Vehicle sliding energy recovery control method and device, electronic equipment and storage medium

Abstract

The invention provides a vehicle sliding energy recovery control method and device, electronic equipment and a storage medium, self-adaptive decision making and safety constraint on sliding deceleration are achieved by mining and learning longitudinal control behaviors of an expert driver in multiple scenes, the sliding working condition energy recovery level is improved, and meanwhile the safety of the vehicle sliding deceleration is improved. The smoothness of the speed reduction process is improved, the mechanical braking dependency degree and the calibration complexity are reduced, and therefore the energy utilization efficiency and the driving experience of the new energy automobile in the real traffic environment are comprehensively improved.

Description

Technical Field

[0001] This disclosure relates to the field of vehicle motion control technology, and more specifically, to a method, device, electronic device, and storage medium for controlling vehicle coasting energy recovery. Background Technology

[0002] As the global automotive industry rapidly moves towards electrification and intelligentization, regenerative braking technology, a key energy-saving technology that distinguishes new energy vehicles from traditional fuel vehicles, can convert the vehicle's kinetic energy into electrical energy and recover it into the power battery during vehicle deceleration by utilizing the generator mode of the drive motor, thereby effectively extending the vehicle's driving range. Among the various application scenarios of regenerative braking, coasting energy recovery is a particularly important condition. Coasting energy recovery refers to the process where, after the driver releases the accelerator pedal, without actively pressing the brake pedal, the vehicle decelerates solely through the braking force generated by the motor's reverse drag, thus recovering energy.

[0003] Currently, a typical approach to coasting energy recovery control involves providing several discrete, fixed coasting recovery intensity levels, which the driver manually selects via the instrument panel or central control interface. While this approach is simple in structure and low in implementation cost, it struggles to automatically adjust deceleration based on external environmental factors such as the status of the vehicle ahead, road gradient, and speed limit changes. This can lead to insufficient energy recovery in real-world conditions like traffic jams, vehicle weaving, and complex inclines, failing to balance comfort and fuel efficiency across different driving styles. Another approach utilizes millimeter-wave radar, cameras, and navigation maps to construct an adaptive coasting recovery control strategy based on expert rules. This involves using a large set of pre-defined rules to select or adjust the coasting recovery level online based on factors such as distance, speed difference, speed limit changes, upcoming intersections, and gradient. However, real-world traffic flow environments are highly complex and random, and human rules cannot cover all edge cases. Sudden decision-making in abnormal conditions can lead to abrupt deceleration experiences and even potential safety risks. Summary of the Invention

[0004] This disclosure provides at least one vehicle coasting energy recovery control method, device, electronic device, and storage medium. By mining and learning the longitudinal control behavior of expert drivers in multiple scenarios, it achieves adaptive decision-making and safety constraints on coasting deceleration. While improving the energy recovery level under coasting conditions, it also improves the smoothness of the deceleration process, reduces the dependence on mechanical braking and calibration complexity, thereby comprehensively improving the energy utilization efficiency and driving experience of new energy vehicles in real traffic environments.

[0005] This disclosure provides a method for controlling vehicle coasting energy recovery, including: The system collects driving logs of vehicles under various road conditions and driving styles, constructs the trajectory of the preceding vehicle based on the driving logs, and uses a pre-set expert driving model to follow the trajectory of the preceding vehicle in a simulation environment, recording the vehicle status and expert decision actions, and generating an expert sample dataset for coasting energy recovery control. Using the temporal features characterizing the vehicle's longitudinal driving state and the relative relationship between the vehicle and the workshop as input, and using the target longitudinal acceleration or deceleration output by the expert driving model in the expert sample dataset as a supervision signal, an imitation ecological driving model for outputting the target deceleration is trained. When the vehicle is detected to be in a condition that meets the conditions for coasting energy recovery, the current vehicle state is collected and the corresponding time-series features are constructed and input into the simulated ecological driving model to obtain the target deceleration. The target deceleration is then converted into the energy recovery control quantity of the drive motor to control the vehicle to recover energy during coasting. Based on the current vehicle state and the target deceleration, at least one safety evaluation index is calculated. When the safety evaluation index does not meet the preset safety constraints, the target deceleration is corrected to a safe deceleration, and the safe deceleration is used as a coasting energy recovery control command.

[0006] In one optional implementation, an expert sample dataset for coasting energy recovery control is generated based on the following steps: The driving logs are collected from drivers of different vehicle types and driving styles in various road scenarios such as urban roads, suburban roads and highways. The driving logs include at least a time series of vehicle speeds collected at a preset sampling frequency. The vehicle speed time series is smoothed and its trajectory reconstructed based on preset kinematic constraints to obtain a set of virtual forward vehicle motion trajectories that meet the requirements of speed and acceleration continuity. In the offline simulation environment, the virtual preceding vehicle's trajectory is used as the excitation input to control the expert driving model to drive the vehicle to follow the simulation. The state update and expert action calculation are performed cyclically at a fixed time step. The state vector at each moment and the corresponding expert deceleration or acceleration are recorded as data pairs to construct the expert sample dataset. The expert driving model is an intelligent driver model. The intelligent driver model calculates the expected longitudinal acceleration based on the vehicle's expected speed, the net distance between the vehicle and the vehicle in front, and the speed difference. The net distance is determined by the difference between the current distance between the vehicles and the preset minimum safe distance when the vehicle is stationary. The intelligent driver model achieves comprehensive constraints on safety and comfort through a distance function that includes the expected headway, comfort acceleration, and comfort deceleration.

[0007] In one alternative implementation, a simulated eco-driving model for outputting the target deceleration is trained based on the following steps: Construct an instantaneous state vector, which includes at least one of the following: vehicle speed, vehicle longitudinal acceleration, net distance between the vehicle and the vehicle in front, relative speed between the vehicle and the vehicle in front, and headway and reciprocal of collision time calculated based on the net distance and relative speed. The instantaneous state vectors of the current moment and several historical moments are stacked in chronological order to obtain a temporal feature vector that characterizes the evolution of the vehicle's longitudinal behavior, and the temporal feature vector is then normalized. The temporal feature vector is used as the input to the simulated eco-driving model, and the expert target deceleration at the corresponding time in the expert sample dataset is used as the output label to supervise the training of the simulated eco-driving model. In one optional implementation, the training of the simulated eco-driving model employs a composite loss function that includes imitation loss, safety loss, and comfort loss, wherein: The imitation loss is used to characterize the deviation between the target deceleration output by the model and the expert target deceleration in the expert sample dataset; The safety loss is based on the current vehicle state and the target deceleration output by the model to predict the collision time at the next moment. When the collision time is less than a preset safety threshold, the model output is penalized to suppress actions that tend to be dangerous following. The comfort loss is calculated as a function of the rate of change of the target deceleration between consecutive time steps, and is used to penalize decisions where the deceleration changes too quickly, so as to improve the smoothness of the coasting deceleration process. The composite loss function is a weighted sum of the above losses, with weighting coefficients used to balance simulation accuracy, safety, and comfort.

[0008] In one alternative implementation, a simulated eco-driving model for outputting the target deceleration is further trained based on the following steps: The initial imitation eco-driving model was trained using the expert sample dataset to obtain the initial policy model; The initial strategy model is deployed in a closed-loop simulation environment to generate a simulated driving process, and the state sequence accessed by the initial strategy model in the simulated driving process is collected. The expert driving model is invoked to label the state sequence, and the corresponding expert target deceleration is obtained; The labeled state-action data is aggregated into the original training dataset, and the simulated ecological driving model is retrained based on the aggregated dataset to obtain an updated policy model. Repeat the steps of generating the simulated driving process until the performance of the simulated ecological driving model converges on the validation set. In one optional implementation, at least one safety evaluation index is calculated based on the current vehicle state and the target deceleration. When the safety evaluation index does not meet preset safety constraints, the target deceleration is corrected to a safe deceleration. Specifically, this includes: The collision time is calculated based on the current vehicle speed, vehicle spacing and relative speed. When the collision time is lower than the preset safety red line, the target deceleration output by the simulated ecological driving model is replaced with the corresponding maximum safe deceleration. Determine whether the target deceleration exceeds the vehicle's physical allowable range. If the target deceleration exceeds the limit, limit the target deceleration to the allowable range. The inference latency and output validity of the simulated ecological driving model are monitored. When an inference timeout or abnormal output is detected, the control mode is switched to a preset fixed deceleration mode or a traditional coasting energy recovery strategy.

