A dynamic experience route generation method

Abstract

This application provides a dynamic empirical route generation method, relating to the field of intelligent scheduling and path planning technology for new energy commercial vehicles. The method includes: collecting multi-dimensional state data across the entire process; filtering valid data from the multi-dimensional state data to generate first state data; constructing a hierarchical energy consumption prediction model, which includes a lightweight model and a high-precision model; employing the NSGA-II algorithm and particle swarm optimization algorithm, with the objective functions of minimizing overall energy cost, minimizing time consumption, and minimizing refueling waiting time, and combining this with the hierarchical energy consumption prediction model to predict energy consumption, thereby generating a candidate route set; determining the total empirical matching score for each candidate route in the candidate route set, and determining the optimal route based on the total empirical matching score. This solution ensures the reliability of range prediction under extreme operating conditions while significantly reducing the load pressure of large-scale concurrent computing in the cloud.

Description

Technical Field

[0001] This invention relates to the field of intelligent scheduling and route planning technology for new energy commercial vehicles, and in particular to a method for generating dynamic experience routes. Background Technology

[0002] In actual transportation scenarios for new energy heavy-duty trucks, vehicle energy consumption is highly sensitive to dynamic factors such as load fluctuations, terrain gradient changes, and ambient temperature. However, existing route planning and energy consumption prediction technologies generally have the following shortcomings, making it difficult to meet the operational requirements of high precision, high efficiency, and strong adaptability.

[0003] First, existing energy consumption prediction models struggle to balance accuracy and efficiency. Current mainstream solutions often employ linear regression or simplified models based on fixed rules for energy consumption estimation. These methods exhibit significant prediction errors under complex operating conditions (such as heavy-load conditions in mountainous areas or low-temperature environments), easily leading to range anxiety or miscalculations in refueling planning. On the other hand, while relying entirely on high-precision, complex models based on deep learning can improve prediction accuracy, it results in wasted cloud computing resources and response delays when facing massive concurrent vehicle requests, especially in simple road scenarios like flat, constant-speed sections. Therefore, current technologies cannot achieve an adaptive balance between accurate measurement in complex scenarios and fast calculation in simple scenarios.

[0004] Second, route planning lacks awareness and response to the dynamic status of energy replenishment facilities. Current route planning for new energy commercial vehicles is largely based on static map data and theoretical energy replenishment site information, failing to consider dynamic disturbances in actual operation, such as sudden equipment failures at charging / swapping / hydrogen refueling stations, station congestion, and real-time fluctuations in energy prices (e.g., time-of-use electricity pricing, regional subsidy differences). This disconnect between "static planning" and "dynamic reality" means that theoretically optimal routes may fail entirely due to single-point failures in actual implementation. More seriously, the system lacks an automatic correction mechanism based on historical deviations, failing to learn from past planning failures, causing the same errors to repeatedly occur in different vehicles or similar scenarios.

[0005] Third, general navigation algorithms are not suitable for the special operating conditions of heavy-duty trucks. Currently, commercial vehicle navigation mostly uses electronic map algorithms from passenger cars, only superimposing basic constraints such as height and weight limits; some electrification solutions simply introduce fixed energy consumption parameters per 100 kilometers and static charging point locations for route recommendation. Such methods have significant drawbacks: they cannot model the nonlinear coupling effect of load and slope on energy consumption, and the error can reach more than 30% in mountainous or heavy-load scenarios; they lack targeted support for the special charging logic of multi-energy vehicles such as hydrogen fuel cell vehicles (such as the scarcity of hydrogen refueling stations and long refueling times); they cannot obtain fine-grained status information such as real-time queuing, faults, and power limitations within charging stations, which can easily lead to vehicles driving into unavailable stations, getting stuck in a "dead end," or waiting for a long time.

[0006] Fourth, the dispatching system lacks real-time closed-loop feedback and collaborative optimization capabilities. Traditional logistics dispatching mostly adopts an offline calculation mode, relying on monthly or quarterly updated historical average data to formulate fixed routes and timetables, requiring a large amount of manual intervention to adjust parameters. Its fundamental defects are: the lack of a real-time closed-loop feedback mechanism, making it unable to dynamically respond to sudden changes in road conditions, abnormal weather, or vehicle status drift; when actual energy consumption deviates from the prediction, the system cannot automatically attribute the cause and correct the model, leading to the long-term accumulation of problems such as overestimation or underestimation; it is difficult to achieve multi-dimensional real-time collaboration between vehicles, stations, roads, and the cloud, and cannot support refined energy cost control (such as utilizing off-peak electricity prices for charging) and dynamic energy replenishment timing optimization.

[0007] In summary, the problems with existing technologies include: the contradiction between accuracy and efficiency, the high computational cost of high-precision models, and the inaccurate prediction of lightweight models; the disconnect between static planning and dynamic reality, making it impossible to perceive and respond to the real-time status of energy replenishment facilities; the mismatch between general algorithms and the characteristics of heavy-duty vehicles, ignoring key nonlinear factors such as load, slope, and multiple energy sources; and the separation between offline scheduling and online operation, lacking closed-loop capabilities such as self-learning, self-correction, and multi-terminal collaboration. Summary of the Invention

[0008] This invention aims to at least solve the aforementioned technical problems existing in the prior art. To this end, this invention proposes a dynamic empirical route generation method, the method comprising:

[0009] Collect multi-dimensional state data across the entire link, filter valid data from the multi-dimensional state data, and generate first state data;

[0010] A hierarchical energy consumption prediction model is constructed, which includes a lightweight model and a high-precision model. The high-precision model is generated by serial fusion of XGBoost, LSTM, and Bayesian network. The lightweight model uses multiple linear regression to predict energy consumption.

[0011] The NSGA-II algorithm and particle swarm optimization algorithm are used to minimize the overall energy cost, the shortest time consumption, and the shortest refueling waiting time as the objective functions. The hierarchical energy consumption prediction model is combined to predict energy consumption and generate a set of candidate routes. The objective function is constructed based on the first state data.

[0012] Determine the total empirical matching score for each candidate route in the candidate route set, and determine the optimal route based on the total empirical matching score.

[0013] Optionally, the NSGA-II algorithm and particle swarm optimization algorithm are used to minimize the overall energy cost, the shortest time consumption, and the shortest refueling waiting time as objective functions, and combined with the hierarchical energy consumption prediction model to predict energy consumption, in order to generate a candidate route set, including:

[0014] Based on the K-shortest path algorithm, multiple feasible paths are determined;

[0015] Based on the scene feature parameters in the multidimensional state data, a target model is selected in the hierarchical energy consumption prediction model to determine the predicted energy consumption of multiple feasible paths; the scene feature parameters include road slope fluctuation, ambient temperature, queuing probability of energy replenishment stations, cargo weight fluctuation, and average vehicle speed fluctuation rate.

[0016] Substitute the multiple feasible paths and their corresponding predicted energy consumptions into the objective function, perform multi-objective optimization based on the NSGA-II algorithm and particle swarm optimization algorithm, and select the Pareto optimal solution set to determine the candidate route set.

[0017] Optionally, after generating the candidate route set, the method further includes:

[0018] For each candidate route in the candidate route set, the corresponding energy replenishment and energy recovery strategies are matched to generate candidate routes that include the energy replenishment and energy recovery strategies.

[0019] Optionally, the matching of corresponding energy replenishment and energy recovery strategies to generate candidate routes including energy replenishment and energy recovery strategies includes:

[0020] If a long downhill or continuous downhill section is detected ahead, reduce the output power of the fuel cell stack or start a controllable auxiliary load.

