Unmanned vehicle fleet energy supplement scheduling system and method based on multi-factor dynamic threshold

Abstract

The application provides a kind of unmanned vehicle fleet energy supplement scheduling system and method based on multi-factor dynamic threshold, by fusing station real-time transport supply and demand state, vehicle power distribution, historical vehicle usage law and charging resource constraint and other multidimensional data, construct multi-factor dynamic threshold model (Tnotify=TbasexF_supplyxF_demandxF_charging_capacity), replace traditional fixed power threshold, realize the dynamic matching of energy supplement trigger timing and operation rhythm;At the same time, introduce dynamic power-on threshold mechanism, quantify the decision basis of "whether it is meaningful to power on now";After energy supplement task generation is simulated and verified by digital twin system, it is issued to real RMS system for execution after verification, and the threshold model parameters are continuously optimized through closed-loop feedback mechanism, realize the intelligent energy supplement control of "multi-dimensional perception-dynamic decision-simulation verification-closed loop evolution".

Description

Technical Field

[0001] This invention relates to the field of energy replenishment scheduling technology, specifically to an unmanned vehicle fleet energy replenishment scheduling system and method based on multi-factor dynamic thresholds. Background Technology

[0002] The current power replenishment and dispatching scheme for the unmanned vehicle fleet is as follows: Method 1 (Passive Response): When the vehicle's battery level drops below a fixed percentage (e.g., 40%), a charging notification is triggered, and the dispatcher notifies the battery swapper offline to perform the task. This method is common practice in the autonomous vehicle industry and is disclosed in numerous charging scheduling patents in the logistics and delivery field.

[0003] Method 2 (Timed Polling): The system periodically scans the battery levels of all vehicles, sorts them from lowest to highest, and assigns charging tasks when a preset threshold is reached. These tasks are then processed by battery swappers on a first-come, first-served basis. This is the industry standard practice.

[0004] Method 3 (existing electric vehicle charging scheduling patent): This method focuses on optimizing queuing at public charging stations for passenger vehicles, addressing the path and queuing problem from "user to charging station," and employing genetic algorithms or reinforcement learning to solve for the optimal charging station allocation.

[0005] Common drawbacks of existing technologies: (1) Both Method 1 and Method 2 use a fixed power threshold to trigger charging, without considering dynamic factors such as the real-time supply and demand status of the station, the fluctuation of demand during the time period, and the constraints of charging resources. This results in insufficient vehicles available to accept orders during peak periods or premature charging during off-peak periods, which wastes manpower.

[0006] (2) The existing charging module and scheduling module operate independently. The charging decision is unaware of the upcoming scheduling demand, and the scheduling decision is unaware of the real-time distribution of charging resources. The information asymmetry between the two leads to task conflicts.

[0007] (3) The battery swappers passively wait for charging notifications and cannot predict the workload in advance, resulting in a backlog of tasks during peak hours and idle waste during off-peak hours.

[0008] (4) The timing of power-on depends entirely on human experience and cannot quantify whether "powering on now is meaningful". This may result in premature power-on when there is no demand for vehicles at the station, which will increase energy consumption.

[0009] (5) Although Method 3 uses an optimization algorithm, it is aimed at the selection of charging piles and route planning in the passenger car scenario. Its core contradiction is "users finding charging piles"; while the core contradiction in the unmanned delivery fleet scenario is "dynamic balance between charging behavior and operational efficiency" - vehicles are both production tools and charging demanders. Every charging decision directly affects the supply of transportation capacity. The problem domains of the two are completely different.

[0010] Patent application CN119758991A discloses a scheduling method, device, and cloud control platform for unmanned vehicles. The method includes: when an unmanned vehicle needs recharging, obtaining the number of unmanned vehicles currently performing recharging tasks and the capacity of a target recharging area; if the number of unmanned vehicles performing recharging tasks is greater than or equal to the capacity of the target recharging area, determining, based on the recharging task execution status of the unmanned vehicles in the target recharging area, whether joining the target recharging area after a first time period would exceed the capacity limit of the target recharging area; the first time period is the time taken for the current unmanned vehicle to travel from its current location to the target recharging area; if the capacity limit is not exceeded, controlling the current unmanned vehicle to perform the recharging task; if the capacity limit is exceeded, controlling the current unmanned vehicle to continue performing the current task. However, this patent cannot completely solve the existing technical problems, nor can it meet the needs of this invention. Summary of the Invention

[0011] To address the shortcomings of existing technologies, the purpose of this invention is to provide an unmanned vehicle fleet refueling scheduling system and method based on multi-factor dynamic thresholds.

[0012] The unmanned vehicle fleet replenishment scheduling system based on multi-factor dynamic threshold provided by the present invention includes: a capacity perception module, a dynamic threshold calculation module, a replenishment task generation and priority ranking module, a digital twin simulation verification module, and a closed-loop feedback and model adaptation module. The capacity sensing module collects multi-dimensional capacity status data of each charging and swapping station in real time. The multi-dimensional capacity status data includes the number of vehicles currently available for orders, the number of vehicles charging, the estimated available time distribution of vehicles charging, and the predicted vehicle demand within a future preset time window. The capacity sensing module is connected to the dynamic threshold calculation module and sends the multi-dimensional capacity status data to the dynamic threshold calculation module. The dynamic threshold calculation module receives the multi-dimensional capacity status data and calculates the dynamic charging notification threshold and dynamic power-on threshold for each station and time period. The dynamic threshold calculation module is connected to the energy replenishment task generation and priority sorting module and sends the calculated dynamic charging notification threshold and dynamic power-on threshold to the energy replenishment task generation and priority sorting module. The energy replenishment task generation and priority sorting module generates energy replenishment tasks when the vehicle's battery level reaches the dynamic charging notification threshold, and sorts them according to a preset priority rule; the energy replenishment task generation and priority sorting module shares capacity status data with the scheduling system and performs conflict detection when generating energy replenishment tasks; the energy replenishment task generation and priority sorting module is connected to the digital twin simulation verification module and sends the generated energy replenishment tasks to the digital twin simulation verification module. The digital twin simulation verification module receives the power replenishment task and synchronizes it to the digital twin system for simulation verification; the digital twin simulation verification module connects to the real RMS system and sends the verified power replenishment task to the RMS system for execution; The closed-loop feedback and model adaptation module is connected to the dynamic threshold calculation module and the digital twin simulation verification module. It records the decision context data and execution results of each energy replenishment decision to construct a feedback dataset, and periodically uses the accumulated feedback dataset to optimize the model parameters of the dynamic threshold calculation module.