[0009] In one alternative implementation, after deploying the simulated eco-driving model in the vehicle, the method further includes: The vehicle's actual driving status during operation, the target deceleration output by the simulated ecological driving model, and the actual deceleration and energy recovery data executed by the vehicle control system are transmitted back to the cloud data platform. In the cloud data platform, scenarios with significant deviations between the decision-making of the simulated ecological driving model and the decision-making or safety objectives of human driving are identified, and the data corresponding to the scenarios are used to construct an incremental training dataset. The simulated eco-driving model is retrained or fine-tuned based on the incremental training dataset, and after passing the cloud regression test, the updated simulated eco-driving model is delivered to the vehicle via over-the-air update.

[0010] This disclosure also provides a vehicle coasting energy recovery control device, including: The expert sample construction module is used to collect driving logs of vehicles under various road conditions and driving styles, construct the trajectory of the preceding vehicle based on the driving logs, and follow the trajectory of the preceding vehicle in a simulation environment using a pre-set expert driving model, record the vehicle status and expert decision actions, and generate an expert sample dataset for coasting energy recovery control. The simulated ecological driving model training module is used to take the temporal features representing the vehicle's longitudinal driving state and the relative relationship between the vehicle and the workshop as input, and the target longitudinal acceleration or deceleration output by the expert driving model in the expert sample dataset as a supervision signal to train an simulated ecological driving model for outputting the target deceleration. The coasting energy recovery control module is used to collect the current vehicle state and construct the corresponding time-series features when the vehicle is detected to be in a condition that meets the coasting energy recovery conditions. The input is then fed into the simulated ecological driving model to obtain the target deceleration. The target deceleration is then converted into the energy recovery control quantity of the drive motor to control the vehicle to recover energy during coasting. The safety assessment module is used to calculate at least one safety evaluation index based on the current vehicle state and the target deceleration. When the safety evaluation index does not meet the preset safety constraints, the target deceleration is corrected to a safe deceleration, and the safe deceleration is used as a coasting energy recovery control command.

[0011] This disclosure also provides an electronic device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the steps of the above-described vehicle coasting energy recovery control method, or any possible implementation of the above-described vehicle coasting energy recovery control method.

[0012] This disclosure also provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of the above-described vehicle coasting energy recovery control method, or any possible implementation thereof.

[0013] This disclosure also provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the above-described vehicle coasting energy recovery control method, or the steps in any possible implementation of the above-described vehicle coasting energy recovery control method.

[0014] This disclosure provides a vehicle coasting energy recovery control method, device, electronic device, and storage medium. By mining and learning the longitudinal control behavior of expert drivers in multiple scenarios, it achieves adaptive decision-making and safety constraints on coasting deceleration. While improving the energy recovery level under coasting conditions, it also improves the smoothness of the deceleration process, reduces the dependence on mechanical braking and calibration complexity, thereby comprehensively improving the energy utilization efficiency and driving experience of new energy vehicles in real traffic environments.

[0015] To make the above-mentioned objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings used in the embodiments will be briefly described below. These drawings are incorporated in and constitute a part of this specification. They illustrate embodiments conforming to this disclosure and, together with the specification, serve to explain the technical solutions of this disclosure. It should be understood that the following drawings only show some embodiments of this disclosure and should not be considered as limiting the scope. Those skilled in the art can obtain other related drawings based on these drawings without creative effort.

[0017] Figure 1 A flowchart of a vehicle coasting energy recovery control method provided in an embodiment of this disclosure is shown; Figure 2 A flowchart illustrating a method for training an eco-driving model provided by an embodiment of this disclosure is shown; Figure 3 A schematic diagram of a vehicle coasting energy recovery control device provided in an embodiment of this disclosure is shown; Figure 4 A schematic diagram of an electronic device provided in an embodiment of the present disclosure is shown. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. The components of the embodiments of this disclosure described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this disclosure provided in the accompanying drawings is not intended to limit the scope of the claimed disclosure, but merely represents selected embodiments of this disclosure. All other embodiments obtained by those skilled in the art based on the embodiments of this disclosure without inventive effort are within the scope of protection of this disclosure.

[0019] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0020] In this document, the term "and / or" merely describes a relationship, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0021] Research has found that current coasting energy recovery control systems struggle to automatically adjust deceleration based on external environmental factors such as the status of the vehicle ahead, road gradient, and speed limit changes. This results in insufficient energy recovery in real-world conditions such as traffic jams, vehicle weaving, and complex inclines. Furthermore, these systems cannot balance comfort and energy efficiency for different driving styles. Real-world traffic flow environments are highly complex and random, and human-defined rules cannot cover all edge scenarios. In abnormal conditions, sudden decision-making can easily occur, leading to abrupt deceleration experiences and even potential safety risks.

[0022] Based on the above research, this disclosure provides a vehicle coasting energy recovery control method, device, electronic device and storage medium. By mining and learning the longitudinal control behavior of expert drivers in multiple scenarios, it realizes adaptive decision-making and safety constraints on coasting deceleration. While improving the energy recovery level in coasting conditions, it also improves the smoothness of the deceleration process, reduces the dependence on mechanical braking and calibration complexity, thereby comprehensively improving the energy utilization efficiency and driving experience of new energy vehicles in real traffic environments.

[0023] To facilitate understanding of this embodiment, a detailed description of the vehicle coasting energy recovery control method disclosed in this disclosure is provided first. The executing entity of the vehicle coasting energy recovery control method provided in this disclosure is generally a computer device with certain computing capabilities. This computer device may include, for example, a terminal device, a server, or other processing devices. The terminal device may be a user equipment (UE), mobile device, user terminal, terminal, cellular phone, cordless phone, personal digital assistant (PDA), handheld device, computing device, in-vehicle device, wearable device, etc. In some possible implementations, the vehicle coasting energy recovery control method can be implemented by a processor calling computer-readable instructions stored in memory.

[0024] See Figure 1 The diagram shows a flowchart of a vehicle coasting energy recovery control method according to an embodiment of this disclosure. The method includes steps S101 to S104, wherein: S101. Collect driving logs of the vehicle under various road conditions and driving styles, construct the trajectory of the preceding vehicle based on the driving logs, and use a pre-set expert driving model to follow the trajectory of the preceding vehicle in a simulation environment, record the vehicle status and expert decision actions, and generate an expert sample dataset for coasting energy recovery control.

[0025] In practice, the first step is to collect and construct expert sample data for imitation learning. Based on real vehicle driving data in various road scenarios and under different driving styles, the motion trajectory of a virtual preceding vehicle is first constructed. Then, a pre-set expert driving model is used to follow the trajectory of the preceding vehicle in the simulation environment. During the simulation, the vehicle's state and expert decision-making actions are recorded simultaneously, thus forming an expert sample dataset for coasting energy recovery control.

[0026] In the data acquisition phase, this embodiment preferably collects a large number of driving logs under real road conditions. A sufficient number of drivers can be selected (e.g., more than 100), including drivers with different driving styles, such as mild-mannered, average, and aggressive drivers; the corresponding vehicle types can include different models such as sedans and SUVs. The collected road environments cover various typical scenarios such as urban roads, suburban roads, and highways, enabling the subsequently constructed expert samples to reflect diverse traffic conditions.

