Truck stop point prediction method based on directed graph spatial-temporal feature fusion

Abstract

The invention belongs to the technical field of intelligent transportation, and provides a truck stop point prediction method based on directed graph spatial-temporal feature fusion, comprising the following steps: S1, preprocessing truck trajectory data through a multi-modal trajectory preprocessing subsystem; s2, constructing a spatial-temporal characteristic matrix with the stay point as a core; s3, encoding the data by using a space-time directed graph encoder, and extracting related features of a stay point; s4, constructing a graph neural network architecture to predict a track and a truck stop point; the step S1 comprises dynamic interruption detection, semantic stay identification and intention analysis, adaptive interpolation and sparsification processing, and stay point determination by using a space-time window function. According to the method, refined staying point classification prediction facing business requirements can be provided. And effective early warning of future key staying behaviors of the truck can be realized.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation technology, specifically relating to a method for predicting truck stopping points based on directed graph spatiotemporal feature fusion. Background Technology

[0002] In the field of intelligent transportation, although some progress has been made in predicting truck stopping points or overall trajectories, there are still several bottlenecks that limit their application effectiveness in complex real-world scenarios.

[0003] Traditional statistical methods typically rely on preset speed and time thresholds to determine whether a vehicle has stopped, or on historical data to predict the frequency of stops at a location. However, this approach has the following problems: a. Poor scenario adaptability: Simple thresholds are insufficient to adapt to stopping behavior under different road conditions and business scenarios. For example, waiting at a red light in the city and resting at a highway service area may have similar appearances, but their business implications are completely different.

[0004] b. Single feature dimension: It mainly relies on speed and time, ignoring multi-dimensional information such as geographical location context (POI information), historical behavior patterns, vehicle status, etc., resulting in low prediction accuracy and weak generalization ability.

[0005] c. Ignoring spatiotemporal correlation: Failed to effectively capture the spatiotemporal dependencies between stopping points, as well as the intrinsic connection between stopping behavior and road network structure.

[0006] d. Sensitive to noise and poor data processing: It is susceptible to interference from noise data such as GPS drift and signal loss, and has limited ability to process large-scale dynamic data.

[0007] Therefore, existing technologies have also developed trajectory prediction methods based on conventional machine learning or deep learning, which can also predict vehicle stopping points. Deep learning models such as Long Short-Term Memory (LSTM) networks and Gated Recurrent Units (GRUs), as well as some trajectory clustering algorithms, have been applied to vehicle trajectory prediction. However, these prediction methods still have the following problems.

[0008] a. Insufficient focus on the core event of "stopping": Most general trajectory prediction models treat all points on the trajectory equally, predicting the entire sequence. This leads to key "stopping" features (such as the start, end, and duration of a stop) being easily smoothed out or obscured by a large amount of continuous driving trajectory data, failing to fully reflect the core position of stopping points in logistics operations. This directly affects the model's sensitivity and accuracy in identifying stopping behavior.

[0009] b. Vague or simplistic definitions and classifications of stopover points: Existing research often defines stopover points too broadly or lacks refined classification standards tailored to business needs. For example, it fails to effectively distinguish between short-term traffic waits, routine loading and unloading operations, and long-term driver rest or abnormal delays. This ambiguity limits the business guidance value of the prediction results and makes it difficult to meet the decision-making needs of different scenarios.

[0010] c. Insufficient Fusion of Spatiotemporal Features: Many existing models have a rudimentary approach to fusing spatial information (such as POI distribution, road network topology, and regional functional attributes) with time-series information (such as historical trajectories and speed changes), or they simply model the two separately and then stitch them together. Most existing methods capture and model the spatial and temporal characteristics of multivariate time-series data separately, failing to fully consider the spatiotemporal unity of multivariate time series, leading to suboptimal modeling. This separate processing approach makes it difficult to capture the deep spatiotemporal interaction patterns of truck dwelling behavior in complex road network environments, thus hindering effective early warning of future dwelling behavior.

[0011] d. Insufficient model learning for critical minority samples: In truck trajectory data, stop events with clear business significance (especially long stops) are usually a minority of samples compared to continuous driving data points. Standard model training methods may lead to insufficient model learning for such critical but sparse events, resulting in weak predictive ability. Summary of the Invention

[0012] The purpose of this invention is to provide a truck stopping point prediction method based on directed graph spatiotemporal feature fusion, in order to solve the technical problems in the optimization process of monitoring vibration and collision thresholds in the prior art, which have the problems of single optimization adjustment parameters, lack of self-optimization mechanism of the model during operation, and limitations in data transmission and processing.

