for neural network training.

[0006] The technical solution adopted to achieve the purpose of the present invention is:

[0007] A method for adaptive predictive energy-saving control of a heavy truck slope with LLM-assisted data collection includes the following steps:

[0008] Step 1: Establish a longitudinal dynamic model of a heavy truck on a slope road;

[0009] Step 2: Establish an engine fuel consumption characteristic model to obtain brake specific fuel consumption (BSFC);

[0010] Step 3: Establish an NMPC controller for the heavy truck powertrain to precisely control the engine torque and provide feedback to a control-oriented model based on the longitudinal dynamics model from Step 1 and the engine fuel consumption characteristic model from Step 2. A multi-objective cost function is constructed to balance speed tracking and fuel consumption. The multi-objective cost function includes a speed traceability weight w1, a fuel economy weight w2, and a torque variation weight w3.

[0011] Step 4: construct a neural network that captures the mapping relationship between the slope sequence and the fuel economy weight w2 in step 2, and the neural network outputs the optimal fuel economy weight w2 to the NMPC controller in step 3;

[0012] Step 5: construct a feature data set for neural network training, i.e., a long-scale slope sequence;

[0013] Step 6: Establish an automated simulation scheduling system consisting of an execution agent, an analysis agent, and a simulation engine. The execution agent generates execution commands for the w2 scanning direction based on the long-scale slope sequence in step 5 and the discrete scanning range of the manually set fuel economy weight w2. The simulation engine converts the execution commands into low-level executable commands for the simulation software and the specific w2 scanning simulation process, generating multimodal simulation results and sending them to the analysis agent. The analysis agent analyzes the multimodal simulation results and generates formatted feedback to help adjust the execution agent's actions in the next round.

[0014] Step 7: The automated simulation scheduling system generates a label, where the label is the optimal fuel economy weight w2 corresponding to each long-scale slope sequence;

[0015] Step 8: Use the training data set and the test data set consisting of the labels obtained in step 7 and the long-scale slope sequence obtained in step 5 to train and test the neural network.

[0016] In the above technical solution, the longitudinal dynamic model in step 1 is:

[0017]

[0018] Where m is the vehicle mass, θ is the road slope, ρ is the air density, C d is the air resistance coefficient, A a is the frontal area of ​​the vehicle, v is the vehicle speed, f is the rolling resistance coefficient, F drive is the driving force that drives the vehicle to move, F aero is the air resistance, F roll is the rolling resistance, F slope is the slope resistance.

[0019] In the above technical solution, in step 2, the engine fuel consumption characteristic model is:

[0020]

[0021] Among them, α is the fuel consumption characteristic fitting coefficient, T e is the engine torque, n e is the engine speed.

[0022] In the above technical solution, in step 3, the process of establishing the NMPC controller of the heavy truck power system is as follows:

[0023]

[0024] in, is the time constant of the torque inertia link, is the control variable of the NMPC optimal control problem, i.e., the target engine torque;

[0025] The state variables of the optimal control problem are vehicle speed v and travel distance s d :

[0026]

[0027] Where i0 is the engine differential ratio, R w is the wheel radius of heavy trucks.

[0028] In the above technical solution, in step 3, the multi-objective cost function is:

[0029] J=J v +J fuel +J change (5)

[0030] J v For speed tracking metrics:

[0031] J v =w1·(v ref -v) 2 (6)

[0032] w1 is the speed traceability weight, v ref is the vehicle cruising reference speed, v is the vehicle speed;

[0033] J fuel Fuel economy index:

[0034] J fuel =w2·f(T e ,n e ) (7)

[0035] w2 is the fuel economy weight;

[0036] J change is the stability indicator:

[0037] J change =w3·ΔT e (8)

[0038] Where w3 is the torque change weight, ΔT e is the torque variation.

[0039] In the above technical solution, in step 3, the engine torque variation range is constrained as follows:

[0040] T emin ≤T e ≤T emax (n e ) (10)

[0041] Where T emin is a preset constant, T emax Calculated based on engine characteristics, it is related to engine speed n e Related functions.

