Vehicle-Road-Cloud Collaborative Energy-Saving Speed ​​Planning Method, Device, Equipment and Storage Medium

Abstract

This application provides a vehicle-road-cloud collaborative energy-saving speed planning method, device, equipment, and storage medium, relating to the field of intelligent connected vehicle technology. The method includes: acquiring map data of a target path and adaptively segmenting the target path based on terrain elevation information to generate multiple road units; employing a preset prompt word construction strategy, constructing structured prompt words based on road features, vehicle parameters, and preset energy-saving optimization targets of the road units, inputting them into a preset large language model, inferring and outputting an initial speed sequence, smoothing the result to obtain the target speed sequence, and sending it to the vehicle terminal. This method constructs structured prompt words through terrain adaptive segmentation and prompt word engineering, and infers the energy-saving speed sequence using a lightweight large language model combined with a low-latency scheduling strategy, achieving a balance between intelligent energy-saving decision-making and real-time control, thereby improving the rationality, accuracy, and scenario adaptability of vehicle-road-cloud collaborative energy-saving speed planning.

Description

Technical Field

[0001] This application relates to the field of intelligent connected vehicle technology, and in particular to a vehicle-road-cloud collaborative energy-saving speed planning method, device, equipment and storage medium. Background Technology

[0002] With the development of vehicle-road-cloud collaborative technology, predictive cruise control (PCC) has become an important means of improving vehicle driving economy by acquiring maps and beyond-line-of-sight information from the cloud to plan energy-saving speeds.

[0003] Current vehicle-road-cloud collaborative energy-saving speed planning schemes typically involve first acquiring vehicle navigation paths and map data, then segmenting the map; next, inputting the segmented information into a predefined fixed rule algorithm to calculate the speed sequence based on a dynamic model, speed limit constraints, and gradient parameters; and finally, sending the data to the vehicle terminal for execution via the cloud.

[0004] However, traditional map segmentation methods only divide the map evenly at fixed distances, completely ignoring core energy-saving factors such as terrain elevation and slope changes. This results in the loss of key road features, making it impossible to provide accurate input for energy-saving algorithms. At the same time, existing hard-coded algorithms based on fixed rules lack the ability to learn autonomously and optimize dynamically. When faced with different vehicle types, loads, and complex road conditions, the strategies become rigid, making it difficult to achieve optimal control. Therefore, the current vehicle-road-cloud collaborative energy-saving speed planning scheme has poor planning rationality, low accuracy, and weak scenario adaptability. Summary of the Invention

[0005] The main purpose of this application is to propose a vehicle-road-cloud collaborative energy-saving speed planning method, device, equipment and storage medium, which aims to improve the rationality, accuracy and scenario adaptability of vehicle-road-cloud collaborative energy-saving speed planning.

[0006] In a first aspect, the present invention provides a vehicle-road-cloud collaborative energy-saving speed planning method, comprising: Obtain map data of the target path, and adaptively segment the target path based on the terrain elevation information in the map data to generate multiple road units; A pre-defined prompt word construction strategy is adopted to construct structured prompt words based on the road characteristics, vehicle parameters, and pre-defined energy-saving optimization targets of the multiple road units; The structured prompt words are input into a preset large language model, and inference is performed through a preset low-latency scheduling strategy to obtain the initial velocity sequence output by the preset large language model. The initial velocity sequence is smoothed to obtain the target velocity sequence, and the target velocity sequence is sent to the vehicle terminal.

[0007] In an optional implementation, the step of adaptively segmenting the target path based on the terrain elevation information in the map data to generate multiple road units includes: Extract the elevation data of each path point on the target path and calculate the elevation difference between adjacent path points; Calculate the rate of change of terrain elevation between adjacent path points based on the elevation difference and the road segment length between adjacent path points; If the terrain elevation change rate is greater than a preset change rate threshold, then segmentation is performed between the corresponding adjacent path points to generate the road unit with uniform terrain features.

[0008] In an optional implementation, the preset prompt word construction strategy includes: preset identity definition, terrain segmentation information, vehicle parameters, energy consumption model, preset energy-saving optimization target, driving constraints, and output format specifications; the road features include: the unit length and average slope information of each road unit; the driving constraints include: driving speed constraints and driving acceleration constraints; Before constructing structured prompts based on the road characteristics, vehicle parameters, and preset energy-saving optimization targets of the multiple road units using a preset prompt word construction strategy, the method further includes: The terrain segmentation information is obtained based on the unit length and the average slope information; Based on the energy consumption model and the global total energy consumption minimization objective, the preset energy-saving optimization objective is determined; The method employs a preset prompt word construction strategy, which constructs structured prompt words based on the road characteristics, vehicle parameters, and preset energy-saving optimization targets of the multiple road units, including: The preset identity definition, the terrain segmentation information, the vehicle parameters, the driving speed constraint, the driving acceleration constraint, the energy consumption model, the preset energy-saving optimization target, and the output format specification are combined into the structured prompt words.

[0009] In an optional implementation, the preset large language model is a lightweight large language model; the preset low-latency scheduling strategy includes: an edge cache pre-computation strategy, a model lightweighting strategy, and a post-processing localization strategy. The step of inputting the structured prompt words into a preset large language model, performing inference through a preset low-latency scheduling strategy, and obtaining the initial velocity sequence output by the preset large language model includes: The structured prompt words are input into a preset large language model, and according to the edge cache pre-computation strategy, a speed sequence matching the structured prompt words is queried in the local cache. If it exists, the cached result is used as the initial velocity sequence; If it does not exist, then according to the model lightweighting strategy, the lightweight large language model is invoked to infer the structured prompt words and obtain the initial speed sequence. Wherein, the total delay for obtaining the initial velocity sequence is less than a preset delay threshold.

[0010] In an optional implementation, before inputting the structured prompt words into a preset large language model, the method further includes: Obtain historical vehicle driving data, which includes: terrain elevation information of historical routes, historical vehicle parameters, and corresponding historical best speed sequences; Extract sample path features, sample vehicle parameters, and corresponding optimal speed sequences from the vehicle's historical driving data; A training dataset is constructed based on the sample path features, the sample vehicle parameters, and the corresponding optimal speed sequence of the samples. According to the preset parameter freezing rules, some parameters of the initial large language model are frozen to obtain a partially frozen large language model; Based on the training dataset, the low-rank matrix of the attention layer in the partially frozen large language model is subjected to low-rank adaptive fine-tuning training to obtain the lightweight large language model.

[0011] In an optional implementation, smoothing the initial velocity sequence to obtain the target velocity sequence includes: The initial velocity sequence is smoothed using a first-order inertial filtering algorithm to eliminate abrupt velocity changes, thereby obtaining the target velocity sequence.

[0012] In an optional implementation, after sending the target speed sequence to the vehicle-mounted terminal for execution, the method further includes: Obtain the real-time vehicle driving status data and energy consumption data uploaded by the vehicle terminal; Based on the actual vehicle driving status data and energy consumption data, the preset prompt word construction strategy and / or the preset large language model are optimized and updated.

