Low-altitude airspace capacity dynamic evaluation method and system based on time sequence diagram neural network

Abstract

The invention discloses a low-altitude airspace capacity dynamic evaluation method and system based on a time sequence diagram neural network, and belongs to the technical field of low-altitude traffic management, and the method comprises a data collection and fusion module which is used for accessing a multi-source heterogeneous data source, and carrying out the cleaning, standardization and fusion; the dynamic space-time diagram construction module is used for constructing a dynamic space-time diagram evolved along with time according to the processed multi-source heterogeneous data; a time sequence prediction engine module which is embedded with a time sequence diagram neural network model and is used for receiving the time sequence of the dynamic space-time diagram and outputting a predicted operation risk score; and the capacity evaluation and quantification module is used for calculating the dynamic operation capacity based on the predicted operation risk score, and providing services for an upper management platform through an application program interface. According to the method, real-time calculation and accurate prediction of a core decision variable, namely the airspace operation capacity, are realized, so that the most fundamental scientific basis is provided for refined and intelligent management of the airspace, and the safety and efficiency of urban low-altitude traffic are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of low-altitude traffic management, and particularly relates to a low-altitude airspace capacity dynamic evaluation method and system based on a time series graph neural network. BACKGROUND

[0002] With the rapid development of unmanned aerial vehicle (UAV) technology and the continuous expansion of commercial applications, urban low-altitude airspace is gradually becoming a new transportation network after ground transportation and high-altitude civil aviation. Diversified urban air mobility (UAM) and advanced air mobility (AAM) operation modes such as logistics drones, air taxis, and emergency rescue aircraft indicate that future urban low-altitude will present a high-density, high-dynamic, and high-complexity "meshed" operation situation. However, the existing air traffic management (ATM) system and its derivative technologies face significant technical bottlenecks in coping with this emerging challenge, including:

[0003] 1. Limitations of traditional air traffic management (ATM) models:

[0004] The traditional air traffic management system mainly relies on programmed flight plans, fixed route structures, and rule-based separation standards. The core of this system is to manage based on static, pre-declared airspace capacity. This capacity is usually defined as a fixed value, for example, the maximum number of aircraft that a control sector can safely handle in an hour, which is mainly calculated based on historical data, sector geometry, and empirical assessment of air traffic controller (ATCO) workload. Although there is a concept of "tactical capacity" (Operational Capacity) that considers dynamic factors, its adjustment relies on the on-site judgment of human controllers rather than real-time prediction based on models. For the rapidly changing urban low-altitude environment, this management method based on static or semi-static capacity assumptions has a serious lag in response and cannot effectively handle the rapid changes in airspace physical capacity caused by sudden conditions (such as extreme weather, communication interruptions), resulting in low utilization of airspace resources and difficulty in supporting the safety and efficiency needs of large-scale commercial operations.

[0005] 2. Limitations of traditional machine learning (ML) methods:

[0006] To improve the prediction capability, some studies turned to statistical or traditional machine learning models based on historical data, such as autoregressive integrated moving average (ARIMA), support vector machine, or random forest, etc. These methods have certain effect in dealing with specific time series data or classification tasks, but there are two core defects. First, they rely heavily on manual feature engineering, making it difficult to automatically capture the complex nonlinear spatiotemporal correlations in urban low-altitude traffic flow. Second, these models usually treat the airspace state as a set of isolated feature vectors, losing the crucial topological structure and spatial dependence between aircraft and between aircraft and the environment. More importantly, many machine learning models, especially deep learning models, are criticized for their "black box" characteristics, and their decision-making process lacks transparency and interpretability, which is a major obstacle to their application in the safety-oriented aviation field.

[0007] 3. Technical gap in cutting-edge graph neural network (GNN) research:

[0008] In recent years, graph neural networks (GNN) have been widely used in the field of air traffic analysis due to their excellent performance in handling spatiotemporal data with graph structure, representing the current technical frontier. Researchers use GNN to construct interaction graphs between aircraft or control sectors and successfully apply them to predict proxy metrics highly related to controller workload. For example, the endpoint of a large number of cutting-edge research is to predict "airspace complexity," the "number of clearances" that controllers will soon issue, or directly predict the "ATCO workload" level. These studies have greatly improved the level of intelligence in airspace state perception.

