Urban Traffic Flow Prediction With Multi-Granularity Spatio-Temporal Graphs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for predicting urban regional traffic flow fail to effectively capture both spatial and temporal dependencies, particularly in complex topological structures, neglecting user travel patterns and land use characteristics, and struggle with gradient vanishing issues and non-stationary data characteristics.
Innovation Solution
A method that constructs distance and semantic graphs to capture spatial correlations and incorporates non-stationary transformers to handle temporal correlations at multiple granularities, organizing traffic flow data into recent, daily, and weekly patterns, and fuses spatio-temporal representations for improved predictions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If convolutional neural networks are used to capture spatial dependencies, then spatial feature extraction is improved, but the model cannot be applied to complex topological structures and regular grids only
Solution Approach 1:
The patent introduces graph neural networks as an intermediary between the input data and the prediction output, enabling the model to handle complex topological structures. The GNN components (spatial attention mechanism, graph convolutional layers) serve as mediators that can process irregular spatial relationships and complex networks, overcoming the limitation of CNNs which only work on regular grids.
2Measurement precision
If LSTM or GRU models are used to capture temporal dependencies, then temporal feature extraction is improved, but gradient vanishing issues occur and long-term dependencies cannot be modeled effectively
Solution Approach 1:
The patent replaces the recurrent neural network mechanism (LSTM/GRU) with a transformer-based temporal attention mechanism. This substitution eliminates the sequential processing bottleneck and gradient vanishing issues by using self-attention to directly model long-range temporal dependencies across the entire time series, enabling effective capture of long-term patterns without the mechanical constraints of recurrent architectures.
3Ease of manufacture
If predefined graph structures considering only distance relationships are used, then spatial dependency modeling is simplified, but user travel patterns and land use characteristics are neglected
Solution Approach 1:
The patent constructs a composite graph structure that integrates multiple types of spatial relationships: distance-based connectivity, land use characteristics, and POI (Points of Interest) features. This composite graph combines heterogeneous information sources to create a more comprehensive representation of urban spatial structure, enabling the model to capture both distance relationships and functional characteristics that influence travel patterns.
4Device complexity
If single-granularity temporal models are used, then model complexity is reduced, but multiple temporal patterns (recent, daily, weekly) cannot be captured simultaneously
Solution Approach 1:
The patent segments the temporal modeling task into multiple granularity levels: short-term (recent) patterns, daily periodic patterns, and weekly periodic patterns. Each temporal granularity is processed by dedicated attention mechanisms that operate at their appropriate timescales, allowing the model to capture multi-scale temporal dependencies without overwhelming complexity. The segmented approach enables parallel processing of different temporal patterns.
Data Source
AI summary
A method for predicting urban regional traffic flow considering multiple spatio-temporal granularities is provided. In this method, a traffic flow dataset, a regional dataset, a weather dataset, and a Points of Interest (POI) dataset are acquired first. Then, data in the traffic flow dataset is preprocessed, and attribute features as well as flow sub-tensors at three temporal granularities are constructed. Next, two regional association graphs are constructed for each area, including a distance graph and a semantic graph. Finally, spatio-temporal network (STN) blocks are constructed, and based on the distance graph and the semantic graph, spatio-temporal representations at each temporal granularity are obtained through the STN blocks combined with the attribute features and the flow sub-tensors at the three temporal granularities. The spatio-temporal representations of the data at each temporal granularity are fused for flow predictions and back-propagation is performed to obtain the final model.