[0013] Secondly, the present invention provides a vehicle-road-cloud collaborative energy-saving speed planning device, comprising: The acquisition module is used to acquire map data of the target path and adaptively segment the target path according to the terrain elevation information in the map data to generate multiple road units; The construction module is used to construct structured prompts based on the road characteristics, vehicle parameters, and preset energy-saving optimization targets of the multiple road units using a preset prompt word construction strategy. The reasoning module is used to input the structured prompt words into a preset large language model, perform reasoning through a preset low-latency scheduling strategy, and obtain the initial velocity sequence output by the preset large language model. The processing module is used to smooth the initial velocity sequence to obtain the target velocity sequence, and send the target velocity sequence to the vehicle terminal.

[0014] Thirdly, the present invention provides an electronic device, comprising: a processor, a storage medium, and a bus, wherein the storage medium stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor communicates with the storage medium via the bus, and the processor executes the machine-readable instructions to perform the steps of any of the methods described in the foregoing embodiments.

[0015] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the method as described in any of the foregoing embodiments.

[0016] The beneficial effects of this application are: The vehicle-road-cloud collaborative energy-saving speed planning method provided in this application includes: acquiring map data of the target path, and adaptively segmenting the target path according to the terrain elevation information in the map data to generate multiple road units; using a preset prompt word construction strategy, constructing structured prompt words based on the road features, vehicle parameters, and preset energy-saving optimization targets of the multiple road units; inputting the structured prompt words into a preset large language model, and performing inference through a preset low-latency scheduling strategy to obtain the initial speed sequence output by the preset large language model; smoothing the initial speed sequence to obtain the target speed sequence, and sending the target speed sequence to the vehicle terminal. In this embodiment, the method adaptively segments the target path according to the terrain elevation information, constructs road features, vehicle parameters, and preset energy-saving optimization targets into structured prompt words using a preset prompt word construction strategy, and then combines the preset low-latency scheduling strategy to perform inference and smoothing on the preset large language model, thereby realizing intelligent energy-saving speed planning with scene adaptability, thus improving the rationality, accuracy, and scene adaptability of vehicle-road-cloud collaborative energy-saving speed planning. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a vehicle-road-cloud collaborative energy-saving speed planning method provided in one embodiment of this application; Figure 2 A flowchart illustrating a vehicle-road-cloud collaborative energy-saving speed planning method provided in another embodiment of this application; Figure 3 A flowchart illustrating a vehicle-road-cloud collaborative energy-saving speed planning method provided in another embodiment of this application; Figure 4 A schematic diagram of the structure of a vehicle-road-cloud collaborative energy-saving speed planning device provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0020] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0021] 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.

[0022] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0023] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0024] Current vehicle-road-cloud collaborative energy-saving speed planning schemes typically involve first acquiring vehicle navigation paths and map data, then segmenting the map; next, inputting the segmented information into a predefined fixed rule algorithm to calculate the speed sequence based on a dynamic model, speed limit constraints, and gradient parameters; and finally, sending the data to the vehicle terminal for execution via the cloud.

[0025] However, traditional map segmentation methods only divide the map evenly at fixed distances, completely ignoring core energy-saving factors such as terrain elevation and slope changes. This results in the loss of key road features, making it impossible to provide accurate input for energy-saving algorithms. At the same time, existing hard-coded algorithms based on fixed rules lack the ability to learn autonomously and optimize dynamically. When faced with different vehicle types, loads, and complex road conditions, the strategies become rigid, making it difficult to achieve optimal control. Therefore, the current vehicle-road-cloud collaborative energy-saving speed planning scheme has poor planning rationality, low accuracy, and weak scenario adaptability.

[0026] To address the aforementioned issues, the main objective of this application is to propose a vehicle-road-cloud collaborative energy-saving speed planning method, aiming to improve the rationality, accuracy, and scenario adaptability of vehicle-road-cloud collaborative energy-saving speed planning.

[0027] Figure 1 This is a flowchart illustrating a vehicle-road-cloud collaborative energy-saving speed planning method according to an embodiment of this application. The executing entity of this method can be, for example, a device with computing power such as a vehicle-mounted system, but is not limited thereto. Please refer to... Figure 1 The method includes: S101. Obtain map data of the target path, and based on the terrain elevation information in the map data, adaptively segment the target path to generate multiple road units.

[0028] For example, the aforementioned target path can refer to the planned driving route of the vehicle from the starting point to the destination. The aforementioned map data can be obtained from a cloud-based map service. In addition to containing the spatial coordinate information of the target path (such as the latitude and longitude sequence of each path point), the map data may also contain the corresponding elevation data. Specifically, the aforementioned terrain elevation information can refer to the elevation values ​​of each location point on the target path, used to characterize the degree of terrain undulation of the road.

[0029] The aforementioned adaptive segmentation can refer to dynamically dividing the target path into segments based on the intensity of changes in terrain elevation information, rather than uniformly splitting it according to a fixed distance or length. By adaptively segmenting the target path, continuous road segments with similar terrain undulation characteristics can be grouped into the same road unit, ensuring that each road unit has uniform and stable slope and elevation attributes. For example, when the elevation change rate of a certain road segment exceeds a preset change rate threshold (e.g., 5%), segmentation is performed at that location to generate a new road unit, but the specific segmentation method is not limited to this.

[0030] By adaptively segmenting based on terrain elevation information, core features that are strongly related to vehicle energy consumption, such as road slope, slope length, and elevation undulation, can be fully preserved. This eliminates the problem of terrain feature loss caused by traditional fixed-distance segmentation methods from the data source, providing a precise and high-fidelity terrain input foundation for subsequent energy-saving speed planning.

[0031] S102. Using a preset prompt word construction strategy, structured prompt words are constructed based on the road characteristics, vehicle parameters, and preset energy-saving optimization targets of the above-mentioned multiple road units.

[0032] For example, the aforementioned preset prompt construction strategy may refer to a predefined set of rules for organizing domain-specific knowledge into a standard prompt format. This preset prompt construction strategy may specify, for example, which components should be included in the prompt, and the format and content requirements for each component.

[0033] The road features of the aforementioned road units may refer to terrain information related to energy-saving speed planning extracted from each road unit, such as the unit length and average slope information of each road unit. The vehicle parameters mentioned above refer to parameters related to vehicle energy consumption characteristics, such as vehicle type and load capacity. The preset energy-saving optimization target refers to the globally optimal goal to be achieved by the energy-saving speed planning, such as minimizing the total energy consumption throughout the journey, but the specific target can be adjusted and determined according to actual needs.

[0034] The aforementioned structured prompts can refer, for example, to standardized text input that can be understood and executed by a large language model, which is organized according to the aforementioned preset prompt construction strategy and includes information such as road features, vehicle parameters, preset energy-saving optimization targets, and driving constraints in a preset format.

[0035] By injecting knowledge from professional fields such as vehicle dynamics models, energy consumption calculation models, and traffic constraints into structured prompt words, a general-purpose large language model can directly possess professional energy-saving speed planning capabilities without fine-tuning training. This replaces traditional hard-coded algorithms based on fixed rules to complete intelligent decision-making, breaking through the bottlenecks of traditional algorithms' rigid logic and poor scenario generalization.