[0009] However, a thorough analysis of these prior arts reveals a fundamental technical gap: none of them has addressed how to translate the predicted abstract, qualitative or human-cognition-related proxy indicators into a physical, quantitative "airspace capacity" value that can be directly used for traffic management decisions. "airspace complexity" or "ATCO workload" is essentially a measure of how challenging the airspace state is to the cognitive ability of human controllers. A high complexity or high workload scenario can be fundamentally caused by a decrease in airspace physical capacity due to bad weather, which requires controllers to exert more effort to maintain safe separation. The existing GNN models successfully predict the "symptom" (high workload), but do not diagnose the "cause" and give a quantitative indicator (the capacity has decreased by what percentage). For future automated or semi-automated UAM traffic management systems, a "high complexity" prediction result itself is ambiguous and not directly executable; the system needs a clear answer: "how many aircraft can this airspace safely accommodate per minute under the current conditions?" Solving the problem of translating from predicted abstract risks to quantitative physical capacity is a key step to achieve fine-grained and intelligent airspace resource management, which is the gap that the existing technologies have failed to fill.

[0010] Therefore, the technical problem to be solved in the field is that there is a lack of a system and method that can systematically integrate the multi-source heterogeneous physical influencing factors of urban low-altitude three-dimensional environment and directly quantify and predict the dynamic operation capacity of airspace using advanced artificial intelligence models. Therefore, there is an urgent need for a technical solution that can dynamically and accurately evaluate airspace capacity. SUMMARY

[0011] The purpose of the present application is to overcome the technical problem in the prior art that the actual operation capacity of a complex three-dimensional low-altitude environment cannot be accurately and dynamically quantitatively evaluated, and to provide a low-altitude airspace capacity dynamic evaluation method and system based on a time series graph neural network.

[0012] To solve the above technical problems, the present application provides the following technical solutions:

[0013] On the one hand, a low-altitude airspace capacity dynamic evaluation method based on a time series graph neural network is disclosed, comprising the following steps:

[0014] S1: discretizing a target urban low-altitude airspace into a plurality of three-dimensional voxels, defining the three-dimensional voxels, key infrastructure and aircraft in the airspace as static nodes, infrastructure nodes and dynamic nodes of a graph respectively, and defining preset routes, spatial adjacency relationships between nodes, dynamic interaction relationships between aircraft and ownership relationships as edges of the graph;

[0015] S2: Real-time acquisition and fusion of multi-source heterogeneous data, which includes at least three-dimensional meteorological data and communication, navigation and surveillance (CNS) performance data, and the processed multi-source heterogeneous data is used as dynamic feature vectors of nodes and edges to construct a dynamic spatiotemporal graph that evolves over time.

[0016] S3: Input the time series of the dynamic spatiotemporal graph into a preset time series graph neural network model. The time series graph neural network model predicts the spatial state within a future time window and outputs a predicted operational risk score that includes at least each of the static nodes.

[0017] S4: Based on the predicted operational risk score, calculate the dynamic operational capacity of each static node within the future time window using a preset capacity assessment algorithm.

[0018] By adopting the above technical solution, an end-to-end assessment framework that can comprehensively reflect the physical environment of the airspace is constructed, enabling real-time calculation and accurate prediction of the core decision variable of airspace operation capacity. This provides the most fundamental scientific basis for the refined and intelligent management of airspace and significantly improves the safety and efficiency of urban low-altitude traffic.

[0019] As a preferred embodiment of the present invention, the multi-source heterogeneous data mentioned in step S2 further includes: real-time aircraft status data, and geographic information and airspace constraint data.

[0020] As a preferred embodiment of the present invention, the temporal graph neural network model described in step S3 includes a feature-level attention module. The feature attention module is configured to generate a weight vector for the input multidimensional feature vector containing multiple heterogeneous feature dimensions with different physical meanings based on the global or local spatial context, and to weight the original feature vector.

[0021] As a preferred embodiment of the present invention, step S3 includes: receiving one or more predicted operational risk scores for the low-altitude airspace volume of the target city at one or more future time points from a trained time-series graph neural network model whose input is a dynamic spatiotemporal graph.

[0022] As a preferred embodiment of the present invention, the temporal graph neural network model described in step S3 is a gated temporal graph convolutional network (A-GTGCN).

[0023] As a preferred embodiment of the present invention, the gated temporal graph convolutional network includes:

[0024] The spatial feature aggregation module employs a graph convolutional network (GCN).

[0025] The time-based dynamic evolution module employs a gated loop unit (GUR) to effectively capture temporal dependencies through update and reset gates.

[0026] The feature-level attention module employs an additive attention mechanism. It calculates the importance scores of each feature dimension through a single hidden layer feedforward network and normalizes them into a weight vector using the Softmax function, ultimately obtaining weighted features for adaptive identification of dominant risk factors.

[0027] As a preferred embodiment of the present invention, the capacity assessment algorithm in step S4 includes: multiplying the baseline capacity of the static node by a dynamic attenuation coefficient determined by the predicted operational risk score, thereby calculating the dynamic operational capacity of the target city's low-altitude airspace volume at the future time point.