[0042] In the above technical solution, the neural network in step 4 is a CNN-LSTM network.

[0043] In the above technical solution, the architecture of the neural network in step 4 includes an input layer for receiving the future slope sequence, two CNN layers for feature extraction, an LSTM layer for capturing time dependency, and a fully connected layer for classifying the output controller parameter w2.

[0044] In the above technical solution, in step 7, the automated simulation scheduling system performs simulation according to the discrete scanning range of the fuel economy w2 input by the user, and the discrete scanning range of the fuel economy weight w2 is manually preset based on experience;

[0045] w 2,i ∈{w 2,1 ,w 2,2 ,w 2,3 ,…,w 2,10}(11)

[0046] After completing the parameter scan (11) for each slope, the performance evaluation is performed using indicators such as fuel consumption and vehicle stability, as shown below:

[0047] E total =αE fuel +βE stability (12)

[0048] Among them, α and β are the weights of fuel consumption and vehicle stability index respectively. fuel and E stability The magnitudes of the indicators are different, so the normalization method is used to ensure that the value of each indicator is between 0 and 1;

[0049] Fuel consumption is measured in terms of consumption per 100 kilometers, recorded as

[0050]

[0051] Where T is the total duration of a single simulation, is the fuel consumption rate per unit time, ρ fuel is the fuel density, s is the total driving distance in the simulation;

[0052] The vehicle stability index is expressed as follows:

[0053]

[0054] Where v is the real-time vehicle speed during the simulation.

[0055] In the above technical solution, in step 8, the cross entropy loss function is used to train the neural network.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] 1. This paper proposes a slope-adaptive predictive energy-saving control method for heavy trucks using LLM-assisted data collection, aiming to improve fuel efficiency on variable slopes. This method utilizes a CNN-LSTM network to dynamically adjust the fuel economy weight w2 based on long-scale slope information, overcoming the limitation of traditional NMPC methods that rely solely on short-scale slope information.

[0058] 2. To address the high time and labor costs associated with collecting network training data, this paper designs an automated simulation scheduling system based on LLM agents to automatically execute simulations and record data. This innovative LLM application enables the agents to leverage expert knowledge to guide the parameter scanning simulation process, effectively focusing on key areas and avoiding unnecessary calculations. This significantly reduces the time and labor required for training data collection.

[0059] In summary, the proposed LLM-assisted data collection method for adaptive, predictive, and energy-saving control of heavy trucks based on slope gradients demonstrates excellent overall performance. Dynamically adjusting the NMPC fuel economy weight w2 through a CNN-LSTM network enables the controller to account for long-range slope information, improving control performance closer to the global optimum and ultimately achieving fuel savings. Furthermore, the automated simulation and scheduling system comprised of intelligent agents equipped with LLMs accelerates the collection of neural network training data and significantly reduces human effort. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 It is a flow chart of the slope-adaptive predictive energy-saving control method for heavy trucks with LLM-assisted data collection.

[0061] Figure 2 It is a longitudinal dynamic model analysis.

[0062] Figure 3 It is the fitting result of engine fuel consumption characteristics BSFC.

[0063] Figure 4 It is the architecture of a neural network.

[0064] Figure 5 It is an automated simulation scheduling system framework. DETAILED DESCRIPTION

[0065] The present invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0066] like Figure 1 As shown, a slope adaptive predictive energy-saving control method for heavy trucks with LLM-assisted data collection includes the following steps:

[0067] Step 1: Based on Newton's second law, a longitudinal dynamic model of a heavy truck on a slope road is established. Figure 2 As shown, the longitudinal dynamics model of the vehicle is as follows:

[0068]

[0069] Where m is the vehicle mass (kg), θ is the road slope (rad), and ρ is the air density (kg / m 3 ), C d is the air resistance coefficient, A a is the frontal area of ​​the vehicle (m 2 ), v is the vehicle speed (m / s), f is the rolling resistance coefficient, F drive is the driving force that drives the vehicle to move (N), F aero is the air resistance (N), F roll is the rolling resistance (N), F slope is the slope resistance (N).