[0036] S103. Input the above structured prompt words into the preset large language model, and perform inference through the preset low-latency scheduling strategy to obtain the initial velocity sequence output by the above preset large language model.

[0037] For example, the aforementioned preset large language model may be a pre-deployed large language model with natural language understanding and reasoning capabilities. The aforementioned preset low-latency scheduling strategy may be a scheduling mechanism used to reduce the inference latency of the large language model and meet the real-time control requirements of the vehicle-road-cloud cooperative system.

[0038] Since the inference latency of directly calling the native large language model typically reaches the second level, it cannot meet the millisecond-level (e.g., ≤200ms) real-time control requirements of the vehicle-road-cloud collaborative system. Therefore, the aforementioned preset low-latency scheduling strategy is needed to optimize the inference process. This preset low-latency scheduling strategy can comprehensively utilize various methods such as edge node cache pre-computation, model lightweighting, and vehicle-side post-processing localization to control the total latency of the large language model decision within the automotive-grade requirements. For example, when the cache hits, the response latency can be controlled within 10ms, the inference latency after model lightweighting can be controlled within 100ms, the vehicle-side smoothing latency can be controlled within 50ms, and the overall total latency can be controlled within 160ms, but this is not a limit.

[0039] The aforementioned initial velocity sequence can refer, for example, to the sequence of velocity values ​​output by the aforementioned pre-defined large language model after inference based on the aforementioned structured prompts, which satisfies the global energy consumption optimization objective. The velocity values ​​in this initial velocity sequence are arranged according to the travel order along the target path.

[0040] S104. The initial velocity sequence is smoothed to obtain the target velocity sequence, and the target velocity sequence is sent to the vehicle terminal.

[0041] For example, since the initial velocity sequence directly output by the aforementioned large language model may contain abrupt changes or discontinuities between velocity values, sending this initial velocity sequence directly as a control command to the vehicle terminal may affect driving smoothness and ride comfort. However, by smoothing the initial velocity sequence, abrupt speed changes can be eliminated, resulting in a smooth transition between adjacent velocity values. This ensures that the vehicle's speed changes are continuous and shock-free during actual driving, thus optimizing the driving experience.

[0042] The aforementioned target speed sequence can be a smoothed speed sequence that can be directly executed by the vehicle controller. The aforementioned vehicle terminal can be a vehicle T-BOX (Telematics BOX) terminal or other vehicle-mounted devices with communication and computing capabilities.

[0043] After the target speed sequence is sent to the vehicle terminal, the terminal can parse it into predictive cruise control commands, driving the vehicle to travel at an energy-saving speed. Simultaneously, the vehicle can upload real-time driving status and energy consumption data to the cloud. The cloud can then use this data to optimize the prompt parameters and terrain segmentation thresholds, achieving self-evolving intelligent closed-loop control.

[0044] The vehicle-road-cloud collaborative energy-saving speed planning method provided in this application includes: acquiring map data of the target path, and adaptively segmenting the target path according to the terrain elevation information in the map data to generate multiple road units. A preset prompt word construction strategy is adopted to construct structured prompt words based on the road features, vehicle parameters, and preset energy-saving optimization goals of the multiple road units. The structured prompt words are input into a preset large language model, and inference is performed through a preset low-latency scheduling strategy to obtain the initial speed sequence output by the preset large language model. The initial speed sequence is smoothed to obtain the target speed sequence, and the target speed sequence is sent to the vehicle terminal. In this embodiment, the method adaptively segments the target path according to terrain elevation information, constructs structured prompt words from road features, vehicle parameters, and preset energy-saving optimization goals using a preset prompt word construction strategy, and then combines this with a preset low-latency scheduling strategy to perform inference and smoothing on the preset large language model. This achieves intelligent energy-saving speed planning with scene adaptability, thereby improving the rationality, accuracy, and scene adaptability of vehicle-road-cloud collaborative energy-saving speed planning.

[0045] Figure 2 A flowchart illustrating a vehicle-road-cloud collaborative energy-saving speed planning method provided in another embodiment of this application is shown below. Figure 2 As shown, optionally, in the foregoing Figure 1 Based on the previous embodiment, the target path is adaptively segmented according to the terrain elevation information in the map data to generate multiple road units, including: S201. Extract the elevation data of each path point on the target path and calculate the elevation difference between adjacent path points.

[0046] For example, each of the aforementioned path points can refer to a spatial location point sampled at a preset interval along the target path. Each path point contains corresponding latitude and longitude coordinates and elevation data. The elevation data can be directly extracted from the aforementioned map data and is used to characterize the terrain elevation at each path point location.

[0047] After extracting the elevation data of each path point, the elevation difference between each pair of adjacent path points can be calculated sequentially according to the driving direction of the target path. Specifically, assume the elevation data of the i-th path point is... , No. The elevation data of each path point are Then the elevation difference between the adjacent path points For example, it can be expressed as the following formula:

[0048] The aforementioned elevation difference can be used to characterize the degree of terrain undulation between adjacent path points, with units such as meters. This is achieved by calculating the elevation difference. This can transform the original absolute elevation value into the relative change between adjacent points, providing basic data for subsequent calculation of terrain elevation change rate.

[0049] Furthermore, if necessary, the elevation difference between the aforementioned adjacent path points can be calculated. Then, the road slope change can be further calculated using the following formula:

[0050] Among them, the above This refers to the aforementioned changes in road slope. For the first The length of a road segment.

[0051] S202. Based on the above elevation difference and the road segment length between the above adjacent path points, calculate the terrain elevation change rate between the above adjacent path points.

[0052] For example, the length of the road segment between the aforementioned adjacent path points can be calculated based on the latitude and longitude coordinates of each path point using methods such as geodetic thematic calculation, and is used to characterize the actual ground distance between adjacent path points.

[0053] Based on the aforementioned elevation differences and road segment lengths, the aforementioned terrain elevation change rate can also be calculated, for example. This terrain elevation change rate can, for example, be used to represent the intensity of terrain undulation within a unit distance, objectively reflecting whether significant terrain changes have occurred. Specifically, assuming the first... The path point and the first The elevation difference between the path points is The length of the road segment is as described above. Then the rate of change of terrain elevation of this section of road This can be expressed as the following formula:

[0054] The rate of change of terrain elevation The larger the absolute value, the more dramatic the terrain undulations of that section of road.

[0055] S203. If the above-mentioned terrain elevation change rate is greater than the preset change rate threshold, then segmentation is performed between the corresponding adjacent path points to generate the above-mentioned road units with uniform terrain features.

[0056] For example, the aforementioned preset rate of change threshold may refer to the critical value of terrain elevation change rate used to determine whether to perform segmentation. This preset rate of change threshold can be set according to actual application needs, for example, it can be set to 3%~5%, but for different vehicle types (e.g., passenger cars and heavy trucks), due to differences in vehicle dynamics characteristics and energy consumption sensitivity, this preset rate of change threshold can be adaptively adjusted.