[0021] Optionally, if a long downhill or continuous downhill section is detected ahead, reducing the fuel cell stack output power or activating a controllable auxiliary load includes:

[0022] If a long downhill or multiple consecutive downhill sections are detected ahead, and the cumulative elevation difference of the downhill sections is greater than the preset elevation difference threshold, the maximum regenerative energy generated by the downhill sections is estimated based on the total mass of the vehicle, the current speed, and the expected driving trajectory.

[0023] The required battery rechargeable capacity is determined based on the maximum regenerative electrical energy.

[0024] If the current remaining free capacity of the battery is less than the rechargeable capacity of the battery, it is determined that it needs to be discharged in advance.

[0025] Predicted values ​​of the thermal state and health state of the power battery at the start of the downhill slope;

[0026] Based on the predicted thermal state value, the predicted health state value, and the battery rechargeable capacity, a corresponding safe state of charge upper limit is generated.

[0027] If the safe state of charge upper limit is less than the preset safe state of charge upper limit threshold, the controllable auxiliary load is actively activated, or the stack output power is reduced.

[0028] Optionally, the matching of corresponding energy replenishment and energy recovery strategies to generate candidate routes including energy replenishment and energy recovery strategies includes:

[0029] The load status of the target charging station is monitored. If the occupancy rate of the target charging station is greater than a preset occupancy rate threshold, a candidate station is generated and pushed to subsequent vehicles that plan to go to the target charging station. Both the subsequent vehicles and the current vehicle interact with the cloud.

[0030] Optionally, the matching of corresponding energy replenishment and energy recovery strategies to generate candidate routes including energy replenishment and energy recovery strategies includes:

[0031] Traffic flow is predicted for the target road segment based on a traffic flow prediction model, and the average vehicle speed when the vehicle arrives at the target road segment is estimated by combining the real-time traffic flow at the checkpoint upstream of the target road segment; the traffic flow prediction model is a spatiotemporal graph convolutional network.

[0032] Based on the average vehicle speed and the speed limit of the target road segment, determine whether the target road segment is a potentially congested route;

[0033] If the target road segment is determined to be a potentially congested route, the dynamic A* pathfinding algorithm is used to determine the bypass branch with the second smallest comprehensive cost, so as to generate alternative detour routes and push them to vehicles that have not yet departed; the comprehensive cost is determined based on the absolute distance and energy consumption data of the bypass branch and the expected increase in time caused by congestion.

[0034] Optionally, before generating alternative detour routes and pushing them to vehicles that have not yet departed, the process further includes:

[0035] Determine the dynamic saturation threshold of the alternative detour routes;

[0036] The vehicles that have not yet departed are sorted, and virtual reservations are performed according to the alternative detour routes. The marginal energy consumption cost of the alternative detour routes is determined after each virtual reservation. The marginal energy consumption cost is the total increase in energy consumption caused by the reduction in the speed of the vehicles already on the alternative detour routes due to the addition of a new vehicle.

[0037] If the current virtual occupancy causes the number of vehicles in the alternative detour route to exceed the dynamic saturation threshold, or causes the marginal energy consumption cost to exceed the preset congestion worsening threshold, then the alternative detour route will stop being pushed to current and subsequent vehicles, and the departure time will be postponed or a new alternative detour route will be generated; the subsequent vehicles are vehicles that are traveling or have not yet departed and are planned to pass through the alternative detour route.

[0038] Optionally, determining the dynamic saturation threshold of the alternative detour routes includes:

[0039] Sum the average vehicle length with the minimum safe following distance under the current speed limit to obtain the sum value;

[0040] The dynamic saturation threshold is obtained by dividing the total length of the alternative detour routes by the sum.

[0041] Optionally, the lightweight model uses multiple linear regression to predict energy consumption, including:

[0042] Based on the actual driving mileage of discrete sub-segments, the absolute difference between the elevation of the end point and the starting point of the discrete sub-segment, vehicle load, and estimated average speed, the predicted energy consumption is determined by multiple linear regression.

[0043] The technical solution provided in this application can include the following beneficial effects: This application constructs a hierarchical modeling engine that includes a lightweight model and a high-precision model, and introduces an automatic scene complexity switching mechanism. It can intelligently select the prediction model according to road condition characteristics. In simple scenarios, it uses a lightweight model to achieve millisecond-level response, and in complex scenarios, it automatically calls a high-precision model that integrates deep learning to ensure prediction accuracy. This not only ensures the reliability of range prediction under extreme conditions and avoids the risk of vehicle breakdown, but also significantly reduces the load pressure of large-scale concurrent computing in the cloud. This application is not limited to a single shortest path planning, but is based on a multi-objective optimization algorithm to generate a comprehensive strategy covering the selection of refueling time, energy recovery reservation and operation path coordination. This significantly reduces the overall energy cost of the entire journey, and effectively shortens the non-transportation time caused by refueling queuing, thereby improving the overall operational efficiency. Attached Figure Description

[0044] Figure 1 A flowchart of a dynamic experience route generation method provided in an embodiment of the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of this disclosure, unless otherwise stated, "a plurality of" means two or more. Furthermore, the use of "based on" or "according to" implies openness and inclusiveness, because processes, steps, calculations, or other actions "based on" or "according to" one or more of the stated conditions or values ​​may in practice be based on additional conditions or beyond the stated values.

[0047] This invention provides a method for generating dynamic empirical routes, such as... Figure 1 As shown, the method may include the following steps:

[0048] Step 101: Collect multi-dimensional state data across the entire link, filter the valid data in the multi-dimensional state data, and generate the first state data.

[0049] Specifically, the entire data collection process from vehicle to station to road to cloud is initiated. To ensure the accuracy of the subsequent energy consumption model input, the cloud management platform not only passively receives data but also actively or periodically acquires multi-dimensional status data of the new energy heavy-duty trucks in response to their route planning requests. This status data originates from onboard terminals, refueling stations, and road monitoring systems, covering all dimensions from the vehicle's micro-state to the environment's macro-state. The multi-dimensional status data includes vehicle status data, refueling status data, road condition status, and environmental status data. Analyzing the entire multi-dimensional status data allows for the determination of vehicle status, refueling status, road condition status, and environmental status.

[0050] For vehicle status data collection, the on-board terminal reads the vehicle's underlying operating parameters in real time via the CAN bus (Controller Area Network). The cloud management platform can determine the specific vehicle status by parsing the multi-dimensional status data across the entire chain. This vehicle status includes, but is not limited to, multi-energy type identification, battery state of charge (SOC), remaining hydrogen, and energy recovery rate. Regarding multi-energy type identification, this embodiment of the invention is compatible with both pure electric heavy-duty trucks and hydrogen fuel cell heavy-duty trucks. For pure electric heavy-duty trucks, the focus is on collecting data on battery capacity, cell temperature, and voltage consistency; for hydrogen fuel cell heavy-duty trucks, the focus is on collecting hydrogen tank pressure, hydrogen tank temperature, and fuel cell stack output power. In addition, driving behavior parameters are also key data collection targets, including the frequency of rapid acceleration, the intervention efficiency of the braking energy recovery system, and vehicle load signals. Vehicle load signals are acquired in real time through weighing sensors on the axle, because load fluctuations in heavy-duty trucks (the difference between empty and fully loaded loads can be tens of tons) are the most critical variable affecting energy consumption.