[0028] As a preferred embodiment of the present invention, the baseline capacity of the static node is a baseline capacity value of the low-altitude airspace volume of the target city.

[0029] On the other hand, a dynamic assessment system for low-altitude airspace capacity based on a temporal graph neural network is disclosed, including:

[0030] The data acquisition and fusion module is used to access multi-source heterogeneous data sources and to clean, standardize, and fuse the multi-source heterogeneous data.

[0031] The dynamic spatial domain graph construction module is used to construct a dynamic spatiotemporal graph that evolves over time based on the processed multi-source heterogeneous data.

[0032] The temporal graph neural network (GNN) prediction engine module has an embedded temporal graph neural network model, which is used to receive the time series of the dynamic spatiotemporal graph and output the predicted operational risk score.

[0033] The capacity assessment and quantification module is used to calculate the dynamic operating capacity based on the predicted operating risk score and provide services to the upper-level management platform through the application programming interface (API).

[0034] Compared with the prior art, the advantages of the present invention are as follows:

[0035] 1. A breakthrough from abstract indicators to quantifiable capacity: For the first time, a complete technical closed loop was proposed and implemented, which can transform the abstract "operational risk" predicted by the GNN model into a specific, measurable "dynamic operating capacity" value with clear physical meaning. This solves the technical bottleneck of existing academic research that is stuck in predicting proxy indicators such as "complexity" or "workload", so that the prediction results can be directly applied to core businesses such as traffic management and resource planning, which has great practical application value.

[0036] 2. A more comprehensive and accurate airspace model has been constructed: The "dynamic spatiotemporal map" construction method proposed in this invention, by taking airspace voxels, infrastructure and aircraft as nodes and integrating multi-source physical data such as meteorology and CNS, fundamentally surpasses the "traffic interaction map" model in the existing technology that only focuses on the interaction between aircraft. This systematic modeling approach can more comprehensively depict various physical factors affecting the airspace operation status, laying a solid foundation for achieving high-precision prediction.

[0037] 3. Improved model adaptability and robustness: By introducing a feature-level attention mechanism that is closely tied to the UAM operation scenario, the model can adaptively identify and focus on the dominant physical risk factors in the current scenario based on the real-time changing airspace context (such as sudden weather changes and communication interference). This makes the capacity assessment results more robust and adaptable to the complex and ever-changing urban low-altitude environment.

[0038] 4. Provides core decision-making basis for refined management of airspace resources: This invention transforms "airspace capacity" from a static, coarse, rule-based parameter into a dynamic quantitative indicator that can be calculated in real time and accurately predicted. This provides the most core scientific basis for the supply and demand balance, resource optimization and refined management of urban low-altitude transportation, and is a key enabling technology for realizing large-scale and high-efficiency commercial operation of UAM in the future. Attached Figure Description

[0039] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. In the drawings:

[0040] Figure 1 This is a schematic diagram comparing the graph construction method of the low-altitude airspace capacity dynamic assessment method based on the temporal graph neural network described in this embodiment of the invention with existing graph construction methods;

[0041] Figure 2 This is a flowchart illustrating a dynamic assessment method for low-altitude airspace capacity based on a time-series graph neural network as described in Embodiment 1 of the present invention.

[0042] Figure 3 This is a schematic diagram of the dynamic airspace environment state diagram construction method of the low-altitude airspace capacity dynamic assessment method based on a time-series graph neural network as described in Embodiment 1 of the present invention.

[0043] Figure 4 This is a schematic diagram of the A-GTGCN neural network model structure of a dynamic evaluation method for low-altitude airspace capacity based on a time-series graph neural network as described in Embodiment 2 of the present invention.

[0044] Figure 5 This is a schematic diagram of the architecture of a low-altitude airspace capacity dynamic assessment system based on a time-series graph neural network, as described in Embodiment 3 of the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0046] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, the terms "first," "second," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance, or suggesting any such actual relationship or order between these entities or operations. Additionally, the terms "connected," "linked," etc., can refer to a direct connection between components or an indirect connection via other components.

[0047] Traditional GNN models, when applied to ground traffic, typically assume a static graph consisting of roads and intersections. However, in the fluid medium of three-dimensional low-altitude airspace, such a readily available graph structure does not exist. Therefore, how to create a meaningful, dynamic, and information-rich graph representation for this complex environment is the primary problem this invention aims to solve, and a fundamental innovation that distinguishes it from existing technologies.