[0027] The driving log includes at least a high-frequency time series of vehicle speed changes over time, such as a sampling frequency of 10 Hz, meaning the vehicle speed is recorded every 0.1 seconds. In practice, auxiliary fields such as timestamps, lane information, and road speed limits can also be recorded simultaneously. In this way, driving log data over long time scales (e.g., thousands of hours or more) can be accumulated, providing rich raw operating condition fragments for subsequent construction of virtual preceding vehicles.

[0028] Furthermore, after obtaining the raw driving logs, preprocessing and segment selection are performed on the log data to improve data quality and reduce the impact of noise. Specifically, several representative long-term driving segments can be extracted from the complete logs, such as selecting speed curve segments containing different operating conditions such as starting, acceleration, following, deceleration, and congestion, exceeding several thousand segments (e.g., more than 2000 segments). Subsequently, outlier removal, missing data interpolation, and smoothing filtering are performed on each speed time series segment to eliminate sensor jitter and occasional measurement errors.

[0029] Preferably, a smooth spline fitting method can be used to fit the velocity curve. While ensuring that the fitted curve closely matches the original observed velocity, the smoothness of the changes in its first and second derivatives is constrained, thereby obtaining a velocity function that is continuous and smooth at both the velocity and acceleration levels.

[0030] Specifically, based on the smoothed velocity function, the corresponding displacement curve is recovered through integration, and the acceleration curve is obtained through differentiation. Thus, a kinematically consistent set of virtual foreground vehicle motion trajectories is constructed, that is, the velocity, acceleration and position of the virtual foreground vehicle are given at each time point.

[0031] It should be noted that in this embodiment, the virtual preceding vehicle trajectory is reconstructed based on the speed changes of a real vehicle as the "leader" under actual operating conditions. In other words, the collected speed time series originally represents the speed of a real vehicle. Through the aforementioned smoothing and reconstruction process, this vehicle is abstracted into a virtual preceding vehicle, serving as a target vehicle that can be followed in the simulation environment. This approach can better preserve the acceleration and deceleration characteristics of real traffic flow and complex operating conditions such as congestion and slow traffic, avoiding complete reliance on manually set simple speed curves, thereby improving the representativeness and diversity of subsequent expert sample data.

[0032] Furthermore, after constructing the virtual preceding vehicle trajectory, this embodiment introduces a pre-set expert driving model in the offline simulation environment to control the vehicle to follow the virtual preceding vehicle trajectory. The expert driving model can adopt a mature longitudinal car-following model such as the Intelligent Driver Model (IDM). Its inputs include the current speed of the vehicle, the net distance between the vehicle and the preceding vehicle, and the speed difference, and the output is the desired longitudinal acceleration or deceleration.

[0033] Here, the intelligent driver model can be represented as:

[0034] in, An expression representing an intelligent driver model. This represents the maximum acceleration expected by the driver, where v is the current speed of the vehicle. The desired velocity in free flow. The acceleration exponent is usually set to 4, and s is the net distance to the vehicle in front. Desired distance. It is a dynamic function that comprehensively considers safety, comfort, and reaction time, and its expression is:

[0035] In the formula, The speed difference between the vehicle and the vehicle in front ( ), Let T be the minimum safe distance between the vehicle and the vehicle in front when the vehicle is stationary, T be the desired safe headway, a be the driver's preferred comfort acceleration, and b be the driver's preferred comfort deceleration.

[0036] By carefully calibrating the model parameters, the expert driving model exhibits safe, comfortable, and eco-friendly coasting and deceleration behaviors, such as a tendency to decelerate slowly in advance and fully utilize the coasting process to reduce mechanical braking. In subsequent simulations, the expert driving model can act as an ideal driver, providing optimal deceleration decisions for each situation.

[0037] Furthermore, to facilitate implementation and reproduction, this embodiment can build an offline simulation environment on a general numerical simulation platform, such as using Python or other simulation frameworks to implement the numerical integration of the vehicle's longitudinal motion equations. The simulation time step can be set to 0.1 s, consistent with the acquisition frequency, to facilitate alignment between the raw data and the simulation data.

[0038] Here, for each virtual preceding vehicle trajectory, the simulation starts from the initial moment and executes the following cyclical process in each time step: First, the speed and position of the virtual preceding vehicle at the current moment are read, and combined with the current speed and position of the own vehicle, the net distance and relative speed between the two vehicles are calculated; then, the above state variables are input into the expert driving model to obtain the target acceleration or deceleration that the expert believes should be applied in the current state; next, based on the obtained target acceleration, the speed and position of the own vehicle are updated forward using the longitudinal kinematic equation of the vehicle to obtain the state of the own vehicle at the next moment, which is then used for the simulation in the next time step.

[0039] Simultaneously with the aforementioned simulation loop, this embodiment systematically records data for each expert decision-making action. Specifically, within each simulation time step, the complete state vector involved in this decision-making process and the target acceleration (or deceleration) given by the expert model are stored together as a "state-action" data pair. The state vector can include multiple dimensions such as the vehicle's speed, vehicle acceleration, virtual preceding vehicle speed, net distance between the vehicle and the preceding vehicle, relative speed between the two vehicles, and current road speed limit information. Of course, in practical applications, derived features such as headway and collision time estimation can be added as needed. By traversing all reconstructed real-world driving trajectory and executing the above simulation and recording process on each trajectory, a large-scale dataset covering various road scenarios and traffic conditions can be obtained.

[0040] Preferably, in engineering implementation, the number of virtual preceding vehicle trajectory segments to be simulated and the duration of each segment can be counted first, and then the final dataset size can be estimated by comprehensively considering the simulation time step. For example, for more than 2,000 representative driving segments, under the condition of a simulation step of 0.1 s, tens of millions of state-action data pairs can be obtained after a complete simulation. After the simulation is completed, the original recorded data is organized and verified, abnormal samples that do not conform to physical constraints are removed, and the data is divided and labeled according to the needs of subsequent imitation learning tasks, such as scene, vehicle speed range, and vehicle distance range, thereby constructing a well-structured and reliable expert sample dataset.

[0041] Through the above steps, this embodiment constructs a set of virtual forward vehicle motion trajectories based on real driving logs, and uses a pre-set expert driving model to repeatedly "drive" the vehicle to follow these trajectories in a simulation environment, thereby collecting a large number of expert-level deceleration and coasting control behaviors in various complex dynamic scenarios.

[0042] S102. Using the temporal features representing the vehicle's longitudinal driving state and the relative relationship between the vehicle and the workshop as input, and using the target longitudinal acceleration or deceleration output by the expert driving model in the expert sample dataset as a supervision signal, a simulated ecological driving model for outputting the target deceleration is trained.

[0043] In specific implementation, after constructing the aforementioned expert sample dataset, an imitation ecological driving model for outputting the target deceleration is trained based on the expert sample dataset. To this end, this embodiment first performs feature processing and temporal modeling on the original "state-action" data in the expert sample dataset, constructing temporal inputs from features that characterize the vehicle's longitudinal driving state and its relative relationship with the vehicle. Then, the target longitudinal acceleration or deceleration output by the expert driving model at each moment is used as a supervision signal to establish an end-to-end imitation learning training process. This allows the trained model to directly output a target deceleration suitable for coasting energy recovery control, given the current and historical driving states.

[0044] Here, in terms of model structure, the imitation eco-driving model of the present invention can adopt a neural network structure capable of processing temporal inputs, mapping the aforementioned temporal features into a scalar form of target deceleration output.

[0045] In a preferred embodiment, the model includes an input layer, one or more Long Short-Term Memory (LSTM) layers, and several fully connected hidden layers: the input layer receives fixed-length temporal feature segments; the LSTM layers mine the temporal dependencies in the feature sequences to capture the dynamic patterns of vehicle state changes over time; the fully connected hidden layers perform nonlinear transformations and feature fusion on the temporal hidden features extracted by the LSTM; and the output layer is a single-node fully connected layer that outputs the predicted value of the target deceleration. To ensure the physical reasonableness of the output, the output layer can be combined with a limiting activation function or post-processing steps to constrain the model output within a preset deceleration range, such as not exceeding the upper limit of the maximum achievable regenerative braking deceleration and the lower limit of the comfort deceleration.