[0013] The truck stopping point prediction method based on directed graph spatiotemporal feature fusion includes the following steps: S1. Preprocess the truck trajectory data through the multimodal trajectory preprocessing subsystem; S2. Construct a spatiotemporal feature matrix with the dwell point as the core; S3. Encode the data using a spatiotemporal directed graph encoder and extract features related to dwell points; S4. Construct a graph neural network architecture to predict the trajectory and truck stopping point; Step S1 includes: a. Dynamic interruption detection: A composite discrimination model based on velocity change rate and signal-to-noise ratio is used to detect interruption points in the truck trajectory; b. Semantic Dwelling Recognition and Intent Analysis: Combining the POI database and historical dwelling data, perform intent analysis on the identified dwelling points; c. Adaptive interpolation and sparsification: Cubic spline interpolation or Kalman filtering is used for trajectory smoothing and densification, and the features of the stopping points are enhanced by sparsification of intermediate trajectories and accurate determination of stopping points.

[0014] Preferably, the determination of the dwell point in step S1 is performed using a spatiotemporal window function, establishing spatial range constraints, velocity range constraints, and time range constraints, and then establishing a comprehensive determination function.

[0015] Where: σ is the sigmoid function, and the calculation result is normalized to [0,1] to form the dwell probability P. stay w1, w2, and w3 are the weighting coefficients for the corresponding spatial range, velocity range, and time range, respectively; R0, V0, and T0 are the set thresholds for the radius of the spatial range, the average velocity, and the time range, respectively; and a threshold P0 is set to determine the probability of dwell. If P... stay If P ≥ P0, then the corresponding point is determined as a stop point; in the feature encoding of step S2, the stop point feature is assigned a higher weight in the feature matrix.

[0016] Preferably, in step S1, the spatial range constraint is expressed as:

[0017] Where R is the radius of the spatial range, (lat i ,lon i ),(lat j ,lon j Let ) represent the coordinates of points i and j, respectively, and Distance represent the distance between the two points. start ,t end Set the maximum threshold of R for the start and end points of the spatiotemporal window; The speed range constraint is expressed as follows:

[0018] Among them, V avg v is the average velocity, N is the number of points, and v i Let V be the instantaneous measured velocity at point i. avg The maximum threshold; The time range constraint is expressed as follows: Set the minimum threshold for T.

[0019] Preferably, semantic dwelling identification and intent analysis include: spatially matching potential dwelling areas with the POI database, and labeling dwelling semantics by combining dwelling duration and historical dwelling patterns; short-term dwellings that cannot be matched or do not match the pattern are regarded as traffic waiting; intermediate trajectory sparsification includes: sparse processing of intermediate trajectory points that are not dwelling points, retaining key trajectory points, and highlighting the importance of dwelling points.

[0020] Preferably, step S2 includes: a. Spatial feature extraction: Convert the latitude and longitude coordinates of the truck into spatial grid codes, focusing on encoding the spatial location of the stopping points; b. Time Feature Construction: Constructing time codes to capture the temporal patterns of dwell points; c. Contextual Feature Fusion: Integrate multi-source data such as POI context, road slope, weather API, and real-time traffic to construct a feature matrix centered on the stop point; among which, the stop point features include the duration of stay, the hour and weekday attributes of the start / end time of the stay, the type and ID of the POI to which it belongs, and the weather conditions during the stay.

[0021] Preferably, in step S3, the construction of the spatiotemporal directed graph includes: a. A multi-level node system, including basic nodes, semantic nodes, and dynamic virtual road segment nodes; b. Directed edge attribute design, which includes static attributes, dynamic attributes, and directional weights; the directional weights are calculated using cosine similarity to determine the directional activation factor, enhancing the model's focus on driving intention. The formula for calculating cosine similarity is as follows:

[0022] Where, θ vehicle θ represents the vehicle direction vector. edge Represents the road boundary vector. Represents the vehicle's directional velocity vector. The velocity vector represents the direction of the road boundary; the direction excitation factor reflects the suitability of a certain edge for driving relative to the possible entry direction of a truck at a certain node.

[0023] Preferably, in step S4, the direction-sensitive and edge-aware graph attention network constructed in step S3 is applied to construct a graph neural network architecture based on the spatiotemporal directed graph obtained from the data and encoding. The graph neural network architecture is a DEAGAT-TCN-based fusion model, including an input layer, a DEAGAT layer, a TCN layer, and a high-level fusion layer. The DEAGAT layer uses a multi-head graph attention network to aggregate the information of its first-order or second-order neighbor nodes and connected edges for each node, outputting spatial context features. The TCN layer captures the evolution patterns and long-range dependencies of these spatial context features over time through multi-layer causal convolution, outputting temporal aggregation features. The high-level fusion layer uses a collaborative attention module to deeply fuse the spatial context features output by the DEAGAT layer with the temporal aggregation features output by the TCN layer.