[0013] Preferably, the dynamic charging notification threshold Tnotify(s,t) is calculated using the following formula: Tnotify(s,t)=Tbase×F_supply(s,t)×F_demand(s,t)×F_charging_capacity(s,t) Where s is the station, t is the time, Tbase is the baseline charging notification percentage; F_supply(s,t) is the capacity supply factor; F_demand(s,t) is the demand urgency factor; and F_charging_capacity(s,t) is the charging resource constraint factor.

[0014] Preferably, the capacity supply factor F_supply(s,t) is calculated using the following formula: ratio=Navail(s,t) / Dforecast(s,t,1h) F_supply=clip(1+k1×(ratio-1),0.6,1.4) Wherein, ratio is the ratio of the current number of vehicles available for order taking to the predicted demand in the next hour, Navail(s,t) is the number of vehicles available for order taking at station s at time t, Dforecast(s,t,1h) is the predicted demand for vehicles at station s in the next hour time window, clip(1+k1×(ratio-1),0.6,1.4) is the cutoff function, which takes 0.6 when 1+k1×(ratio-1)<0.6, takes 1.4 when 1+k1×(ratio-1)>1.4, and takes 1+k1×(ratio-1) otherwise, where k1 is the sensitivity coefficient.

[0015] Preferably, the demand urgency factor F_demand(s,t) is calculated using the following formula: slope=(Dforecast(4h)-Dforecast(1h)) / max(Dforecast(1h),1) F_demand=clip(1-k2×slope,0.7,1.3) Where slope is the relative slope of the three-window forecast, Dforecast(4h) is the predicted vehicle demand of station s in the next 4-hour time window, Dforecast(1h) is the predicted vehicle demand of station s in the next 1-hour time window, k2 is the sensitivity coefficient, and clip is the cutoff function.

[0016] Preferably, the charging resource constraint factor F_charging_capacity(s,t) is calculated using the following formula: ρ=used_charging_num / total_charging_num λ=max(0,total_charging_num-used_charging_num) / N_staff F_charging_capacity=1+k3×ρ+k4×max(0,λ-λ0) Where ρ is the charging position occupancy rate, used_charging_num is the number of occupied charging piles, total_charging_num is the total number of charging piles, λ is the number of idle charging positions that a single battery swapper needs to oversee, N_staff is the number of battery swappers on duty, λ0 is the upper limit threshold of idle positions within the capacity of a single battery swapper, k3 is the occupancy rate sensitivity coefficient, k4 is the manpower reserve sensitivity coefficient, and max(0,λ-λ0) is the maximum value function.

[0017] Preferably, the dynamic power-on threshold is calculated using the following formula: Tpower_on(s,t)=Eroute_min(s)+Emargin(s,t) Where Tpower_on(s,t) is the dynamic power-on threshold, Eroute_min(s) is the estimated power consumption of the shortest available route from station s, Emargin(s,t) is the dynamic margin, and t is the time.

[0018] Preferably, the energy replenishment task generation and priority sorting module sorts the generated energy replenishment tasks from high to low according to the following priorities: First priority: Vehicle battery level is below Eroute_min(s); Second priority: The current station Navail is less than Dforecast(1h), and the vehicle is a potentially dispatchable vehicle; Third priority: Regular charging tasks, sorted by battery level from low to high; Where Navail is the number of vehicles currently available for order taking, Dforecast(1h) is the predicted vehicle demand within the next 1-hour time window, and Eroute_min(s) is the estimated power consumption of the shortest route among all available routes departing from station s.

[0019] Preferably, the digital twin simulation verification module converts the content of the energy replenishment task into simulation parameters in the digital twin system. The content includes the target vehicle, charging station, and estimated charging time. It then simulates changes in station capacity over a preset time period within the digital twin environment, simulating the recovery of the number of vehicles available for order acceptance at each station after energy replenishment. Finally, it compares the core KPI indicators before and after the simulation. These core KPI indicators include the number of vehicles available for order acceptance, order matching success rate, and average user waiting time. If the simulation results show improvement in the core KPI indicators, the module marks the verification as successful and sends the energy replenishment command to the RMS system for execution.

[0020] Preferably, the decision context data recorded by the closed-loop feedback and model adaptation module includes: the dynamic charging notification threshold Tnotify value at the time of decision-making, the values ​​of each factor function F_supply, F_demand, and F_charging_capacity, and a snapshot of the station's capacity status; the execution results include: whether a charging and scheduling conflict occurs, how long after power-on the vehicle is selected by the scheduling system to execute the delivery task, whether the vehicle fails the task due to low battery, and a comparison of the number of vehicles that can accept orders at the station before and after the charging is completed; The closed-loop feedback and model adaptation module utilizes the feedback dataset to optimize the parameters of each factor function in the dynamic threshold calculation module using machine learning methods. The machine learning method aims to minimize the composite loss function. L = w_a×P1 + w_b×E + w_c×P2 Where L is the composite loss, P1 is the probability of charging and scheduling conflict, E is the expected value of the waiting time from power-on to the first task assignment, P2 is the probability of task failure due to low battery, and w_a, w_b, and w_c are preset weighting coefficients.