[0048] (1) Node Definition: Nodes in a graph represent basic entities in the spatial domain. This invention defines three types of nodes to comprehensively characterize the state of the spatial environment:

[0049] Static Nodes: The entire urban target's low-altitude airspace is discretized into three-dimensional mesh units, or voxels, according to a uniform size (e.g., 100m×100m×30m). Each voxel serves as a static node, forming the background mesh of the airspace and acting as the primary carrier of environmental information (such as meteorological data).

[0050] Infrastructure Nodes: These abstract ground infrastructure such as vertical take-off and landing sites, key communication base stations, and navigation stations into functional nodes.

[0051] Dynamic Nodes: Each UAV operating in the airspace is considered a dynamic node.

[0052] (2) Edge Definition: The edges of a graph represent the relationships or interactions between nodes. This invention defines various types of edges to capture different dependencies:

[0053] Adjacency Edges: Connect adjacent voxel nodes in space, representing connectivity in physical space.

[0054] Route Edges: Connect voxel nodes along the route based on the preset UAM route or flight corridor, representing strong correlation in the planning.

[0055] Interaction Edges: Dynamically connect aircraft nodes whose distance is less than a certain threshold, used to explicitly model potential conflict relationships between aircraft.

[0056] Belonging Edges: Connect dynamic aircraft nodes to their current voxel nodes, and infrastructure nodes to their voxel nodes, representing a subordinate or containment relationship.

[0057] (3) Feature Engineering of Nodes and Edges: This is a crucial step in integrating multi-source heterogeneous data into the graph model. A real-time updated multi-dimensional feature vector is associated with each node and each edge, as shown in Table 1. For example, the feature vector of a voxel node can incorporate wind shear coefficient and turbulence intensity provided by a 3D numerical weather model, as well as traffic density statistics from a monitoring system; the feature vector of an aircraft node can include communication reliability scores provided by a communication link monitoring system and navigation accuracy levels provided by a GNSS monitor. Using the EKF state estimation algorithm described in Section 1, high-fidelity position, velocity, and attitude characteristics can be provided for aircraft nodes, and based on these precise states, dynamic features such as traffic_density and congestion_index of the voxel nodes can be aggregated and calculated.

[0058]

[0059]

[0060]

[0061] Table 1. Examples of data sources and engineering features for figure modeling.

[0062] Through the above steps, this invention successfully transforms a complex, unstructured physical environment problem into a well-defined, information-rich dynamic spatiotemporal graph learning problem.

[0063] like Figure 1 As shown, this diagram visually compares the graph construction method of this invention with existing technologies. The left side displays a common "traffic interaction graph" in existing technologies. In this type of graph, nodes typically only represent aircraft, and edges represent distances or potential interactions between aircraft. This representation can capture the interactions within traffic flow but completely ignores the background environment in which the aircraft are located. The right side displays the "airspace environment state graph" proposed in this invention. This graph, by introducing voxel nodes and infrastructure nodes and endowing them with rich environmental features (meteorology, CNS, etc.), constructs a complex system that can comprehensively reflect the interaction between aircraft and the environment. This intuitive visual comparison greatly enhances the understanding of the novelty of this invention at the model representation level.

[0064] Example 1

[0065] A dynamic assessment method for low-altitude airspace capacity based on a temporal graph neural network, such as... Figure 2 As shown, it includes the following steps:

[0066] S1: As Figure 3 As shown, the low-altitude airspace of the target city is discretized into multiple three-dimensional voxels (i.e., three-dimensional grid cells). The three-dimensional voxels, key infrastructure (such as vertical take-off and landing fields) in the airspace, and each aircraft operating in the airspace are defined as static nodes, infrastructure nodes, and dynamic nodes of the graph, respectively. The preset routes, spatial adjacency relationships between nodes, dynamic interaction relationships between aircraft, and attribution relationships are defined as edges of the graph.

[0067] S2: Acquire and fuse multi-source heterogeneous data in real time through a data interface. The multi-source heterogeneous data includes at least three-dimensional meteorological data and communication, navigation and monitoring performance data. The processed multi-source heterogeneous data is used as the dynamic feature vector of nodes and edges to construct a dynamic spatiotemporal map that can evolve over time and characterize the state of the airspace environment.

[0068] The aforementioned multi-source heterogeneous data also includes: real-time aircraft status data, as well as geographic information and airspace constraint data.

[0069] S3: Input the time series of the dynamic spatiotemporal graph into a preset time series graph neural network model. The time series graph neural network model predicts the spatial state within a future time window and outputs a predicted operational risk score that includes at least each of the static nodes.

[0070] Specifically, the temporal graph neural network model includes a feature-level attention module, which is configured to generate a weight vector for the input multidimensional feature vector containing multiple heterogeneous feature dimensions with different physical meanings based on the global or local spatial context, and to weight the original feature vector.