[0024] Preferably, the DEAGAT layer includes: 1) The historical trajectory point sequence of trucks is first mapped onto graph nodes, with the graph node v where the truck is located at each time step in the sequence as the reference. k (k=1... T hist Centered on ), run k layers of DEAGAT; 2) Apply the attention mechanism of the DEAGAT layer, as shown in the following expression:

[0025]

[0026]

[0027] in, and For node feature vectors, For the edge e ij The edge feature vectors, For directional alignment, the coefficient is... For the learnable weight matrix, the subscripts are... For node, , or , For attention parameter vectors, For nodes The neighborhood group, Indicates a vector concatenation operation; Representing edge e ij Alignment between the corresponding driving direction and the side direction. For edge e ij Attention weights; 3) The formula for calculating the weight of direction-sensitive attention is as follows: , Here, concat represents the concatenation operation, and Attention represents the attention mechanism. Let be the directional weight of the edge connecting point i and point j.

[0028] Preferably, the TCN layer includes: 1) The T output from the DEAGAT layer hist A spatial context representation sequence As input to TCN, Corresponding to T hist Contextual representation; 2) The TCN layer captures the evolution patterns and long-range dependencies of these spatial representations over time through multiple layers of causal convolution; 3) The output of the TCN layer is a temporal aggregation feature z of the entire historical trajectory. tcn .

[0029] Preferably, the collaborative attention module computes spatial context features. and time aggregation features The affinity matrix between them is then used to generate time-aware weighted and space-aware weighted matrices, respectively, as shown in the following formulas:

[0030]

[0031]

[0032]

[0033]

[0034]

[0035] in, For spatial context features, This is a time-aggregated feature; The learnable weight matrix is ​​the affinity matrix. Through and The affinity matrix obtained after the nonlinear activation following the linear transformation represents the similarity between spatial and temporal features. For the corresponding activation function, and These are the weighted vectors for spatial attention and temporal attention, respectively. and These are the features after spatial perception and time perception weighting, respectively; the final fused feature calculation formula is as follows:

[0036]

[0037] Here, concat represents the concatenation operation. This represents the fusion features enhanced by the collaborative attention mechanism; MLP (Multilayer Perceptron). This represents the final output feature of the advanced fusion layer.

[0038] The technical advantages of this invention are as follows: 1) This invention significantly improves the accuracy and robustness of predicting the core logistics event of truck "stop points" by focusing on dwell times, sparse trajectories, fine-grained classification, and reinforcement learning; and provides refined dwell point classification and prediction based on business needs. It can also achieve effective early warning of future key dwell behaviors of trucks, providing strong decision support for logistics scheduling, resource planning, and risk management.

[0039] 2) Deep understanding of "directionism" and complex topology: Compared with traditional methods and simple undirected graph applications, this invention, through spatiotemporal directed graph modeling and direction-sensitive graph attention mechanism, can more deeply understand the key constraints of road networks such as one-way traffic and turning restrictions, as well as the impact of complex topological structures such as intersections and ramps on truck route selection.

[0040] 3) Achieving true "deep spatiotemporal fusion": Breaking through the limitations of simple feature splicing or shallow fusion, it achieves deep, non-linear interactive learning between spatial and temporal features through the collaborative work of multi-layer graph neural networks (such as DEAGAT) and temporal networks (such as TCN), as well as specially designed advanced fusion layers (such as collaborative attention). The exploration of spatiotemporal integrated graph neurons aims to fundamentally unify the processing of spatiotemporal information.

[0041] 4) Adaptive integration of heterogeneous information from multiple sources: Through dynamic embedding and gating / attention fusion mechanisms, it can more intelligently integrate heterogeneous information from different sources (road network, trajectory, weather, traffic events, vehicle status) and dynamically adjust its impact according to its relative importance to the current prediction task.

[0042] 5) Generate richer semantic representations: The output high-dimensional spatiotemporal semantic embedding vectors provide downstream prediction tasks with inputs that are far more powerful and discriminative than the original features or shallow features. Attached Figure Description

[0043] Figure 1 This is a basic flowchart of a truck stopping point prediction method based on directed graph spatiotemporal feature fusion according to the present invention. Detailed Implementation

[0044] The following detailed description of the embodiments, with reference to the accompanying drawings, will further illustrate the specific implementation of the present invention, in order to help those skilled in the art to have a more complete, accurate, and in-depth understanding of the inventive concept and technical solution of the present invention.