[0046] Here, the training of the simulated ecological driving model employs a composite loss function that includes imitation loss, safety loss, and comfort loss. Specifically: the imitation loss characterizes the deviation between the target deceleration output by the model and the target deceleration of experts in the expert sample dataset; the safety loss predicts the collision time at the next moment based on the current vehicle state and the target deceleration output by the model, penalizing the model output when the collision time is less than a preset safety threshold to suppress actions that tend towards dangerous following states; the comfort loss is a function that calculates the rate of change of the target deceleration between consecutive time steps, used to penalize decisions with excessively rapid deceleration changes to improve the smoothness of the coasting deceleration process; the composite loss function is a weighted sum of the above losses, with weight coefficients used to balance imitation accuracy, safety, and comfort.

[0047] For example, the composite loss function It can be represented as: +

[0048] Imitation loss This is the standard mean squared error loss, used to penalize the model output. Expert actions in the dataset The differences between them form the basis for model learning expert behavior.

[0049]

[0050] Safety loss This loss term is designed to penalize any decisions that could lead to a dangerous catapult state. In each training step, a single-step forward prediction model is used, based on the current state and the acceleration output by the model. Estimate the collision time at the next moment. .when Below the preset safety threshold At that time, the security function is activated:

[0051] Loss of comfort This loss term is used to ensure the smoothness of the model output, and is achieved by penalizing the square of the rate of change of acceleration (i.e., the impact Jerk).

[0052] Weighting coefficient , These are hyperparameters that need to be adjusted to balance the model's performance across the three objectives of accuracy, safety, and comfort.

[0053] As one possible implementation method, see Figure 2 The diagram shows a flowchart of a method for training an eco-driving model according to an embodiment of this disclosure. The method includes steps S1021 to S1023, wherein: S1021. Construct an instantaneous state vector, wherein the instantaneous state vector includes at least one of the following: vehicle speed, vehicle longitudinal acceleration, net distance between the vehicle and the vehicle in front, relative speed between the vehicle and the vehicle in front, and headway and reciprocal of collision time calculated based on the net distance and relative speed.

[0054] S1022. Stack the instantaneous state vectors of the current moment and several historical moments in chronological order to obtain a temporal feature vector for characterizing the evolution of the vehicle's longitudinal behavior, and normalize the temporal feature vector.

[0055] S1023. The time-series feature vector is used as the input of the simulated ecological driving model, and the expert target deceleration at the corresponding time in the expert sample dataset is used as the output label to supervise the training of the simulated ecological driving model.

[0056] In this specific implementation, this embodiment extracts instantaneous feature vectors from each record in the aforementioned expert sample dataset to describe the longitudinal running state and interrelationship of the vehicle and the vehicle in front at the current moment. The instantaneous feature vectors may include at least: the vehicle's current speed, the vehicle's current longitudinal acceleration, the net distance between the vehicle and the vehicle in front (i.e., the actual distance between vehicles minus the stationary safety distance), the relative speed between the vehicle and the vehicle in front (the difference between the vehicle's speed and the vehicle's speed), the vehicle in front's current speed, and current road speed limit information.

[0057] In a preferred embodiment, derived features calculated based on the above-mentioned original quantities can be further introduced, such as the headway calculated based on the net distance and the vehicle speed, the collision time or the reciprocal of the collision time estimated based on the net distance and relative speed, the road gradient level ahead, lane type markings, etc., to more comprehensively depict the risk level and driving environment in the following scenario.

[0058] Here, in order to enable the simulated ecological driving model to utilize the temporal evolution information of driving behavior, rather than just the static state at a single moment, this embodiment adopts a temporal feature construction scheme, which stacks the instantaneous feature vectors of several consecutive moments in chronological order to form a temporal feature segment of fixed length.

[0059] Specifically, a series of moments can be selected from the sampling points within the previous few seconds, with the current moment as the center or end point. For example, under the condition of a sampling period of 0.1 s, the instantaneous feature vectors corresponding to the most recent N moments (such as N=20~50) can be selected, and these features can be arranged and concatenated in chronological order to form a high-dimensional temporal feature tensor, which can be used as the input of the simulated ecological driving model.

[0060] By using this sliding time window method, the changing trend of the vehicle's longitudinal state, the evolution of the speed difference between the front and rear vehicles, and the process of narrowing or widening the distance between vehicles are explicitly encoded into the input, enabling the model to learn the dynamic decision-making patterns of expert drivers who can anticipate and slow down in advance.

[0061] Furthermore, after the temporal features are constructed, this embodiment performs uniform normalization on all feature dimensions to improve the numerical stability of model training. Specifically, linear normalization or standardization methods based on training set statistics can be used. For example, the mean and standard deviation of each feature dimension on the training set can be calculated, and then the original feature values ​​can be converted into standardized feature values ​​with zero mean and unit variance. For discrete features with limited value ranges (such as road type, number of lanes, etc.), one-hot encoding or other methods can be used to map them into numerical vectors. At the same time, to avoid extreme outliers interfering with model training, reasonable cutoff ranges can be set for some sensitive features (such as acceleration, reciprocal of collision time, etc.), and samples that are obviously inconsistent with physical meaning or extremely rare can be pruned or removed.

[0062] Here, corresponding to the input temporal features, this embodiment uses the target longitudinal acceleration or deceleration given by the expert driving model in the expert sample dataset as a supervision signal to guide the training of the imitation ecological driving model. Specifically, for each temporal feature segment, the expert target longitudinal acceleration or deceleration corresponding to its last moment is used as the label value of that segment.

[0063] In coasting energy recovery applications, the focus is typically on deceleration control. Therefore, label preprocessing is possible: when the expert target acceleration is positive (i.e., acceleration), it can be considered a sample not involved in coasting energy recovery control and filtered out, or the label value can be truncated to zero. When the expert target acceleration is negative (i.e., deceleration), its absolute value or the value after sign transformation is used as the target deceleration label, ensuring the final trained model focuses on outputting the magnitude of coasting deceleration. Simultaneously, samples under different operating conditions can be appropriately resampled or weighted to fully reflect key scenarios such as near-collision risk and congested following during training.

[0064] Preferably, in terms of training method, this embodiment uses an expert sample dataset to train an initial imitation ecological driving model to obtain an initial policy model; the initial policy model is deployed in a closed-loop simulation environment to generate a simulated driving process, and the state sequences accessed by the initial policy model during the simulated driving process are collected; the expert driving model is called to annotate the state sequences to obtain the corresponding expert target deceleration; the annotated state-action data is aggregated into the original training dataset, and the imitation ecological driving model is retrained based on the aggregated dataset to obtain an updated policy model; the steps of generating the simulated driving process are repeated until the performance of the imitation ecological driving model on the validation set converges.

[0065] In practice, this embodiment does not rely solely on the teaching data of experts on the original sample distribution for one-time training. Instead, it allows the currently trained model to drive itself in a closed-loop simulation environment. Then, the experts drive the model to re-label the states it actually visits, and continuously add these newly labeled data to the training set, forming a closed-loop iterative process of model exploration, expert correction, data aggregation, and retraining.

[0066] Specifically, in the first stage of this embodiment, the initial imitation eco-driving model is first trained using the aforementioned constructed expert sample dataset to obtain an initial policy model. The training process can follow the input temporal feature construction method and supervised learning method described above, using each temporal feature segment and its corresponding expert target longitudinal acceleration or deceleration as training samples. Regression loss functions such as mean squared error and Huber loss are used to optimize the parameters of the imitation eco-driving model. The initial policy model obtained after training already possesses the ability to generate relatively reasonable target decelerations in typical following and coasting scenarios. However, the policies it learns are mainly limited to the state distribution appearing in the expert sample dataset. For atypical states that the model may encounter in actual operation, its decision quality remains uncertain.