[0057] During the segmentation process, all adjacent path point pairs on the target path can be traversed, and the relationship between the terrain elevation change rate and the preset change rate threshold can be compared one by one. If the terrain elevation change rate of an adjacent path point pair is greater than or equal to the preset change rate threshold, the location is determined to be a terrain abrupt change point, and segmentation is performed between the adjacent path points to generate a new road unit. If the terrain elevation change rate is less than the preset change rate threshold, the terrain at the location is determined to be stable, and the adjacent path point pair is merged into the current road unit.

[0058] After traversing all adjacent path pairs, a series of road units with uniform terrain features can be generated. Each road unit can contain the following information: total unit length, average unit slope, unit starting elevation, unit ending elevation, and the number of path points contained in the unit. For example, suppose the target path is generated after adaptive segmentation. The nth road unit, then the nth Road Units It can be represented in the following form:

[0059] Among the above For the first The total length of each road unit For the first The average slope of each road unit and The first The elevations of the start and end points of each road unit. For the first The number of path points contained in a road unit.

[0060] The above adaptive segmentation method can achieve a strong binding between map segmentation logic and terrain undulation intensity, rather than a strong binding with fixed distance. This eliminates the terrain feature loss problem caused by traditional fixed-distance segmentation methods from the data source, and provides a high-precision, high-matching input foundation for energy-saving speed planning of subsequent large language models.

[0061] In addition, in the aforementioned Figure 1 Based on the embodiments, the aforementioned preset prompt word construction strategy includes: preset identity definition, terrain segmentation information, vehicle parameters, energy consumption model, preset energy-saving optimization target, driving constraints, and output format specifications. The aforementioned road features include: the unit length and average slope information of each of the aforementioned road units. The aforementioned driving constraints include: driving speed constraints and driving acceleration constraints.

[0062] The above-mentioned method, which employs a preset prompt word construction strategy, further includes the following steps before constructing structured prompt words based on the road characteristics, vehicle parameters, and preset energy-saving optimization targets of the aforementioned multiple road units: Based on the above unit length and the above average slope information, the above terrain segmentation information is obtained.

[0063] Based on the above energy consumption model and the goal of minimizing global total energy consumption, the above-mentioned preset energy-saving optimization goals are determined.

[0064] For example, the above-mentioned preset prompt word construction strategy may specify the components that the structured prompt words should include. Specifically, the above-mentioned preset identity definition can be used to set a professional role for the above-mentioned preset large language model. For example, it can be set as "You are a professional vehicle-road-cloud collaborative energy-saving speed planning expert, generating the globally energy-optimal vehicle speed sequence based on input information, and strictly adhering to all constraints."

[0065] The terrain segmentation information described above can be used to describe the terrain features of each road unit after adaptive segmentation of the target path.

[0066] The vehicle parameters mentioned above can be used to describe parameters related to vehicle energy consumption characteristics, such as vehicle type and full load weight.

[0067] The energy consumption model mentioned above can refer to the vehicle energy consumption calculation model per unit mileage obtained by calibration based on actual vehicle test data.

[0068] The aforementioned preset energy-saving optimization targets can be used to represent the global optimization goals to be achieved by the energy-saving speed planning.

[0069] The above driving constraints can be used to limit the range of speed sequence values ​​and ride comfort requirements.

[0070] The above output format specification can be used to define the format of the output results of the above preset large language model. For example, it can be set to "output only speed sequence, unit km / h, comma-separated, no extra content".

[0071] However, it is understandable that the above content is just one possible example. The actual content, specific form and function of the preset prompts can be adjusted according to the actual situation and are not limited to the content in the examples above.

[0072] The terrain segmentation information obtained based on the unit length and average slope information can be described as follows: For example, it can refer to the unit length and average slope information of each road unit generated after adaptive segmentation of the target path, organized into a text description in the form of a road segment sequence according to a preset format. For instance, it can be described as "Road segment sequence: M1, M2, ..., MK, where K is the total number of terrain segmentation units. Road segment k: length Lk=XX m, average slope αk=XX °, road segment speed limit Vlimit_k=XX km / h," where XX is only for illustrative purposes and each XX can be a different value.

[0073] The above energy consumption model can be calibrated from real vehicle test data and is used to characterize the quantitative relationship between vehicle speed and energy consumption per unit mileage. For example, this energy consumption model can be expressed as the following formula: E = a·v³ + b·v² + c·v + d· + e, Where E represents the vehicle's energy consumption per unit distance, and the unit can be, for example, kWh / m. v represents the vehicle's speed, and the unit can be, for example, m / s. This refers to the road slope, specifically the aforementioned changes in road slope. a, b, c, d, and e are the calibration coefficients for the actual vehicle.

[0074] The aforementioned goal of minimizing total global energy consumption refers to minimizing the total energy consumption along the entire target path. For example, this optimization objective can be expressed as the following formula:

[0075] Among the above This represents the total energy consumption for the entire path. For the first The speed of each road unit Energy consumption per unit distance For the first The length of each road unit, as described above Total number of terrain segmentation units The above-mentioned strategy of constructing pre-defined prompts uses road features, vehicle parameters, and pre-defined energy-saving optimization targets for multiple road units to construct structured prompts, including: The aforementioned preset identity definition, terrain segmentation information, vehicle parameters, driving speed constraints, driving acceleration constraints, energy consumption model, preset energy-saving optimization target, and output format specifications are combined into the aforementioned structured prompt words.

[0076] For example, after obtaining the above-mentioned terrain segmentation information and determining the above-mentioned preset energy-saving optimization goals, the components specified in the above-mentioned preset prompt word construction strategy can be combined into a complete structured text input that can be directly understood and executed by the above-mentioned preset large language model, namely the above-mentioned structured prompt words.

[0077] For example, the constructed structured prompts can include the following structure: (1) Preset identity definition paragraph, used to clarify the model's role and task. (2) Terrain segmentation information paragraph, used to describe the terrain features of each road unit of the target path. (3) Vehicle parameter paragraph, used to provide basic parameters related to vehicle energy consumption. (4) Energy consumption model paragraph, used to embed the unit mileage energy consumption calculation formula calibrated by the actual vehicle. (5) Optimization target paragraph, used to clarify the global total energy consumption minimization target. (6) Driving constraint condition paragraph, used to limit driving speed constraints and driving acceleration constraints, etc. (7) Output format specification paragraph, used to specify the format requirements of the output speed sequence.

[0078] By injecting knowledge from professional fields such as vehicle dynamics models, energy consumption calculation models, and traffic constraints into the aforementioned structured prompts in a standardized structure, the pre-defined large language model can accurately understand the physical logic and optimization objectives of energy-saving speed planning without fine-tuning training. It can directly output the globally optimal energy-saving speed sequence that conforms to vehicle dynamics characteristics, terrain features, and traffic rules, replacing the traditional hard-coded algorithm based on fixed rules to complete intelligent decision-making, thus giving the algorithm high flexibility and scenario generalization ability.

