A low-altitude dynamic airspace partitioning and task scheduling method and system

Abstract

This invention belongs to the field of low-altitude dynamic airspace partitioning, and discloses a method and system for low-altitude dynamic airspace partitioning and task scheduling. The method includes: S1, configuring basic low-altitude airspace information and constructing a hierarchical grid airspace with parent-child nesting relationships; S2, acquiring the state features of the hierarchical grid airspace, and having the reinforcement learning policy model output hold, split, or merge actions for each active airspace unit, and performing legality verification on the split or merge actions; S3, updating the set of active airspace units and their adjacency relationships, and constructing an active airspace topology graph; S4, receiving task scheduling requests and performing a time-dimensional path search on the active airspace topology graph; S5, determining whether a complete path from the task start point to the end point exists. If it exists, the task is accepted; otherwise, the task is rejected. It is evident that this invention enables the airspace granularity to adjust according to the reservation and occupancy status and task pressure changes.

Description

Technical Field

[0001] This invention relates to the field of low-altitude dynamic airspace delineation technology, specifically a method and system for low-altitude dynamic airspace delineation and task scheduling. Background Technology

[0002] With the rapid development of the low-altitude economy, urban drone delivery, low-altitude inspection, and regional air traffic organization, flight missions in low-altitude airspace exhibit the following characteristics: First, flight missions are typically generated continuously around several fixed or semi-fixed take-off and landing points. Second, flight activities are mainly concentrated in the near-ground and low-to-medium altitude regions, with limited available vertical altitude. Third, permanent obstacles such as buildings, facilities, and no-fly zones exist within the airspace. Fourth, the intensity of mission arrival varies across different areas and at different times, easily leading to local hotspots and bottlenecks.

[0003] In the above scenarios, using a fixed-granularity grid spatial domain typically presents significant limitations. When the spatial domain is divided too coarsely, although the number of managed objects is smaller and the overall management cost is lower, there is a lack of sufficient degrees of freedom to maneuver around obstacles, making it difficult to manage local hotspots in detail and reducing task acceptance capacity. When the spatial domain is divided too finely, although local control capabilities are enhanced, the number of active spatial domain units increases significantly, management costs rise, and unnecessary refinement may occur in low-task-pressure scenarios. Summary of the Invention

[0004] To address the technical problems existing in the prior art, this invention provides a method and system for low-altitude dynamic airspace partitioning and task scheduling, which enables the airspace granularity to be adjusted according to the reservation and occupancy status and task pressure.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A method for low-altitude dynamic airspace allocation and task scheduling includes the following steps: S1. Configure basic information for low-altitude airspace and construct a hierarchical grid airspace with parent-child nesting relationships; determine the set of active airspace units in the hierarchical grid airspace and establish a spatiotemporal reservation and occupancy table that records the number of active airspace units reserved and occupied per time slice. S2. Obtain the state features of the hierarchical grid spatial domain, input the state features into the trained reinforcement learning policy model, and have the reinforcement learning policy model output maintain, split or merge actions for each active spatial domain unit, and perform legality verification on the split or merge actions. S3. After executing the action that passes the legality check, update the set of active spatial units and their adjacency relationships, and construct the active spatial topology graph; S4. Receive a task scheduling request, construct a spatiotemporal search state based on the activity airspace unit and arrival time, and perform a path search with time dimension on the activity airspace topology map based on the spatiotemporal reservation occupancy table. S5. Determine if a complete path exists from the start point to the end point of the task. If it exists, accept the task and update the spatiotemporal reservation occupancy table based on the time slice occupancy corresponding to the task; otherwise, reject the task.

[0006] Preferably, the basic information of the low-altitude airspace includes three-dimensional airspace boundaries, permanent obstacle constraints, take-off and landing point networks and their weights; the hierarchical grid airspace includes at least two granularity levels, and the same parent unit has a spatial inclusion relationship with multiple child units of its next lower granularity level.

[0007] Preferably, the state features include the number of reservations occupied per time slice, the three-dimensional spatial state tensor, the pressure context features, the active spatial unit features, and the action legality mask of the active spatial unit; The stress context features are constructed based on the long-window exponential moving average and the short-window exponential moving average of the number of newly generated tasks in the current period.

[0008] Preferably, the reinforcement learning strategy model includes a 3D convolutional feature extraction module, a context encoding module, an active spatial unit feature encoding module, an action decision branch, and a value evaluation branch. The 3D spatial state tensor is processed by the 3D convolutional feature extraction module to form global spatial features, the pressure context features are processed by the context encoding module to form context features, and the active spatial unit features, action legality encoding, and the context features together form unit structure features. The action decision branch outputs the probability distribution of three types of actions: maintain, split, and merge, based on the global spatial features and unit structure features, and the value evaluation branch outputs the current spatial state value.

[0009] Preferably, the reinforcement learning policy model is trained using a shared parameter Actor-Critic training method based on PPO cut-off targets. State features, actions, log probabilities of old policy actions, immediate training rewards, termination flags, and state values ​​are collected at continuous time steps in multiple parallel environments. Generalized advantage estimates and bootstrap rewards are calculated. The action decision branch is updated using PPO cut-off targets, and the value evaluation branch is updated using mean squared error loss. Multiple candidate policy models are saved during training, and the target model is selected based on evaluation items consistent with the final experimental evaluation: task acceptance benefit, airspace management cost, and cross-regional handover cost.

[0010] Preferably, the legality verification of the splitting action specifically includes: The entry and exit sub-units are determined based on the geometric positions of existing tasks entering and leaving the parent unit; local sub-paths connecting the entry and exit sub-units are determined in the sub-unit adjacency graph; the parent unit reservation occupancy is mapped to the sub-units corresponding to the local sub-paths according to the time sequence of existing tasks; if the local sub-path does not exist, or if the reservation occupancy of any sub-unit exceeds the capacity limit of that sub-unit after migration in any time slice, the verification is deemed to fail; otherwise, the verification passes.

[0011] Preferably, the legality verification of the merging action specifically includes: Confirm that all child units corresponding to the same parent unit are in the current active airspace unit set; aggregate the reservation occupancy of all child units in the same time slot; if the aggregated reservation occupancy of any time slot exceeds the parent unit's capacity limit, or if the geometric boundary of the parent unit includes a permanent obstacle area into the flyable airspace, the merging action verification is deemed to fail; otherwise, the verification passes.

