Causal Gated-Low-Pass Graph Convolution for Traffic Flow Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural networks for short-term traffic flow prediction face challenges in accurately capturing spatial features due to the amplification of high-frequency data, leading to reduced prediction accuracy, and are computationally inefficient for real-time applications.
Innovation Solution
A causal gated-low-pass graph convolutional network is employed, utilizing a causal gated-linear unit for temporal feature extraction, a low-pass graph convolutional block to suppress high-frequency information, and a fully-connected output layer for prediction, which combines temporal and spatial features effectively.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If Chebyshev polynomials are used as convolution kernels to extract spatial features, then spatial feature extraction is performed, but high-frequency data is amplified which affects prediction accuracy
Solution Approach 1:
The patent extracts and removes the harmful high-frequency components from the spatial feature extraction process by replacing Chebyshev polynomial-based graph convolution with a low-pass filter mechanism. This extraction principle isolates the problematic high-frequency amplification and eliminates it through the low-pass filtering operation, allowing only the beneficial low-frequency spatial features to contribute to prediction accuracy.
Solution Approach 2:
The patent changes the fundamental parameter of the convolution kernel from Chebyshev polynomials (which amplify high frequencies) to a low-pass filter design (which suppresses high frequencies). This parameter change in the convolution mechanism fundamentally alters the frequency response characteristics, transforming the system from one that amplifies harmful high-frequency noise to one that naturally filters it out while preserving accurate spatial feature extraction.
2Reliability
If LSTM is used for temporal feature extraction, then temporal dependencies are captured, but the model becomes redundant and computationally time-consuming
Solution Approach 1:
The patent replaces the complex, computationally expensive LSTM architecture with a simpler, more efficient temporal feature extraction mechanism. This substitution uses a lightweight temporal convolutional approach that achieves the same temporal dependency capture without the redundant computational overhead of LSTM's gate mechanisms, effectively using a 'cheaper' computational approach that maintains reliability while improving productivity.
Solution Approach 2:
The patent substitutes the mechanical LSTM gate system (with its multiple gate operations and state transitions) with a more efficient temporal convolution mechanism. This substitution replaces the complex mechanical-like operations of LSTM (forget gate, input gate, output gate) with a streamlined temporal filtering approach that achieves equivalent temporal feature extraction with significantly reduced computational complexity and faster processing speed.
3Measurement precision
If standard graph convolution is used, then spatial features are extracted, but high-frequency information dominates which reduces prediction accuracy
Solution Approach 1:
The patent converts the harmful effect of high-frequency dominance in standard graph convolution into a benefit by deliberately designing a low-pass filter mechanism. Instead of trying to suppress high frequencies reactively, the invention proactively designs the convolution operation to inherently emphasize low-frequency components, thereby converting the information loss of high-frequency noise into a beneficial enhancement of low-frequency spatial patterns that are crucial for accurate traffic flow prediction.
Data Source
AI summary
A short-term traffic flow prediction method based on a causal gated-low-pass graph convolutional network can include constructing a causal gated-low-pass graph convolutional network, where the causal gated-low-pass graph convolutional network includes a causal gated-low-pass convolutional block. The causal gated-low-pass convolutional block is connected to a fully-connected output layer, the causal gated-low-pass convolutional block includes two causal gated linear units and a low-pass graph convolutional block, and the low-pass graph convolutional block is set between the causal gated linear units. The method can further include obtaining a traffic flow network diagram and a traffic flow value based on traffic flow data, using the traffic flow network diagram as input, and performing short-term traffic flow prediction by using the causal gated-low-pass graph convolutional network. The method can predict short-term traffic flow with high accuracy.