[0079] Optionally, in the foregoing Figure 1 Based on the embodiments, the aforementioned preset large language model is a lightweight large language model. The aforementioned preset low-latency scheduling strategy includes: an edge caching pre-computation strategy, a model lightweighting strategy, and a post-processing localization strategy.

[0080] The aforementioned structured prompts are input into a preset large language model, and inference is performed using a preset low-latency scheduling strategy to obtain the initial velocity sequence output by the preset large language model, including: The structured prompts are input into a preset large language model. Based on the edge caching pre-computation strategy, the system queries the local cache to see if there is a velocity sequence that matches the structured prompts.

[0081] For example, the aforementioned lightweight large language model can refer to a model obtained by compressing a general large language model, which has a faster inference speed and lower computational consumption. This lightweight large language model significantly reduces inference latency while ensuring controllable accuracy loss, making it better suited to the real-time requirements of vehicle-road-cloud collaborative systems.

[0082] The aforementioned edge caching pre-computation strategy could refer to a caching mechanism that pre-stores energy-saving speed sequences corresponding to high-frequency driving routes and fixed routes at the system's edge nodes. Since there are numerous repetitive driving route requests (such as commuter routes and fixed bus routes) in the vehicle-road-cloud cooperative system, by pre-compiling and caching the inference results corresponding to these requests at the edge nodes, the cached results can be directly returned when the same or similar structured prompts are received again, thus avoiding repeated calls to the large language model for inference and significantly reducing response latency. When querying the local cache, key fields such as terrain segmentation information and vehicle parameters in the current structured prompts can be matched with the corresponding fields in the cache.

[0083] If it exists, the cached result will be used as the initial velocity sequence mentioned above.

[0084] For example, when a velocity sequence matching the structured hints mentioned above exists in the local cache, this velocity sequence can be directly read from the cache as the initial velocity sequence. In this case, the system's response latency depends only on the cache lookup latency. For instance, when a cache hit occurs at an edge node, the cache lookup latency can be controlled to within 10ms. The cache hit rate can be calculated using the following formula:

[0085] Among them, the above This is the cache hit rate mentioned above. To cache the hit count, This represents the total number of requests. Through the aforementioned edge caching pre-computation strategy, the response latency of high-frequency requests can be compressed to an extremely low level, thus reserving sufficient latency margin for the smooth processing of subsequent post-processing localization strategies.

[0086] If it does not exist, then according to the above lightweight model strategy, the above lightweight large language model is called to infer the above structured prompt words and obtain the above initial speed sequence.

[0087] The total time delay for obtaining the initial velocity sequence is less than a preset time delay threshold.

[0088] For example, if no velocity sequence matching the structured prompt words exists in the local cache, the lightweight large language model needs to be invoked to perform real-time inference on the structured prompt words. The lightweight model strategy can comprehensively utilize model distillation and quantization techniques to compress the original general-purpose large language model, significantly reducing inference latency while ensuring controllable accuracy loss (e.g., less than 1%). Specifically, the inference latency after model lightweighting can be controlled to within 100ms, but is not limited to this.

[0089] The total latency for obtaining the initial velocity sequence refers to the end-to-end latency from the input of the structured prompt words to the acquisition of the initial velocity sequence. This total latency can be decomposed into three parts: cache lookup latency, model inference latency, and subsequent smoothing processing latency. For example, it can be expressed as the following formula:

[0090] Among them, the above For the total system latency, To cache query latency, The inference latency of the aforementioned lightweight large language model, This is to smooth out the latency at the vehicle-mounted end. The preset latency threshold refers to the upper limit of the latency required for real-time control by the vehicle-road-cloud cooperative system, for example, it can be set to 200ms.

[0091] Through the three-level collaborative scheduling of the above edge caching pre-computation strategy, the above model lightweighting strategy, and the above post-processing localization strategy, the total latency of obtaining the above initial speed sequence can be controlled within the above preset latency threshold (for example, the overall total latency can be controlled within 160ms), thereby meeting the engineering implementation requirements of millisecond-level real-time control of the vehicle-road-cloud collaborative system and solving the technical problem that the native call latency of the general large language model is too high and cannot meet the real-time control requirements.

[0092] Figure 3 A flowchart illustrating a vehicle-road-cloud collaborative energy-saving speed planning method provided in another embodiment of this application is shown below. Figure 3 Optionally, based on the above embodiments, before inputting the structured prompt words into the preset large language model, the method further includes: S301. Obtain historical vehicle driving data, which includes: terrain elevation information of historical routes, historical vehicle parameters, and corresponding historical best speed sequences.

[0093] For example, the aforementioned vehicle historical driving data may refer to historical data related to energy-saving speed planning collected and recorded during the vehicle's previous actual driving process.

[0094] The aforementioned terrain elevation information of historical routes may refer to, for example, the elevation data of various points along the vehicle's historical driving route, used to characterize the terrain undulation features of the historical route.

[0095] The aforementioned historical vehicle parameters may refer to parameters related to energy consumption characteristics of the vehicle during its historical driving, such as vehicle type and full load weight.

[0096] The aforementioned historical optimal speed sequence can refer, for example, to a speed sequence that achieves optimal global energy consumption in the historical driving scenario, determined through offline optimization calculation or manual annotation based on terrain elevation information and historical vehicle parameters of the historical path.

[0097] It is understandable that the aforementioned historical vehicle driving data can originate from the long-term accumulation of real-vehicle operating data or from simulated data generated by a simulation platform; no specific restrictions are imposed here. Obtaining this historical vehicle driving data can provide a rich sample base for subsequent model fine-tuning and training.

[0098] S302. Extract sample path features, sample vehicle parameters, and corresponding optimal speed sequences from the above-mentioned historical vehicle driving data.

[0099] For example, the aforementioned sample path features may be terrain features related to energy-saving speed planning extracted from the terrain elevation information of each historical path in the aforementioned vehicle historical driving data. For example, they may include the unit length and average slope information of each road unit after adaptive segmentation of each historical path.

[0100] The aforementioned sample vehicle parameters may refer, for example, vehicle energy consumption-related parameters extracted from the historical driving data of the aforementioned vehicles.

[0101] The aforementioned optimal speed sequence for corresponding samples can, for example, refer to the historical optimal speed sequence used as training labels, corresponding to the aforementioned sample path features and sample vehicle parameters. By extracting structured sample features and corresponding labels from the original historical driving data, the original data can be transformed into a format suitable for model training.

[0102] S303. Based on the above sample path features, the above sample vehicle parameters, and the above corresponding optimal speed sequence, construct a training dataset.

[0103] For example, the training dataset mentioned above may be a dataset used for model training, which is formed by taking the sample path features and the sample vehicle parameters as input features, taking the corresponding sample optimal speed sequence as training labels, and dividing it according to a preset ratio (e.g., 80% as training set and 20% as validation set).

[0104] When constructing the training dataset, the data can be preprocessed, such as normalizing the features of each sample to eliminate the influence of dimensional differences on the training process. Constructing this training dataset provides standardized data input for subsequent low-rank adaptive fine-tuning training.

[0105] S304. According to the preset parameter freezing rules, freeze some parameters of the initial large language model to obtain a partially frozen large language model.