[0012] Preferably, the time-dimensional path search adopts a heuristic priority search, with active spatial units as spatial nodes and discrete time slices for reaching active spatial units as time states; the cost of candidate state transitions includes the number of discrete time steps required for movement and the additional cost of crossing active spatial units, and the candidate state transition is allowed when the number of time slices reserved for the target active spatial unit is less than the upper limit of the capacity of the active spatial unit.

[0013] This invention also proposes a low-altitude dynamic airspace allocation and task scheduling system, comprising: The airspace configuration module configures basic low-altitude airspace information and constructs a hierarchical grid airspace with parent-child nesting relationships; it determines the set of active airspace units in the hierarchical grid airspace and establishes a spatiotemporal reservation and occupancy table that records the number of active airspace units reserved and occupied per time slice. The state acquisition module acquires the state features of the hierarchical grid spatial domain, inputs the state features into the trained reinforcement learning policy model, and the reinforcement learning policy model outputs hold, split or merge actions for each active spatial domain unit, and performs legality verification on the split or merge actions. The airspace reconstruction module, after performing actions that pass the legality verification, updates the set of active airspace units and their adjacency relationships, and constructs the active airspace topology graph; The task scheduling module receives task scheduling requests, establishes a spatiotemporal search state based on active spatial domain units and arrival times, performs a time-dimensional path search on the active spatial domain topology map, and queries the spatiotemporal reservation occupancy table during candidate state transitions. If the number of reserved occupancy slots for a target active spatial domain unit in its corresponding arrival time slice is less than the capacity limit of that active spatial domain unit, the candidate state transition is allowed. The module determines whether a complete path exists from the task's origin to its destination; if it does, the task is accepted, and the spatiotemporal reservation occupancy table is updated based on the time-slice occupancy corresponding to the task; otherwise, the task is rejected.

[0014] The present invention also proposes a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the low-altitude dynamic airspace partitioning and task scheduling method.

[0015] Compared with the prior art, the beneficial effects of the present invention are: (1) The present invention uses the active spatial domain unit with parent-child nesting relationship as the runtime reconstruction object, and adjusts the current active spatial domain topology as needed through the actions of maintaining, splitting and merging, so that the spatial domain granularity can be adjusted according to the reservation occupancy status and task pressure. (2) When the parent unit splits, the present invention maps and migrates the existing task trajectory and its time slice reservation to the corresponding sub-unit, and aggregates the reservation of the sibling sub-units when merging, so as to maintain the spatiotemporal execution relationship of the existing task during the process of spatial structure change; (3) Before performing splitting or merging operations, the present invention verifies the time-slice capacity, sub-unit integrity and permanent obstacle constraints to avoid airspace reconfiguration causing capacity overrun or re-incorporating permanent obstacle areas into flyable airspace. (4) This invention integrates global airspace features, pressure context features and active airspace unit features as policy inputs, enabling the policy to simultaneously perceive the overall airspace reservation pressure and the local unit status. (5) This invention improves the feasibility of task acceptance by introducing a time dimension into path planning and querying the time and space reservation occupancy table, taking into account spatial connectivity, time reservation occupancy and unit capacity limitations. (6) The present invention adopts the PPO algorithm combined with the shared parameter Actor-Critic structure to optimize the spatial domain reconstruction strategy, so that each active spatial domain unit can make parallel decisions based on the shared global state information. Attached Figure Description

[0016] Figure 1 This is a flowchart of the low-altitude dynamic airspace partitioning and task scheduling method proposed in an embodiment of the present invention; Figure 2 This is a schematic diagram of a hierarchical grid spatial structure proposed in another embodiment of the present invention; Figure 3 This is a static partitioning diagram under a weighted terminal design proposed in another embodiment of the present invention; Figure 4 This is a block diagram of a low-altitude dynamic airspace partitioning and task scheduling system proposed in another embodiment of the present invention. Detailed Implementation

[0017] 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.

[0018] Example 1

[0019] Please see Figure 1 This embodiment provides a method for low-altitude dynamic airspace partitioning and task scheduling, including the following steps: S1. Configure basic information for low-altitude airspace and construct a hierarchical grid airspace with parent-child nesting relationships; determine the set of active airspace units in the hierarchical grid airspace and establish a spatiotemporal reservation and occupancy table that records the number of active airspace units reserved and occupied per time slice. The basic information of the low-altitude airspace includes three-dimensional airspace boundaries, permanent obstacle constraints, take-off and landing point networks and their weights.

[0020] In this embodiment, the three-dimensional spatial boundary can be set as a low-altitude three-dimensional space with a horizontal range of 16×16 and a height range of 8, i.e., a 16×16×8 spatial domain. In this embodiment, permanent obstacles are represented in the form of axis-aligned cuboid boxes, and each obstacle is described by six numbers in the following order:

[0021] in and This represents the lower and upper boundaries of the obstacle in the x-direction. and This represents the lower and upper boundaries of the obstacle in the y-direction. and The lower and upper boundaries of the obstacle in the height direction are represented, and their coordinate values ​​have the same range as the spatial domain range, i.e., the x-direction coordinate range is 0≤x≤16, the y-direction coordinate range is 0≤y≤16, and the z-direction coordinate range is 0≤z≤8. In the current embodiment, the obstacle is represented using the finest-grained grid coordinates and uses a half-open interval, i.e. [ ), [ ), [ ).

[0022] For example: M1=(4,8,4,8,0,4) indicates that the obstacle covers a three-dimensional spatial region x∈[4,8), y∈[4,8), z∈[0,4). In simpler terms, it is an axis-aligned obstacle region from x=4 to x=8, from y=4 to y=8, and from z=0 to z=4.

[0023] Specifically, permanent obstacles can be set as follows: M1=(4,8,4,8,0,4); M2=(12,14,12,14,0,4); M3=(10,12,4,6,0,4); M4=(4,6,12,14,0,4); M5=(12,14,8,10,0,2).

[0024] The takeoff and landing point network setup and its weight sampling mechanism can be as follows. The current embodiment uses 6 terminal takeoff and landing anchor points: T1=(2,2,2), weight 1.2; T2=(1,15,1), weight 1.0; T3=(15,3,1), weight 1.0; T4=(10,10,2), weight 1.2; T5=(15,15,1), weight 0.9; T6=(3,9,1), weight 1.0.