[0071] The dynamic spatiotemporal graph sequence constructed in the previous step, which contains a period of time in the past, is used as the model input. The model learns complex spatiotemporal dependencies through its internal modules and finally outputs a comprehensive state prediction vector for each spatial node (especially voxel nodes) in one or more future time steps. This vector includes at least a predicted operational risk score for each of the static nodes.

[0072] S4: Based on the predicted operational risk score, calculate the dynamic operational capacity of each static node within the future time window using a preset capacity assessment algorithm.

[0073] Specifically, it receives one or more predicted operational risk scores for the low-altitude airspace volume of a target city at one or more future time points from a trained temporal graph neural network model that takes a dynamic spatiotemporal graph as input.

[0074] Specifically, the capacity assessment algorithm multiplies a preset baseline capacity value by a dynamic attenuation coefficient determined by the predicted operational risk score, thereby calculating the dynamic operational capacity of the airspace volume at a future time point. This capacity represents the maximum traffic flow that the airspace unit can accommodate under the premise of ensuring safety.

[0075] Example 2

[0076] This embodiment is a preferred implementation of the low-altitude airspace capacity dynamic assessment method based on a time-series graph neural network described in Embodiment 1;

[0077] Preferably, the data fusion implementation includes: to address the heterogeneity of multi-source data in terms of update rate, accuracy, and noise, this step preferably employs an Extended Kalman Filter (EKF) to estimate the state of dynamic nodes (aircraft). The EKF effectively fuses nonlinear and noisy data from sensors such as GPS, inertial measurement units (IMU), and barometric altimeters through recursive prediction and update steps, generating a smooth and high-precision nine-dimensional state vector (three-dimensional position, three-dimensional velocity, and three-dimensional attitude). This high-fidelity state vector and its covariance matrix constitute the direct source of aircraft node features and are used to accurately calculate aggregate features such as traffic density and congestion index of the voxel node to which it resides, thereby providing high-quality input for subsequent neural network models.

[0078] While standard time series diagram models (such as STGCN) can combine spatial and temporal information, the importance of different factors changes dynamically in low-altitude environments with numerous influencing factors under different scenarios. For example, in clear weather, traffic density is the dominant risk factor; while during storms, the weights of wind shear and turbulence should be significantly increased.

[0079] Preferably, the temporal graph neural network model constructed in this invention is a gated temporal graph convolutional network, such as... Figure 4 As shown, the gated temporal graph convolutional network includes:

[0080] The spatial feature aggregation module models the spatial dependencies in the graph at each time step, employing a graph convolutional network. Its layer-by-layer propagation mathematical expression is as follows:

[0081]

[0082] Among them, H (l) For node features, W (l) For learnable weights, To incorporate self-loops into the adjacency matrix, a time-dynamic evolution module is used. This module receives a sequence of feature representations of a node changing over time and employs a gated loop unit, updating the gate z. t =σ(W z ·[h t-1′ ,h t ]+b z ) and reset door r t =σ(W r ·[h t-1′ ,h t ]+b r This method can be used to solve the gradient vanishing problem in long sequence training and can effectively capture long-term dependencies and short-term patterns in time series.

[0083] The feature-level attention module dynamically assigns weights to each dimension of the input features. It employs an additive attention mechanism, using a single hidden layer feedforward network to generate a weight vector α for the input multi-dimensional feature vector F based on the current spatial context. The calculation process is as follows: first, the attention score e = v is calculated. T tanh(W·F+b w Then, the weights α = softmax(e) are obtained by normalizing the original feature vectors using the Softmax function. The original feature vectors are then element-wise multiplied with these weight vectors to obtain the weighted feature vector F. attended =α⊙F. In this way, the model can learn during training to automatically "pay more attention" to related feature dimensions (such as weather or communication dimensions) when certain specific conditions are detected, so as to achieve adaptive identification of the dominant physical risk factors in different operating scenarios.

[0084] Model training scheme: The goal of model training is to minimize the mean squared error (MSE) loss function between the predicted risk and the actual risk.

[0085]

[0086] in, To represent the loss function itself, its calculation result is a scalar value used to measure the difference between the model's prediction and the actual result, also known as the mean squared error (MSE); n represents the batch size, i.e., the number of samples processed by the model in a single training iteration; i represents the index of the samples in the batch, from 1 to n; y i This represents the true value or true label of the i-th sample. Specifically, it refers to the true operational risk score obtained through post-hoc analysis or simulation. Let represent the model's predicted value for the i-th sample. Specifically, this refers to the predicted operational risk score output by the A-GTGCN model.