[0070] Step 2: Based on the engine dynamometer test, the engine fuel consumption characteristics are obtained and the brake specific fuel consumption (BSFC) is obtained by fitting a fifth-order polynomial, as shown in Formula 2.

[0071]

[0072] Among them, α is the fuel consumption characteristic fitting coefficient, T e is the engine torque (N·m), n e is the engine speed (r / min).

[0073] Figure 3 The figure shows the fitting result of the engine fuel consumption characteristic (BSFC). The red line is the engine external characteristic torque curve, which is the maximum torque curve at different engine speeds. The green line represents the constant power line, and the blue line represents the optimal fuel economy line, also known as the E line.

[0074] Step 3: Build an NMPC controller for the heavy truck powertrain to achieve precise control of the engine torque. To ensure the accuracy of the heavy truck powertrain, the inertia of the powertrain needs to be considered:

[0075]

[0076] in, is the time constant of the torque inertia link (s), is the control variable of the NMPC optimal control problem, namely the target engine torque (N·m).

[0077] The state variables of the optimal control problem are vehicle speed v and travel distance s d The state space equation of the optimal control problem can be expressed as:

[0078]

[0079] Where i0 is the engine differential ratio, R w is the wheel radius of heavy truck (m).

[0080] In the predictive cruise control system, the two performance indicators of speed tracking and fuel consumption conflict with each other, forming a multi-objective optimization problem. The multi-objective cost function is:

[0081] J=J v +J fuel +J change (5)

[0082] Among them, the designed speed tracking index is

[0083] J v =w1·(v ref -v) 2 (6)

[0084] Where w1 is the speed traceability weight, v ref is the vehicle cruising reference speed (m / s).

[0085] In vehicle cruise control, the evaluation of fuel economy requires consideration of both fuel consumption and vehicle dynamic performance, with the above-mentioned effective fuel consumption rate being used as an indicator of economy rather than using fuel consumption alone.

[0086] J fuel =w2·f(T e ,n e ) (7)

[0087] Where w2 is the fuel economy weight, f(T e ,n e ) is the effective fuel consumption function (g / kW·h) related to engine torque and engine speed, which is fitted by equation (2). This function reduces fuel consumption by keeping the engine operating point close to the optimal operating point while ensuring good vehicle power performance.

[0088] In terms of stability, in order to limit the fluctuation of vehicle acceleration caused by frequent changes in torque, a stability index is added to constrain the frequent changes of the power system.

[0089] J change =w3·ΔT e (8)

[0090] Where w3 is the torque change weight, ΔT e is the torque change (N·m).

[0091] Due to computational limitations, the predicted slope information for the first 150 meters ahead, provided by a digital map, was selected as the NMPC prediction field. To accommodate varying slopes over a wider range, the proposed CNN-LSTM network sets the fuel economy weight w2 as a dynamic weight that is updated based on long-scale slope information.

[0092] w2=f(s long ) (9)

[0093] For this optimal control problem, there is a constraint: torque variation range constraint,

[0094] T emin ≤T e ≤T emax (n e ) (10)

[0095] Where T emin is a preset constant, T emax Calculated based on engine characteristics, it is a function related to the speed. This constraint provides boundary conditions for the optimal control problem in this paper, ensuring that the algorithm finds the optimal solution within the feasible region.

[0096] Step 4: Build a neural network that captures the mapping relationship between the slope sequence and the NMPC fuel consumption weight:

[0097] The NMPC algorithm in step 3 is limited in its predictive power, with a prediction range of only 150 meters. This limitation results in the NMPC algorithm essentially only being able to obtain local optimal solutions. Iteratively applying these local optimal solutions through rolling updates does not guarantee convergence to the global optimum. Therefore, the present invention considers using a neural network to capture longer-scale slope information, thereby making the results closer to the global optimum.

[0098] like Figure 4 The neural network architecture consists of an input layer for receiving the future slope sequence, two CNN layers for feature extraction, an LSTM layer for capturing temporal dependencies, and a fully connected layer for classifying and outputting controller parameters. It receives long-scale slope information within a range of 600 meters ahead and outputs the optimal fuel economy weight w2 for that slope.