[0025] All terminal anchor points share a unified sampling radius r=1. Let the weight of the m-th terminal anchor point be... The sampling probability of selecting this terminal anchor point as the starting or ending point. for:

[0026] Where M represents the total number of available terminal anchors. is the sampling weight of the q-th terminal anchor point.

[0027] In this application, the hierarchical grid spatial domain includes at least two granularity levels, and the same parent cell has a spatial containment relationship with multiple child cells of its next granularity level.

[0028] In a preferred embodiment of the present invention Figure 2 This is a schematic diagram of a hierarchical grid spatial structure, such as... Figure 2 As shown, the current embodiment employs three granularity levels: L0: Coarse-grained layer; L1: Medium-grained layer; L2: Fine-grained layer. Taking L1 and L2 granularity levels as examples, the same parent unit and multiple child units in its next granularity level have a spatial containment relationship.

[0029] Specifically, the spacetime reservation and occupancy table is used to record the number of tasks occupied by the active airspace unit at the current time and at several future times. Establishing a spacetime reservation and occupancy table for the active airspace unit can facilitate subsequent task planning and the next round of airspace reconfiguration decisions.

[0030] S2. Obtain the state features of the hierarchical grid spatial domain, input the state features into the trained reinforcement learning policy model, and have the reinforcement learning policy model output maintain, split or merge actions for each active spatial domain unit, and perform legality verification on the split or merge actions.

[0031] The state features specifically include: The number of reservations occupied per time slice, the three-dimensional spatial state tensor and pressure context features; the activity spatial unit features and the action legality mask.

[0032] Specifically, the reservation occupancy quantity per time slice refers to the number of tasks reserved by each grid cell in the current gridded airspace at the current moment and at several future moments.

[0033] In the current embodiment, the operating space is 16×16×8, and the state tensor is denoted as... Where C is the number of channels, C = 2 + K, and K is the number of future occupancy prediction steps. In the current embodiment, K = 5, therefore... Its dimensions are 7×16×16×8. Wherein: The first channel represents the granularity scale of the current active spatial unit on the finest granular grid unit; the second channel represents the occupancy rate of this location at the current moment, defined as the ratio of the current number of reservations to the capacity of the active spatial unit; the third to seventh channels represent the occupancy rate information for the next 5 time steps calculated based on the current reservation structure.

[0034] For locations not covered by the current active spatial unit, the tensor value is set to 0.

[0035] Corresponding to the three-dimensional spatial state tensor, the system also constructs a pressure context feature vector. This describes the change in task arrival intensity over time. Here, task arrival intensity represents the number of newly generated tasks, *n*, within a unit decision period. Let the number of newly generated tasks in the *t*-th decision period be... In the current embodiment, the pressure context feature vector It can be constructed based on the long-window exponential moving average and the short-window exponential moving average of task arrival intensity, with a dimension of 2.

[0036] Specifically, the long-window exponential moving average and the short-window exponential moving average can be expressed as follows:

[0037] in, and Let represent the exponential sliding update coefficients for the long window and the short window, respectively, and satisfy 0 < 0. < <1, making the short-window exponential moving average more sensitive to recent changes in task arrival, while the long-window exponential moving average reflects the overall task pressure level over a longer time scale. In the current embodiment, 0.2 is acceptable. 0.5 is acceptable.

[0038] Therefore, we can first obtain the unnormalized stress statistics vector:

[0039] In the current embodiment, before inputting the stress statistics into the reinforcement learning policy model, a preset task-reaching intensity reference value is used. The long-window exponential moving average and the short-window exponential moving average are normalized, and the normalization result is restricted to a preset interval, as shown below:

[0040] Accordingly, the stress context feature vector of the actual input context encoding module is represented as:

[0041] in, The pre-set task intensity reference value can be determined based on the task pressure range corresponding to the training scenario; The preprocessing stage inputs fixed configuration values ​​to the model and do not participate in the parameter updates of the reinforcement learning policy model; in this embodiment, Take the task pressure parameters corresponding to the training scenario . Used to characterize the normalized long-term average task arrival intensity This is used to characterize the recent task arrival intensity after normalization. By normalizing and limiting the interval, the influence of different task pressure levels on the feature numerical scale can be reduced, and the excessively large amplitude of the context feature caused by the number of abnormal task arrivals can be avoided. The dimension of the pressure context feature vector is 2.

[0042] Obtain the features and action validity masks of each active spatial unit. Let the set of active spatial units at the current moment be... Then each active spatial unit Corresponding to a set of active spatial unit features This includes, in sequence: hierarchical normalized value, x-coordinate normalized value, y-coordinate normalized value, z-coordinate normalized value, current reservation occupancy rate, peak reservation occupancy rate over several future time steps, and average reservation occupancy rate of adjacent activity spatial units. The three-dimensional position normalized value is obtained by dividing the center coordinate of the activity spatial unit by the spatial scale of the corresponding direction; the current reservation occupancy rate is the ratio of the number of reservations occupied in the current time slice to the capacity of that unit; the peak future occupancy rate and the average adjacent occupancy rate are calculated based on the spatiotemporal reservation occupancy table and the activity spatial topology, respectively. Simultaneously, an action legality mask is constructed for each activity spatial unit. This corresponds to whether the keep, split, and merge actions are allowed to be executed, respectively.

[0043] The state features are input into a reinforcement learning policy model, which includes an action decision branch and a value evaluation branch. The value evaluation branch outputs the current spatial state value. The action decision branch outputs the action decisions and action probability distributions of active spatial units in the hierarchical grid spatial domain, and performs legality verification on the action decisions.

[0044] Specifically, this step mainly involves making decisions on reinforcement learning strategies.

[0045] The three-dimensional spatial state tensor is input into the three-dimensional convolutional feature extraction module of the reinforcement learning policy model, and the global spatial features are output. The pressure context features are input into the context encoding module of the reinforcement learning policy model, and the context features are output. The local features of the active spatial unit and the two-dimensional legality signals related to splitting and merging in the action legality mask are concatenated with the context features to obtain the structural features of the active spatial unit.