[0087] The optimization process employs the Adam optimizer and sets a set of baseline hyperparameters, such as a learning rate of 1×10⁻³ and a batch size of 64, and applies Dropout regularization to prevent overfitting. Preferably, the capacity assessment algorithm in step S4 includes: multiplying the baseline capacity of the static node by a dynamic decay coefficient determined by the predicted operational risk score, thereby calculating the dynamic operational capacity of the target city's low-altitude airspace volume at the future time point.

[0088] Furthermore, the input to the time-series graph neural network model is a three-dimensional tensor X with dimensions (T, N, F), where T is the number of historical time steps, N is the total number of nodes in the graph, and F is the feature dimension of each node. The output of the model is a three-dimensional tensor Y with dimensions (T′, N, O), where T′ is the number of predicted future time steps, and O is the output dimension of each node. In this embodiment, O can be 1, representing the predicted runtime risk score of the node (e.g., a voxel) at a future time.

[0089] Furthermore, the baseline capacity of the static node is a baseline capacity value of the low-altitude airspace volume of the target city.

[0090] Specifically, the baseline capacity (C base The calculation of the baseline capacity: The baseline capacity is not a single value, but is determined by a multi-factor model: C base (v)=C vol ×F rules ×F perf Among them, C vol It is the basic volume capacity based on voxel geometry and safety interval standards; F rules It is an adjustment factor associated with GIS data and reflecting static airspace usage rules; F perf It is an adjustment factor that reflects the performance of the region's communication, navigation, and surveillance infrastructure.

[0091] Specifically, the implementation of the dynamic decay function: The output of the GNN model is a high-dimensional prediction vector containing a quantitative score of future operational risk. However, as described in the background section, simply obtaining a risk score is insufficient for practical airspace management. One of the core inventive steps of this invention is to design an algorithm that transforms this predicted risk into an intuitive and executable dynamic capacity value.

[0092] This module transforms the prediction results of the GNN into an intuitive capacity metric. For each voxel node v and a future time t, its dynamic operating capacity Capacity(v,t) can be calculated using the following algorithm:

[0093] Capacity(v,t) = C base (v)×f(Risk(v,t))

[0094] Where: C base (v) is the baseline capacity of voxel v. This is a static value pre-calculated based on factors such as its geometry, static airspace constraints (e.g., whether it is within an airway), and CNS performance, for example, "8 sorties per minute under nominal conditions." It represents the maximum throughput under ideal conditions. Its calculation formula is: C base (v)=C vol×F rules ×F perf C vol Based on volume capacity, F rules F is the airspace rule adjustment factor. perf This is a CNS performance adjustment factor.

[0095] Risk(v,t) is a comprehensive scalar operational risk score predicted by the GNN model for voxel v at future time t.

[0096] This score can be obtained by weighted summation of the multidimensional prediction vectors output by the model:

[0097]

[0098] Where Risk(v,t) represents the comprehensive scalar operational risk score for voxel v at future time t. This is the final risk value obtained by weighted summation of the multidimensional prediction vectors output by the model; v represents a voxel, i.e., a three-dimensional grid cell (static node) in the spatial domain; t represents a future time point or time step; O represents the number of dimensions in the output vector of the time-series graph neural network model; i represents the dimension index in the output vector, from 1 to 0; w i Let represent the weight coefficients of the i-th output dimension. These weights can be pre-set according to the operating strategy to adjust the importance of different predictive factors (such as weather risk, congestion risk, etc.) in the final comprehensive risk score.

[0099] Y i (v,t) represents the value of the i-th dimension in the multidimensional prediction vector Y output by the time-series graph neural network model for voxel v at future time t.

[0100] f is a dynamic decay function with a range between (0,1). It transforms the risk score into a discount factor on the baseline capacity. This invention preferably employs a generalized logic function (Sigmoid function) for implementation, with the following form: Where R0 is the risk inflection point and k is the decay steepness. This S-shaped decay function can realistically simulate the system's resilience in the low-risk region and the saturation effect in the high-risk region, and is more physically plausible than simple linear or exponential decay.

[0101] Through this algorithm, the present invention successfully connects the predictive capabilities of GNN with the actual needs of airspace management, transforming airspace capacity from a fixed number into a dynamic value that changes in real time with environmental and traffic conditions, thus providing a foundation for refined management.

[0102] Example 3

[0103] A dynamic assessment system for low-altitude airspace capacity based on a temporal graph neural network, such as...Figure 5 As shown, it includes:

[0104] The Data Ingestion & Fusion Layer is used to access multi-source heterogeneous data sources and clean, standardize, and fuse the multi-source heterogeneous data.