[0051] For collecting charging status data, the cloud management platform connects with major charging operators and hydrogen refueling station operation platforms via API interfaces. The collected charging status includes the power level of charging piles or hydrogen refueling stations along the route (e.g., 120 kW fast charging or 35 MPa hydrogen refueling), real-time electricity or hydrogen prices (e.g., time-of-use pricing), and estimated queue waiting times. The estimated queue waiting time is not solely based on station reports but also estimates considering the number of currently connected charging guns at the station and the remaining charging time for vehicles on site. Station dynamic data also includes fault status and estimated recovery time to prevent vehicles from visiting faulty stations.

[0052] The collection of road condition and environmental data relies not only on general navigation map data but also integrates data from weather stations and road IoT devices. This data includes road slope sequences, real-time traffic flow, and ambient temperature data. The road slope sequence forms the basis for refined energy consumption prediction, corrected by combining high-precision map elevation data with measured data from vehicle-mounted gyroscopes. Real-time traffic flow includes congestion index and average vehicle speed. Environmental sensing data such as temperature, humidity, precipitation probability, wind speed, and altitude are used to adjust battery discharge efficiency and drag coefficient under different operating conditions.

[0053] The road slope sequence is constructed using a fusion method combining static map data and dynamic vehicle-mounted correction. Specifically, in the static layer, based on high-precision map GIS data, the elevation of nodes along the planned route at regular intervals (e.g., every 10 meters) is extracted. The theoretical slope is calculated. In the dynamic correction layer: using the vehicle terminal's built-in three-axis gyroscope and accelerometer, the pitch angle of the vehicle during travel is collected in real time. Combined with the vehicle speed signal to filter out acceleration and deceleration interference, the measured slope is obtained. The path is discretized into a set of road segments. The corresponding slope sequence is represented as ,in For the length of the road segment, The corrected average slope value is given by n, which represents the number of road segments.

[0054] In the static layer, after extracting the elevation data of adjacent road segment nodes based on the high-precision map, the theoretical slope is calculated using the following expression:

[0055]

[0056] In the formula, For theoretical slope, and The elevation of adjacent road segment nodes. For example, the fixed horizontal distance between two road segment nodes can be set to 10 meters.

[0057] In the dynamic correction layer, vehicle acceleration and deceleration cause the vehicle body to pitch up or down, resulting in a deviation in the directly collected pitch angle. Therefore, the calculation of the measured slope needs to eliminate this interference. Specifically, the measured slope is obtained through the following expression:

[0058]

[0059] In the formula, Indicates the measured slope. α is the vehicle pitch angle directly acquired by the three-axis gyroscope, α is the vehicle longitudinal acceleration obtained by differentiating the vehicle speed signal, and g is the gravitational acceleration.

[0060] Corrected average slope value It is the spatial average of the merged slope at multiple points within a certain road segment. The merged slope at a single point... Using a weighted average, , , This represents the corresponding weight, and the sum of these two weights is 1. The average slope value is obtained by averaging the combined slopes of all data collection points within the road segment. .

[0061] After collecting massive amounts of data, directly inputting the raw data into the cloud model would result in significant computational waste and noise interference. Therefore, this embodiment of the invention preprocesses the entire data chain after acquiring multidimensional state data. After preprocessing, the data is filtered based on empirical confidence levels, retaining only data with an empirical confidence level greater than 90% as valid empirical data. Preprocessing includes real-time cleaning and standardization. The vehicle terminal is equipped with an edge computing module to clean and standardize the collected data in real time, such as removing distorted data caused by sensor jitter and transmitting only incremental difference data to the cloud to save bandwidth. Standardization can be achieved through Z-Score standardization or normalization. The cloud management platform classifies and stores the received data and uses an empirical confidence level labeling mechanism for filtering. The empirical confidence level filtering process is extremely rigorous. This embodiment of the invention cross-validates each uploaded energy consumption data, such as comparing it with the average energy consumption of the same vehicle model on the same road segment. Only records with complete data, normal sensor status, and deviations from statistical regularities within a reasonable range are labeled as high confidence. In this embodiment, only data with an empirical confidence level greater than 90% are retained as valid empirical data for subsequent model training and updates.

[0062] For example, after receiving status data from the vehicle terminal, the cloud management platform determines the empirical confidence level of the status data using the following expression:

[0063]

[0064] in, Indicates empirical confidence level. This indicates the sensor status score, which is either 0 or 1, determined by the OBD fault code. This represents the actual energy consumption collected in this instance, specifically the actual energy consumption value corresponding to the single complete transportation task or the specific planned route segment that has just been completed. This represents the historical average energy consumption of the same vehicle type on the same road segment. The data integrity ratio is indicated by dividing the number of fields collected by the number of fields that should be collected. , , This represents the weighting coefficient, which can take values ​​of 0.4, 0.4, and 0.2 respectively.

[0065] Step 102: Construct a hierarchical energy consumption prediction model, which includes a lightweight model and a high-precision model; the high-precision model is generated by serial fusion of XGBoost, LSTM and Bayesian network; the lightweight model uses multiple linear regression to predict energy consumption.

[0066] Specifically, a lightweight hierarchical modeling engine is constructed, which is a hierarchical energy consumption prediction model. This model is used to automatically match the energy consumption prediction model based on the current scenario complexity after feasible paths are generated, predicting the energy consumption data of new energy heavy trucks under different candidate routes. The initial design goal of this lightweight hierarchical modeling engine is to resolve the contradiction between "high accuracy" and "low latency." The lightweight hierarchical modeling engine includes a basic lightweight model and an enhanced high-precision model. In this embodiment, the model switches between the basic lightweight model and the enhanced high-precision model based on the scenario complexity in the multi-dimensional state data across the entire link. The switching mechanism relies on real-time judgment of scenario feature parameters. If the scenario complexity in the multi-dimensional state data meets a preset complex scenario threshold, the enhanced high-precision model is invoked for energy consumption prediction. If the scenario complexity does not meet the preset complex scenario threshold, the current scenario is determined to be a simple scenario, such as a plain, uniform speed, no congestion, and normal refueling stations, in which case the basic lightweight model is invoked for energy consumption prediction.

[0067] The high-precision model of the enhancement layer adopts a fusion architecture of XGBoost, LSTM, and Bayesian network, generated through serial fusion. XGBoost is used to process discrete features, LSTM is used to process temporal features, and Bayesian network is used to handle uncertain inference. Specifically, the input of the feature extraction layer XGBoost is discrete features, including but not limited to vehicle type, road surface type, weather labels, and discrete driving style rating labels; the output is a ranking of feature importance and a preliminary energy consumption baseline value. The input of the temporal prediction layer LSTM is a continuous time series and the output of the feature extraction layer XGBoost. The continuous time series includes but is not limited to speed curves, slope sequences, current changes, and temporal terrain matrix sequences, thereby capturing the dynamic trend of energy consumption changes with time / road conditions and outputting energy consumption prediction values. When the model is in online prediction mode, the starting point of the continuous time series is the current time, and the ending time is the time when the vehicle is expected to reach the end of the target route or the next refueling node according to the planned speed. The uncertainty calibration layer (Bayesian network) receives the prediction results of LSTM, i.e., the output of LSTM, and combines them with external uncertainties, such as queuing probability distribution, to output the final energy consumption prediction range and confidence level, correcting errors caused by random disturbances. For example, the final energy consumption prediction range is 100 kWh-110 kWh with a confidence level of 95%, meaning there is a 95% probability that the energy consumption will be between 100 kWh and 110 kWh. The input features of the enhanced high-precision model can include dynamic environmental data and multi-energy switching rules. Furthermore, tire wear and tire pressure can be introduced as input features, and the energy consumption prediction can be optimized by dynamically adjusting the rolling resistance coefficient. For example, when tire wear increases by 10%, the rolling resistance coefficient is increased by 5%; when tire pressure is 10% below the recommended value, the rolling resistance coefficient is increased by 8%.