[0067] Furthermore, in the second stage, this embodiment deploys the aforementioned initial strategy model into the closed-loop simulation environment established earlier to generate a simulated driving process. Specifically, in the simulation environment, the virtual preceding vehicle trajectory reconstructed from real driving logs is still used as the external stimulus, and the initial strategy model is used as the vehicle's longitudinal control strategy: at each simulation time step, the vehicle's longitudinal state and relative relationship with the vehicle at the current time and several historical times are extracted first, constructing corresponding temporal feature segments, which are then input into the initial strategy model. The initial strategy model outputs the target deceleration at the current time; then, the vehicle's speed and position are updated according to this target deceleration to achieve the vehicle following the virtual preceding vehicle. In this way, within a closed-loop control framework similar to real-world applications, the initial strategy model can generate a complete simulated driving process, and record the state sequence actually accessed by the model during this process.

[0068] It should be noted that during closed-loop operation, the model may enter state regions that are less common or even absent in the original samples of the expert driving model, such as abnormal vehicle distances or special speed combinations caused by accumulated small deviations. These states are precisely the objects that need to be corrected.

[0069] Furthermore, in the third stage, this embodiment calls the expert driving model to re-label the state sequences accessed by the initial strategy model during the simulated driving process, obtaining the expert target deceleration corresponding to each state. Specifically, for each recorded state vector or temporal feature segment, the vehicle speed, inter-vehicle clearance, relative speed, and other state quantities are input into the expert driving model, which calculates the reference target longitudinal acceleration or deceleration in the current state according to its inherent ecological driving strategy.

[0070] Subsequently, the longitudinal acceleration or deceleration of the target is used as a label value and paired with the corresponding state feature to form a new "state-action" sample. Through this labeling process, a batch of new expert teaching data covering the actual access distribution of the model can be obtained. This data reflects how an expert driver would make deceleration decisions if in these states, thus providing a basis for correcting possible biases in the initial strategy model under these states.

[0071] Furthermore, in the fourth stage, this embodiment aggregates the newly labeled state-action data into the original training dataset to form an expanded aggregated dataset. Based on this aggregated dataset, the simulated ecological driving model is retrained to obtain an updated policy model. Specifically, the original expert sample dataset and the newly acquired simulated scenario labeled data can be merged in a certain proportion, maintaining coverage of typical operating conditions while increasing the sample density in frequently accessed state regions. Regarding the training strategy, a unified loss function can be applied to both the original and new samples, or higher weights can be given to the newly labeled samples to more quickly correct decision-making biases in the new state distribution. Through retraining the aggregated dataset, the parameters of the simulated ecological driving model are updated, enabling it to better approximate the decision-making behavior of the expert driving model in a state space close to its own operating distribution.

[0072] Here, to further improve the model's stability in complex scenarios, this embodiment sets the process of generating the simulated driving process—expert annotation—data aggregation—retraining as an iterative execution. That is, after completing one training iteration based on the aggregated dataset and obtaining an updated policy model, the updated policy model is redeployed back into the closed-loop simulation environment to generate a new simulated driving process. The state sequences accessed in this new round of operation are collected, annotated by the expert driving model, and then merged again with the existing dataset for the next round of training. Through multiple iterations, the model continuously receives expert correction under its actual access distribution, and its decision-making strategy gradually stabilizes and converges towards the expert driving model.

[0073] Regarding the iteration termination condition, this embodiment can set a performance convergence criterion based on the validation set. Specifically, an independent validation dataset can be pre-defined. This validation set can include real driving condition samples not used in training and special scenario samples constructed from virtual preceding vehicle trajectories. After each round of training, the current version of the simulated ecological driving model is tested on the validation set, and its mean squared error, mean absolute error, and safety and comfort-related indicators (such as the smoothness of predicted deceleration, collision time constraint violation rate, etc.) are calculated in the target deceleration prediction. When the performance improvement on the validation set is less than a preset threshold in multiple consecutive iterations, or when the change of the loss function on the validation set tends to stabilize, the model is considered to have reached convergence on the target task, the iteration can be terminated, and the current version of the simulated ecological driving model is solidified as the final strategy model.

[0074] As one possible implementation, after the training of the simulated ecological driving model is completed, the trained floating-point precision model is converted into a general intermediate representation; a quantization algorithm is used to compress the model's weights and activation values ​​from floating-point precision to low-bit integer precision to reduce model storage space and reduce inference computation; the quantized model is deployed to the vehicle's on-board electronic control unit, which periodically receives the vehicle status and outputs the target deceleration under preset real-time constraints.

[0075] S103. When the vehicle is detected to be in a condition that meets the conditions for coasting energy recovery, the current vehicle state is collected and the corresponding time-series features are constructed and input into the simulated ecological driving model to obtain the target deceleration. The target deceleration is converted into the energy recovery control quantity of the drive motor to control the vehicle to recover energy during coasting.

[0076] In practice, once the vehicle enters actual driving conditions, the onboard control system, based on the simulated eco-driving model trained and deployed in the onboard controller, performs real-time control of the energy recovery process during coasting. Specifically, during the vehicle control cycle, the vehicle controller (VCU) or brake energy recovery control unit continuously monitors the vehicle's current operating status. When a condition that meets the preset coasting energy recovery conditions is detected, a timing feature is constructed according to rules consistent with the model training phase. This timing feature is then input into the simulated eco-driving model to obtain the target deceleration at the current moment. Subsequently, this target deceleration is transformed into an energy recovery control quantity for the drive motor (e.g., target regenerative braking torque or target regenerative braking current) via vehicle dynamics and sent to the motor controller (MCU) to control the vehicle to achieve energy recovery during coasting.

[0077] Specifically, this embodiment determines whether the coasting energy recovery conditions are met by judging the current operating conditions of the vehicle. Preferably, the coasting energy recovery conditions include at least one or more of the following constraints: the driver's accelerator pedal opening is less than a first threshold or close to zero, indicating that the driver has actively released the accelerator pedal; the brake pedal is not depressed or the brake pedal opening is lower than a second threshold, to avoid conflict with manual brake pedal control; the vehicle's current speed is within a preset speed range, for example, higher than a low-speed threshold and lower than a high-speed threshold, to ensure that the regenerative braking effect is obvious and does not affect the comfort of low-speed crawling.

[0078] Here, after determining that the vehicle has entered the coasting energy recovery mode, the vehicle controller collects the current vehicle state information in each control cycle and constructs temporal feature inputs consistent with those in the model training phase. Specifically, the controller obtains instantaneous state quantities such as the vehicle's current speed, current longitudinal acceleration, distance or net distance between the vehicle and the vehicle in front, relative speed between the vehicle and the vehicle in front, current speed of the vehicle in front, road speed limit information, and road longitudinal slope information from the vehicle speed sensor, motor controller, radar / camera perception module, and navigation system. It can further calculate derived features such as headway and estimated collision time or reciprocal of collision time.

[0079] Then, the onboard controller inputs the constructed temporal feature fragments into the locally deployed simulated ecological driving model. Within each control cycle, the model outputs the target deceleration value for the current moment based on the input temporal features. To ensure physical plausibility and compatibility with the vehicle's hardware capabilities, the deceleration output by the model is first limited to ensure it does not exceed the pre-set maximum regenerative braking deceleration limit. Simultaneously, a deceleration rate constraint can be set according to comfort requirements to prevent excessive jumps in target deceleration between adjacent control cycles, which could cause abrupt changes in ride comfort. The target deceleration after these constraints is used as the reference value for coasting deceleration in that control cycle.

[0080] Furthermore, after obtaining the target deceleration, this embodiment converts the deceleration measurement into an energy recovery control quantity for the drive motor to realize the specific execution of coasting energy recovery. Specifically, the vehicle controller converts the target deceleration into the corresponding total braking force requirement based on parameters such as vehicle mass, effective wheel radius, current vehicle speed, and transmission system efficiency, according to the vehicle's longitudinal dynamics.