[0099] Because the network processes long-term slope information hidden in the future (sampled every 10 meters), the input layer size is set to 61. The CNN layers (both CNN layers are one-dimensional, denoted as Conv1D) are responsible for extracting local and global features from the slope sequence. The first layer uses a convolution kernel of size 1 to capture initial slope changes, maintaining the sequence length with padding; subsequent layers use larger convolution kernels to identify broader trends in the data. Specifically, the second CNN layer doubles the number of output channels to 64, uses a convolution kernel of size 3, and uses padding of 1. Furthermore, each CNN layer is followed by batch normalization to stabilize the learning process, and a ReLU activation function is applied to introduce nonlinearity. After the CNN layers process the slope data, the output is passed to an LSTM layer. This layer is crucial for capturing the temporal dynamics in the slope sequence, contributing to accurate prediction of fuel consumption weights. This LSTM component, consisting of 32 hidden units, operates on the features output by the CNN layers, capturing both short-term and long-term dependencies in the slope sequence. A dropout rate of 0.3 is set in the LSTM layer to mitigate overfitting and ensure good generalization of the model to unseen data. The model concludes with a fully connected network (FC) layer, which compresses the high-dimensional features extracted by the LSTM layer into a more manageable form for classification, specifically selecting the optimal fuel economy weight. This FC layer uses a softmax activation function at the output layer to output the probabilities corresponding to 10 different preset fuel economy weights, thereby enabling precise selection of fuel economy weights based on the slope sequence data.

[0100] Step 5: Generate feature data for neural network training, namely slope sequence:

[0101] Construct a dataset of slope sequences for multiple 600m-long highway sections. Considering the relatively flat nature of highways, the slope generation problem is simplified to a combination of straight line segments. Because this paper focuses on predictive energy-saving strategies during uphill driving, all slopes are designed to be upward. To ensure broad coverage of the generated data, slope angle and horizontal distance are set as variables. By combining these variables, a sufficiently diverse set of straight line segments can be generated.

[0102] In this example, the first section is a 150m flat road section with a zero slope to allow the truck to enter a stable initial state. The second section begins after 150m and the slope increases at a certain angle. The height is calculated using the following formula:

[0103] h(x)=x·tanθ (11)

[0104] Where h(x) represents the slope height (m), θ is the slope angle (rad), and its maximum value is limited to 0.04 rad to meet typical highway design standards. The variable x represents the horizontal distance (m) from the starting point of the slope (150 m), ranging from 0 to 450 m.

[0105] The third section is a flat section that starts from the point where the grade ends and extends to a total length of 600 meters, maintaining the maximum height reached at the end of the grade.

[0106] By taking the slope angle and horizontal distance as variables, discretizing them and combining their values, a set of slope sequences covering a sufficiently wide range of uphill scenarios can be generated.

[0107] Step 6: Establish an automated simulation scheduling system:

[0108] The automated simulation scheduling system consists of three parts: execution agent, analysis agent and simulation engine. The framework is as follows: Figure 5 As shown in the figure. Among them, the execution agent is responsible for generating the corresponding parameter scanning direction execution command based on the user instructions and the feedback of the analysis agent to guide the simulation engine (that is, inputting the next parameter w2 to be simulated into the simulation engine to obtain the multimodal results under this parameter). The analysis agent analyzes the multimodal simulation results and generates formatted feedback to help adjust the execution agent's actions in the next round. The simulation engine plays a bridging role between the two. It converts the execution command into the underlying executable command of the simulation software, the specific parameter scanning simulation process, and sends the simulation results to the analysis agent. The user only needs to provide the execution agent with the slope sequence (the long-scale slope sequence data set in step 5) and the parameter scanning range (that is, the manually preset range of w2) to start the simulation process. During the simulation process, the execution agent determines the direction of the parameter scanning based on the feedback text of the analysis agent and the prior knowledge in the prompt word.

