Demand Prediction Model Using GNNs for Fine-Interval People Flow
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing demand prediction systems, such as those using convolutional neural networks, struggle to accurately predict transportation demands in specific areas like event venues or buildings due to difficulties in capturing the relationship between area features, time-based demands, and traveler preferences, limiting accuracy for fine time intervals.
Innovation Solution
A demand prediction device utilizing a graph neural network trained with area information, travel route data, and traveler preferences to generate highly accurate demand predictions by employing a spatial index and deep learning techniques.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a convolutional neural network is used to predict transportation demand, then the system can process OD data and generate predictions, but it cannot accurately capture relationships between area features, time-based demands, and traveler preferences, limiting prediction accuracy for fine time intervals
Solution Approach 1:
The patent segments the prediction task into multiple components: area feature extraction, temporal pattern extraction, and traveler preference modeling. Each component is handled by specialized modules (graph neural network for spatial relationships, temporal convolutional network for time patterns), allowing the system to achieve high prediction accuracy while maintaining manageable complexity through modular architecture.
Solution Approach 2:
The patent introduces multiple dimensions to the prediction model: spatial dimension (area features and relationships), temporal dimension (time-based patterns and intervals), and attribute dimension (traveler preferences and characteristics). This multi-dimensional approach enables the model to capture complex relationships that a single-dimensional CNN cannot handle, thereby improving prediction accuracy for fine time intervals.
2Measurement precision
If a graph neural network is trained with comprehensive area information and travel route data, then prediction accuracy for specific areas is improved, but the system complexity and data processing requirements increase
Solution Approach 1:
The patent performs preliminary actions by pre-processing and structuring area information, travel route data, and traveler preference attributes before feeding them into the graph neural network. The data is organized into standardized formats with pre-defined relationships, which reduces the complexity of the actual training process and enables the system to handle comprehensive data without proportionally increasing system complexity.
Solution Approach 2:
The patent introduces intermediary structures such as graph data structures and feature embedding layers that mediate between the raw comprehensive data and the final prediction output. These intermediaries simplify the relationship between input data and model parameters, allowing the graph neural network to process comprehensive area information and travel route data more efficiently without directly increasing system complexity.
3Measurement precision
If the system processes fine time interval data (15-30 minutes), then prediction detail and accuracy are improved, but the computational load and processing time increase
Solution Approach 1:
The patent employs periodic action through temporal convolutional networks that process time-series data in periodic intervals. The model learns temporal patterns at different resolutions and can efficiently aggregate information across multiple time steps, enabling accurate predictions for fine time intervals (15-30 minutes) while reducing computational load through efficient temporal processing techniques that avoid brute-force computation.
Data Source
AI summary
A demand prediction device includes an index management unit configured to apply a predetermined spatial index unit to first area information to generate a first spatial index, a training information management unit configured to analyze the first spatial index and first travel route information to generate, for a first period, first people flow information indicating a people flow in a first area, a model training unit configured to train a graph neural network using people flow information and the first area information as training information to generate a trained demand prediction model, and a prediction unit configured to process, by the trained demand prediction model, second area information characterizing a target location in a second area and second people flow information indicating a people flow in the second area to generate, for a second period, demand prediction information indicating a demand degree for each target location in the second area.