[0068] Among them, driving style rating Determined by the following expression:

[0069]

[0070] in, , , These represent the number of times each of rapid acceleration, rapid deceleration, and sharp turns is triggered. For driving mileage, , , These correspond to the deduction weights. Rapid acceleration and deceleration are determined by comparing the vehicle's real-time longitudinal acceleration with corresponding preset thresholds; sharp turns do not rely on longitudinal acceleration but are determined by lateral acceleration and steering wheel operation. A sharp turn is recorded when the lateral acceleration exceeds 0.4g, or the steering wheel angular velocity (steering wheel angular acceleration) exceeds the set warning threshold. Data in the driving style score can be derived from a long-term statistical period for the driver or vehicle, such as the past week or the last complete long-distance transport mission.

[0071] Temporal terrain matrix sequence ,in, Indicates length, Indicates the slope value. The curvature is represented by a temporal vector containing the length, slope, and curvature of each road segment within the temporal terrain matrix sequence. The data in the temporal terrain matrix sequence is taken from the target candidate path to be evaluated, i.e., spatial road condition data slices from the route's starting point to its end point.

[0072] In one possible implementation, the lightweight model uses multiple linear regression to predict energy consumption, including:

[0073] Based on the actual driving mileage of discrete sub-segments, the absolute difference between the elevation of the end point and the starting point of the discrete sub-segment, vehicle load, and estimated average speed, the predicted energy consumption is determined by multiple linear regression.

[0074] Specifically, the lightweight model is determined using multiple linear regression combined with a rule lookup table. The multiple linear regression uses the following expression:

[0075]

[0076] In the formula, This represents the energy consumption predicted by the lightweight model. arrive All are multiple linear regression coefficients. This represents the actual mileage traveled on the currently calculated discrete sub-road segment. Indicates load capacity. This indicates the change in altitude, specifically the absolute difference between the elevation of the current sub-section's endpoint and its starting point. Indicates the estimated average speed. This represents the basic power consumption constant, such as the basic power consumption of an air conditioner.

[0077] The rule lookup table is a pre-defined empirical mapping matrix that records fixed energy consumption penalties that cannot be covered by a linear model under specific, highly discrete operating conditions. It contains key-value pairs such as {Scenario: Severe congestion, Penalty value: +5kWh}. The penalty value is a constant obtained through historical statistics. For example, in extreme crawling scenarios where the average vehicle speed is less than 10 km / h, the results of previous regression calculations can be directly applied... A fixed energy consumption bias is added to it. The energy consumption bias is the energy consumption penalty. In other words, under specific, highly discrete operating conditions, the final predicted energy consumption of the lightweight model will still be subject to a certain penalty. Add energy consumption bias to the basis The lightweight model in this embodiment of the invention has an extremely fast inference speed, typically less than fifty milliseconds, and the energy consumption prediction error in simple scenarios can be controlled within 10%.

[0078] Step 103: Using the NSGA-II algorithm and particle swarm optimization algorithm, with the objective function of minimizing the overall energy cost, the least time consumption, and the shortest refueling waiting time, and combining the hierarchical energy consumption prediction model to predict energy consumption, a set of candidate routes is generated; the objective function is constructed based on the first state data.

[0079] In one possible implementation, the NSGA-II algorithm and particle swarm optimization algorithm are used to minimize the overall energy cost, the shortest time consumption, and the shortest refueling waiting time as objective functions, and combined with the hierarchical energy consumption prediction model to predict energy consumption, in order to generate a candidate route set, including:

[0080] Based on the K-shortest path algorithm, multiple feasible paths are determined;

[0081] Based on the scene feature parameters in the multidimensional state data, a target model is selected in the hierarchical energy consumption prediction model to determine the predicted energy consumption of multiple feasible paths; the scene feature parameters include road slope fluctuation, ambient temperature, queuing probability of energy replenishment stations, cargo weight fluctuation, and average vehicle speed fluctuation rate.

[0082] Substitute the multiple feasible paths and their corresponding predicted energy consumptions into the objective function, perform multi-objective optimization based on the NSGA-II algorithm and particle swarm optimization algorithm, and select the Pareto optimal solution set to determine the candidate route set.

[0083] Specifically, based on the K-shortest path algorithm, K feasible paths are searched in the road network. For example, K can be set to 5, meaning 5 feasible paths are searched. For each path, the predicted energy consumption is calculated by substituting it into the energy consumption prediction model. Specifically, the choice between a lightweight model and a high-precision model is determined by the scene feature parameters in the multi-dimensional state data. If the road slope fluctuation exceeds 15 degrees, the ambient temperature is below -10 degrees Celsius or above 40 degrees Celsius, the queuing probability of the refueling station exceeds 60%, or the cargo weight fluctuation exceeds 20%, the enhanced layer high-precision model is called for energy consumption prediction. Otherwise, the basic layer lightweight model is called for energy consumption prediction.

[0084] Within the predicted detection window, such as 2 kilometers ahead or 3 minutes ahead, the difference between the maximum and minimum slope values ​​at the sampling point is the road slope fluctuation.

[0085] Queueing probability at charging stations Determined by the following expression:

[0086]

[0087] In the formula, This represents the sum of the number of guns used at the current station and the number of vehicles that have entered the site. Indicates the total number of stations. This represents the average queuing probability of the current station during the same historical period. This represents the real-time weighting factor, for example, 0.7 can be taken.

[0088] The objective function is solved based on the NSGA-II algorithm. Population initialization is performed for the selection of energy replenishment points, i.e., at which station to replenish and by how much. Then, the Pareto optimal solution set is selected through non-dominated sorting to form the final candidate route recommendation list.

[0089] objective function as follows:

[0090]

[0091]

[0092] In the formula, Indicates the overall cost of energy. Indicates the estimated power consumption. This indicates the electricity price for the refueling station. This indicates the estimated hydrogen consumption. This indicates the price of hydrogen at hydrogen refueling stations.

[0093]

[0094] In the formula, Indicates the total time elapsed. This indicates the estimated time for pure driving. This indicates the estimated time for physical insertion / refilling.

[0095]

[0096] In the formula, This indicates the waiting time for refueling. This represents the estimated number of vehicles queuing in front of the station upon arrival. This number is determined through dynamic queuing probability extrapolation. Indicates the number of guns available at the site. This indicates the historical average refueling time for a single vehicle.

[0097] This invention employs the NSGA-II algorithm combined with particle swarm optimization as the core optimization engine. The objective function is not limited to minimizing distance but is set as multiple objectives, including minimizing overall energy cost, minimizing time consumption, and minimizing refueling waiting time. The formula for calculating overall energy cost encompasses electricity costs, hydrogen costs, and the wear and tear costs of batteries or fuel cells.

[0098] In one possible implementation, after generating the candidate route set, the method further includes:

[0099] For each candidate route in the candidate route set, the corresponding energy replenishment and energy recovery strategies are matched to generate candidate routes that include the energy replenishment and energy recovery strategies.

[0100] In one possible implementation, the matching of corresponding energy replenishment and energy recovery strategies to generate candidate routes that include energy replenishment and energy recovery strategies includes:

[0101] If a long downhill or continuous downhill section is detected ahead, reduce the output power of the fuel cell stack or start a controllable auxiliary load.