[0021] The unmanned vehicle fleet replenishment scheduling method based on multi-factor dynamic thresholds provided by the present invention performs the following steps: Step 1: Collect multi-dimensional capacity status data for each charging and swapping station in real time. The multi-dimensional capacity status data includes the number of vehicles currently available for order taking, the number of vehicles currently charging, the estimated available time distribution of vehicles currently charging, and the predicted vehicle demand within a future preset time window. Step 2: Receive the multi-dimensional capacity status data and calculate the dynamic charging notification threshold and dynamic power-on threshold for each station and time period; Step 3: When the vehicle's battery level reaches the dynamic charging notification threshold, a charging task is generated and sorted according to a preset priority rule; when generating the charging task, the capacity status data is shared with the scheduling system and conflict detection is performed; Step 4: Synchronize the generated power replenishment task to the digital twin system for simulation verification, and send the verified power replenishment task to the real RMS system for execution; Step 5: Record the decision context data and execution results of each energy replenishment decision to build a feedback dataset, and periodically use the accumulated feedback dataset to optimize the model parameters in the dynamic threshold calculation step.

[0022] Compared with the prior art, the present invention has the following beneficial effects: (1) By adopting a product model of capacity supply factor, demand urgency factor and charging resource constraint factor to dynamically calculate the charging notification threshold, the traditional fixed threshold method is replaced, which solves the problem that the fixed threshold cannot adapt to the fluctuation of the operation scenario, and makes the timing of energy replenishment always match the actual operation demand. (2) By adopting a dynamic power-on threshold, the decision-making basis for "whether the vehicle should be immediately restored to the order-accepting state after charging is completed" is quantified, which solves the problem that the power-on timing judgment relies entirely on human experience and cannot be quantified. (3) By sharing capacity status data with the scheduling system and performing conflict detection in the energy replenishment task generation module, the task conflict problem of "being scheduled as soon as charging is dispatched" caused by the separation of charging decision and scheduling decision is solved. (4) By simulating and verifying the energy replenishment decision in the digital twin system before issuing it to the real system for execution, the problem of "the effect is not as expected only after the energy replenishment decision is issued" is solved, and the decision risk is moved from the real operating environment to the simulation environment for digestion. (5) By recording the decision context and execution results through a closed-loop feedback mechanism and periodically optimizing the threshold model parameters, the problem that traditional energy replenishment strategies cannot adapt to evolution is solved, and the system can continuously adapt to changes in the operating environment. Attached Figure Description

[0023] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a diagram of the overall system architecture. Figure 2 Here is a flowchart for dynamic threshold calculation; Figure 3 Flowchart for avoiding conflicts in energy replenishment scheduling; Figure 4 This is a flowchart for digital twin simulation verification. Detailed Implementation

[0024] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0025] Example 1 This invention provides an unmanned vehicle fleet refueling scheduling system based on multi-factor dynamic thresholds, comprising the following modules: (1) Capacity perception module This module collects the following multi-dimensional capacity status data for each charging and battery swapping station in real time: Navail(s,t): The number of vehicles that can accept orders at station s at time t, which is obtained by querying the vehicle table for vehicles that are currently at station s, have no ongoing tasks, and whose real-time battery level is not lower than the current charging notification threshold. Ncharging(s,t): The number of vehicles charging at station s at time t, obtained from the records in the charging record table associated with station s and whose status is "charging"; Echarging(s,t): The estimated available time distribution of vehicles currently charging, calculated based on the initial charge amount, current charging duration, and historical charging duration statistics; Dforecast(s,t,Δt): The predicted demand for vehicles at station s within the future time window Δt, obtained by time-series prediction based on historical order data according to the dimensions of (station, time period, week).

[0026] Of the data mentioned above, Navail and Ncharging are obtained directly from the RMS business database in real time, Echarging is estimated through the initial charge amount and historical charging duration distribution in the charging record table, and Dforecast is obtained through a time-series prediction model. This module sends the above four types of data to the dynamic threshold calculation module.

[0027] (2) Dynamic threshold calculation module This module receives multi-dimensional status data sent by the capacity sensing module and calculates the dynamic charging notification threshold and dynamic power-on threshold for each station at each time period.