[0109] Specifically:

[0110] In the execution agent, the function call method plays a key role in achieving simulation automation. Specifically, the semantic information of the parameter sweep simulation process is converted into a standardized function, in which the input and output of the simulation and its detailed Schema description are clearly defined and stored in the function library. Then, the large language model GLM-4.0 is deployed to identify the most appropriate simulation function by comparing user instructions and Schema descriptions. Here, the function for parameter sweep simulation is identified from the function library, and GLM-4.0 generates the execution command and passes it to the rule-based simulation engine. The following shows the system prompt words of the execution agent, defining the functions and roles:

[0111] System prompt words:

[0112] You are the executive agent, responsible for managing and executing simulation tasks. Your primary responsibility is to process user commands and analyze agent feedback, translating them into executable commands for the simulation engine.

[0113] You will receive the data required for the simulation from the user, including the path to the ramp sequence file and the parameter sweep range. Your task is to execute the simulation and adjust subsequent operations based on feedback from previous results.

[0114] Because traversing the entire parameter sweep space is too time-consuming, this paper uses a hint-based approach to incorporate expert controller tuning experience into the large language model. By leveraging the expert knowledge contained in these hints, the LLM can prioritize exploring specific parameter regions that are more likely to produce optimal results, thereby reducing the need to search the entire parameter space. The hint design is as follows:

[0115] Threshold detection cues: These cues enable the model to detect critical thresholds and automatically halt further parameter exploration. For example, when the fuel consumption weight reaches a certain threshold, the engine may begin to oscillate, causing vehicle instability. By promptly identifying oscillating regions, the agent skips parameter ranges that no longer offer optimization potential, focusing computational resources on more valuable areas.

[0116] Examples of threshold detection prompt words are as follows:

[0117] If the engine speed oscillates, stop further parameter sweeps.

[0118] If the vehicle speed stability deteriorates rapidly, stop increasing the weight parameter.

[0119] If the vehicle speed drops to 0, stop increasing the fuel economy weighting.

[0120] Jump parameter adjustment hints: When simulation results show minimal differences between adjacent parameters, these hints instruct the model to skip unnecessary parameter sweeps and adjust directly to a larger parameter step size. This jump adjustment strategy speeds up the optimization process and avoids redundant computations.

[0121] Examples of prompts for adjusting jump parameters are as follows:

[0122] If the adjacent weight w 2,1 , w 2,2 , and w 2,3 The difference between them is small, so the parameter scanning step size is increased, such as the next step weight is set to w 2,5 .

[0123] If no significant improvement is observed in the first few steps, skip the parameter w 2,i .

[0124] The executing agent utilizes GPT4-o, a framework with exceptional multimodal capabilities. Its primary responsibility is to analyze the results generated by each simulation run, generate textual feedback, and send it to the executing agent. The overall prompt, including standardized requirements for controlling result analysis and expert guidance, is as follows:

[0125] System prompt words:

[0126] You are a vehicle control expert specializing in analyzing and optimizing simulation results. You receive data generated by a controller parameter sweep simulation, including the slope sequence ID, current fuel consumption weight, engine torque and speed, vehicle speed, and acceleration. You need to thoroughly analyze this data and its corresponding graphs.

[0127] Standardization requirement prompt words:

[0128] You need to analyze each simulation result one by one to determine whether the key performance indicators of the system meet the expected values. Please format the analysis report as follows:

[0129] Simulation ID (insert simulation ID in the format: [slope sequence ID - current fuel economy weight]).

[0130] Fuel consumption analysis (briefly describe the fuel consumption, focusing on the value at the end of the simulation).

[0131] Engine oscillation analysis (indicates whether torque and speed oscillate).

[0132] Speed ​​and acceleration analysis (describing the smoothness of the speed and acceleration curves, noting any sharp changes).

[0133] Overall recommendations (parameter sweep suggestions based on analysis results).

[0134] Expert guidance and suggestion words:

[0135] Pay particular attention to the fuel consumption values ​​at the end of the simulation and report any unusual spikes.

[0136] Monitor the engine torque and speed during simulation for oscillations.

[0137] Carefully evaluate the smoothness of the vehicle's speed and acceleration curves, checking for sharp rises or falls.

[0138] If any anomalies are found, provide detailed explanations and suggest how to adjust the simulation parameters.