[0081] As a possible implementation, between step S103 and step S104, the simulated ecological driving model can be evaluated and verified using the following method: offline evaluation based on a reserved test set, calculating the mean square error and mean absolute error between the model output deceleration and the expert deceleration to evaluate the simulation accuracy; running the simulated ecological driving model in a closed-loop simulation environment, constructing multiple test scenarios including emergency braking, frequent start-stop, and vehicle entry and exit, and statistically analyzing the root mean square of impact, standard deviation of deceleration, proportion of regenerative braking energy recovery, and frequency of mechanical braking activation during the driving process to evaluate the model performance from the dimensions of comfort and economy.

[0082] S104. Calculate at least one safety evaluation index based on the current vehicle state and the target deceleration. When the safety evaluation index does not meet the preset safety constraints, correct the target deceleration to a safe deceleration and use the safe deceleration as a coasting energy recovery control command.

[0083] In practical implementation, to further ensure driving safety and avoid potential risks in complex traffic environments, this invention sets up an independent safety assessment and deceleration correction module in the energy recovery control link, targeting the target deceleration output by the simulated eco-driving model under coasting conditions. This module calculates at least one safety evaluation index based on the current vehicle state information and the target deceleration output by the simulated eco-driving model, and verifies the target deceleration according to preset safety constraints. When it is determined that a certain safety evaluation index does not meet the corresponding constraints, the target deceleration is automatically corrected to a safe deceleration that meets the safety requirements, and this safe deceleration is issued as the final coasting energy recovery control command, thereby prioritizing the longitudinal safety control needs of the vehicle while ensuring the eco-coasting effect.

[0084] Specifically, within each control cycle, the onboard controller first collects current vehicle state information, including but not limited to: the vehicle's current speed, current longitudinal acceleration, net distance between the vehicle and the vehicle in front, relative speed between the vehicle and the vehicle in front, current road speed limit, road longitudinal slope, and estimated road surface adhesion coefficient; simultaneously, it acquires the target deceleration output by the simulated ecological driving model for this cycle. Based on the above state variables and target deceleration, the safety assessment module can predict the vehicle's speed changes, distance evolution, and potential collision risks over several future time steps under the target deceleration, and construct one or more safety evaluation indicators accordingly. For example, it can calculate the estimated collision time or its reciprocal based on the current net distance and relative speed when the target deceleration is applied; it can also calculate the minimum deceleration required to meet a certain safe headway requirement and compare it with the target deceleration output by the model.

[0085] In a preferred embodiment, the safety evaluation indicators include at least a collision time indicator and a longitudinal comfort-related indicator. The collision time can be estimated based on the current vehicle clearance, relative speed, and relative motion trend under the target deceleration. When there is a risk of rear-end collision, the remaining time before the longitudinal distance between the two vehicles approaches zero is calculated by extrapolating the vehicle's deceleration trajectory under the target deceleration. If this time is less than a preset safety threshold (e.g., below a certain number of seconds), the current target deceleration scheme is deemed to have a potential collision risk. The longitudinal comfort indicator can be calculated based on the absolute value of the target deceleration and the rate of change of deceleration. When the target deceleration is too large or changes too rapidly, although it may meet the hard safety constraints in terms of collision time, it will lead to a deterioration in the riding experience and even affect vehicle stability; therefore, corresponding soft constraint control is required. Furthermore, road speed limit information can be combined with the target deceleration for judgment. When the vehicle speed is significantly higher than the speed limit and the distance to the vehicle ahead is small, but the target deceleration is too small or not aggressive enough, it can also be considered that the safety evaluation indicators do not meet the requirements.

[0086] Furthermore, after forming the safety evaluation indicators, this embodiment judges these indicators according to preset safety constraints. When all safety evaluation indicators are within the safe range, it indicates that the target deceleration output by the simulated eco-driving model satisfies both longitudinal safety and comfort and energy recovery requirements under the current operating conditions. In this case, there is no need to adjust the target deceleration, and it can be directly used for subsequent braking force distribution and motor energy recovery control. If one or more safety evaluation indicators do not meet the corresponding safety constraints, it is determined that the current target deceleration has safety or comfort risks, and a new safe deceleration needs to be obtained through a correction mechanism.

[0087] As one possible implementation, the collision time is calculated based on the current vehicle speed, vehicle spacing, and relative speed. When the collision time is lower than the preset safety red line, the target deceleration output by the simulated ecological driving model is replaced with the corresponding maximum safe deceleration. It is determined whether the target deceleration exceeds the vehicle's physical allowable range. When the target deceleration exceeds the limit, the target deceleration is limited to the allowable range. The inference delay and output validity of the simulated ecological driving model are monitored. When inference timeout or abnormal output is detected, the control mode is switched to the preset fixed deceleration mode or the traditional coasting energy recovery strategy.

[0088] Here, after the correction is completed, the obtained safe deceleration is used as the coasting energy recovery control command to participate in the subsequent braking force distribution and energy recovery control process. That is, the safe deceleration is used to replace the original target deceleration for total braking force calculation and regenerative braking / mechanical braking distribution, and the motor energy recovery torque or power command is determined accordingly.

[0089] It should be noted that after the simulated ecological driving model is deployed in the vehicle, the driving status of the vehicle during actual operation, the target deceleration output by the simulated ecological driving model, and the actual deceleration and energy recovery data executed by the vehicle control system are transmitted back to the cloud data platform. In the cloud data platform, scenarios with large deviations between the simulated ecological driving model's decisions and human driving decisions or safety objectives are identified, and the data corresponding to these scenarios are used to construct an incremental training dataset. Based on the incremental training dataset, the simulated ecological driving model is retrained or fine-tuned, and after passing the cloud regression test, the updated simulated ecological driving model is delivered to the vehicle via over-the-air upgrade.

[0090] Specifically, after the simulated eco-driving model is deployed in the vehicle's electronic control unit (such as the vehicle controller VCU or a dedicated coasting energy recovery control unit), key state and decision information related to coasting control is recorded during each control cycle of vehicle operation. This information includes at least the following categories: First, actual vehicle driving state data, such as vehicle speed, vehicle longitudinal acceleration, brake pedal and accelerator pedal opening, power battery state of charge (SOC), battery temperature, road speed limit, road longitudinal slope information, distance and relative speed between the vehicle and the vehicle in front, etc.; Second, the target deceleration output by the simulated eco-driving model at that moment, and the safe deceleration (if any) after correction by the safety assessment and constraint module; Third, the deceleration ultimately actually executed by the vehicle control system, such as the actual deceleration that can be calculated by wheel speed sensors or inertial measurement units, and the corresponding regenerative braking torque, regenerative electric power, and energy recovery per unit time.

[0091] In the cloud data platform, this invention uses big data analysis and scenario mining modules to systematically clean, classify, and cluster the returned vehicle operation data, identifying model deviation scenarios. On one hand, it compares the target deceleration output by the simulated ecological driving model with the actual deceleration. When there is a persistent deviation, it indicates a difference between the current vehicle control system or the driver's actual operation and the model's expectations. On the other hand, under certain conditions, it compares the target deceleration output by the model with human driving decisions (such as the deceleration curve formed when the driver presses the brake pedal) or preset safety targets (such as maintaining the minimum safe headway, meeting collision time thresholds, etc.). When the deviation exceeds the preset threshold, the condition is marked as a "high deviation scenario".

[0092] After identifying the aforementioned high-deviation scenarios, this invention extracts the corresponding time-series data to construct an incremental training dataset. Specifically, for each selected scenario segment, the cloud replays the vehicle's state evolution process within that scenario and re-invokes the expert driving model or preset safety control strategy to re-label the vehicle's state at the corresponding moment, generating a reference target deceleration sequence. Based on this incremental training dataset, the cloud retrains or fine-tunes the simulated ecological driving model to achieve online optimization of model parameters. In specific implementations, a fine-tuning strategy can be adopted that maintains the original model structure unchanged and only makes minor adjustments to the model parameters to avoid disrupting the learned good control behavior; alternatively, if certain input features or structural design deficiencies are found, the network structure can be appropriately adjusted before retraining.