[0028] Dynamic charging notification threshold calculation: Tnotify(s,t)=Tbase×F_supply(s,t)×F_demand(s,t)×F_charging_capacity(s,t) Where Tbase is the baseline charging notification battery percentage (system initial default value, such as 40%), which serves as the baseline anchor point for threshold calculation; F_supply(s,t) is the capacity supply factor, calculated by the function f(Navail / Dforecast). When there are enough vehicles available to accept orders, the factor is greater than 1 (increase the threshold and charge in advance), and when the capacity is tight, the factor is less than 1 (decrease the threshold and delay charging to preserve capacity). The function f takes the form of a linear mapping plus truncation: ratio=Navail(s,t) / Dforecast(s,t,1h) F_supply=clip(1+k1×(ratio-1),0.6,1.4) Wherein, F_supply is the capacity supply factor (dimensionless), with a value range of 0.6~1.4, used to adjust the charging notification threshold according to the capacity supply-demand ratio; ratio is the ratio of the number of vehicles currently available for order acceptance to the predicted demand in the next hour, i.e., ratio=Navail(s,t) / Dforecast(s,t,1h); k1 is the sensitivity coefficient, with a default value of 0.3, updated by closed-loop learning in step (5); clip(x,0.6,1.4) is the cutoff function, which takes 0.6 when x<0.6, 1.4 when x>1.4, and x itself otherwise, to prevent extreme ratios from causing threshold jumps; F_supply maps the "capacity supply-demand ratio" to the threshold multiplier, making the energy replenishment behavior a tool for peak shaving and valley filling of capacity, rather than an isolated power management action; F_demand(s,t) is the demand urgency factor, calculated by the function g(Dforecast gradient change in the next 1h / 2h / 4h). The factor is less than 1 during the demand rise period (postponing charging to ensure transport capacity) and greater than 1 during the demand fall period (charging in advance to utilize off-peak resources). The relative slope of function g based on three-window prediction: slope=(Dforecast(4h)-Dforecast(1h)) / max(Dforecast(1h),1) F_demand=clip(1-k2×slope,0.7,1.3) Wherein, F_demand is the demand urgency factor (dimensionless), with a value range of 0.7~1.3, used to adjust the charging notification threshold according to the demand trend; slope is the relative slope of the three-window prediction, slope=(Dforecast(4h)-Dforecast(1h)) / max(Dforecast(1h),1), representing the upward or downward trend of future demand; k2 is the sensitivity coefficient, with a default of 0.4, updated by closed-loop learning in step (5); clip is the truncation function, used to prevent extreme slopes from causing threshold jumps; F_demand allows the algorithm to not only look at "whether there is a shortage of cars now", but also "the trend in the next few hours", upgrading from passive response to active planning.

[0029] F_charging_capacity(s,t) is the charging resource constraint factor, which is calculated by the function h(total number of charging piles, number of occupied charging piles, number of battery swappers on duty). When charging spaces are scarce, the factor is greater than 1 (encouraging early charging to avoid queuing), and when there are plenty of charging spaces, the factor approaches 1.

[0030] Define two intermediate quantities and then linearly superimpose them: ρ=used_charging_num / total_charging_num λ=max(0,total_charging_num-used_charging_num) / N_staff F_charging_capacity=1+k3×ρ+k4×max(0,λ-λ0) F_charging_capacity is the charging resource constraint factor (dimensionless), used to adjust the charging notification threshold by a multiplier according to the charging resource shortage; ρ is the charging position occupancy rate, ρ=used_charging_num / total_charging_num, with a value of 0~1; λ is the number of idle charging positions that a single battery swapper needs to take care of, λ=max(0,total_charging_num-used_charging_num) / N_staff; λ0 is the upper limit threshold of the number of idle positions that a single battery swapper can "take over", with a default value of 2; k3 is the occupancy rate sensitivity coefficient, with a default value of 0.5; k4 is the manpower reserve sensitivity coefficient, with a default value of 0.1; max(0,λ-λ0) is the maximum value function, which takes 0 when λ≤λ0 and (λ-λ0) when λ>λ0; k3, k4, and λ0 are all updated by closed-loop learning in step (5); the actual meaning is to encode the dual constraints of "charging resources + human resources" into a single multiplier.

[0031] Dynamic power-on threshold calculation: Tpower_on(s,t)=Eroute_min(s)+Emargin(s,t) Wherein: Eroute_min(s) is the estimated power consumption of the shortest route among all available routes departing from station s (average estimated power consumption from the route operation statistics table); Emargin(s,t) is a dynamic margin weighted based on demand urgency. The margin decreases when demand is urgent (the vehicle is put into use when powered on after reaching the minimum executable power), and increases when demand is loose (the vehicle is charged for a longer time to reduce the charging frequency). This module sends the calculated Tnotify and Tpower_on to the energy replenishment task generation module.

[0032] (3) Energy replenishment task generation and priority sorting module This module receives the Tnotify value sent by the dynamic threshold calculation module, automatically generates an energy replenishment task when the vehicle power reaches the threshold, and pushes it to battery swap operators according to the following priority sorting: First priority P0 (urgent): the vehicle power is already lower than Eroute_min(s), which means it cannot perform any delivery tasks, and must be replenished immediately; Second priority P1 (high): Navail of the current station < Dforecast(1h), that is, short-term prediction shows that the number of available vehicles for receiving orders is insufficient, and this vehicle is a potential schedulable vehicle; Third priority P2 (regular): a regular charging task, sorted from lower power to higher power.

[0033] In addition, this module shares capacity status data with the scheduling system, checks whether the vehicle has been selected by the scheduling system to perform a delivery task when generating an energy replenishment task. If it has been selected, it will mark the conflict and postpone the energy replenishment, so as to avoid the contradiction of "being scheduled immediately after being assigned a charging task" from the source. This module sends the generated energy replenishment task to the digital twin simulation verification module.

[0034] (4) Digital twin simulation verification module Before the energy replenishment task is issued to the real RMS system, this module synchronizes it to the digital twin system for simulation verification. The specific steps are as follows: Convert the content of the energy replenishment task (target vehicle, charging station, estimated charging duration) into simulation parameters in the digital twin system; Deduce the change of station capacity within a future T period (e.g., 1 hour) in the digital twin environment, and simulate the recovery of the number of available vehicles for receiving orders at each station after energy replenishment is completed; Compare the core KPI indicators before and after simulation, including the number of available vehicles for receiving orders, order matching success rate, and average user waiting time; If the simulation results show that the core KPI indicators have improved, it will be marked as "verification passed" and the power replenishment command will be sent to the RMS system for execution; if the simulation results do not meet the standards, the power replenishment task will be returned to the power replenishment task generation module for re-optimization or the power replenishment scope will be expanded.

[0035] This module addresses the problem of "disappointing results only after a power replenishment decision has been issued" by absorbing decision-making risks in a simulated environment.