[0102] In one possible implementation, the step of reducing the fuel cell stack output power or activating a controllable auxiliary load if a long downhill or continuous downhill section is detected ahead includes:

[0103] If a long downhill or multiple consecutive downhill sections are detected ahead, and the cumulative elevation difference of the downhill sections is greater than the preset elevation difference threshold, the maximum regenerative energy generated by the downhill sections is estimated based on the total mass of the vehicle, the current speed, and the expected driving trajectory.

[0104] The required battery rechargeable capacity is determined based on the maximum regenerative electrical energy.

[0105] If the current remaining free capacity of the battery is less than the rechargeable capacity of the battery, it is determined that it needs to be discharged in advance.

[0106] Predicted values ​​of the thermal state and health state of the power battery at the start of the downhill slope;

[0107] Based on the predicted thermal state value, the predicted health state value, and the battery rechargeable capacity, a corresponding safe state of charge upper limit is generated.

[0108] If the safe state of charge upper limit is less than the preset safe state of charge upper limit threshold, the controllable auxiliary load is actively activated, or the stack output power is reduced.

[0109] This invention also includes an energy recovery reservation mechanism for long downhill sections. Specifically, based on high-precision maps, real-time positioning, and real-time load information, a forward-looking analysis of the slope sequence of the road ahead is performed before entering a potentially high-recovery section. This allows for the identification of whether a long or continuous downhill section will occur. When a long or continuous downhill section is identified ahead, and the cumulative elevation difference exceeds a preset threshold, the state of charge (SOC) of the power battery is adjusted in advance. This is achieved by suppressing unnecessary power input (e.g., reducing fuel cell output power) or actively activating controllable auxiliary loads to consume power, ensuring that the battery retains sufficient rechargeable capacity before entering the downhill section. This maximizes the energy recovery capability and regenerative braking safety during the downhill phase.

[0110] Specifically, when a long or continuous downhill section is detected ahead, and the cumulative elevation difference exceeds a preset threshold, the maximum regenerative electrical energy that can be generated during the downhill process is estimated based on the vehicle's total mass, current speed, and expected trajectory, and the required battery rechargeable capacity is assessed accordingly. After obtaining the predicted thermal state and health state values, the Battery Management System (BMS) generates a corresponding safe state of charge (SOC) upper limit based on these values ​​and the battery's rechargeable capacity. Specifically, generating the safe SOC upper limit is a dynamic process of establishing safety boundaries and reserving recycling space. First, based on the predicted thermal and health states, the BMS queries its built-in safety boundary mapping table to determine the maximum physical SOC threshold allowed under that operating condition. This threshold can be lowered to prevent thermal runaway when high temperatures are predicted. Then, the maximum physical SOC threshold is subtracted from the battery's rechargeable capacity percentage to obtain the final safe SOC upper limit, i.e., the safe SOC upper limit. This safe SOC upper limit includes thermal safety margin, charging power limits, and voltage window constraints. Wherein, the percentage of battery rechargeable capacity is expressed as a percentage of the battery's rechargeable capacity. In this embodiment of the invention, the predicted thermal state value can be a predicted battery temperature value.

[0111] After obtaining the upper limit of the safe state of charge (SOC), if this upper limit is lower than the preset threshold, the upper limit of SOC is used as the target for SOC control. For pure electric heavy-duty trucks, controllable auxiliary loads (such as PTC heaters, air compressors, or cooling pumps) are activated to actively consume electrical energy under permissible operating conditions. For hydrogen fuel cell heavy-duty trucks, the stack output power is reduced in advance to suppress further SOC increases, or the coasting strategy is optimized to avoid unnecessary regeneration. For example, the upper limit threshold of the safe state of charge can be set to 90%. This ensures that the battery has sufficient spare capacity and thermal management margin before the vehicle enters a long downhill section to support efficient and stable regenerative braking, balancing energy recovery efficiency and driving safety.

[0112] Among them, the largest regenerative power Determined by the following expression:

[0113]

[0114] In the formula, Indicates the total mass of the vehicle. Represents gravitational acceleration. Indicates the elevation difference of the road section. This indicates the overall efficiency of the regenerative braking system, which includes the generator efficiency and inverter efficiency, and is generally between 0.6 and 0.8. This represents the energy consumed by rolling resistance and wind resistance.

[0115] The required rechargeable battery capacity equals the regenerative energy plus a safety margin. Specifically, the required rechargeable battery capacity... Determined by the following expression:

[0116]

[0117] In the formula, Indicates the current battery voltage. To represent the safety factor, for example, it can be... Set it to 0.1.

[0118] The thermal state prediction value is the battery temperature prediction value, which refers to the average predicted temperature of the cells inside the battery pack, i.e., the lumped node temperature corresponding to the system's equivalent heat capacity. This is obtained by collecting the battery's current real-time temperature, combining it with the predicted current sequence required before the vehicle reaches the downhill starting point, calculating the heat generation over a future period using Joule's law, and then subtracting the cooling system's heat dissipation before summing the results. The predicted current sequence is derived by inversely from the road gradient and vehicle speed planning. Specifically, it is based on the planned vehicle speed v and gradient... Calculate the total driving force of wheel end demand Then, the total driving force of wheel-end demand. Multiplying the planned vehicle speed by the wheel-end mechanical power, and then combining this with the transmission system efficiency and motor efficiency, the calculated output power demanded by the battery pack is obtained. Dividing the battery pack output power demanded by the current battery platform voltage yields the predicted current sequence. The expression for the total driving force demanded at the wheel ends is as follows:

[0119]

[0120] In the formula, This represents the total driving force of wheel-end demand, where v represents the planned vehicle speed. Indicates slope, Let g represent the total mass of the vehicle, and g represent the acceleration due to gravity. Indicates the tire rolling resistance coefficient. Indicates air density, Indicates the drag coefficient. Indicates the vehicle's frontal area. This indicates the longitudinal acceleration of the vehicle.

[0121] The predicted state of health (SOH) is calculated based on the battery's current static health (SOH) by looking up the corresponding temperature correction factor in a table using the predicted temperature. The predicted SOH is the product of the SOH and the temperature correction factor, representing the battery's actual usable capacity retention rate under specific future temperature conditions. It's important to note that the table used here refers to the temperature correction factor table, which pre-records the actual usable capacity retention rate values ​​for different ambient / cell temperature ranges.

[0122] The expression for predicting temperature is as follows:

[0123]

[0124] In the formula, Indicates the predicted temperature. Indicates the current temperature. Represents the predicted current sequence. Indicates the battery's internal resistance. Indicates the heat dissipation power of the cooling system. This indicates the equivalent heat capacity of the battery system.

[0125] Through the aforementioned potential energy feedback reservation mechanism for long downhill sections, the energy supply, storage and recovery capabilities are coordinated and matched in the spatiotemporal dimensions, which not only ensures the continuous effectiveness of regenerative braking and driving safety on long downhill sections, but also maximizes the energy utilization efficiency of the entire journey.

[0126] In one possible implementation, the matching of corresponding energy replenishment and energy recovery strategies to generate candidate routes that include energy replenishment and energy recovery strategies includes:

[0127] The load status of the target charging station is monitored. If the occupancy rate of the target charging station is greater than a preset occupancy rate threshold, a candidate station is generated and pushed to subsequent vehicles that plan to go to the target charging station. Both the subsequent vehicles and the current vehicle interact with the cloud.