[0093] After completing training and verification in the cloud, this invention distributes the updated simulated ecological driving model to the vehicle via over-the-air (OTA) updates. Specifically, the cloud-based model management module assigns a unique version number and metadata to each verified model version. Subsequently, the model is packaged into a deployable model file or software update package through the vehicle-to-everything (V2X) backend and pushed to the target vehicle via a secure communication channel. Upon receiving the update package, the vehicle completes the model update according to the preset upgrade process, either with the engine off or in a safe state, and retains the old version model locally as a rollback option to prevent functional abnormalities in extreme situations.

[0094] This disclosure provides a vehicle coasting energy recovery control method that, by mining and learning the longitudinal control behavior of expert drivers in multiple scenarios, achieves adaptive decision-making and safety constraints on coasting deceleration. While improving the energy recovery level under coasting conditions, it also improves the smoothness of the deceleration process, reduces the dependence on mechanical braking and calibration complexity, thereby comprehensively improving the energy utilization efficiency and driving experience of new energy vehicles in real traffic environments.

[0095] The above scheme will now be described in conjunction with specific implementation methods.

[0096] First, driving logs totaling 5000 hours were collected from over 100 drivers with different driving styles, under mixed driving conditions in Chinese cities, suburbs, and highways. Time series data of vehicle speeds (sampling frequency 10Hz) were extracted from these logs, and over 2000 representative long-term driving segments were extracted. For each raw speed data segment, data preprocessing was performed to obtain a set of fully kinematically consistent and noise-free virtual preceding vehicle trajectories.

[0097] To simulate a smooth, coasting-skilled expert driver and perform parameter calibration and adaptation, an offline simulation environment based on Python was built, with a simulation step size Δt set to 0.1 s. The aforementioned 2000+ virtual preceding vehicle trajectories were used as stimuli, allowing the calibrated IDM expert model to perform closed-loop following. During the simulation, state-action data pairs were fully recorded for each effective deceleration decision, ultimately generating a high-quality expert dataset containing approximately 20 million data points.

[0098] Instantaneous state features are extracted from the dataset. The 6-dimensional feature vectors of the current time step and the past four time steps (a total of five steps) are stacked to form a 30-dimensional input vector, which serves as the input to the model. The output of the model is the target deceleration. The model is built using the PyTorch framework. The model input is a tensor of shape (BatchSize, 5, 6), which passes through the following layers in sequence: an LSTM layer with 128 hidden units; a 256-node ReLU-activated fully connected layer; a 128-node ReLU-activated fully connected layer; and finally, a single-node Tanh-activated output layer. The output is then linearly mapped to the physical range of [-3, 0].

[0099] The DAgger algorithm was used for 5 rounds of iterative training. In each round, the trained model was run in a simulation environment across 1000 scenarios, collecting approximately 2 million new state data points. These data points were then labeled by IDM experts and incorporated into the main training set. The AdamW optimizer was used with an initial learning rate of 0.001, employing a cosine annealing strategy for decay. The batch size was 512. During training, the validation set loss was monitored; if it did not decrease for three consecutive epochs, an early stopping mechanism was triggered to prevent overfitting.

[0100] The finally trained model was evaluated through closed-loop simulation in an independent test environment containing 200 new and challenging scenarios. The evaluation focused on economy, comfort, and energy recovery. The trained PyTorch model was first converted to the common ONNX format, and then optimized and quantized using the NVIDIA TensorRT toolchain for the target automotive computing platform. The optimized model achieved an average inference latency of less than 5ms on the target hardware, fully meeting the real-time requirements of vehicle control. This model was encapsulated as a standalone software module and integrated into the vehicle's domain controller along with a parallel safety monitoring module with a TTC hard-coded safety threshold of 1.5 seconds.

[0101] Those skilled in the art will understand that, in the above-described method of the specific implementation, the order in which each step is written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined by its function and possible internal logic.

[0102] Based on the same inventive concept, this disclosure also provides a vehicle coasting energy recovery control device corresponding to the vehicle coasting energy recovery control method. Since the principle of the device in this disclosure for solving the problem is similar to the above-mentioned vehicle coasting energy recovery control method in this disclosure, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be described again.

[0103] Please see Figure 3 , Figure 3 This is a schematic diagram of a vehicle coasting energy recovery control device provided as an embodiment of this disclosure. Figure 3 As shown in the figure, the vehicle coasting energy recovery control device 300 provided in this embodiment includes: The expert sample construction module 310 is used to collect driving logs of the vehicle under various road conditions and driving styles, construct the trajectory of the preceding vehicle based on the driving logs, and follow the trajectory of the preceding vehicle in a simulation environment using a pre-set expert driving model, record the vehicle status and expert decision actions, and generate an expert sample dataset for coasting energy recovery control.

[0104] The simulated ecological driving model training module 320 is used to take the temporal features representing the longitudinal driving state of the vehicle and the relative relationship between the vehicle and the workshop as input, and the target longitudinal acceleration or deceleration output by the expert driving model in the expert sample dataset as a supervision signal to train an simulated ecological driving model for outputting the target deceleration.

[0105] The coasting energy recovery control module 330 is used to collect the current vehicle state and construct the corresponding time-series features when the vehicle is detected to be in a condition that meets the coasting energy recovery conditions. The input is then fed into the simulated ecological driving model to obtain the target deceleration. The target deceleration is then converted into the energy recovery control quantity of the drive motor to control the vehicle to recover energy during coasting.

[0106] The safety assessment module 340 is used to calculate at least one safety evaluation index based on the current vehicle state and the target deceleration. When the safety evaluation index does not meet the preset safety constraints, the target deceleration is corrected to a safe deceleration, and the safe deceleration is used as a coasting energy recovery control command.

[0107] The processing flow of each module in the device and the interaction flow between each module can be referred to the relevant descriptions in the above method embodiments, and will not be detailed here.

[0108] This disclosure provides a vehicle coasting energy recovery control device that, by mining and learning the longitudinal control behavior of expert drivers in multiple scenarios, achieves adaptive decision-making and safety constraints on coasting deceleration. While improving the energy recovery level under coasting conditions, it also improves the smoothness of the deceleration process, reduces the dependence on mechanical braking and calibration complexity, thereby comprehensively improving the energy utilization efficiency and driving experience of new energy vehicles in real traffic environments.

[0109] Corresponding to Figure 1 In the vehicle coasting energy recovery control method, this disclosure also provides an electronic device 400, such as... Figure 4 The diagram shown is a structural schematic of an electronic device 400 provided in an embodiment of this disclosure, including: Processor 41, memory 42, and bus 43; memory 42 is used to store execution instructions, including main memory 421 and external memory 422; main memory 421, also called internal memory, is used to temporarily store the computational data in processor 41, as well as the data exchanged with external memory 422 such as hard disk. Processor 41 exchanges data with external memory 422 through main memory 421. When electronic device 400 is running, processor 41 and memory 42 communicate through bus 43, enabling processor 41 to execute... Figure 1 The steps of the vehicle coasting energy recovery control method.

[0110] This disclosure also provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of the vehicle coasting energy recovery control method described in the above-described method embodiments. The storage medium may be a volatile or non-volatile computer-readable storage medium.

[0111] This disclosure also provides a computer program product, which includes computer instructions. When the computer instructions are executed by a processor, they can perform the steps of the vehicle coasting energy recovery control method in the above method embodiments. For details, please refer to the above method embodiments, which will not be repeated here.

[0112] Finally, it should be noted that the above embodiments are merely specific implementations of this disclosure, used to illustrate the technical solutions of this disclosure, and not to limit it. The protection scope of this disclosure is not limited thereto. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this disclosure. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure, and should all be covered within the protection scope of this disclosure. Therefore, the protection scope of this disclosure should be determined by the protection scope of the claims.