[0045] like Figure 1 As shown, the present invention provides a truck stopping point prediction method based on directed graph spatiotemporal feature fusion, which includes the following steps.

[0046] S1. The truck trajectory data is preprocessed through the multimodal trajectory preprocessing subsystem.

[0047] This step includes the following: a. Dynamic interruption detection: A composite discrimination model based on velocity mutation rate (Δv / Δt) and signal-to-noise ratio (SNR) is used to accurately detect interruption points in the truck trajectory and improve the accuracy of trajectory segmentation.

[0048] b. Semantic Dwelling Identification and Intent Analysis: Combining POI (Point of Interest) databases (such as logistics parks, warehouses, gas stations, highway service areas, and large supermarket delivery points) and historical dwelling data, intent analysis is performed on the identified dwelling points.

[0049] Potential stopping areas are spatially matched with the POI database (e.g., the centroid of the stopping area is within the POI polygon or the distance is less than 100 meters). The semantics of stopping are then labeled by combining the duration of the stop (e.g., a stop in a logistics park for more than 30 minutes but less than 8 hours may be for loading / unloading; a stop in a service area for 15-60 minutes may be for rest / refueling) and historical stopping patterns (common stopping behaviors of this vehicle / similar vehicles in this POI). Short stops that cannot be matched or do not match the pattern are considered traffic waiting. This helps to more accurately understand truck behavior.

[0050] For example, if the centroid of a potential stopping area is within the POI polygon or less than 100 meters away, a stop in a logistics park for more than 30 minutes but less than 8 hours may be for loading and unloading; a stop in a service area for 15-60 minutes may be for rest / refueling. If the vehicle / similar vehicle's common stopping behavior in this POI is found in the historical stopping patterns, it is judged as a regular stop.

[0051] c. Adaptive interpolation and sparsification: To address the issues of uneven or missing GPS data sampling rates, cubic spline interpolation or Kalman filtering is used for trajectory smoothing and densification. Furthermore, feature enhancement of stop points is achieved through intermediate trajectory sparsification and stop point weight enhancement.

[0052] Intermediate trajectory sparsification: The intermediate trajectory points that are not stopping points are sparsified, while key trajectory points (such as turning points and points with significant speed changes) are retained to reduce the amount of computation, while highlighting the importance of stopping points.

[0053] Accurate determination of dwell points: Dwell points are determined by using a spatiotemporal window function, which makes it easier to assign higher feature weights to dwell points later and ensures that the model focuses on the spatiotemporal characteristics of dwell points.

[0054] The determination of the dwell point is performed using a spatiotemporal window function, as detailed below: a. Establish spatial constraints, expressed as:

[0055] Where R is the radius of the spatial range, (lat i ,lon i ),(lat j ,lon j Let ) represent the coordinates of points i and j, respectively, and Distance represent the distance between the two points. start ,t end The start and end points of the spatiotemporal window are defined; the maximum threshold of R is set as needed, such as requiring R ≤ 2km.

[0056] b. Establish a speed range constraint, expressed as:

[0057] Among them, V avg v is the average velocity, N is the number of points, and v i The instantaneous velocity at point i; V is set as needed. avg The maximum threshold, such as requiring V avg ≤70km / h.

[0058] c. Establish time range constraints, the expression is: The minimum threshold for T can be set as needed, such as requiring T≥2h. It can also be classified as needed, such as judging T≥2h as a long-term dwell point and T≥4h as an ultra-long-term dwell point.

[0059] d. Establish a comprehensive decision function, with the following expression:

[0060] Where: σ is the sigmoid function, and the calculation result is normalized to [0,1] to form the dwell probability P. stayw1, w2, and w3 are the weighting coefficients for the corresponding spatial range, velocity range, and time range, respectively (w1=0.4, w2=0.3, w3=0.3); R0, V0, and T0 are the thresholds for the radius of the spatial range, the average velocity, and the time range, respectively, with the first two being upper limit thresholds and the time range threshold being the lower limit threshold. A threshold P0 is set for determining the probability of dwell. If P... stay If the value is greater than or equal to P0, then the corresponding point is determined as the dwell point. In this embodiment, P0 can be set to 0.8.

[0061] S2. Construct a spatiotemporal feature matrix with the dwell point as the core.

[0062] a. Spatial feature extraction: The latitude and longitude coordinates of the truck are converted into spatial grid codes using the Geohash algorithm, with a focus on encoding the spatial location of the stopping points.