[0036] (5) Closed-loop feedback and model adaptation module This module records the context data and execution results of each energy replenishment decision, constructing a feedback dataset for continuous optimization of the threshold model parameters. The recorded content includes: Decision context: dynamic threshold Tnotify value at the time of decision, value of each factor function (F_supply, F_demand, F_charging_capacity), and site capacity status snapshot (Navail, Ncharging, Dforecast); Execution results: whether charging and scheduling conflicts occurred, how long after power-on the vehicle was selected by the scheduling system to execute the delivery task, whether the vehicle failed the task due to low battery, and a comparison of the number of vehicles that can accept orders at the station before and after the power replenishment was completed.

[0037] This module periodically utilizes accumulated feedback data and optimizes the parameters of each factor function f, g, and h in the dynamic threshold calculation module through machine learning methods (such as gradient boosting regression). This enables the system to evolve adaptively and adapt to changes in the operating environment, such as the addition of new stations, seasonal changes, and fleet expansion, without the need for manual recalibration of the threshold.

[0038] Minimize the composite loss function: L = w_a×P1 + w_b×E + w_c×P2 w_a, w_b, w_c represent the business-side trade-offs (default 0.5 / 0.3 / 0.2), satisfying w_a+w_b+w_c=1; P1 is the probability of a charging and scheduling conflict (value 0~1), i.e., the probability that a task will be selected by the scheduling system after being dispatched for charging; P2 is the probability of a task failing due to low battery (value 0~1), i.e., the proportion of tasks that are not fully completed due to insufficient battery power; E is the expected value of the waiting time from power-on to the first dispatch of a task (unit: minutes).

[0039] Figure 1 The module demonstrates the data flow and collaborative relationships between the capacity perception module, dynamic threshold calculation module, energy replenishment task generation module, digital twin simulation verification module, and closed-loop feedback and model adaptation module.

[0040] Figure 2The calculation process of Tnotify and Tpower_on is demonstrated, including the integration method of capacity supply factor, demand urgency factor and charging resource constraint factor.

[0041] Figure 3 The system demonstrates the conflict detection logic during the generation of energy replenishment tasks and its linkage mechanism with the scheduling system.

[0042] Figure 4 The simulation verification demonstrates the input / output / judgment logic and feedback iteration closed loop. TH1~TH4 are configured by the operation of each site, with a default 1-hour window. KPI comparison = baseline before execution vs. feedback capacity prediction model for continuous optimization.

[0043] Example 2 This invention provides a method for scheduling refueling of unmanned vehicle fleets based on multi-factor dynamic thresholds, applicable to unmanned vehicle fleets in urban last-mile delivery scenarios, specifically including the following steps: Step 1: Capacity perception and data collection.

[0044] The capacity awareness module polls the status data of each charging and swapping station at fixed time intervals (e.g., 30 seconds). Specifically, for each station s and the current time t, the module queries the vehicle management database to obtain a list of vehicles at station s that are in an "idle" state and whose real-time battery level is not lower than the current dynamic charging notification threshold, and calculates the number of vehicles available for order taking, Navail(s,t). By querying the charging record table for vehicle records associated with station s and in a "charging" state, the module calculates the number of vehicles currently charging, Ncharging(s,t). For each vehicle currently charging, based on its initial charge amount, charging duration, and the historical charging rate distribution of that vehicle model (fitted from charging records over the past 30 days), the module calculates its estimated available time distribution, Echarging(s,t), which is output as, for example, "estimated remaining charging time". Simultaneously, this module invokes a time-series forecasting engine. Based on historical order data from the same time period over the past 60 days (e.g., Monday morning 9-10 AM), and combined with special event factors of the day (such as holiday promotions), it predicts the vehicle demand for station s within the next 1-hour, 2-hour, and 4-hour time windows, denoted as Dforecast(s,t,1h), Dforecast(s,t,2h), and Dforecast(s,t,4h), respectively. All of the above data is packaged into a multi-dimensional state vector and sent to the dynamic threshold calculation module.

[0045] Step 2: Dynamic threshold calculation.

[0046] After receiving the multidimensional state vector, the dynamic threshold calculation module calculates the dynamic charging notification threshold Tnotify(s,t) and the dynamic power-on threshold Tpower_on(s,t) respectively.

[0047] The dynamic charging notification threshold is calculated as follows: Tnotify(s,t)=Tbase×F_supply(s,t)×F_demand(s,t)×F_charging_capacity(s,t) Where Tbase is the baseline charging notification battery percentage, initially set by operators based on the fleet's average range and delivery radius, for example, 40%. F_supply(s,t) is the capacity supply factor, calculated as follows: ratio=Navail(s,t) / Dforecast(s,t,1h) F_supply=clip(1+k1×(ratio-1),0.6,1.4) For example, if there are currently 5 vehicles available for order taking, and the predicted demand for the next hour is 10 vehicles, then ratio = 0.5, k1 takes the default value of 0.3, so 1 + 0.3 × (0.5 - 1) = 0.85. After clipping, F_supply = 0.85. This value is less than 1, indicating that capacity is tight, and the charging notification threshold should be lowered to delay charging in order to preserve capacity.

[0048] F_demand(s,t) is the demand urgency factor, which is calculated as follows: slope=(Dforecast(4h)-Dforecast(1h)) / max(Dforecast(1h),1) F_demand=clip(1-k2×slope,0.7,1.3) If the demand for the next hour is 10 vehicles and the demand for the next 4 hours is 18 vehicles, then slope = (18-10) / 10 = 0.8. Taking the default value of k2 as 0.4, then 1 - 0.4 × 0.8 = 0.68. After clipping, F_demand = 0.7. This value is less than 1, indicating that demand is rising, and charging should be postponed to ensure subsequent capacity.