Claims

1. A method for controlling energy recovery during vehicle coasting, characterized in that, include: The system collects driving logs of vehicles under various road conditions and driving styles, constructs the trajectory of the preceding vehicle based on the driving logs, and uses a pre-set expert driving model to follow the trajectory of the preceding vehicle in a simulation environment, recording the vehicle status and expert decision actions, and generating an expert sample dataset for coasting energy recovery control. Using the temporal features characterizing the vehicle's longitudinal driving state and the relative relationship between the vehicle and the workshop as input, and using the target longitudinal acceleration or deceleration output by the expert driving model in the expert sample dataset as a supervision signal, an imitation ecological driving model for outputting the target deceleration is trained. When the vehicle is detected to be in a condition that meets the conditions for coasting energy recovery, the current vehicle state is collected and the corresponding time-series features are constructed and input into the simulated ecological driving model to obtain the target deceleration. The target deceleration is then converted into the energy recovery control quantity of the drive motor to control the vehicle to recover energy during coasting. Based on the current vehicle state and the target deceleration, at least one safety evaluation index is calculated. When the safety evaluation index does not meet the preset safety constraints, the target deceleration is corrected to a safe deceleration, and the safe deceleration is used as a coasting energy recovery control command.

2. The method according to claim 1, characterized in that, Specifically, an expert sample dataset for coasting energy recovery control is generated based on the following steps: The driving logs are collected from drivers of different vehicle types and driving styles in various road scenarios such as urban roads, suburban roads and highways. The driving logs include at least a time series of vehicle speeds collected at a preset sampling frequency. The vehicle speed time series is smoothed and its trajectory reconstructed based on preset kinematic constraints to obtain a set of virtual forward vehicle motion trajectories that meet the requirements of speed and acceleration continuity. In the offline simulation environment, the virtual preceding vehicle's trajectory is used as the excitation input to control the expert driving model to drive the vehicle to follow the simulation. The state update and expert action calculation are performed cyclically at a fixed time step. The state vector at each moment and the corresponding expert deceleration or acceleration are recorded as data pairs to construct the expert sample dataset. The expert driving model is an intelligent driver model. The intelligent driver model calculates the expected longitudinal acceleration based on the vehicle's expected speed, the net distance between the vehicle and the vehicle in front, and the speed difference. The net distance is determined by the difference between the current distance between the vehicles and the preset minimum safe distance when the vehicle is stationary. The intelligent driver model achieves comprehensive constraints on safety and comfort through a distance function that includes the expected headway, comfort acceleration, and comfort deceleration.

3. The method according to claim 1, characterized in that, Specifically, the following steps are used to train an eco-driving model that mimics the target deceleration: Construct an instantaneous state vector, which includes at least one of the following: vehicle speed, vehicle longitudinal acceleration, net distance between the vehicle and the vehicle in front, relative speed between the vehicle and the vehicle in front, and headway and reciprocal of collision time calculated based on the net distance and relative speed. The instantaneous state vectors of the current moment and several historical moments are stacked in chronological order to obtain a temporal feature vector that characterizes the evolution of the vehicle's longitudinal behavior, and the temporal feature vector is then normalized. The temporal feature vector is used as the input to the simulated eco-driving model, and the expert target deceleration at the corresponding time in the expert sample dataset is used as the output label to supervise the training of the simulated eco-driving model.

4. The method according to claim 1, characterized in that, The training of the simulated eco-driving model employs a composite loss function that includes simulation loss, safety loss, and comfort loss, where: The imitation loss is used to characterize the deviation between the target deceleration output by the model and the expert target deceleration in the expert sample dataset; The safety loss is based on the current vehicle state and the target deceleration output by the model to predict the collision time at the next moment. When the collision time is less than a preset safety threshold, the model output is penalized to suppress actions that tend to be dangerous following. The comfort loss is calculated as a function of the rate of change of the target deceleration between consecutive time steps, and is used to penalize decisions where the deceleration changes too quickly, so as to improve the smoothness of the coasting deceleration process. The composite loss function is a weighted sum of the above losses, with weighting coefficients used to balance simulation accuracy, safety, and comfort.

5. The method according to claim 1, characterized in that, Specifically, the following steps are used to train an eco-driving model that mimics the target deceleration: The initial imitation eco-driving model was trained using the expert sample dataset to obtain the initial policy model; The initial strategy model is deployed in a closed-loop simulation environment to generate a simulated driving process, and the state sequence accessed by the initial strategy model in the simulated driving process is collected. The expert driving model is invoked to label the state sequence, and the corresponding expert target deceleration is obtained; The labeled state-action data is aggregated into the original training dataset, and the simulated ecological driving model is retrained based on the aggregated dataset to obtain an updated policy model. Repeat the steps of generating the simulated driving process until the performance of the simulated ecological driving model converges on the validation set.

6. The method according to claim 1, characterized in that, Based on the current vehicle state and the target deceleration, at least one safety evaluation index is calculated. When the safety evaluation index does not meet the preset safety constraints, the target deceleration is corrected to a safe deceleration, specifically including: The collision time is calculated based on the current vehicle speed, vehicle spacing and relative speed. When the collision time is lower than the preset safety red line, the target deceleration output by the simulated ecological driving model is replaced with the corresponding maximum safe deceleration. Determine whether the target deceleration exceeds the vehicle's physical allowable range. If the target deceleration exceeds the limit, limit the target deceleration to the allowable range. The inference latency and output validity of the simulated ecological driving model are monitored. When an inference timeout or abnormal output is detected, the control mode is switched to a preset fixed deceleration mode or a traditional coasting energy recovery strategy.

7. The method according to claim 1, characterized in that, After deploying the simulated eco-driving model in the vehicle, the method further includes: The vehicle's actual driving status during operation, the target deceleration output by the simulated ecological driving model, and the actual deceleration and energy recovery data executed by the vehicle control system are transmitted back to the cloud data platform. In the cloud data platform, scenarios with significant deviations between the decision-making of the simulated ecological driving model and the decision-making or safety objectives of human driving are identified, and the data corresponding to the scenarios are used to construct an incremental training dataset. The simulated eco-driving model is retrained or fine-tuned based on the incremental training dataset, and after passing the cloud regression test, the updated simulated eco-driving model is delivered to the vehicle via over-the-air update.

8. A vehicle coasting energy recovery control device, characterized in that, include: The expert sample construction module is used to collect driving logs of vehicles under various road conditions and driving styles, construct the trajectory of the preceding vehicle based on the driving logs, and follow the trajectory of the preceding vehicle in a simulation environment using a pre-set expert driving model, record the vehicle status and expert decision actions, and generate an expert sample dataset for coasting energy recovery control. The simulated ecological driving model training module is used to take the temporal features representing the vehicle's longitudinal driving state and the relative relationship between the vehicle and the workshop as input, and the target longitudinal acceleration or deceleration output by the expert driving model in the expert sample dataset as a supervision signal to train an simulated ecological driving model for outputting the target deceleration. The coasting energy recovery control module is used to collect the current vehicle state and construct the corresponding time-series features when the vehicle is detected to be in a condition that meets the coasting energy recovery conditions. The input is then fed into the simulated ecological driving model to obtain the target deceleration. The target deceleration is then converted into the energy recovery control quantity of the drive motor to control the vehicle to recover energy during coasting. The safety assessment module is used to calculate at least one safety evaluation index based on the current vehicle state and the target deceleration. When the safety evaluation index does not meet the preset safety constraints, the target deceleration is corrected to a safe deceleration, and the safe deceleration is used as a coasting energy recovery control command.

9. An electronic device, characterized in that, include: The device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the steps of the vehicle coasting energy recovery control method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the vehicle coasting energy recovery control method as described in any one of claims 1 to 7.

Images