[0128] This invention also implements a multi-vehicle collaborative peak-avoidance mechanism based on the occupancy rate of the target charging station, preventing other vehicles in the fleet from going to busy charging stations. For example, the occupancy rate threshold can be set to 80%.

[0129] In one possible implementation, the matching of corresponding energy replenishment and energy recovery strategies to generate candidate routes that include energy replenishment and energy recovery strategies includes:

[0130] Traffic flow is predicted for the target road segment based on a traffic flow prediction model, and the average vehicle speed when the vehicle arrives at the target road segment is estimated by combining the real-time traffic flow at the checkpoint upstream of the target road segment; the traffic flow prediction model is a spatiotemporal graph convolutional network.

[0131] Based on the average vehicle speed and the speed limit of the target road segment, determine whether the target road segment is a potentially congested route;

[0132] If the target road segment is determined to be a potentially congested route, the dynamic A* pathfinding algorithm is used to determine the bypass branch with the second smallest comprehensive cost, so as to generate alternative detour routes and push them to vehicles that have not yet departed; the comprehensive cost is determined based on the absolute distance and energy consumption data of the bypass branch and the expected increase in time caused by congestion.

[0133] Specifically, the multi-vehicle cooperative peak-avoidance mechanism in this embodiment of the invention also includes identifying potential congested routes based on real-time location data of multiple vehicles, generating alternative detour routes, and pushing them to new energy heavy trucks that have not yet departed, in order to alleviate traffic pressure. If the expected average vehicle speed is less than 30% of the speed limit of the target road segment, the target road segment is determined to be a potentially congested route. Then, using the dynamic A* (A-Star) pathfinding algorithm, the side impedance parameter of the aforementioned congested road segment is increased in the road network topology map, triggering the algorithm to automatically search for the bypass branch with the second smallest comprehensive cost as an alternative detour route.

[0134] In one possible implementation, before generating alternative detour routes and pushing them to vehicles that have not yet departed, the process further includes:

[0135] Determine the dynamic saturation threshold of the alternative detour routes;

[0136] The vehicles that have not yet departed are sorted, and virtual reservations are performed according to the alternative detour routes. The marginal energy consumption cost of the alternative detour routes is determined after each virtual reservation. The marginal energy consumption cost is the total increase in energy consumption caused by the reduction in the speed of the vehicles already on the alternative detour routes due to the addition of a new vehicle.

[0137] If the current virtual occupancy causes the number of vehicles in the alternative detour route to exceed the dynamic saturation threshold, or causes the marginal energy consumption cost to exceed the preset congestion worsening threshold, then the alternative detour route will stop being pushed to current and subsequent vehicles, and the departure time will be postponed or a new alternative detour route will be generated; the subsequent vehicles are vehicles that are traveling or have not yet departed and are planned to pass through the alternative detour route.

[0138] Specifically, the dynamic saturation threshold for heavy-duty trucks using alternative detour routes is calculated. This threshold is the maximum vehicle capacity calculated based on the minimum safe following distance, which is the minimum safe following distance required for heavy-duty trucks to maintain an economical speed. In this embodiment of the invention, a virtual occupancy operation is performed in a cloud-based model. When it is determined that alternative detour routes should no longer be pushed to current and subsequent vehicles, a time-domain traffic diversion strategy is activated, generating a time window instruction to postpone departure or a guidance instruction pointing to a new route, in order to achieve precise distribution of traffic pressure.

[0139] The marginal energy cost is determined by the following expression:

[0140]

[0141] In the formula, This represents the marginal energy cost. The energy consumption of a newly added vehicle refers to the energy consumption of the new vehicle after entering the road section, at its current average speed. The energy consumed by the vehicle while driving; the current average speed refers to the average speed after the traffic volume has decreased. This indicates the current number of vehicles, that is, the total number of vehicles that were already present and were currently in motion on this road segment; This represents the vehicle index, indicating the first vehicle in the existing vehicle set. A vehicle. The low flow rate after congestion refers to the decrease in average vehicle speed on a road segment due to increased traffic density after the addition of new vehicles. This indicates the original vehicle speed, that is, the average vehicle speed on the road segment before the new vehicle joined; Indicates the first The existing vehicles slowed down to The energy consumption value after that; Indicates the first The existing vehicle at its original speed The energy consumption value below.

[0142] In one possible implementation, determining the dynamic saturation threshold of the alternative detour routes includes:

[0143] Sum the average vehicle length with the minimum safe following distance under the current speed limit to obtain the sum value;

[0144] The dynamic saturation threshold is obtained by dividing the total length of the alternative detour routes by the sum.

[0145] Maximum vehicle capacity The expression is as follows:

[0146]

[0147] In the formula, Indicates the total length of the detour route. This indicates the average length of the heavy truck. This indicates the minimum safe following distance under the current speed limit v.

[0148] The multi-vehicle collaborative peak avoidance mechanism and traffic anti-oscillation control based on virtual occupancy in the embodiments of the present invention are intended to solve the secondary congestion caused by the "herding effect" that may be caused by simply pushing the same "optimal" alternative route to multiple vehicles.

[0149] For different road types, the strategies for coordinated energy replenishment and energy recovery exhibit significant differences. In urban areas, the strategy prioritizes off-peak charging, combining real-time charging station queue status, time-of-use pricing, and vehicle task scheduling information to plan for replenishment during off-peak hours when electricity prices are low, queues are shorter, and operational efficiency is not affected, thereby reducing electricity costs and improving depot turnover efficiency. In highway transportation scenarios, the timing and depth of replenishment are dynamically adjusted based on the distance between replenishment stations along the route, the vehicle's remaining range, and traffic flow forecasts to ensure sufficient range redundancy and resolutely avoid stoppages due to insufficient battery or hydrogen. In mountainous or hilly areas, the strategy shifts its focus to ensuring the safety of downhill energy recovery: for pure electric heavy-duty trucks, it is not recommended to charge the battery to an excessively high SOC before going uphill, to reserve sufficient capacity for subsequent regenerative braking on downhill slopes; for hydrogen fuel cell heavy-duty trucks, priority is given to hydrogen refueling stations with no queues and stable hydrogen supply, and the fuel cell stack power is actively adjusted before entering continuous downhill sections, taking into account the fuel cell output characteristics, to prevent the battery from being passively fully charged.

[0150] This invention introduces a terrain-based forward-looking energy recovery collaborative control logic. In conventional strategies, if a vehicle is in a high-charge state (e.g., SOC > 90%) after previous downhill energy recovery or charging, and dynamic SOC management is not implemented based on the road conditions ahead, the battery may trigger the charging limit of the battery management system (BMS) when entering a subsequent long downhill section. This results in a significant reduction or even complete disabling of regenerative braking torque. In this situation, the vehicle has to rely more on mechanical braking or auxiliary retarding devices, which not only wastes a large amount of gravitational potential energy but also easily leads to brake overheating and thermal fade, significantly increasing the safety risks of driving on long downhill sections.

[0151] Step 104: Determine the total empirical matching score for each candidate route in the candidate route set, and determine the optimal route based on the total empirical matching score.

[0152] Total score of experience matching Determined by the following expression:

[0153]

[0154] In the formula, Indicates the weight of battery life safety. This indicates a habitual fit weight. Indicates cost economic weight. Indicates the battery life and safety rating. The score indicates the compatibility of driving habits. This indicates the economic cost score.

[0155]

[0156] In the formula, This indicates the predicted remaining battery power upon arrival at the destination. This represents the set absolute safety lower limit; for example, it can be... Set to 15%.