[0046] In this application, the strategy model adopts a shared trunk plus an Actor-Critic dual-branch structure. Three-dimensional spatial state tensor. First, the module proceeds to the 3D convolutional feature extraction module. In this embodiment, this module consists of three stacked 3D convolutional layers, with the number of channels increasing sequentially from 7 to 16 to 32 to 64. The kernel size is 3×3×3, the stride is 1, and the padding is 1. Each convolutional layer is followed by ReLU and 2×2×2 max pooling. After flattening the convolutional output, a 256-dimensional global spatial feature is obtained through linear projection. Stress context features The context encoding module then generates the corresponding context feature representation. For each active spatial unit, the local features are... The two-dimensional legality signals related to splitting and merging in the action legality mask are concatenated with contextual features to form an 11-dimensional structured input, which is then encoded through two layers of fully connected layers to obtain a 64-dimensional unit structure feature. .

[0047] The global spatial domain features are concatenated with the structural features of each active spatial domain unit and then input into the action decision branch to obtain the action decision and action probability distribution corresponding to the active spatial domain unit.

[0048] The action decision for each active spatial unit is output, including keep, split, and merge actions. Specifically, the global spatial feature ft is concatenated with the structural feature zi of each active spatial unit and input into the action decision branch to obtain the 3D action logits corresponding to that active spatial unit, which correspond to keep, split, and merge, respectively. During the policy training phase, sampling can be performed based on this action distribution; during the application phase, the action with the highest probability can be selected as the execution result. For actions deemed illegitimate by the action validity mask, their logits are masked and do not participate in the final decision.

[0049] At the same time, the legality of the action decision needs to be verified.

[0050] Semantic validity verification includes verifying splitting and merging actions. Specifically, it verifies the feasibility of migrating existing tasks and time slice capacity constraints under splitting actions.

[0051] Specifically, let the parent unit be... The set of sub-units corresponding to its subdivision is When a task trajectory segment already traverses the parent unit, the corresponding entry and exit sub-units are first determined based on the task's geometric positions upon entering and leaving the parent unit. Then, based on the adjacency relationships between sub-units, a local sub-path is searched at the sub-unit level to allow the existing task to continue being represented at the sub-unit level. If the local sub-path does not exist, the splitting action is deemed invalid. If the local sub-path exists, the existing reservations of the parent unit in each time slice are migrated and allocated to the sub-units. For any sub-unit... For any time slice τ, it must satisfy the following condition. ,in Indicates the number of time slices occupied after migration. This indicates the upper limit of the capacity of the spatial domain cell at the same level as the sub-cell.

[0052] Verify the integrity of sibling units, obstacle containment, and whether the occupancy limit is exceeded during the merging operation.

[0053] Specifically, the sibling unit integrity check is performed only if the units to be merged constitute a complete set of siblings under the same parent unit. The occupancy limit check is also performed, assuming the sibling set is... Its parent unit is denoted as For any time slice τ, the merged result must satisfy... , This indicates the upper limit of the capacity of the spatial domain cells at the same level as the parent cell. Obstacle containment verification: Further verification is required. The geometric boundary of the parent unit will not reintegrate permanent obstacles into the flyable airspace; if the boundary of the parent unit is inconsistent with the boundary of the permanent obstacle, causing the parent unit to include part of the permanent obstacle in the flyable range, then the merging action is illegal.

[0054] In the current embodiment, a complete sibling set refers to a set of all child units corresponding to the same parent unit that are in the current active spatial domain unit set. When the number of child units that output merging actions reaches a preset merging trigger number, capacity and obstacle constraint checks are uniformly performed on the complete sibling set. In the current embodiment, the preset merging trigger number can be 4.

[0055] S3. After executing the action that passes the legality check, update the set of active spatial units and their adjacency relationships, and construct the active spatial topology graph.

[0056] After each active spatial unit performs a hold, split, or merge action, the hierarchical grid spatial domain is reconstructed. Let the set of active spatial units after reconstruction be... The set of adjacency relationships is This forms the activity space topology map. In the current embodiment, active spatial domain units are used as graph nodes, spatial adjacency relationships are used as graph edges, and a time dimension is further introduced to construct the spatiotemporal search state n=(v,τ), where τ represents the current active airspace cell, and τ represents the relative time step to reach that active airspace cell.

[0057] S4. Receive the task scheduling request, construct the spatiotemporal search state with the activity airspace unit and arrival time, and perform a path search with time dimension on the activity airspace topology map based on the spatiotemporal reservation occupancy table.

[0058] Specifically, the system executes A on this spacetime graph. The search (heuristic-first search algorithm) defines the node evaluation cost as:

[0059] f(n) = g(n) + h(n)

[0060] Where g(n) represents the cumulative spatiotemporal transition cost from the starting point to the current spatiotemporal node n=(v,τ), and for the initial state have If from Extended to ,but

[0061] For from arrive The single-step cost of a state transition is denoted as .

[0062] in, Indicates the activity airspace unit Exercise to The required number of discrete time steps can be determined based on the spatial distance and velocity constraints between active spatial domain units; This represents the cost of a single cross-regional handover during path search, used to account for the additional operational costs incurred when traversing different activity airspace units. In the current embodiment, A value of 0.05 is acceptable. In one specific implementation, It can be determined by the Euclidean distance between the center points of adjacent active spatial units. Divide by the corresponding speed constraint Rounding up gives the result, i.e.: It should be noted that, It is used only for calculating the additional path cost of cross-active spatial domain cell transfers during the spatiotemporal path search process; described later. The two methods are used to determine the penalty strength of the number of cross-regional handovers in immediate training rewards, candidate model selection, and final experimental evaluation; their roles are different.

[0063]

[0064] in, Take the smaller of the upper velocity limits of the two adjacent active spatial units at their respective hierarchical levels. The heuristic cost can be written as: , in, This represents the Manhattan distance between the center of the current active spatial unit and the center of the target terminal's neighborhood. This indicates the maximum permissible flight speed in the current airspace. This represents the side length of the largest active spatial unit in the current embodiment; under the current three-layer structure, That is, the side length of the L0 layer spatial unit. For the current spatiotemporal state... Pointing to candidate spacetime state For any candidate state transition, the system needs to query the target activity airspace unit in the spatiotemporal reservation occupancy table. At the time slice reservation occupancy number Only when And the target activity airspace unit The candidate state transition is allowed to be added to the search front only if it does not belong to a permanent obstacle region.

[0065] S5. Determine if a complete path exists from the start point to the end point of the task. If it exists, accept the task and update the spatiotemporal reservation occupancy table based on the time slice occupancy corresponding to the task; otherwise, reject the task.