[0106] For example, the aforementioned initial large language model can refer to a general large language model selected before fine-tuning training. For instance, the LLaMA-7B model with 7 billion parameters can be selected as the initial large language model. This model balances inference accuracy and latency performance and supports low-rank adaptive fine-tuning. The aforementioned preset parameter freezing rule can refer to a predefined rule used to determine which parameters in the model remain frozen and which parameters participate in fine-tuning training.

[0107] For example, the core logic of Low-Rank Adaptation (LoRA) could be to freeze most of the parameters of the initial large language model and fine-tune only the low-rank matrix of the attention layer in the model, thereby significantly reducing computational costs while adapting the model to specific tasks.

[0108] Specifically, the aforementioned preset parameter freezing rule can freeze all parameters in the initial large language model except for the query projection layer and value projection layer of the attention layer, resulting in the partially frozen large language model. By fine-tuning only a small number of parameters, targeted optimization of the model can be achieved without significantly increasing the number of parameters, but this is not the only limitation.

[0109] S305. Based on the above training dataset, perform low-rank adaptive fine-tuning training on the low-rank matrix of the attention layer in the partially frozen large language model to obtain the above lightweight large language model.

[0110] For example, the low-rank matrix of the attention layer mentioned above refers to the low-rank matrix obtained by decomposing the weight matrix used to calculate the query vector and value vector within the attention layer of the partially frozen large language model. During the low-rank adaptive fine-tuning training process, the low-rank matrix of the attention layer can be trained using a preset optimizer (e.g., the AdamW optimizer) and a preset loss function (e.g., the cross-entropy loss function) based on the training dataset.

[0111] For example, the specific parameter configuration for the aforementioned low-rank adaptive fine-tuning training can include: a rank of 8, used to control the dimension of the low-rank matrix and balance fine-tuning accuracy with the number of parameters; a LoRA scaling factor of 32, used to increase the weight ratio of fine-tuning; the target fine-tuning modules are the query projection layer and value projection layer of the attention layer; a LoRA dropout of 0.05, used to prevent overfitting; a learning rate of 5e-5, using a cosine annealing learning rate scheduling strategy; a batch size of 8; and 5 training epochs. During training, the training loss and validation loss can be monitored in real time. If the validation loss does not decrease for two consecutive epochs, an early stopping strategy can be used to terminate training early to avoid overfitting.

[0112] After training, the fine-tuned low-rank matrix weights can be merged with the base weights of the initial large language model to generate the lightweight large language model. To further reduce inference latency, the lightweight large language model can be quantized (e.g., INT8 quantization) to further compress the model size. Based on this, the inference latency of the lightweight large language model is further reduced. It can then be expressed as the following formula:

[0113] Among them, the above This refers to the inference latency of the full model, which is also the inference latency of the initial large language model. To reduce the number of model parameters, This refers to the total number of parameters in the full model.

[0114] Through the aforementioned low-rank adaptive fine-tuning training, the lightweight large language model can accurately grasp the professional knowledge of energy-saving speed planning while maintaining its general language understanding ability, thus achieving a balance between high precision, low cost, and low latency.

[0115] Optionally, in the foregoing Figure 1 Based on the embodiments, the above-mentioned initial velocity sequence is smoothed to obtain the target velocity sequence, including: The initial velocity sequence is smoothed using a first-order inertial filtering algorithm to eliminate abrupt velocity changes, thus obtaining the target velocity sequence.

[0116] For example, the aforementioned first-order inertial filtering algorithm can be a low-pass filtering algorithm that smooths the signal by weighted averaging of the current input value and the output value from the previous moment. Since the initial velocity sequence directly output by the aforementioned preset large language model may contain jumps or discontinuities between adjacent velocity values, directly issuing it as a control command to the vehicle terminal could lead to sudden acceleration or deceleration during vehicle operation, affecting driving smoothness and ride comfort. Therefore, the aforementioned first-order inertial filtering algorithm can be used to smooth the initial velocity sequence.

[0117] For example, the first-order inertial filtering algorithm described above can be expressed as the following formula:

[0118] Among them, the above Let i be the i-th velocity value in the above initial velocity sequence. This is the target velocity value output after smoothing from the previous moment. This is the target velocity value output after smoothing at the current moment. This is the smoothing coefficient. The above smoothing coefficient... The value range of this smoothing coefficient can be, for example, between 0.6 and 0.8. This smoothing coefficient can be used to adjust the weight of the current input velocity value in the output result: the larger the smoothing coefficient, the faster the target velocity sequence follows the initial velocity sequence. The smaller the smoothing coefficient, the smoother the change in the target velocity sequence. By performing the above smoothing process on each velocity value in the initial velocity sequence one by one, a target velocity sequence with continuous velocity changes and no abrupt changes can be obtained.

[0119] The latency consumed by the aforementioned vehicle-side smoothing process can be controlled to within 50ms, for example. Combined with the cache query latency of the aforementioned edge caching pre-computation strategy (e.g., within 10ms) and the inference latency of the model lightweighting strategy (e.g., within 100ms), the total system latency can be controlled to within 160ms, meeting the engineering implementation requirements of millisecond-level real-time control for vehicle-road-cloud cooperative systems.

[0120] By employing the first-order inertial filtering algorithm to smooth the initial velocity sequence, it is possible to ensure that the vehicle's speed changes are continuous, smooth, and shock-free during actual driving, thus achieving energy-saving speed planning while also considering driving smoothness and ride comfort.

[0121] In addition, in the aforementioned Figure 1 Based on the embodiments, after sending the target speed sequence to the vehicle terminal for execution, the method further includes: Obtain the real-time vehicle driving status data and energy consumption data uploaded by the aforementioned vehicle-mounted terminals.

[0122] Based on the above-mentioned real vehicle driving status data and energy consumption data, the above-mentioned preset prompt word construction strategy and / or the above-mentioned preset large language model are optimized and updated.

[0123] For example, the aforementioned vehicle-mounted terminal may refer to a device with computing capabilities, such as the vehicle infotainment system. The aforementioned real-vehicle driving status data may be data related to the driving status collected or calculated during actual vehicle operation through sensors and vehicle infotainment system readings, such as actual driving speed, actual acceleration, and driving trajectory. The aforementioned energy consumption data may be data related to energy consumption collected or calculated during actual vehicle operation through sensors and vehicle infotainment system readings, such as actual energy consumption per unit distance and total energy consumption. The aforementioned vehicle-mounted terminal can upload the aforementioned real-vehicle driving status data and energy consumption data to a cloud server via a wireless communication network.

[0124] After obtaining the real-vehicle driving status data and energy consumption data uploaded by the aforementioned in-vehicle terminal, these can be used as feedback data to optimize and update the aforementioned preset prompt word construction strategy and / or the aforementioned preset large language model. Specific optimization and update methods may include at least one of the following: Firstly, the parameters in the aforementioned preset prompt word construction strategy can be adjusted and optimized. For example, based on the aforementioned real vehicle driving status data and energy consumption data, the impact of the current terrain elevation change rate threshold on energy-saving effects can be analyzed, and the aforementioned preset change rate threshold can be adaptively adjusted to make the terrain adaptive segmentation results more closely match the actual energy consumption characteristics. Simultaneously, the real vehicle calibration coefficients in the aforementioned energy consumption model can be calibrated based on the aforementioned energy consumption data, continuously improving the prediction accuracy of the energy consumption model.