[0157]

[0158] In the formula, This represents the penalty coefficient for speed deviation. Indicates the planned vehicle speed. This indicates the driver's habitual speed. This represents the penalty coefficient for road type deviation. This represents the quantitative assessment value of the road type currently in the plan. For example, a weight of 1 can be assigned to expressways, and a weight of 2 can be assigned to national highways. This represents a quantitative assessment value indicating the driver's historical road type preferences.

[0159] The driving habit fit scoring formula penalizes points based on the difference between the planned speed and the driver's habitual speed, as well as the difference between road type preferences (such as preference for highways or national roads).

[0160]

[0161] In the formula, This represents the highest estimated overall cost in the candidate route set. Indicates the estimated cost. This represents the lowest estimated overall cost in the candidate route set.

[0162] The above formula uses max-min normalization to map the estimated cost to a percentage score.

[0163] The total empirical matching score is determined using the above formula. Then, the optimal route, i.e., the target dynamic experience route, is determined based on the total experience matching score. Furthermore, a grading system is implemented based on the specific scores. Specifically, if the total experience matching score is greater than or equal to 85 points, it is classified as a green level, indicating a high degree of matching; if the total experience matching score is greater than or equal to 60 points but less than 85 points, it is classified as a yellow level, indicating a partial deviation; and if the total experience matching score is less than 60 points, it is classified as a red level, indicating a complete conflict.

[0164] In one possible implementation, after determining the total experience matching score and classifying it into levels, the method further includes:

[0165] The cloud sends human-machine interaction commands to the in-vehicle terminal to display the experience matching level of the target dynamic route on the terminal. The experience matching level is displayed by energy type and visually differentiated by color. The experience matching level includes a green level (highly matched), a yellow level (partially deviating), and a red level (completely conflicting). A high match means the route perfectly matches the current vehicle status and driving habits; partial deviation requires the driver to be aware of potential risks; and complete conflict means the route has significant energy consumption or refueling risks. It is important to note that, to minimize driver interference, voice and text prompts are only triggered near the end-of-range threshold, refueling points, and when switching scenarios.

[0166] The system receives driver feedback from the in-vehicle terminal. To simplify the process, driver feedback is reduced to "one-click accept" or "one-click reject," with a selected reason. If the driver rejects the target dynamic route, the reason for rejection (e.g., station malfunction, excessive price, route congestion) is recorded and used as negative sample data input into the energy consumption prediction model for subsequent model correction. This embodiment of the invention enhances user experience through human-computer interaction feedback steps.

[0167] The vehicle terminal is also equipped with a three-level caching system to optimize response speed. The vehicle terminal cache stores the driver's personal high-frequency experience, including parameters of habitual driving style and the status of refueling points on frequently used routes. High-frequency routes are cached for seven days, and low-frequency routes are cached for twenty-four hours. The edge node cache stores regional public experience, covering the gradient energy consumption mapping rules within a 50-kilometer radius. The cloud cache stores enterprise-level global experience.

[0168] After the transportation task is completed, the closed-loop optimization phase begins. Based on the actual operational feedback after transportation, the deviation between the actual operational data and the predicted data is calculated. If the deviation exceeds a preset threshold, an online incremental update process for the energy consumption prediction model is triggered. For example, the preset threshold can be set to ±10%.

[0169] The process of triggering an online incremental update of the energy consumption prediction model includes: using the online gradient descent (OGD) algorithm to fine-tune the parameters of the energy consumption prediction model based on actual operation data of a single transport and driver feedback data, without loading all historical data; the deviation value is calculated by comparing the actual and predicted electricity consumption, hydrogen consumption, refueling time, and transport time; after the model parameters are updated, the validated empirical data is stored in a multi-energy experience template library, which is classified and stored according to energy type, scenario type, and transport type. The online gradient descent (OGD) algorithm reduces the model update time to less than five seconds, greatly improving the system's iteration efficiency. When the deviation exceeds a threshold, particle swarm optimization can be used to optimize the weights and input features of the multiple models.

[0170] The deviation between the actual operating data and the predicted data is determined by the following expression:

[0171]

[0172] In the formula, This represents the combined deviation between actual operating data and predicted data. , , These are the weights for energy consumption deviation, time deviation, and SOC deviation, which can be set to 0.5, 0.3, and 0.2 respectively. Indicates predicted energy consumption. Indicates actual energy consumption. Indicates the predicted travel time. Indicates the actual driving time. This indicates a predicted SOC (State of Charge). This indicates the actual arrival at the State of the Occupation (SOC).

[0173] This invention utilizes incremental data to fine-tune model parameters, enabling self-iteration and calibration of individual vehicle characteristics and specific road segment experience during use.

[0174] The process of accumulating experience templates into a multi-energy experience library also includes performing a full-chain quantitative evaluation of experience: calculating an experience value score based on the product of cost savings and the number of times the experience is reused; selecting high-value experiences with an experience value score greater than or equal to 80, including off-peak energy replenishment strategies, energy-saving driving strategies in mountainous areas, and multi-energy switching timing strategies; marking high-value experiences as priority for promotion and updating them to the cloud cache for sharing by the entire fleet. This achieves a collective intelligent evolution where "one person drives, the whole team benefits."

[0175] The above process allows validated experience data to be stored in a multi-energy experience template library. This library is categorized and stored according to energy type (pure electric, hydrogen fuel cell), scenario type (mountainous, plains, urban), and transportation type (heavy load, light load). Templates include core parameters such as slope correction coefficients, temperature compensation coefficients, and refueling priority rules. High-value experiences selected include, but are not limited to, off-peak refueling strategies, energy-saving driving strategies in mountainous areas, and multi-energy switching timing strategies.

[0176] This invention constructs a hierarchical modeling engine that includes lightweight and high-precision models, and introduces an automatic scene complexity switching mechanism. It can intelligently select the prediction model based on road condition characteristics. In simple scenarios, the lightweight model is used to achieve millisecond-level response, while in complex scenarios, the high-precision model fused with deep learning is automatically called to ensure prediction accuracy. This not only ensures the reliability of range prediction under extreme conditions and avoids the risk of vehicle breakdown, but also significantly reduces the load pressure of large-scale concurrent computing in the cloud, achieving a dynamic optimal balance between energy consumption prediction accuracy and computational efficiency. This invention is not limited to a single shortest path planning, but is based on a multi-objective optimization algorithm to generate a comprehensive strategy covering refueling timing selection, energy recovery reservation, and operation path coordination. For pure electric and hydrogen fuel cell heavy trucks, it combines time-of-use electricity pricing, real-time queuing status of refueling / charging stations, and terrain features ahead. In urban areas, it recommends off-peak refueling to reduce energy costs; on highways, it dynamically avoids high-queue-risk stations to reduce waiting time; and in mountainous areas, it pre-adjusts the battery state of charge and reserves regenerative braking capacity for long downhill slopes. This strategy significantly reduces the overall energy cost of the entire journey, while effectively shortening the non-transportation time caused by refueling queues, thus improving overall operational efficiency.

[0177] This invention utilizes a deviation triggering mechanism and an online incremental learning algorithm to establish a complete closed loop from "prediction-execution-feedback" to "model update." Once the deviation between actual operating data and predicted data exceeds a threshold, model fine-tuning is immediately triggered, and validated high-confidence data is accumulated as experience templates. This mechanism enables the system to become increasingly accurate with prolonged operation, transforming individual driving deviations into shared collective experience, thus addressing the pain point of traditional navigation systems where outdated data leads to repeated errors.