[0049] F_charging_capacity(s,t) is the charging resource constraint factor, which is calculated as follows: ρ=used_charging_num / total_charging_num λ=max(0,total_charging_num-used_charging_num) / N_staff F_charging_capacity=1+k3×ρ+k4×max(0,λ-λ0) If a charging station has 4 charging spots, 3 of which are already occupied, and there are 2 battery swappers on duty, then ρ = 0.75, λ = (4-3) / 2 = 0.5. If λ0 = 2, then max(0, 0.5-2) = 0, and F_charging_capacity = 1 + 0.5 × 0.75 + 0.1 × 0 = 1.375. This value is greater than 1, indicating that charging resources are scarce, and users should charge in advance to avoid queuing.

[0050] Considering the three factors mentioned above, assuming Tbase = 40%, then Tnotify = 40% × 0.85 × 0.7 × 1.375 ≈ 32.7%. That is, in this scenario of tight capacity, rising demand, and scarce charging resources, the system will dynamically lower the charging notification threshold from 40% to approximately 32.7% to avoid premature charging that would occupy scarce charging resources.

[0051] The dynamic power-on threshold is calculated as follows: Tpower_on(s,t)=Eroute_min(s)+Emargin(s,t) Here, Eroute_min(s) is the estimated power consumption of the shortest delivery route from station s, derived from historical average energy consumption data in the route operation statistics table. Emargin(s,t) is the dynamic margin, calculated based on demand urgency. Specifically, Emargin(s,t) = Emargin_base × F_demand(s,t), where Emargin_base is the baseline margin, with a default value of 5%. In the example above, if F_demand = 0.7, then Emargin = 5% × 0.7 = 3.5%. Assuming Eroute_min(s) = 15%, then Tpower_on = 15% + 3.5% = 18.5%. That is, the power-on judgment is triggered only when the vehicle's battery level is below 18.5%, and due to the urgency of demand, the margin is compressed to allow the vehicle to quickly return to an order-accepting state.

[0052] Step 3: Energy replenishment task generation and priority sorting.

[0053] The recharge task generation module continuously monitors the real-time battery level of each autonomous vehicle in the fleet. When a vehicle's battery level first falls below the dynamic charging notification threshold Tnotify(s,t) of its current station, the system automatically generates a recharge task for that vehicle. Before generating the task, the module initiates a quick query with the scheduling system to confirm whether the vehicle has already been assigned a delivery task. If it has been assigned, it is marked as "conflicted," the recharge task's status is set to "suspended," and the reason for the conflict is recorded; if it has not been assigned, the module continues to generate the task.

[0054] The generated energy replenishment tasks are sorted according to the following priority rules and then pushed to the battery swapper's terminal: Priority P0 (highest): The vehicle's current battery level is below Eroute_min(s), meaning it cannot complete any delivery tasks and must be recharged immediately. These tasks are highlighted in red and accompanied by a voice prompt.

[0055] Priority P1 (High): The number of vehicles available for order taking at the current station, Navail, is less than the predicted demand for the next hour, Dforecast(1h), and the vehicle's current battery level is lower than Tnotify but higher than Eroute_min(s). This type of vehicle is classified as "schedulable but with low battery". These tasks are highlighted in yellow.

[0056] Priority P2 (Normal): Other routine charging tasks, sorted in order of vehicle's current battery level from low to high.

[0057] Step 4: Digital twin simulation verification.

[0058] Before the power replenishment task is sent to the real RMS system, the digital twin simulation verification module first performs a simulation in the twin environment. The specific process is as follows: First, the module transforms the list of charging tasks to be verified into input parameters for the digital twin system, including: target vehicle ID, target charging station ID, estimated charging time (calculated based on the vehicle's current battery level and the time required to fully charge), and planned start time for charging.

[0059] Secondly, in the digital twin environment, the system uses the current real-world state as the initial state and extrapolates the capacity changes over the next 60 minutes. During the extrapolation, it simulates the charging completion time of each vehicle currently charging, the estimated return time of each vehicle performing a delivery task, and the impact of newly generated refueling tasks on the charging queue.

[0060] Then, after the simulation, the system compares the three sets of core KPI indicators before and after the simulation: (1) the time-series curve of the number of vehicles that can accept orders at each station in the next 60 minutes; (2) the order matching success rate, that is, the proportion of successfully delivered orders to the total number of orders; (3) the average user waiting time, that is, the average interval from when the order is placed to when the unmanned vehicle departs.

[0061] Finally, if the simulation results show that after executing the replenishment task, the trough value of the number of vehicles available for order taking is improved at at least two time points, the order matching success rate is improved by no less than 1%, and the average user waiting time is increased by no more than 5%, then it is judged as "verification passed". Otherwise, the replenishment task will be returned to the replenishment task generation module, and the priority will be recalculated or the execution will be postponed.

[0062] Step 5: Closed-loop feedback and model adaptation.

[0063] The closed-loop feedback module establishes a persistent record for each power replenishment decision. The record consists of two parts: the decision context and the execution result.

[0064] The decision context includes: the Tnotify value when the decision is triggered, the Tpower_on value, the specific values ​​of each factor F_supply, F_demand, and F_charging_capacity, the site capacity status snapshot (including Navail, Ncharging, and Dforecast for each time window), and the current timestamp (containing week and hour information).

[0065] The results are collected within 24 hours after the recharge task is completed, including: whether the vehicle was selected by the dispatch system to perform a delivery task after being powered on and the specific waiting time (minutes); whether a conflict event of "being dispatched immediately after being sent to charge" occurred; whether the vehicle failed the task or returned to the station midway due to low battery during the delivery task; and the change in the number of vehicles available to accept orders at the station within 1 hour before and after the recharge task was executed.

