Vehicle Flow Prediction With LSTM–Transformer Road Modeling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for determining vehicle flow rates on road networks are limited by the need for expensive and static data, are computationally intensive, and struggle with spatial extrapolation, especially in urban areas with frequent intersections and bifurcations, lacking accuracy and continuity.
Innovation Solution
A method using a combination of short-term and long-term memory neural networks (LSTM) and transformer networks to process static and dynamic descriptors from fixed traffic sensors, enabling spatial and temporal prediction of vehicle flow rates on any road section without continuous measurement data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional travel modeling tools are used to determine vehicle flow rates, then comprehensive traffic information can be obtained, but the implementation becomes difficult and expensive requiring population survey data and significant calibration effort
Solution Approach 1:
The patent uses graph neural networks to create a virtual model of the road network that copies the topological structure and traffic patterns. Instead of requiring actual population surveys, the system learns from available traffic data and creates a simplified computational representation that replicates complex traffic behaviors, thereby reducing implementation complexity while maintaining accuracy
Solution Approach 2:
The patent replaces the mechanical calibration process of traditional travel modeling tools with an automated machine learning approach. The graph neural network automatically learns traffic patterns from data without requiring manual calibration efforts, substituting the labor-intensive mechanical calibration process with an automated computational system
2Reliability
If population survey data is collected every 5 to 10 years to update travel models, then the models remain relevant, but the updates are too slow to capture rapid changes in mobility
Solution Approach 1:
The patent implements a dynamic updating mechanism where the graph neural network continuously learns from new traffic data as it becomes available. Unlike static population surveys conducted every 5-10 years, the system adapts in real-time to changing mobility patterns, allowing frequent updates that capture rapid changes in traffic behavior without requiring complete model recalibration
Solution Approach 2:
The system maintains continuous learning and adaptation of traffic patterns through the neural network architecture. Instead of discrete periodic updates, the model continuously processes new data streams, ensuring that traffic flow predictions remain current and relevant without interruption, thereby maintaining reliability while enabling rapid response to changing conditions
3Speed
If floating car data is used to dynamically track vehicle journeys, then real-time mobility information is available, but the data is not available on all road sections of the urban network
Solution Approach 1:
The patent introduces graph neural networks as an intermediary that connects sparse floating car data with the complete road network structure. The GNN propagates information from measured road sections to unmeasured sections through the learned topological relationships, acting as a mediator that fills data gaps without requiring direct measurements on every road segment
Solution Approach 2:
The graph neural network model serves multiple functions simultaneously: it processes available floating car data, infers traffic patterns on unmeasured sections, and provides comprehensive network-wide predictions. This multi-functional approach allows the system to work effectively with partial data coverage while delivering complete network information
4Loss of information
If extrapolation and spatio-temporal correlation methods are used to estimate vehicle flows without complete measurement coverage, then flow estimation on unmeasured sections becomes possible, but the computation time and memory requirements increase exponentially with the road network size
Solution Approach 1:
The patent segments the road network into a graph structure where intersections are nodes and road segments are edges. The graph neural network processes this segmented structure efficiently, computing predictions for each node based on local neighborhood information rather than requiring global matrix operations. This segmentation reduces computational complexity from exponential to polynomial scaling with network size
Solution Approach 2:
The patent changes the computational parameters from traditional matrix-based spatio-temporal correlations to neural network-based feature representations. By transforming the problem into a learning-based framework with optimized parameters, the system achieves comparable estimation accuracy with significantly reduced computational power and memory requirements
5Measurement precision
If deep learning models are used to predict traffic flow with better temporal correlation, then prediction accuracy improves, but the models are limited to temporal predictions without spatial extrapolation capability
Solution Approach 1:
The patent merges temporal and spatial processing capabilities into a unified graph neural network architecture. The GNN combines the temporal sequence processing strengths of deep learning models with the spatial relationship modeling of graph structures, enabling simultaneous temporal prediction and spatial extrapolation across the road network
Solution Approach 2:
The patent creates a composite model that integrates multiple computational approaches: the temporal prediction capabilities of deep learning sequences are combined with the spatial reasoning of graph neural networks. This composite architecture leverages the strengths of both approaches to achieve accurate temporal predictions while extending spatial extrapolation to unmeasured road sections
Data Source
Figure 1A~1B
Figure 1C
Figure 2A~2B
AI summary
The present invention relates to a method for determining a vehicle flow rate on a road section of a road network. A model is constructed for determining a vehicle flow rate on a road section in the following manner: for each road section of a learning road network, an ordered sequence is determined according to a traffic direction comprising static descriptors of the learning road section and of its neighbors; a short-term and long-term memory neural network is trained on the ordered sequences and the result is concatenated with dynamic descriptors of the learning road sections; a transformer neural network is trained on these concatenations. Then the vehicle flow rate on a road section of a road network of interest is determined by applying the constructed model to static and dynamic descriptors of the road section and to static descriptors of the neighboring road sections.