[0139] Finally, in the simulation engine design, deep integration with simulation platforms (such as Simulink) is first achieved through an API (such as the MATLAB API). This integration process involves configuring the API to enable seamless communication with the simulation platform, ensuring that instructions generated by the execution agent are correctly transmitted to the target simulation platform. The instructions generated by the execution agent are then converted into a command format executable by the simulation platform and transmitted through the API. Furthermore, result logging and forwarding are required to ensure that the simulation platform outputs the execution status or results to the analysis agent after executing the instructions.

[0140] Step 7: Generate labels for neural network training, i.e., the optimal fuel economy weight w2 corresponding to each slope sequence:

[0141] The label is determined by running simulations in the simulation engine to find the optimal fuel economy weight w2 for each slope sequence. The automated simulation scheduling system in step 6 then performs simulations based on the parameter sweep range entered by the user. The discrete sweep range for the fuel economy weight is pre-set based on experience to reduce computational burden.

[0142] w 2,i ∈{w 2,1 ,w 2,2 ,w 2,3 ,…,w 2,10}(11)

[0143] After completing the parameter scan (11) for each slope, the performance evaluation is performed using indicators such as fuel consumption and vehicle stability, as shown below:

[0144] E total =αE fuel +αE stability (12)

[0145] Among them, α and β are the weights of fuel consumption and vehicle stability index respectively. fuel and E stability To ensure that the value of each indicator is between 0 and 1, a normalization method is used.

[0146] Fuel consumption is measured in terms of consumption per 100 kilometers, recorded as

[0147]

[0148] Where T is the total duration of a single simulation (s), is the fuel consumption rate per unit time (kg / s), ρ fuel is the fuel density (kg / L), and s is the total driving distance in the simulation (km).

[0149] The vehicle stability index is expressed as follows:

[0150]

[0151] Where v is the real-time vehicle speed (m / s) during the simulation.

[0152] Step 8: Training of the neural network:

[0153] The neural network was trained and tested using a training dataset and a test dataset consisting of the labels obtained in step 7 and the long-scale slope sequence obtained in step 5. During training, the network used the cross-entropy loss function, a common and standard choice in multi-classification tasks, to measure the difference between the predicted class and the actual label. The Adam optimizer was used for the training process due to its efficiency in handling sparse gradients and adaptive learning rates.

[0154] In this embodiment, the batch size of the training scheme is set to 128, with a total of 500 epochs, which provides sufficient time for the network to capture the complex dependency between the slope sequence data and the optimal fuel economy weight.

[0155] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A heavy truck slope adaptive predictive energy-saving control method with LLM-assisted data collection, characterized in that: The following steps are involved: Step 1: Establish a longitudinal dynamic model of a heavy truck on a slope road; Step 2: Establish an engine fuel consumption characteristic model to obtain brake specific fuel consumption (BSFC); Step 3: Establish an NMPC controller for the heavy truck powertrain to precisely control the engine torque and provide feedback to a control-oriented model based on the longitudinal dynamics model from Step 1 and the engine fuel consumption characteristic model from Step 2. A multi-objective cost function is constructed to balance speed tracking and fuel consumption. The multi-objective cost function includes a speed traceability weight w1, a fuel economy weight w2, and a torque variation weight w3. Step 4: construct a neural network that captures the mapping relationship between the slope sequence and the fuel economy weight w2 in step 2, and the neural network outputs the optimal fuel economy weight w2 to the NMPC controller in step 3; Step 5: construct a feature data set for neural network training, i.e., a long-scale slope sequence; Step 6: Establish an automated simulation scheduling system consisting of an execution agent, an analysis agent, and a simulation engine. The execution agent generates execution commands for the w2 scanning direction based on the long-scale slope sequence in step 5 and the discrete scanning range of the manually set fuel economy weight w2. The simulation engine converts the execution commands into low-level executable commands for the simulation software and the specific w2 scanning simulation process, generating multimodal simulation results and sending them to the analysis agent. The analysis agent analyzes the multimodal simulation results and generates formatted feedback to help adjust the execution agent's actions in the next round. Step 7: The automated simulation scheduling system generates a label, where the label is the optimal fuel economy weight w2 corresponding to each long-scale slope sequence; Step 8: Use the training data set and the test data set consisting of the labels obtained in step 7 and the long-scale slope sequence obtained in step 5 to train and test the neural network.