[0105] Specifically, the data acquisition and fusion module serves as the system's perception foundation. This module, through standardized data interfaces, is responsible for accessing information from multiple real-time or near-real-time data sources. These data sources include, but are not limited to: surveillance data sources (such as ADS-B), meteorological data sources (such as 3D numerical weather prediction (NWP)), communication and navigation data sources (such as C2 link surveillance systems), and geographic and constraint data sources (such as GIS databases). This module also includes a data preprocessing unit responsible for cleaning, denoising, coordinate transformation, and time alignment of the raw data. In this module, state estimation algorithms such as the Extended Kalman Filter (EKF) are preferably used to fuse noisy data from different sensors to generate a high-precision aircraft state vector.

[0106] The Dynamic Graph ConstructorModule is used to construct a dynamic spatiotemporal graph that evolves over time based on processed multi-source heterogeneous data.

[0107] Specifically, this module is responsible for transforming unstructured multi-source data into a structured spatiotemporal graph. It discretizes the low-altitude airspace of the target city into multiple three-dimensional voxels (i.e., three-dimensional grid cells), defining the three-dimensional voxels, key infrastructure within the airspace (such as vertical takeoff and landing fields), and each aircraft operating in the airspace as static nodes, infrastructure nodes, and dynamic nodes of the graph, respectively. Preset routes, spatial adjacency relationships between nodes, dynamic interaction relationships between aircraft, and attribution relationships are defined as edges of the graph. Through a data interface, it acquires and fuses multi-source heterogeneous data in real time. This multi-source heterogeneous data includes at least three-dimensional meteorological data and communication, navigation, and surveillance performance data. The processed multi-source heterogeneous data is used as dynamic feature vectors for nodes and edges, thereby periodically (e.g., every 30 seconds) generating a graph snapshot representing the current airspace state, i.e., generating a dynamic spatiotemporal graph.

[0108] The Temporal GNN Prediction Engine module embeds a temporal graph neural network model, which is used to receive the time series of the dynamic spatiotemporal graph and output the predicted operational risk score.

[0109] Specifically, the time-series prediction engine module is the core of the system's artificial intelligence. It embeds a time-series graph neural network model that has been trained offline and fine-tuned online. The engine receives a series of continuous graph snapshots generated by the dynamic graph construction module as input, predicts the spatial state over a future period, and outputs a prediction vector containing information such as operational risk scores.

[0110] The Capacity Evaluation & Quantification Module is used to calculate the dynamic operating capacity based on the predicted operating risk score and provide services to the upper-level management platform through the application programming interface.

[0111] Specifically, the capacity assessment and quantification module is responsible for transforming the abstract risk score output by the GNN prediction engine into a concrete, actionable capacity value. It includes the capacity quantification algorithm detailed in Section 4 below and pushes the final dynamic capacity value to the upper-level Unmanned Aerial Vehicle Traffic Management (UTM) platform or airspace planning system through a standardized application programming interface (API).

[0112] The system's data flow is as follows: After multi-source heterogeneous data is collected and fused, it is sent to the dynamic graph construction module to generate a time-series graph data stream; this data stream is input into the GNN prediction engine to generate predictions of future operational risks; finally, the capacity assessment and quantification module analyzes these predictions to generate dynamic capacity values, thereby achieving accurate assessment of spatial capacity.

[0113] The technical solution proposed in this invention is fundamentally different from the prior art described in the background section. In particular, compared with the latest academic research that also uses the GNN model, the innovative value of this invention lies in its solution to different and more fundamental technical problems.

[0114] Existing GNN models typically aim to predict "airspace complexity" or "ATCO workload." These metrics are essentially simulations and measurements of the cognitive load on human controllers. For example, studies predicting the number of instructions required for ATCO may have input characteristics similar to those in this invention (such as aircraft status), but their ultimate purpose is to assist or evaluate human controllers. This human-centered modeling paradigm is valuable within the traditional ATM framework, but it has fundamental shortcomings for future-oriented, highly automated UAM / AAM systems. Future UAM systems may involve large-scale traffic management by automated systems, with humans playing a more supervisory and anomaly handling role. In this new paradigm, simply predicting a metric that simulates human workload has limited significance for automated decision-making systems. Automated systems do not need to know whether a virtual "controller" is "overworked"; they need to know the physical boundaries and constraints of the airspace.

[0115] The method for directly assessing and quantifying "capacity" proposed in this invention aims to solve this fundamental problem. While the model input of this invention also includes traffic conditions, it focuses more on objective factors affecting the physical capacity of airspace, such as three-dimensional wind fields, communication signal quality, and navigation accuracy. Its output, "dynamic operational capacity," is an objective, physically meaningful quantitative indicator (e.g., "X flights per minute"), which can be directly used by automated decision-making systems as a hard constraint for flow control, route planning, and resource allocation. Therefore, this invention does not simply improve the algorithm for prediction accuracy, but proposes a completely new airspace assessment paradigm for the automation era, shifting from measuring "human capacity" to measuring "airspace capacity." This shift solves the fundamental problem that existing technologies cannot provide direct, quantitative, and actionable decision-making basis for automated air traffic control systems.