[0125] Secondly, incremental fine-tuning can be performed on the aforementioned pre-set large language model. For example, newly added real-world vehicle driving status data and energy consumption data can be collected periodically (e.g., every 3 months). Sample path features, sample vehicle parameters, and corresponding optimal speed sequences can be extracted from the new data and added to the aforementioned training dataset (e.g., adding 1000 new samples). The same low-rank adaptive fine-tuning parameters as mentioned above can be used to incrementally train the aforementioned lightweight large language model to continuously optimize the prediction accuracy and scene adaptability of the aforementioned lightweight large language model.

[0126] Through the above optimizations and updates, a self-evolving intelligent closed-loop control system can be formed, consisting of "prompt word construction → model reasoning → smoothing processing → execution → data feedback → strategy and model optimization". This enables the vehicle-road-cloud collaborative energy-saving speed planning method to continuously adapt to dynamic factors such as vehicle model updates and road condition changes, maintaining long-term planning accuracy and energy-saving effects.

[0127] Figure 4This is a schematic diagram of the structure of a vehicle-road-cloud collaborative energy-saving speed planning device according to an embodiment of this application. This device can execute the aforementioned vehicle-road-cloud collaborative energy-saving speed planning method. The device can be integrated into the aforementioned vehicle-mounted system or other devices with computing capabilities, such as… Figure 4 As shown, the device may include: The acquisition module 410 is used to acquire map data of the target path and, based on the terrain elevation information in the map data, adaptively segment the target path to generate multiple road units.

[0128] The construction module 420 is used to construct structured prompts based on the road characteristics, vehicle parameters, and preset energy-saving optimization targets of the above-mentioned multiple road units using a preset prompt word construction strategy.

[0129] The reasoning module 430 is used to input the above-mentioned structured prompt words into a preset large language model, and to perform reasoning through a preset low-latency scheduling strategy to obtain the initial velocity sequence output by the above-mentioned preset large language model.

[0130] The processing module 440 is used to smooth the initial speed sequence to obtain the target speed sequence and send the target speed sequence to the vehicle terminal.

[0131] The vehicle-road-cloud collaborative energy-saving speed planning method provided in this application includes: acquiring map data of the target path, and adaptively segmenting the target path according to the terrain elevation information in the map data to generate multiple road units. A preset prompt word construction strategy is adopted to construct structured prompt words based on the road features, vehicle parameters, and preset energy-saving optimization goals of the multiple road units. The structured prompt words are input into a preset large language model, and inference is performed through a preset low-latency scheduling strategy to obtain the initial speed sequence output by the preset large language model. The initial speed sequence is smoothed to obtain the target speed sequence, and the target speed sequence is sent to the vehicle terminal. In this embodiment, the method adaptively segments the target path according to terrain elevation information, constructs structured prompt words from road features, vehicle parameters, and preset energy-saving optimization goals using a preset prompt word construction strategy, and then combines this with a preset low-latency scheduling strategy to perform inference and smoothing on the preset large language model. This achieves intelligent energy-saving speed planning with scene adaptability, thereby improving the rationality, accuracy, and scene adaptability of vehicle-road-cloud collaborative energy-saving speed planning.

[0132] Optionally, the acquisition module 410 is specifically used to extract the elevation data of each path point on the target path and calculate the elevation difference between adjacent path points. Based on the elevation difference and the road segment length between adjacent path points, the terrain elevation change rate between adjacent path points is calculated. If the terrain elevation change rate is greater than a preset change rate threshold, segmentation is performed between the corresponding adjacent path points to generate the road units with uniform terrain features.

[0133] Optionally, the aforementioned preset prompt word construction strategy includes: preset identity definition, terrain segmentation information, vehicle parameters, energy consumption model, preset energy-saving optimization target, driving constraints, and output format specifications. The aforementioned road features include: the unit length and average slope information of each of the aforementioned road units. The aforementioned driving constraints include: driving speed constraints and driving acceleration constraints.

[0134] The aforementioned acquisition module 410 can also be used to acquire the aforementioned terrain segmentation information based on the aforementioned unit length and the aforementioned average slope information. Based on the aforementioned energy consumption model and the global total energy consumption minimization objective, the aforementioned preset energy-saving optimization objective is determined.

[0135] The aforementioned construction module 420 is specifically used to combine the aforementioned preset identity definition, the aforementioned terrain segmentation information, the aforementioned vehicle parameters, the aforementioned driving speed constraints, the aforementioned driving acceleration constraints, the aforementioned energy consumption model, the aforementioned preset energy-saving optimization target, and the aforementioned output format specifications into the aforementioned structured prompt words.

[0136] Optionally, the aforementioned preset large language model is a lightweight large language model. The aforementioned preset low-latency scheduling strategies include: an edge caching pre-computation strategy, a model lightweighting strategy, and a post-processing localization strategy.

[0137] The aforementioned inference module 430 is specifically used to input the structured prompt words into a preset large language model, and according to the aforementioned edge caching pre-computation strategy, query whether a velocity sequence matching the structured prompt words exists in the local cache. If it exists, the cached result is used as the initial velocity sequence. If it does not exist, according to the aforementioned model lightweighting strategy, the aforementioned lightweight large language model is invoked to infer the structured prompt words and obtain the initial velocity sequence. The total latency for obtaining the initial velocity sequence is less than a preset latency threshold.

[0138] Optionally, the acquisition module 410 can also be used to acquire historical vehicle driving data, which includes: terrain elevation information of historical paths, historical vehicle parameters, and corresponding historical optimal speed sequences. From the historical vehicle driving data, sample path features, sample vehicle parameters, and corresponding sample optimal speed sequences are extracted.

[0139] The aforementioned construction module 420 can also be used to construct a training dataset based on the aforementioned sample path features, the aforementioned sample vehicle parameters, and the aforementioned corresponding optimal speed sequence of the samples.

[0140] The aforementioned vehicle-road-cloud collaborative energy-saving speed planning device may further include: a training module, used to freeze some parameters of the initial large language model according to a preset parameter freezing rule, to obtain a partially frozen large language model. Based on the aforementioned training dataset, low-rank adaptive fine-tuning training is performed on the low-rank matrix of the attention layer in the partially frozen large language model to obtain the aforementioned lightweight large language model.

[0141] Optionally, the processing module 440 is specifically used to smooth each velocity value in the initial velocity sequence using a first-order inertial filtering algorithm to eliminate velocity abrupt changes and obtain the target velocity sequence.

[0142] Optionally, the acquisition module 410 can also be used to acquire the real vehicle driving status data and energy consumption data uploaded by the vehicle terminal. Based on the real vehicle driving status data and energy consumption data, the preset prompt word construction strategy and / or the preset large language model are optimized and updated.