[0063] b. Time Feature Construction: Construct time codes, including timestamps, days of the week, seasons, etc., to capture the time patterns of stop points.

[0064] c. Contextual Feature Fusion: Integrate POI (Point of Interest) context, road slope, weather API, real-time traffic and other multi-source data to construct a feature matrix with the stop point as the core.

[0065] For each travel segment or semantic dwell segment, the extracted features include the following features.

[0066] a. Characteristics of the travel segment: travel time, distance, average / maximum / minimum speed, mean and variance of speed / acceleration / direction angle change rate, and statistics of trajectory point curvature.

[0067] b. Stop characteristics: duration of stay, hour and weekday attributes of start / end time of stay, type and ID of POI, and weather conditions during the stay.

[0068] c. General characteristics: Geohash encoded sequence (for driving segments) or center point Geohash encoding (for stopping segments), associated start / end segment IDs (obtained through map matching).

[0069] The techniques used in feature encoding also include prioritizing stop point features and sparsifying intermediate trajectories.

[0070] Point-of-stay features are prioritized: Point-of-stay features (such as stay duration and POI type) are assigned higher weights in the feature matrix (e.g., the weighting factor is twice that of non-point-of-stay features).

[0071] Intermediate trajectory sparsification: Downsampling of trajectory points in the driving segment, retaining only key points (such as turning points and speed change points), reducing the dimension of the feature matrix and improving computational efficiency.

[0072] S3. Encode the data using a spatiotemporal directed graph encoder and extract features related to dwell points.

[0073] The construction of a spacetime directed graph includes the following.

[0074] a. Multi-level node system.

[0075] Basic nodes: Key intersections in the road network are designated as basic nodes.

[0076] Semantic nodes: Important stop points identified by the preprocessing module are modeled as semantic nodes in the graph to explicitly carry key activity information of the trucks. Important stop points include target delivery points, frequently visited gas stations, and large transit hubs. Important semantic stop points are added to the graph as special nodes.

[0077] Dynamic Virtual Road Segment Nodes: For critical road segments with drastic traffic flow changes, dynamic virtual road segment nodes are created within specific time slots. Their attributes can be dynamically updated to represent time-varying characteristics. The specific time slots include the morning and evening rush hours on weekdays.

[0078] b. Design of directed edge properties.

[0079] Static attributes include: road grade, legal speed limit, number of lanes, road segment length, gradient information, and whether specific types of trucks are allowed to pass.

[0080] Dynamic attributes: historical traffic characteristics, real-time / predictive characteristics, congestion index / level, etc.

[0081] Some dynamic edge properties are as follows: Historical traffic characteristics: For this road section, the average travel time and speed distribution are calculated based on historical data from the same period last year (e.g., Tuesday morning between 10-11 am) and similar weather conditions.

[0082] Real-time / predictive passage characteristics: If an API is available, obtain the predicted passage time for the current or next 15 minutes; otherwise, it can be predicted using a simple model (such as time series ARIMA) based on recent (e.g., the past 1 hour) observation data.

[0083] Congestion Index / Level: Calculated from traffic information APIs or based on the ratio of real-time speed to free-flow speed.

[0084] Directional weights: Directional excitation factors are calculated using cosine similarity to enhance the model's focus on driving intentions. The formula for calculating cosine similarity is as follows:

[0085] Where, θ vehicle θ represents the vehicle direction vector. edgeRepresents the road boundary vector. Represents the vehicle's directional velocity vector. This represents the velocity vector in the direction of the road boundary.

[0086] The directional excitation factor reflects the direction of a certain edge e ij Relative to the possible entry direction dir of the truck at node i in The suitability of driving. For example, if a truck has just entered an intersection from a certain direction, then going straight or making a small right turn at a small angle usually has a higher prior probability of exiting the intersection than making a sharp turn or a U-turn. The directional excitation factor is calculated based on historical steering frequency statistics or geometric angles.

[0087] c. Graph construction scope and granularity control.

[0088] This step involves constructing a local dynamic spatiotemporal directed graph for each truck trajectory to be predicted. The spatial range is dynamically determined as the region that may be reached within the prediction time window. The time window covers historical trajectories and the current moment, and the dynamic attributes of the graph are updated periodically.

[0089] The node feature vectors and edge feature vectors in the constructed spatiotemporal directed graph are expressed as follows: Node feature vector: [Node type encoding (intersection / semantic point), POI category encoding (if semantic point), Geohash encoding, historical traffic flow / stop frequency of the node, node centrality index (such as betweenness centrality)].

[0090] Edge feature vector: [segment length, speed limit, congestion index, normalized historical travel time, normalized real-time travel time, directional excitation factor value].