2. The heavy truck slope adaptive predictive energy-saving control method according to claim 1, characterized in that: The longitudinal dynamic model in step 1 is: Where m is the vehicle mass, θ is the road slope, ρ is the air density, C d is the air resistance coefficient, A a is the frontal area of ​​the vehicle, v is the vehicle speed, f is the rolling resistance coefficient, F drive is the driving force that drives the vehicle to move, F aero is the air resistance, F roll is the rolling resistance, F slope is the slope resistance.

3. The heavy truck slope adaptive predictive energy-saving control method according to claim 1, characterized in that: In step 2, the engine fuel consumption characteristic model is: Among them, α is the fuel consumption characteristic fitting coefficient, T e is the engine torque, n e is the engine speed.

4. The heavy truck slope adaptive predictive energy-saving control method according to claim 1, characterized in that: In step 3, the process of establishing the NMPC controller of the heavy truck power system is as follows: in, is the time constant of the torque inertia link, is the control variable of the NMPC optimal control problem, i.e., the target engine torque; The state variables of the optimal control problem are vehicle speed v and travel distance s d : Where i0 is the engine differential ratio, R w is the wheel radius of heavy trucks.

5. The heavy truck slope adaptive predictive energy-saving control method according to claim 1, characterized in that: In step 3, the multi-objective cost function is: J=J v +J fuel +J change (5) J v For speed tracking metrics: J v =w1·(v ref -v) 2 (6) w1 is the speed traceability weight, v ref is the vehicle cruising reference speed, v is the vehicle speed; J fuel Fuel economy index: J fuel =w2·f(T e ,n e ) (7) w2 is the fuel economy weight; J change is the stability indicator: J change =w3·ΔT e (8) Where w3 is the torque change weight, ΔT e is the torque variation.

6. The heavy truck slope adaptive predictive energy-saving control method according to claim 1, characterized in that: In step 3, the engine torque variation range is constrained as follows: T emin ≤T e ≤T emax (n e ) (10) Where T emin is a preset constant, T emax Calculated based on engine characteristics, it is related to engine speed n e Related functions.

7. The heavy truck slope adaptive predictive energy-saving control method according to claim 1, characterized in that: The neural network in step 4 is a CNN-LSTM network.

8. The heavy truck slope adaptive predictive energy-saving control method according to claim 7, characterized in that: The architecture of the neural network in step 4 includes an input layer for receiving the future slope sequence, two CNN layers for feature extraction, an LSTM layer for capturing time dependencies, and a fully connected layer for classifying the output controller parameter w2.

9. The heavy truck slope adaptive predictive energy-saving control method according to claim 1, characterized in that: In step 7, the automated simulation scheduling system performs simulation based on the discrete scanning range of the fuel economy w2 input by the user. The discrete scanning range of the fuel economy weight w2 is manually preset based on experience. In 2,i ∈{in 2,1 ,In 2,2 ,In 2,3 ,…,In 2,10 } (11) After completing the parameter sweep (11) for each slope, the performance evaluation is performed using fuel consumption and vehicle stability indicators as shown below: E total =αE fuel +βE stability (12) Among them, α and β are the weights of fuel consumption and vehicle stability index respectively. fuel and E stability The magnitudes of the indicators are different, so the normalization method is used to ensure that the value of each indicator is between 0 and 1; Fuel consumption is measured in terms of consumption per 100 kilometers, recorded as Where T is the total duration of a single simulation, is the fuel consumption rate per unit time, ρ fuel is the fuel density, s is the total driving distance in the simulation; The vehicle stability index is expressed as follows: Where v is the real-time vehicle speed during the simulation.

10. The heavy truck slope adaptive predictive energy-saving control method according to claim 1, characterized in that: In step 8, the neural network is trained using a cross entropy loss function.

Images