[0143] The above-described device is used to execute the method provided in the foregoing embodiments, and its implementation principle and technical effect are similar, so they will not be described again here.

[0144] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device can be a device with computing power, such as the aforementioned vehicle infotainment system. Figure 5 As shown, the device 500 includes: The processor 510, storage medium 520, and bus 530 are connected in communication via bus 530.

[0145] The storage medium 520 stores machine-readable instructions that can be executed by the processor 510. When the electronic device is running, the processor 510 executes the machine-readable instructions to execute the vehicle-road-cloud collaborative energy-saving speed planning method.

[0146] It should be understood that, Figure 5 The structure shown is only a schematic diagram of an electronic device; the electronic device may also include components that are larger than those shown. Figure 5 The more or fewer components shown, or having the same Figure 5 The different configurations shown. Figure 5 The components shown can be implemented using hardware, software, or a combination thereof.

[0147] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the vehicle-road-cloud collaborative energy-saving speed planning method described in the above method embodiments.

[0148] Computer-readable storage media can be electronic storage devices such as flash memory, EEPROM (Electrically Erasable Programmable Read-Only Memory), EPROM, hard disk, or ROM. Optionally, computer-readable storage media includes non-transitory computer-readable storage medium. The computer-readable storage medium has storage space for program code that performs any of the method steps described above. This program code can be read from or written to one or more computer program exhibits. The program code can be compressed, for example, in a suitable form.

[0149] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program exhibits according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0150] In addition, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0151] If the functionality is implemented as a software module and sold or used as an independent exhibit, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software exhibit. This computer software exhibit is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0152] The above description is merely a preferred embodiment of this application and does not limit the patent scope of this application. Any equivalent structural transformations made based on the inventive concept of this application and the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included within the patent protection scope of this application.

Claims

1. A vehicle-road-cloud cooperative energy-saving speed planning method, characterized in that, include: Obtain map data of the target path, and adaptively segment the target path based on the terrain elevation information in the map data to generate multiple road units; A pre-defined prompt word construction strategy is adopted to construct structured prompt words based on the road characteristics, vehicle parameters, and pre-defined energy-saving optimization targets of the multiple road units; The structured prompt words are input into a preset large language model, and inference is performed through a preset low-latency scheduling strategy to obtain the initial velocity sequence output by the preset large language model. The initial velocity sequence is smoothed to obtain the target velocity sequence, and the target velocity sequence is sent to the vehicle terminal.

2. The method of claim 1, wherein, The step involves adaptively segmenting the target path based on the terrain elevation information in the map data to generate multiple road units, including: Extract the elevation data of each path point on the target path and calculate the elevation difference between adjacent path points; Calculate the rate of change of terrain elevation between adjacent path points based on the elevation difference and the road segment length between adjacent path points; If the terrain elevation change rate is greater than a preset change rate threshold, then segmentation is performed between the corresponding adjacent path points to generate the road unit with uniform terrain features.

3. The method of claim 1, wherein, The preset prompt word construction strategy includes: preset identity definition, terrain segmentation information, vehicle parameters, energy consumption model, preset energy-saving optimization target, driving constraints, and output format specifications; the road features include: the unit length and average slope information of each road unit; the driving constraints include: driving speed constraints and driving acceleration constraints; Before constructing structured prompts based on the road characteristics, vehicle parameters, and preset energy-saving optimization targets of the multiple road units using a preset prompt word construction strategy, the method further includes: The terrain segmentation information is obtained based on the unit length and the average slope information; Based on the energy consumption model and the global total energy consumption minimization objective, the preset energy-saving optimization objective is determined; The method employs a preset prompt word construction strategy, which constructs structured prompt words based on the road characteristics, vehicle parameters, and preset energy-saving optimization targets of the multiple road units, including: The preset identity definition, the terrain segmentation information, the vehicle parameters, the driving speed constraint, the driving acceleration constraint, the energy consumption model, the preset energy-saving optimization target, and the output format specification are combined into the structured prompt words.

4. The method of claim 1, wherein, The preset large language model is a lightweight large language model; the preset low-latency scheduling strategy includes: edge caching pre-computation strategy, model lightweighting strategy, and post-processing localization strategy. The step of inputting the structured prompt words into a preset large language model, performing inference through a preset low-latency scheduling strategy, and obtaining the initial velocity sequence output by the preset large language model includes: The structured prompt words are input into a preset large language model, and according to the edge cache pre-computation strategy, a speed sequence matching the structured prompt words is queried in the local cache. If it exists, the cached result is used as the initial velocity sequence; If it does not exist, then according to the model lightweighting strategy, the lightweight large language model is invoked to infer the structured prompt words and obtain the initial speed sequence. Wherein, the total delay for obtaining the initial velocity sequence is less than a preset delay threshold.

5. The method of claim 4, wherein, Before inputting the structured prompt words into the preset large language model, the method further includes: Obtain historical vehicle driving data, which includes: terrain elevation information of historical routes, historical vehicle parameters, and corresponding historical best speed sequences. Extract sample path features, sample vehicle parameters, and corresponding optimal speed sequences from the vehicle's historical driving data. A training dataset is constructed based on the sample path features, the sample vehicle parameters, and the corresponding optimal speed sequence of the samples. According to the preset parameter freezing rules, some parameters of the initial large language model are frozen to obtain a partially frozen large language model; Based on the training dataset, the low-rank matrix of the attention layer in the partially frozen large language model is subjected to low-rank adaptive fine-tuning training to obtain the lightweight large language model.

6. The method of claim 1, wherein, The smoothing process of the initial velocity sequence to obtain the target velocity sequence includes: The initial velocity sequence is smoothed using a first-order inertial filtering algorithm to eliminate abrupt velocity changes, thereby obtaining the target velocity sequence.

7. The method of claim 1, wherein, After sending the target speed sequence to the vehicle terminal for execution, the method further includes: Obtain the real-time vehicle driving status data and energy consumption data uploaded by the vehicle terminal; Based on the actual vehicle driving status data and energy consumption data, the preset prompt word construction strategy and / or the preset large language model are optimized and updated. 8.A vehicle-road-cloud cooperative energy-saving speed planning device, characterized in that, include: The acquisition module is used to acquire map data of the target path and adaptively segment the target path according to the terrain elevation information in the map data to generate multiple road units. The construction module is used to construct structured prompts based on the road characteristics, vehicle parameters, and preset energy-saving optimization targets of the multiple road units using a preset prompt word construction strategy. The reasoning module is used to input the structured prompt words into a preset large language model, perform reasoning through a preset low-latency scheduling strategy, and obtain the initial velocity sequence output by the preset large language model. The processing module is used to smooth the initial velocity sequence to obtain the target velocity sequence, and send the target velocity sequence to the vehicle terminal.

9. An electronic device, comprising: include: The device includes a processor, a storage medium, and a bus, wherein the storage medium stores machine-readable instructions executable by the processor, and when the electronic device is in operation, the processor communicates with the storage medium via the bus, and the processor executes the machine-readable instructions to perform the steps of the method as described in any one of claims 1-7.

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 method as described in any one of claims 1-7.

Images