[0091] S4. Construct a graph neural network architecture to predict the trajectory and truck stopping point.

[0092] The orientation-sensitive and edge-aware graph attention network constructed in step S3 is used to build a graph neural network architecture based on the spatiotemporal directed graph obtained from the data and encoding. The resulting graph neural network architecture includes a fusion model based on DEAGAT-TCN, specifically comprising the following components.

[0093] a. Input layer: The data involved specifically includes the following.

[0094] 1) The target truck passed T hist The sequence of trajectory points at each time step has been preprocessed and matched with the map, and associated with graph nodes or edges.

[0095] 2) The local spatiotemporal directed graph G at the current moment tIt includes node features and edge features. The node feature vector includes node type (intersection / semantic stop point), POI information, historical traffic flow clustering degree, Geohash encoding, etc.

[0096] 3) Other auxiliary information: such as weather forecasts and truck destination information.

[0097] b. Local Spatial Context Encoding (DEAGAT Layer): For each node (or the node associated with the current location of the truck), a multi-head GAT (Graph Attention Network) is first used to aggregate information about its first- or second-order neighbor nodes and connected edges, outputting spatial context features. Specifically, this includes the following:

[0098] 1) The historical trajectory point sequence of trucks is first mapped onto graph nodes. The graph node v where the truck is located at each time step in the sequence is used as the reference. k (k=1... T hist Centered on ), the k-layer DEAGAT layer runs.

[0099] 2) Apply the attention mechanism of the DEAGAT layer.

[0100]

[0101]

[0102]

[0103] in, and For node feature vectors, For the edge e ij The edge feature vectors, For directional alignment, the coefficient is... For the learnable weight matrix, the subscripts are... For node, , or , For attention parameter vectors, For nodes The neighborhood group, Indicates a vector concatenation operation; Representing edge e ij Alignment between the corresponding driving direction and the side direction. For edge e ij Attention weights.

[0104] 3) The formula for calculating the weight of direction-sensitive attention is as follows: , Here, concat represents the concatenation operation, and Attention represents the attention mechanism. The directional weights of the edge connecting points i and j are calculated using cosine similarity:

[0105] 4) Message passing function:

[0106] Here, MsgFunc is the message passing function, and the coefficients are... Let be the learnable weight matrix of the message passing function, which makes the message depend not only on the node state, but also on the state and direction matching degree of the edges.

[0107] Example parameters: DEAGAT layer number K=2 or 3; number of attention heads=4; node / edge feature dimension=64.

[0108] Improvements in local spatial context encoding enable messages to depend not only on node states but also on edge states and direction matching. Therefore, this method possesses both direction sensitivity and edge awareness, addressing the problem of insufficient utilization of directed edge information in traditional GAT.

[0109] c. Temporal Dynamic Evolutionary Coding (TCN Layer): The TCN layer captures the evolution patterns and long-range dependencies of these spatial context features over time through multiple layers of causal convolution, outputting temporally aggregated features. Specifically, it includes the following:

[0110] 1) The T output from the DEAGAT layer hist A spatial context representation sequence As input to TCN, Corresponding to T hist Contextual representation.

[0111] 2) The TCN layer captures the evolution patterns and long-range dependencies of these spatial representations over time through multi-layer causal convolution (kernel_size=3, dilations=[1,2,4,8]).

[0112] 3) The output of the TCN layer is a temporal aggregation feature z of the entire historical trajectory. tcn .

[0113] Example parameters: Number of TCN residual blocks = 4; kernel size = 3; number of channels = 128.

[0114] d. Advanced fusion layer: Deeply fuses the spatial context features output by the DEAGAT layer with the temporal aggregation features output by the TCN layer.

[0115] This step employs a co-attention module, allowing spatial and temporal features to mutually "attention" to each other's key components, generating a joint, more information-rich spatiotemporal fusion vector z. st .

[0116] 1) Input: The most relevant spatial context features output by the DEAGAT layer at the last time step. (or all) Weighted pooling) and temporal aggregation features of TCN layer output .

[0117] 2) Calculate spatial context features and time aggregation features The affinity matrix between them is then used to generate time-aware weighted and space-aware weighted matrices, respectively, as shown in the following formulas:

[0118]

[0119]

[0120]

[0121]

[0122]

[0123] in, For spatial context features, This is a time-aggregated feature; The learnable weight matrix is ​​the affinity matrix. Through and The affinity matrix obtained after the linear transformation and nonlinear activation (tanh) represents the similarity between spatial and temporal features. For the corresponding activation function, and These are the weighted vectors for spatial attention and temporal attention, respectively. and These are the features after spatial perception and time perception weighting, respectively. The final fused feature calculation formula is as follows:

[0124]

[0125] Here, concat represents the concatenation operation. This represents the fusion features enhanced by the collaborative attention mechanism; the MLP (Multilayer Perceptron) is essentially a dimensionality reduction mechanism achieved through an MLP. This represents the final output feature of the advanced fusion layer.