[0066] Furthermore, it is necessary to determine whether a feasible path exists that satisfies connectivity, capacity, and obstacle avoidance constraints. In the current embodiment, path search simultaneously considers spatial connectivity, time reservation occupancy, cell capacity limitations, obstacle avoidance constraints, velocity constraints, and the additional costs associated with switching across activity spatial domain cells. The task is deemed acceptable under the current reconstructed spatial structure only if a complete spatiotemporal path exists from the starting terminal neighborhood to the ending terminal neighborhood.

[0067] If a feasible path exists, the task is accepted and written to the spacetime reservation table. Specifically, the activity airspace unit occupancy corresponding to the task at each time step is written to the reservation table, and the occupancy statistics at the current time and several future times are updated synchronously for subsequent task planning and the next round of airspace reconfiguration decisions.

[0068] If no feasible path exists, the task is rejected and the unaccepted result is recorded. The unaccepted result and the accepted result together constitute the task acceptance performance at that time step, and serve as the basis statistics for subsequent policy training and external evaluation.

[0069] Update the task acceptance status, airspace occupancy status, and time reservation structure based on the path planning results, and generate the next round of state features accordingly.

[0070] The offline training and candidate model selection process for reinforcement learning policy models includes the following steps.

[0071] This training process is used to optimize the reinforcement learning policy model, save multiple candidate policy models, and select the target model based on the comprehensive validation score.

[0072] In this embodiment, the reinforcement learning policy model adopts a shared backbone and Actor-Critic dual-branch structure. The shared backbone receives a 7×16×16×8 three-dimensional spatial state tensor and outputs 256-dimensional global spatial features; the action decision branch receives global spatial features, active spatial unit structure features, and action legitimacy information, and outputs the policy distribution of each active spatial unit, including maintaining, splitting, or merging; the value evaluation branch receives global spatial features, context features, and the normalized value of the current number of active spatial units. Output the current spatial state value. Let the set of active spatial units at the current moment be... The maximum number of parallel decision objects preset in the strategy model is The normalized value of the number of currently active spatial units is then expressed as:

[0073] in, This indicates the number of active airspace units currently participating in airspace reconfiguration decisions. This represents the upper limit of the number of active spatial units that the policy model can accommodate in a single forward computation, and It is not less than the number of currently active airspace units. The normalized value of the number of currently active airspace units is used to provide the value assessment branch with a quantitative characteristic of the complexity of the current airspace structure.

[0074] (1) Collect state features, action decisions, log probability of old strategy actions, immediate training rewards, termination flags and state values ​​within consecutive time steps; During the rollout sampling phase, the reinforcement learning policy model is based on the current state. Output the action probability distribution, and sample or determine the actual action to be performed based on the action probability distribution. The policy network used in this sampling phase can be referred to as the old policy. During the sampling process, the system not only records the state and actions It will also record the log probability of the action under the old strategy. Simultaneously record real-time training rewards Termination Mark and state value .

[0075] (2) Calculate the advantage estimate and bootstrap return; Based on the immediate training reward, termination flag, and state value estimate obtained in step (1), the discount factor γ and generalized dominance estimation parameters are used. Calculate the original advantage estimate and bootstrap reward, and normalize the original advantage estimate before policy update: Let the immediate training reward at step t be... The state value is The termination identifier is ,in This indicates that the current round ends after this time step. If the current round has not ended, then the time difference residual is... It can be represented as:

[0076] Original advantage estimate The generalized dominance estimation method is used to calculate the result recursively from back to front, as follows:

[0077] Where γ is the discount factor. For generalized dominance estimation parameters; for the end of the rollout segment, the recursive boundary can be used. Set to 0, and estimate the value of the final state. Participate in the calculation of the terminal time difference residual; Based on the original advantage estimate Calculate the bootstrap reward value used for supervising updates of the valuation branch:

[0078] in, The return value for Bootstrap.

[0079] The calculation of immediate training rewards includes: The instant training reward is calculated based on task acceptance benefits, airspace management costs, and penalties for cross-regional handovers. In this embodiment, a normalized reward format is used: ; in, This represents the immediate training reward that the reinforcement learning algorithm receives from the environment at time step t. This indicates the task acceptance rate at the current time step; This represents the normalized airspace management cost; This represents the number of cross-region handovers at the current time step, normalized by the number of tasks. , and These represent the weights for mission acceptance benefits, airspace management costs, and cross-regional handover frequency penalties, respectively. In the current embodiment, =4.0, =0.3, =0.0125. The above weights are pre-configured evaluation parameters, determined by the comprehensive evaluation results in the verification scenario, and maintained consistent in training rewards, candidate model selection, and final experimental evaluation. The normalized airspace management cost... It can be represented as:

[0080] in The set of active spatial units at the current moment. The hierarchical management weights are represented as follows:

[0081] in This represents the side length of the spatial cell corresponding to level l. This represents the management density coefficient of the finest-grained grid cell at this level; The This is a pre-configured hierarchical management density coefficient, used to characterize the relative management complexity corresponding to the unit finest-grained grid volume at different granularity levels. In the current embodiment, =0.25, =0.50, =0.75; The above values ​​are pre-set based on the relative management complexity of different levels and calibrated through verification scenarios, and are not output online by the reinforcement learning policy model.

[0082] This is a normalization constant used to normalize airspace management costs to a uniform dimension, avoiding direct comparisons of management costs under different airspace sizes and hierarchical management weight settings. It can be determined by the total number of the finest-grained flyable grid cells in the current scenario and the maximum value of the hierarchical management density coefficient, expressed as:

[0083] in, = This indicates that the management density coefficient is taken across all levels. The maximum value. This represents the total number of the finest-grained mesh cells that can fly in the current scene after deducting permanent obstacles.

[0084] (3) Based on the advantage estimate and bootstrap reward, the action decision branch is updated using PPO shearing target, and the value assessment branch is updated using mean squared error loss.