[0178] This invention, by introducing experience confidence labels and filtering mechanisms during the data acquisition phase and combining a three-level caching system of vehicle terminals, edge nodes, and the cloud, effectively eliminates one-off or accidental noise data (such as abnormal energy consumption caused by temporary accidents), ensuring that all data entering the core computing engine has passed high-standard screening, enhancing the real usability and anti-interference ability of the data, and significantly improving the feasibility and reference value of the generated route suggestions and energy replenishment strategies in practical applications.

[0179] This invention, through a streamlined human-computer interaction feedback mechanism and tiered experience-based recommendation display, provides scientific data support while respecting the personal experience of experienced drivers. It only provides prompts at key decision points (such as near-limited range or scene transitions), avoiding information overload that could interfere with the driver. Furthermore, by recording the driver's reasons for rejecting commands, it optimizes the recommendation logic in reverse, achieving an organic integration of algorithmic intelligence and human experience, thus improving human-computer collaboration efficiency and the driving experience.

[0180] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions within the technical scope disclosed in the present invention should be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for generating dynamic empirical routes, characterized in that, include: Collect multi-dimensional state data across the entire link, filter valid data from the multi-dimensional state data, and generate first state data; A hierarchical energy consumption prediction model is constructed, which includes a lightweight model and a high-precision model. The high-precision model is generated by serial fusion of XGBoost, LSTM, and Bayesian network. The lightweight model uses multiple linear regression to predict energy consumption. The NSGA-II algorithm and particle swarm optimization algorithm are used to minimize the overall energy cost, the shortest time consumption, and the shortest refueling waiting time as the objective functions. The hierarchical energy consumption prediction model is combined to predict energy consumption and generate a set of candidate routes. The objective function is constructed based on the first state data. Determine the total empirical matching score for each candidate route in the candidate route set, and determine the optimal route based on the total empirical matching score.

2. The dynamic experience route generation method according to claim 1, characterized in that, The algorithm employs NSGA-II and particle swarm optimization, with the objective functions of minimizing overall energy cost, minimizing time consumption, and minimizing refueling waiting time. It also combines these with the hierarchical energy consumption prediction model to predict energy consumption and generate a candidate route set, including: Based on the K-shortest path algorithm, multiple feasible paths are determined; Based on the scene feature parameters in the multidimensional state data, a target model is selected in the hierarchical energy consumption prediction model to determine the predicted energy consumption of multiple feasible paths; the scene feature parameters include road slope fluctuation, ambient temperature, queuing probability of energy replenishment stations, cargo weight fluctuation, and average vehicle speed fluctuation rate. Substitute the multiple feasible paths and their corresponding predicted energy consumptions into the objective function, perform multi-objective optimization based on the NSGA-II algorithm and particle swarm optimization algorithm, and select the Pareto optimal solution set to determine the candidate route set.

3. The dynamic experience route generation method according to claim 1, characterized in that, After generating the candidate route set, the process further includes: For each candidate route in the candidate route set, the corresponding energy replenishment and energy recovery strategies are matched to generate candidate routes that include the energy replenishment and energy recovery strategies.

4. The dynamic experience route generation method according to claim 3, characterized in that, The matching of corresponding energy replenishment and energy recovery strategies, and the generation of candidate routes that include energy replenishment and energy recovery strategies, include: If a long downhill or continuous downhill section is detected ahead, reduce the output power of the fuel cell stack or start a controllable auxiliary load.

5. The dynamic experience route generation method according to claim 4, characterized in that, If a long downhill or continuous downhill section is detected ahead, reducing the fuel cell output power or activating a controllable auxiliary load includes: If a long downhill or multiple consecutive downhill sections are detected ahead, and the cumulative elevation difference of the downhill sections is greater than the preset elevation difference threshold, the maximum regenerative energy generated by the downhill sections is estimated based on the total mass of the vehicle, the current speed, and the expected driving trajectory. The required battery rechargeable capacity is determined based on the maximum regenerative electrical energy. If the current remaining free capacity of the battery is less than the rechargeable capacity of the battery, it is determined that it needs to be discharged in advance. Predicted values ​​of the thermal state and health state of the power battery at the start of the downhill slope; Based on the predicted thermal state value, the predicted health state value, and the battery rechargeable capacity, a corresponding safe state of charge upper limit is generated. If the safe state of charge upper limit is less than the preset safe state of charge upper limit threshold, the controllable auxiliary load is actively activated, or the stack output power is reduced.

6. The dynamic experience route generation method according to claim 3, characterized in that, The matching of corresponding energy replenishment and energy recovery strategies, and the generation of candidate routes that include energy replenishment and energy recovery strategies, include: The load status of the target charging station is monitored. If the occupancy rate of the target charging station is greater than a preset occupancy rate threshold, a candidate station is generated and pushed to subsequent vehicles that plan to go to the target charging station. Both the subsequent vehicles and the current vehicle interact with the cloud.

7. The dynamic experience route generation method according to claim 3, characterized in that, The matching of corresponding energy replenishment and energy recovery strategies, and the generation of candidate routes that include energy replenishment and energy recovery strategies, include: Traffic flow is predicted for the target road segment based on a traffic flow prediction model, and the average vehicle speed when the vehicle arrives at the target road segment is estimated by combining the real-time traffic flow at the checkpoint upstream of the target road segment; the traffic flow prediction model is a spatiotemporal graph convolutional network. Based on the average vehicle speed and the speed limit of the target road segment, determine whether the target road segment is a potentially congested route; If the target road segment is determined to be a potentially congested route, the dynamic A* pathfinding algorithm is used to determine the bypass branch with the second smallest comprehensive cost, so as to generate alternative detour routes and push them to vehicles that have not yet departed; the comprehensive cost is determined based on the absolute distance and energy consumption data of the bypass branch and the expected increase in time caused by congestion.

8. The dynamic experience route generation method according to claim 7, characterized in that, Before generating alternative detour routes and pushing them to vehicles that have not yet departed, the process also includes: Determine the dynamic saturation threshold of the alternative detour routes; The vehicles that have not yet departed are sorted, and virtual reservations are performed according to the alternative detour routes. The marginal energy consumption cost of the alternative detour routes is determined after each virtual reservation. The marginal energy consumption cost is the total increase in energy consumption caused by the reduction in the speed of the vehicles already on the alternative detour routes due to the addition of a new vehicle. If the current virtual occupancy causes the number of vehicles in the alternative detour route to exceed the dynamic saturation threshold, or causes the marginal energy consumption cost to exceed the preset congestion worsening threshold, then the alternative detour route will stop being pushed to current and subsequent vehicles, and the departure time will be postponed or a new alternative detour route will be generated; the subsequent vehicles are vehicles that are traveling or have not yet departed and are planned to pass through the alternative detour route.

9. The dynamic experience route generation method according to claim 8, characterized in that, Determining the dynamic saturation threshold of the alternative detour routes includes: Sum the average vehicle length with the minimum safe following distance under the current speed limit to obtain the sum value; The dynamic saturation threshold is obtained by dividing the total length of the alternative detour routes by the sum.

10. The dynamic experience route generation method according to claim 1, characterized in that, The lightweight model uses multiple linear regression to predict energy consumption, including: Based on the actual driving mileage of discrete sub-segments, the absolute difference between the elevation of the end point and the starting point of the discrete sub-segment, vehicle load, and estimated average speed, the predicted energy consumption is determined by multiple linear regression.

Images