[0126] Example parameters: Feature dimension after fusion = 256. e. Dynamic embedding and fusion of multi-source heterogeneous information: In addition to the road network and trajectory itself, external features are also crucial for prediction. External information includes the output of real-time weather forecast APIs, traffic event broadcasts, and sensor data from the truck itself, such as fuel consumption and tire pressure, which are all heterogeneous information.

[0127] One approach is to encode these heterogeneous information into vectors and integrate them at the node and edge feature level of the graph; another approach is to add a fusion module based on a gated recurrent unit or attention mechanism between the output features of local spatial context encoding and temporal dynamic evolution encoding and these heterogeneous information, and dynamically adjust the influence weight of external features on the final prediction through the fusion module.

[0128] The present invention has been described above by way of example with reference to the accompanying drawings. Obviously, the specific implementation of the present invention is not limited to the above-described manner. Any non-substantial improvements made using the inventive concept and technical solution of the present invention, or the direct application of the inventive concept and technical solution of the present invention to other occasions without modification, are all within the protection scope of the present invention.

Claims

1. A method for predicting truck stopping points based on directed graph spatiotemporal feature fusion, characterized in that: Includes the following steps: S1. Preprocess the truck trajectory data through the multimodal trajectory preprocessing subsystem; S2. Construct a spatiotemporal feature matrix with the dwell point as the core; S3. Encode the data using a spatiotemporal directed graph encoder and extract features related to dwell points; S4. Construct a graph neural network architecture to predict the trajectory and truck stopping point; Step S1 includes: a. Dynamic interruption detection: A composite discrimination model based on velocity change rate and signal-to-noise ratio is used to detect interruption points in the truck trajectory; b. Semantic Dwelling Recognition and Intent Analysis: Combining the POI database and historical dwelling data, perform intent analysis on the identified dwelling points; c. Adaptive interpolation and sparsification: Cubic spline interpolation or Kalman filtering is used for trajectory smoothing and densification, and the features of the stopping points are enhanced by sparsification of intermediate trajectories and accurate determination of stopping points.

2. The truck stopping point prediction method based on directed graph spatiotemporal feature fusion according to claim 1, characterized in that: In step S1, the determination of the dwell point is performed using a spatiotemporal window function, establishing spatial range constraints, velocity range constraints, and time range constraints, and then establishing a comprehensive determination function. Where: σ is the sigmoid function, and the calculation result is normalized to [0,1] to form the dwell probability P. stay w1, w2, and w3 are the weighting coefficients for the corresponding spatial range, velocity range, and time range, respectively; R0, V0, and T0 are the set thresholds for the radius of the spatial range, the average velocity, and the time range, respectively; and a threshold P0 is set to determine the probability of dwell. If P... stay If P ≥ P0, then the corresponding point is determined as a stop point; in the feature encoding of step S2, the stop point feature is assigned a higher weight in the feature matrix.

3. The truck stopping point prediction method based on directed graph spatiotemporal feature fusion according to claim 2, characterized in that: In step S1, the spatial extent constraint is expressed as follows: Where R is the radius of the spatial range, (lat i ,lon i ),(lat j ,lon j Let ) represent the coordinates of points i and j, respectively, and Distance represent the distance between the two points. start ,t end Set the maximum threshold of R for the start and end points of the spatiotemporal window; The speed range constraint is expressed as follows: Among them, V avg v is the average velocity, N is the number of points, and v i Let V be the instantaneous velocity measured at point i. avg The maximum threshold; The time range constraint is expressed as follows: Set the minimum threshold for T.

4. The truck stopping point prediction method based on directed graph spatiotemporal feature fusion according to claim 1, characterized in that: Semantic dwell recognition and intent analysis include: spatially matching potential dwell areas with the POI database, and labeling dwell semantics by combining dwell duration and historical dwell patterns; short-term dwells that cannot be matched or do not match the pattern are regarded as traffic waiting; intermediate trajectory sparsification includes: sparse processing of intermediate trajectory points that are not dwell points, retaining key trajectory points, and highlighting the importance of dwell points.