[0085] During the parameter update phase, the sampled rollout data remains constant. The system will collect the state... Re-enter the policy network currently being optimized and recalculate the actions already performed by the current policy during the sampling phase. probability It should be noted here that the current strategy This is the policy network being optimized during parameter updates; the old policy... This is the policy network used to generate this batch of rollout data. This occurs immediately after a rollout has ended and before the first gradient update has occurred. and The parameters are usually the same; however, after performing multiple small batches or multiple rounds of updates on the same batch of rollout data, It will change with gradient updates, and The probability of the action is always represented by the logarithmic probability recorded at the time of sampling. Therefore, during the PPO update phase, the two are distinguished as the "current policy" and the "old policy".

[0086] When updating the strategy, set the probability ratio between the old and new strategies. for:

[0087] In practical implementation, since the sampling phase typically stores the logarithmic probability of the action, the above formula can also be written as:

[0088] in, The log probabilities of actions saved during the rollout sampling phase. This is then recalculated by the current policy network during the parameter update phase. Therefore, the state... Used to re-enter the policy network, actions This is used to specify which action probability in the current policy output distribution should be read. The action log probability is used to represent the probability of that action under the old policy, and together they constitute the policy probability ratio of PPO.

[0089] Subsequently, the original advantage estimate was... Normalize the values ​​to the mean and standard deviation within the current rollout batch to obtain the normalized dominance estimate. Based on the normalized advantage estimate Ratio of strategy probability The objective function for PPO shearing in the action decision branch is expressed as:

[0090] in, The objective function for optimizing the action decision branch is θ, which is the objective function used to update the Actor branch parameters in the PPO algorithm. θ represents the policy network parameters currently being optimized. This objective function calculates the ratio of the probability between the old and new policies and updates the action decision branch based on the sampled state, action, the log probability of actions already executed under the old policy, and the normalized advantage estimate. This aims to increase the probability of actions with positive advantage and decrease the probability of actions with negative advantage, while limiting the magnitude of a single policy update.

[0091] In this context, E represents the expectation operator. In the PPO formula, it represents calculating the average target value over a batch of sampled training samples. Theoretically, reinforcement learning optimization objectives are often written in expectation form because policy updates are not based solely on a single time step or action, but rather aim to improve policy performance across the overall state-action distribution resulting from the interaction between the policy and the environment. Therefore, This can be understood as: statistically averaging the samples obtained from multiple time steps during the rollout sampling process.

[0092] This represents the shearing function in PPO, used to limit the probability ratio between the old and new policies. The range of variation. Let the PPO shear threshold be ε, then:

[0093] Indicates will The policy probability ratio is limited to the range [1-ε, 1+ε]. When the current policy changes too much compared to the old policy, the pruning function will cut off the change, thereby avoiding training instability caused by excessively large single update magnitude. In the current embodiment, ε can be 0.10. Therefore, in the pruning branch of the PPO pruning target, the policy probability ratio is limited to the range [0.90, 1.10]. The original probability ratio that is not pruned is still retained in the other target branch, and the magnitude of a single policy update is limited by taking the smaller value of the two target values.

[0094] On the other hand, state It will also input a value assessment branch to obtain a state value estimate under the current parameters. And compared with the bootstrap reward value obtained based on the instant training reward sequence, termination flag, and state value estimation. Construct the mean squared error loss:

[0095] in, The bootstrap reward value is obtained based on the instant training reward sequence, termination flag, and state value estimation. Used to update the value assessment branch, making the state value estimate Approximate Bootstrap Return Value To maintain policy exploration capabilities and avoid premature convergence, an entropy regularization term can be added to the policy objective, and the entropy coefficient can be annealed during training.

[0096] It can be seen that the state Simultaneously participate in the updating of both the action decision branch and the value assessment branch; Action The logarithmic probability of actions is mainly used in calculating the ratio of the old and new strategies' probabilities in the PPO (Probability of Opportunity Point) branch; while the immediate training reward, termination flag, and state value estimation are mainly used to calculate the advantage estimate. and Bootstrap reward value These data together constitute the complete sample information for the PPO / Actor-Critic training process.

[0097] In a preferred embodiment of the invention, the optimizer uses Adam, with a discount factor γ = 0.99, and the generalized dominance estimation parameters are... =0.97, PPO shearing threshold ε=0.10, number of parallel environments is 6, and the number of optimization rounds per parameter update is 3. As an example training parameter, the policy-side learning rate can be 3×10. -5 Up to 4×10 -5 The learning rate of the value evaluation branch can be 5×10^-4; the rollout step size can be 48 to 64; the mini-batch size can be 144 to 192; and the total number of update rounds can usually be 40 to 80. For scenarios that are easy to converge or have low to medium pressure transition, an early stopping or early optimal model selection mechanism can be adopted based on the comprehensive score of the validation set. The entropy coefficient can be annealed from 0.004 to 0.008 to 0.001 to 0.002 during training.

[0098] (4) Save multiple candidate policy models during training and select the target model based on the comprehensive validation score.

[0099] During the model selection process, candidate models are selected based on the overall validation score. In this embodiment, the overall validation score in the model selection phase can be expressed as:

[0100] in, This represents the average task acceptance rate in the verification scenario. This represents the average normalized airspace management cost in the verification scenario. This represents the average number of cross-region handovers for a single task in the verification scenario. Indicates the weight of task acceptance benefits; Indicates the weight of airspace management costs; This indicates the penalty weight for cross-regional handover items.

[0101] In the current embodiment, the model selection adopts =4.0、 =0.3、 =0.0125, consistent with the effective weight configuration in the immediate training reward.

[0102] It should be noted that, The validation comprehensive score during the model selection phase serves as a reward for single-step instant training. The overall score in the final static comparison experiment is a comprehensive evaluation result based on statistics from multiple verification or test rounds. The three are not the same numerical object, but they all use task acceptance benefits, normalized airspace management costs, and cross-area handover penalty terms as evaluation items, and use consistent weighting parameters. , and Configuration.

[0103] After training and model selection are completed, the target model is loaded to perform the airspace reconstruction decision, active airspace topology update, spatiotemporal path search and mission acceptance steps described in S2 to S5 above.

[0104] The target model outputs actions to maintain, split, or merge active airspace units based on the real-time airspace status, and combines the reconstructed active airspace topology, spatiotemporal reservation and occupancy table, and path planning results to complete task acceptance and the next round of status updates.