[0066] Every 7 days, the closed-loop feedback module triggers model parameter optimization. This module reads all accumulated feedback records from the past 7 days and, with the goal of minimizing the composite loss function, refits parameters such as k1, k2, k3, k4, λ0, and Emargin_base using the Gradient Boosting Regression Tree (GBRT) algorithm. The composite loss function is defined as follows: L = w_a×P1 + w_b×E + w_c×P2 Wherein, P1 is the probability of charging and scheduling conflicts occurring within the week (number of conflicts / total number of charging attempts), E is the average waiting time (minutes) from power-on to the first task assignment, P2 is the probability of task failure due to low battery within the week (number of failures / total number of deliveries), and w_a, w_b, and w_c are set to 0.5, 0.3, and 0.2, respectively. After optimization, the new parameters are automatically loaded into the dynamic threshold calculation module without manual intervention.

[0067] Through the above closed-loop mechanism, the system continuously adapts during operation: when the conflict rate P1 is found to be too high, the optimization algorithm will automatically adjust k1 and k2 to shift the charging notification threshold to a lower direction, reducing the probability of "being scheduled immediately after charging"; when the task failure rate P2 is found to be too high, the optimization algorithm will automatically adjust each factor to shift the charging notification threshold to a higher direction, charging in advance to ensure capacity.

[0068] This embodiment addresses the technical challenge of fixed thresholds failing to adapt to fluctuating operational scenarios by combining dynamic threshold calculation with digital twin verification. Compared to existing technologies, Tnotify in this embodiment can dynamically match the timing of energy replenishment triggering with the operational rhythm by changing in real time with the capacity supply-demand ratio, demand trends, and charging resource occupancy. Simultaneously, through digital twin simulation verification, decision-making risks are mitigated in the simulated environment, avoiding capacity losses due to improper energy replenishment decisions in real-world scenarios. Furthermore, the closed-loop feedback mechanism enables the system to continuously evolve adaptively, adapting to changes in the operational environment such as new stations, seasonal variations, and fleet expansion without requiring manual threshold recalibration.

[0069] Those skilled in the art will understand that, in addition to implementing the system, apparatus, and their modules provided by this invention in purely computer-readable program code, the same program can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system, apparatus, and their modules provided by this invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; alternatively, modules for implementing various functions can be considered both software programs implementing the method and structures within the hardware component.

[0070] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A refueling scheduling system for unmanned vehicle fleets based on multi-factor dynamic thresholds, characterized in that, include: The module includes a capacity perception module, a dynamic threshold calculation module, a capacity replenishment task generation and priority ranking module, a digital twin simulation verification module, and a closed-loop feedback and model adaptation module. The capacity sensing module collects multi-dimensional capacity status data of each charging and swapping station in real time. The multi-dimensional capacity status data includes the number of vehicles currently available for orders, the number of vehicles charging, the estimated available time distribution of vehicles charging, and the predicted vehicle demand within a future preset time window. The capacity sensing module is connected to the dynamic threshold calculation module and sends the multi-dimensional capacity status data to the dynamic threshold calculation module. The dynamic threshold calculation module receives the multi-dimensional capacity status data and calculates the dynamic charging notification threshold and dynamic power-on threshold for each station and time period. The dynamic threshold calculation module is connected to the power replenishment task generation and priority sorting module, and sends the calculated dynamic charging notification threshold and dynamic power-on threshold to the power replenishment task generation and priority sorting module. The energy replenishment task generation and priority sorting module generates energy replenishment tasks when the vehicle's battery level reaches the dynamic charging notification threshold, and sorts them according to a preset priority rule; the energy replenishment task generation and priority sorting module shares capacity status data with the scheduling system and performs conflict detection when generating energy replenishment tasks; the energy replenishment task generation and priority sorting module is connected to the digital twin simulation verification module and sends the generated energy replenishment tasks to the digital twin simulation verification module. The digital twin simulation verification module receives the power replenishment task and synchronizes it to the digital twin system for simulation verification; the digital twin simulation verification module connects to the real RMS system and sends the verified power replenishment task to the RMS system for execution; The closed-loop feedback and model adaptation module is connected to the dynamic threshold calculation module and the digital twin simulation verification module. It records the decision context data and execution results of each energy replenishment decision to construct a feedback dataset, and periodically uses the accumulated feedback dataset to optimize the model parameters of the dynamic threshold calculation module. The dynamic charging notification threshold Tnotify(s,t) is calculated using the following formula: Tnotify(s,t)=Tbase×F_supply(s,t)×F_demand(s,t)×F_charging_capacity(s,t) Where s is the station, t is the time, Tbase is the baseline charging notification percentage; F_supply(s,t) is the capacity supply factor; F_demand(s,t) is the demand urgency factor; and F_charging_capacity(s,t) is the charging resource constraint factor.

2. The unmanned vehicle fleet replenishment scheduling system based on multi-factor dynamic thresholds according to claim 1, characterized in that, The capacity supply factor F_supply(s,t) is calculated using the following formula: ratio=Navail(s,t) / Dforecast(s,t,1h) F_supply=clip(1+k1×(ratio-1),0.6,1.4) Wherein, ratio is the ratio of the current number of vehicles available for order taking to the predicted demand in the next hour, Navail(s,t) is the number of vehicles available for order taking at station s at time t, Dforecast(s,t,1h) is the predicted demand for vehicles at station s in the next hour time window, clip(1+k1×(ratio-1),0.6,1.4) is the cutoff function, which takes 0.6 when 1+k1×(ratio-1)<0.6, takes 1.4 when 1+k1×(ratio-1)>1.4, and takes 1+k1×(ratio-1) otherwise, where k1 is the sensitivity coefficient.