5. The truck stopping point prediction method based on directed graph spatiotemporal feature fusion according to claim 1, characterized in that: Step S2 includes: a. Spatial feature extraction: Convert the latitude and longitude coordinates of the truck into spatial grid codes, focusing on encoding the spatial location of the stopping points; b. Time Feature Construction: Constructing time codes to capture the temporal patterns of dwell points; c. Contextual Feature Fusion: Integrate multi-source data such as POI context, road slope, weather API, and real-time traffic to construct a feature matrix centered on the stop point; among which, the stop point features include the duration of stay, the hour and weekday attributes of the start / end time of the stay, the type and ID of the POI to which it belongs, and the weather conditions during the stay.

6. The truck stopping point prediction method based on directed graph spatiotemporal feature fusion according to claim 1, characterized in that: In step S3, the construction of the spacetime directed graph includes: a. A multi-level node system, including basic nodes, semantic nodes, and dynamic virtual road segment nodes; b. Directed edge attribute design, which includes static attributes, dynamic attributes, and directional weights; the directional weights are calculated using cosine similarity to determine the directional activation factor, enhancing the model's focus on driving intention. The formula for calculating cosine similarity is as follows: Where, θ vehicle θ represents the vehicle direction vector. edge Represents the road boundary vector. Represents the vehicle's directional velocity vector. The velocity vector represents the direction of the road boundary; the direction excitation factor reflects the suitability of a certain edge for driving relative to the possible entry direction of a truck at a certain node.

7. The truck stopping point prediction method based on directed graph spatiotemporal feature fusion according to claim 1, characterized in that: In step S4, the orientation-sensitive and edge-aware graph attention network constructed in step S3 is applied to construct a graph neural network architecture based on the spatiotemporal directed graph obtained from the data and encoding. The graph neural network architecture is a fusion model based on DEAGAT-TCN, including an input layer, a DEAGAT layer, a TCN layer and a high-level fusion layer. For each node, the DEAGAT layer uses a multi-head graph attention network to aggregate the information of its first-order or second-order neighbor nodes and connected edges, and outputs spatial context features. The TCN layer captures the evolution patterns and long-range dependencies of these spatial context features over time through multiple layers of causal convolution, and outputs temporal aggregate features. The advanced fusion layer uses a collaborative attention module to deeply fuse the spatial context features output by the DEAGAT layer with the temporal aggregation features output by the TCN layer.

8. The truck stopping point prediction method based on directed graph spatiotemporal feature fusion according to claim 7, characterized in that: The DEAGAT layer includes: 1) The historical trajectory point sequence of trucks is first mapped onto graph nodes, with the graph node v where the truck is located at each time step in the sequence as the reference. k (k=1... T hist Centered on ), run k layers of DEAGAT; 2) Apply the attention mechanism of the DEAGAT layer, as shown in the following expression: in, and For node feature vectors, For the edge e ij The edge feature vectors, For directional alignment, the coefficient is... For the learnable weight matrix, the subscripts are... For node, , or , For attention parameter vectors, For nodes The neighborhood group, Indicates a vector concatenation operation; Representing edge e ij Alignment between the corresponding driving direction and the side direction. For edge e ij Attention weights; 3) The formula for calculating the weight of direction-sensitive attention is as follows: , Here, concat represents the concatenation operation, and Attention represents the attention mechanism. Let be the directional weight of the edge connecting point i and point j.

9. The truck stopping point prediction method based on directed graph spatiotemporal feature fusion according to claim 7, characterized in that: The TCN layer includes: 1) The T output from the DEAGAT layer hist A spatial context representation sequence As input to TCN, Corresponding to T hist Contextual representation; 2) The TCN layer captures the evolution patterns and long-range dependencies of these spatial representations over time through multiple layers of causal convolution; 3) The output of the TCN layer is a temporal aggregation feature z of the entire historical trajectory. tcn .

10. The truck stopping point prediction method based on directed graph spatiotemporal feature fusion according to claim 7, characterized in that: The collaborative attention module computes spatial context features. and time aggregation features The affinity matrix between them is then used to generate time-aware weighted and space-aware weighted matrices, respectively, as shown in the following formulas: in, For spatial context features, This is a time-aggregated feature; The learnable weight matrix is ​​the affinity matrix. Through and The affinity matrix obtained after the nonlinear activation following the linear transformation represents the similarity between spatial and temporal features. For the corresponding activation function, and These are the weighted vectors for spatial attention and temporal attention, respectively. and These are the features after spatial perception and time perception weighting, respectively; the final fused feature calculation formula is as follows: Here, concat represents the concatenation operation. This represents the fusion features enhanced by the collaborative attention mechanism; MLP (Multilayer Perceptron). This represents the final output feature of the advanced fusion layer.

Images