[0105] In a preferred embodiment of the invention, the low-altitude dynamic airspace allocation and task scheduling method proposed in this embodiment is compared with the existing fixed static allocation strategy, as follows: In this embodiment, the task pressure parameter λ represents the Poisson arrival intensity during the task generation process. The number of new tasks in each decision cycle follows a Poisson distribution with parameter λ; simultaneously, the task start and end points are sampled according to a preset take-off and landing point network and its weights. Therefore, λ not only affects the total number of tasks but also forms local task clusters of varying degrees near the weighted take-off and landing points. The larger λ is, the higher the task arrival pressure faced by the low-altitude airspace. In this embodiment, a static sweep is performed on the pressure range of λ=10~120 with a step size of 10. The results show that: for the pressure range of 10~30, L0 is optimal; for the pressure range of 40~60, L1 is optimal; and for the pressure range of 70~120, L2 is optimal.

[0106] In a preferred embodiment of the present invention, such as Figure 3 As shown, this is the static partitioning diagram under the current weighted terminal design. Based on this static partitioning result, representative pressure points λ=20, λ=30, λ=50, and λ=90 are selected from the L0 dominant region, L1 dominant region, and L2 dominant region, respectively, to verify the adaptability of the dynamic spatial domain partitioning method under low-pressure near-saturation, medium-low pressure transition, medium pressure, and high pressure conditions. For each representative task pressure, the aforementioned reinforcement learning training framework is used to complete the policy model training and candidate model selection under the same spatial domain modeling method, reward function form, and comprehensive evaluation index, obtaining the target model under the corresponding task pressure condition. Each target model outputs a spatial domain reconstruction action based on the reservation occupancy status and task arrival intensity context within the current decision cycle. In this embodiment, the comprehensive score in the final experimental stage is calculated based on the statistical results of the test scenario and can be written as: .

[0107] Where Acc represents the average task acceptance rate during the testing phase, Grid represents the average normalized airspace management cost, and Handover represents the average number of cross-area handovers per task.

[0108] The final overall experimental score uses the same weighting configuration as the immediate training reward and candidate model selection stages, i.e. =4.0、 =0.3、 =0.0125.

[0109] Table 1 below compares the overall performance of the method of this invention with the optimal fixed-granularity static partitioning strategy. It should be noted that the "average grid cost" listed in Table 1 is the original statistical value before normalization. When calculating the overall score, it is divided by the normalization constant. Substitute the values ​​into the Grid field. In the table, L0, L1, and L2 represent coarse-grained, medium-grained, and fine-grained fixed static partitioning strategies, respectively. For each task pressure, the strategy with the highest overall score among the three fixed static partitioning strategies is selected as the optimal fixed-grained static partitioning strategy and compared with the method described in this invention. The table also lists the task acceptance rate, average grid cost, average number of cross-region handovers per task, and overall score to reflect the comprehensive trade-off between task mobility and operational management complexity of different spatial partitioning methods.

[0110] Table 1: Comparison of the overall performance of the method of the present invention and the optimal fixed-granularity static partitioning strategy

[0111] As shown in Table 1, under low task pressure conditions (λ=20), the static L0 partitioning strategy can achieve near-complete acceptance, and the overall score of the method in this invention is basically close to that of static L0. As the task pressure increases to λ=30, the method in this invention improves the task acceptance rate through local spatial domain reconstruction, achieving a higher overall score than static L0. Furthermore, under medium-to-high task pressure conditions (λ=50 and λ=90), the fixed-granularity static partitioning strategy struggles to simultaneously balance task acceptance capacity and operational management costs. The method in this invention demonstrates a more significant overall performance advantage compared to the corresponding optimal fixed-granularity static partitioning strategy.

[0112] In this application, by using consistent evaluation items and weight configurations for immediate training rewards, validation comprehensive scores in the model selection phase, and comprehensive scores in the final experimental phase, the model selection bias caused by inconsistencies between training objectives, candidate model selection objectives, and final experimental evaluation objectives can be reduced, thereby improving the interpretability and reproducibility of experimental results.

[0113] Example 2

[0114] like Figure 4 As shown in the figure, this embodiment of the invention also proposes a low-altitude dynamic airspace allocation and task scheduling system, including: Airspace Configuration Module 1 configures basic information for low-altitude airspace and constructs a hierarchical grid airspace with parent-child nesting relationships; it determines the set of active airspace units in the hierarchical grid airspace and establishes a spatiotemporal reservation occupancy table that records the number of active airspace units reserved for each time slice. The state acquisition module 2 acquires the state features of the hierarchical grid spatial domain, inputs the state features into the trained reinforcement learning policy model, and the reinforcement learning policy model outputs maintain, split or merge actions for each active spatial domain unit, and performs legality verification on the split or merge actions. After performing the action that passes the legality check, the airspace reconstruction module 3 updates the set of active airspace units and their adjacency relationships, and constructs the active airspace topology graph. Task scheduling module 4 receives task scheduling requests, establishes a spatiotemporal search state based on active spatial domain units and arrival times, performs a time-dimensional path search on the active spatial domain topology map, and queries the spatiotemporal reservation occupancy table during candidate state transitions; when the number of reservations for a target active spatial domain unit in the corresponding arrival time slice is less than the capacity limit of that active spatial domain unit, the candidate state transition is allowed; it determines whether a complete path from the task start point to the destination exists. If it does, the task is accepted, and the spatiotemporal reservation occupancy table is updated based on the time-slice occupancy corresponding to the task; otherwise, the task is rejected.

[0115] Example 3

[0116] This embodiment provides a computer-readable storage medium storing a computer program thereon. When the program is executed by a processor, it implements the steps of the low-altitude dynamic airspace partitioning and task scheduling method as described in Embodiment 1.

[0117] As can be seen, in a preferred embodiment of the present invention, by using the active spatial domain units with parent-child nesting relationships as runtime reconstruction objects and synchronously processing existing task reservations during the splitting and merging process, the spatial domain granularity can be adjusted under different task pressures, while maintaining the existing spatiotemporal execution relationship and capacity safety of tasks.

[0118] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware, and of course, it can also be implemented by special hardware including application-specific integrated circuits, special CPUs, special memory, special components, etc. Generally, any function performed by a computer program can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be diverse, such as analog circuits, digital circuits, or special-purpose circuits. However, for this application, software program implementation is more often the preferred implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium, such as a computer floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk, or optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of this application.