3. The unmanned vehicle fleet replenishment scheduling system based on multi-factor dynamic thresholds according to claim 1, characterized in that, The demand urgency factor F_demand(s,t) is calculated using the following formula: slope=(Dforecast(4h)-Dforecast(1h)) / max(Dforecast(1h),1) F_demand=clip(1-k2×slope,0.7,1.3) Where slope is the relative slope of the three-window forecast, Dforecast(4h) is the predicted vehicle demand of station s in the next 4-hour time window, Dforecast(1h) is the predicted vehicle demand of station s in the next 1-hour time window, k2 is the sensitivity coefficient, and clip is the cutoff function.

4. The unmanned vehicle fleet replenishment scheduling system based on multi-factor dynamic thresholds according to claim 1, characterized in that, The charging resource constraint factor F_charging_capacity(s,t) is calculated using the following formula: ρ=used_charging_num / total_charging_num λ=max(0,total_charging_num-used_charging_num) / N_staff F_charging_capacity=1+k3×ρ+k4×max(0,λ-λ0) Where ρ is the charging position occupancy rate, used_charging_num is the number of occupied charging piles, total_charging_num is the total number of charging piles, λ is the number of idle charging positions that a single battery swapper needs to oversee, N_staff is the number of battery swappers on duty, λ0 is the upper limit threshold of idle positions within the capacity of a single battery swapper, k3 is the occupancy rate sensitivity coefficient, k4 is the manpower reserve sensitivity coefficient, and max(0,λ-λ0) is the maximum value function.

5. The unmanned vehicle fleet refueling scheduling system based on multi-factor dynamic thresholds according to claim 1, characterized in that, The dynamic power-on threshold is calculated using the following formula: Tpower_on(s,t)=Eroute_min(s)+Emargin(s,t) Where Tpower_on(s,t) is the dynamic power-on threshold, Eroute_min(s) is the estimated power consumption of the shortest available route from station s, Emargin(s,t) is the dynamic margin, and t is the time.

6. The unmanned vehicle fleet refueling scheduling system based on multi-factor dynamic thresholds according to claim 1, characterized in that, The power replenishment task generation and priority sorting module sorts the generated power replenishment tasks from high to low according to the following priority: First priority: Vehicle battery level is below Eroute_min(s); Second priority: The current station Navail is less than Dforecast(1h), and the vehicle is a potentially dispatchable vehicle; Third priority: Regular charging tasks, sorted by battery level from low to high; Where Navail is the number of vehicles currently available for order taking, Dforecast(1h) is the predicted vehicle demand within the next 1-hour time window, and Eroute_min(s) is the estimated power consumption of the shortest route among all available routes departing from station s.

7. The unmanned vehicle fleet replenishment scheduling system based on multi-factor dynamic thresholds according to claim 1, characterized in that, The digital twin simulation verification module transforms the content of the energy replenishment task into simulation parameters in the digital twin system. The content includes the target vehicle, charging station, and estimated charging time. It extrapolates the changes in station capacity within a preset time period in the digital twin environment and simulates the recovery of the number of vehicles available for order taking at each station after energy replenishment. It compares the core KPI indicators before and after the simulation. The core KPI indicators include the number of vehicles available for order taking, order matching success rate, and average user waiting time. If the simulation results show that the core KPI indicators have improved, it is marked as verified and the energy replenishment command is sent to the RMS system for execution.

8. The unmanned vehicle fleet refueling scheduling system based on multi-factor dynamic thresholds according to claim 1, characterized in that, The decision context data recorded by the closed-loop feedback and model adaptation module includes: the dynamic charging notification threshold Tnotify value at the time of decision-making, the values ​​of each factor function F_supply, F_demand, and F_charging_capacity, and a snapshot of the station's capacity status; the execution results include: whether a charging and scheduling conflict occurred, how long after power-on the vehicle was selected by the scheduling system to execute the delivery task, whether the vehicle failed the task due to low battery, and a comparison of the number of vehicles that can accept orders at the station before and after the charging is completed; The closed-loop feedback and model adaptation module utilizes the feedback dataset to optimize the parameters of each factor function in the dynamic threshold calculation module using machine learning methods. The machine learning method aims to minimize the composite loss function. L = w_a×P1 + w_b×E + w_c×P2 Where L is the composite loss, P1 is the probability of charging and scheduling conflict, E is the expected value of the waiting time from power-on to the first task assignment, P2 is the probability of task failure due to low battery, and w_a, w_b, and w_c are preset weighting coefficients.

9. A method for scheduling refueling of unmanned vehicle fleets based on multi-factor dynamic thresholds, characterized in that, The unmanned vehicle fleet refueling scheduling system based on multi-factor dynamic thresholds according to any one of claims 1 to 8 shall be used to perform the following steps: Step 1: Collect multi-dimensional capacity status data for each charging and swapping station in real time. The multi-dimensional capacity status data includes the number of vehicles currently available for order taking, the number of vehicles currently charging, the estimated available time distribution of vehicles currently charging, and the predicted vehicle demand within a future preset time window. Step 2: Receive the multi-dimensional capacity status data and calculate the dynamic charging notification threshold and dynamic power-on threshold for each station and time period; Step 3: When the vehicle's battery level reaches the dynamic charging notification threshold, a charging task is generated and sorted according to a preset priority rule; when generating the charging task, the capacity status data is shared with the scheduling system and conflict detection is performed; Step 4: Synchronize the generated power replenishment task to the digital twin system for simulation verification, and then send the verified power replenishment task to the real RMS system for execution; Step 5: Record the decision context data and execution results of each energy replenishment decision to build a feedback dataset, and periodically use the accumulated feedback dataset to optimize the model parameters in the dynamic threshold calculation step.

Images