[0116] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, this disclosure will not describe the various possible combinations separately.

[0117] Furthermore, various different embodiments of this disclosure can be combined in any way, as long as they do not violate the spirit of this disclosure, they should also be regarded as the content disclosed in this disclosure.

Claims

1. A method for dynamic assessment of low-altitude airspace capacity based on a temporal graph neural network, characterized in that, Includes the following steps: S1: Discretize the low-altitude airspace of the target city into multiple three-dimensional voxels. Define the three-dimensional voxels, key infrastructure and aircraft in the airspace as static nodes, infrastructure nodes and dynamic nodes of the graph, respectively. Define the preset routes, spatial adjacency relationships between nodes, dynamic interaction relationships and ownership relationships between aircraft as edges of the graph. S2: Real-time acquisition and fusion of multi-source heterogeneous data, which includes at least three-dimensional meteorological data and communication, navigation and monitoring performance data, and the processed multi-source heterogeneous data is used as dynamic feature vectors of nodes and edges to construct a dynamic spatiotemporal graph that evolves over time. S3: Input the time series of the dynamic spatiotemporal graph into a preset time series graph neural network model. The time series graph neural network model predicts the spatial state within a future time window and outputs a predicted operational risk score that includes at least each of the static nodes. S4: Based on the predicted operational risk score, calculate the dynamic operational capacity of each static node within the future time window using a preset capacity assessment algorithm.

2. The method for dynamic assessment of low-altitude airspace capacity based on a time-series graph neural network according to claim 1, characterized in that, The multi-source heterogeneous data mentioned in step S2 also includes: real-time aircraft status data, as well as geographic information and airspace constraint data.

3. The method for dynamic assessment of low-altitude airspace capacity based on a time-series graph neural network according to claim 1, characterized in that, The temporal graph neural network model described in step S3 includes a feature-level attention module. The feature attention module is configured to generate a weight vector for the input multidimensional feature vector containing multiple heterogeneous feature dimensions with different physical meanings based on the global or local spatial context, and to weight the original feature vector.

4. The method for dynamic assessment of low-altitude airspace capacity based on a time-series graph neural network according to claim 1, characterized in that, Step S3 includes receiving one or more predicted operational risk scores for the low-altitude airspace volume of the target city at one or more future time points from a trained temporal graph neural network model whose input is a dynamic spatiotemporal graph.

5. The method for dynamic assessment of low-altitude airspace capacity based on a time-series graph neural network according to claim 1, characterized in that, The temporal graph neural network model described in step S3 is a gated temporal graph convolutional network.

6. The method for dynamic assessment of low-altitude airspace capacity based on a time-series graph neural network according to claim 5, characterized in that, The gated temporal graph convolutional network includes: The spatial feature aggregation module employs a graph convolutional network. The time-dynamic evolution module uses a gated loop unit to effectively capture temporal dependencies through update and reset gates; The feature-level attention module employs an additive attention mechanism. It calculates the importance scores of each feature dimension through a single hidden layer feedforward network and normalizes them into a weight vector using the Softmax function, ultimately obtaining weighted features for adaptive identification of dominant risk factors.

7. The method for dynamic assessment of low-altitude airspace capacity based on a time-series graph neural network according to claim 1, characterized in that, The capacity assessment algorithm described in step S4 includes: multiplying the baseline capacity of the static node by a dynamic attenuation coefficient determined by the predicted operational risk score, thereby calculating the dynamic operational capacity of the target city's low-altitude airspace volume at the future time point.

8. The method for dynamic evaluation of low-altitude airspace capacity based on a time-series graph neural network according to claim 7, characterized in that, The baseline capacity of the static node is a baseline capacity value of the low-altitude airspace volume of the target city.

9. A dynamic assessment system for low-altitude airspace capacity based on a time-series graph neural network, characterized in that, include: The data acquisition and fusion module is used to access multi-source heterogeneous data sources and to clean, standardize, and fuse the multi-source heterogeneous data. The dynamic spatial domain graph construction module is used to construct a dynamic spatiotemporal graph that evolves over time based on the processed multi-source heterogeneous data. The time series prediction engine module has an embedded time series graph neural network model, which is used to receive the time series of the dynamic spatiotemporal graph and output the predicted operational risk score. The capacity assessment and quantification module is used to calculate the dynamic operating capacity based on the predicted operating risk score and provide services to the upper-level management platform through the application programming interface.

Images