[0119] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can store or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0120] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is only used to illustrate the technical solution of the present invention, and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A method for low-altitude dynamic airspace allocation and task scheduling, characterized in that, Includes the following steps: S1. Configure basic information for low-altitude airspace and construct a hierarchical grid airspace with parent-child nesting relationships; Determine the set of active spatial domain units in the hierarchical grid spatial domain, and establish a spatiotemporal reservation and occupancy table that records the number of active spatial domain units reserved and occupied per time slice; S2. Obtain the state features of the hierarchical grid spatial domain, input the state features into the trained reinforcement learning policy model, and have the reinforcement learning policy model output maintain, split or merge actions for each active spatial domain unit, and perform legality verification on the split or merge actions. S3. After executing the action that passes the legality check, update the set of active spatial units and their adjacency relationships, and construct the active spatial topology graph; S4. Receive a task scheduling request, construct a spatiotemporal search state based on the activity airspace unit and arrival time, and perform a path search with time dimension on the activity airspace topology map based on the spatiotemporal reservation occupancy table. S5. Determine if a complete path exists from the start point to the end point of the task. If it exists, accept the task and update the spatiotemporal reservation occupancy table based on the time slice occupancy corresponding to the task; otherwise, reject the task.

2. The low-altitude dynamic airspace allocation and task scheduling method as described in claim 1, characterized in that, The basic information of the low-altitude airspace includes three-dimensional airspace boundaries, permanent obstacle constraints, take-off and landing point network and its weights; The hierarchical grid spatial domain includes at least two granularity levels, and the same parent unit has a spatial containment relationship with multiple child units of its next granularity level.

3. The low-altitude dynamic airspace allocation and task scheduling method as described in claim 1, characterized in that, The state features include the number of reservations and occupancy per time slice, the three-dimensional spatial state tensor, pressure context features, active spatial unit features, and the action legality mask of the active spatial unit. The stress context features are constructed based on the long-window exponential moving average and the short-window exponential moving average of the number of newly generated tasks in the current period.

4. The low-altitude dynamic airspace allocation and task scheduling method as described in claim 3, characterized in that, The reinforcement learning strategy model includes a three-dimensional convolutional feature extraction module, a context encoding module, an active spatial domain unit feature encoding module, an action decision branch, and a value evaluation branch. The three-dimensional spatial domain state tensor is used by the three-dimensional convolutional feature extraction module to form global spatial domain features, the pressure context features are used by the context encoding module to form context features, and the active spatial domain unit features, action legality encoding, and the context features together form unit structure features. The action decision branch outputs the probability distribution of three types of actions: maintain, split, and merge, based on global airspace characteristics and unit structure characteristics, while the value evaluation branch outputs the current airspace state value.

5. The low-altitude dynamic airspace allocation and task scheduling method according to claim 4, characterized in that, The reinforcement learning policy model is trained using a shared parameter Actor-Critic method based on PPO cut-off targets. State features, actions, log probabilities of old policy actions, immediate training rewards, termination flags, and state values ​​are collected at continuous time steps in multiple parallel environments. Generalized advantage estimates and bootstrap rewards are calculated. The action decision branch is updated using PPO cut-off targets, and the value evaluation branch is updated using mean squared error loss. Multiple candidate policy models are saved during training, and the target model is selected based on evaluation items consistent with the final experimental evaluation: task acceptance benefit, airspace management cost, and cross-regional handover cost.

6. The low-altitude dynamic airspace allocation and task scheduling method according to claim 1, characterized in that, The legality verification of the splitting action specifically includes: The entry and exit sub-units are determined based on the geometric positions of existing tasks entering and leaving the parent unit; local sub-paths connecting the entry and exit sub-units are determined in the sub-unit adjacency graph; the parent unit reservation occupancy is mapped to the sub-units corresponding to the local sub-paths according to the time sequence of existing tasks; if the local sub-path does not exist, or if the reservation occupancy of any sub-unit exceeds the capacity limit of that sub-unit after migration in any time slice, the verification is deemed to fail; otherwise, the verification passes.

7. The low-altitude dynamic airspace allocation and task scheduling method according to claim 1, characterized in that, The specific steps for validating the merger action include: Confirm that all child units corresponding to the same parent unit are in the current active airspace unit set; aggregate the reservation occupancy of all child units in the same time slot; if the aggregated reservation occupancy of any time slot exceeds the parent unit's capacity limit, or if the geometric boundary of the parent unit includes a permanent obstacle area into the flyable airspace, the merging action verification is deemed to fail; otherwise, the verification passes.

8. The low-altitude dynamic airspace allocation and task scheduling method according to claim 1, characterized in that, The path search with time dimension adopts a heuristic priority search, with active spatial domain units as spatial nodes and discrete time slices that reach the active spatial domain units as time states. The cost of a candidate state transition includes the number of discrete time steps required for the movement and the additional cost of crossing activity spatial units, and the candidate state transition is allowed when the number of arrival time slice reservations for the target activity spatial unit is less than the upper limit of the capacity of that activity spatial unit.

9. A low-altitude dynamic airspace allocation and task scheduling system, characterized in that, include: The airspace configuration module configures basic low-altitude airspace information and constructs a hierarchical grid airspace with parent-child nesting relationships. Determine the set of active spatial domain units in the hierarchical grid spatial domain, and establish a spatiotemporal reservation and occupancy table that records the number of active spatial domain units reserved and occupied per time slice; The state acquisition module acquires the state features of the hierarchical grid spatial domain, inputs the state features into the trained reinforcement learning policy model, and the reinforcement learning policy model outputs hold, split or merge actions for each active spatial domain unit, and performs legality verification on the split or merge actions. The airspace reconstruction module, after performing actions that pass the legality verification, updates the set of active airspace units and their adjacency relationships, and constructs the active airspace topology graph; The task scheduling module receives task scheduling requests, constructs a spatiotemporal search state using active airspace units and arrival times, performs a path search with a time dimension on the active airspace topology map, and queries the spatiotemporal reservation and occupancy table during the candidate state transition process. When the number of reservations for a target activity space unit in the corresponding arrival time slot is less than the upper limit of the activity space unit's capacity, the candidate state transition is allowed. Determine if a complete path exists from the start point to the end point of the task. If it exists, accept the task and update the spatiotemporal reservation occupancy table based on the time slice occupancy corresponding to the task; otherwise, reject the task.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the low-altitude dynamic airspace partitioning and task scheduling method as described in any one of claims 1 to 8.

Images