Congestion Prediction Models for Sensor-Sparse Areas

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional congestion prediction systems only provide current congestion levels, failing to predict future congestion, which limits users' ability to avoid congestion, thereby reducing satisfaction for future users.

Innovation Solution

A learning device and method that utilizes past congestion-area data and sensor data to generate models for predicting future congestion levels in areas without sensors, using a first model to predict current congestion levels and a second model to forecast future congestion based on past and current data.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If sensors are installed in all areas to be estimated, then measurement precision of congestion levels is improved, but device complexity and cost increase

Engineering Contradiction:
Improvecongestion level measurement precisionVSAvoidsensor installation complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent creates a virtual copy of the sensor network by training a deep neural network model to replicate the functionality of physical sensors. The model learns from sensor data collected in areas with sensors and generates congestion level predictions for areas without sensors, effectively copying the measurement capability without physical deployment

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent introduces timetable information as an intermediary data source that bridges the gap between areas with sensors and areas without sensors. This intermediary information helps the model infer congestion levels in unsensorized areas by considering the dynamic causal relationship between train schedules and passenger flow patterns

Inventive Principle:
Principle #24Intermediary (Mediator)

2Productivity

If only current congestion levels are predicted, then productivity of prediction system is maintained, but loss of information about future congestion occurs

Engineering Contradiction:
Improveprediction system efficiencyVSAvoidfuture congestion information
Core Design Contradiction:
ProductivityVSLoss of information

Solution Approach 1:

The patent performs preliminary prediction of future congestion levels before users make travel decisions. The system forecasts congestion at multiple future time points (e.g., 5 minutes, 10 minutes, 15 minutes ahead) using the trained deep neural network, enabling users to take preventive action before encountering congestion

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent implements dynamic prediction by continuously updating congestion forecasts as new sensor data arrives and time progresses. The system adapts to changing conditions by retraining the model with latest data and generating updated predictions, maintaining accuracy while providing forward-looking information

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP4401013B1Training device, prediction device, training prediction device, program, training method, prediction method, and training prediction method
Publication Date: 2025.10.22 MITSUBISHI ELECTRIC CORP
  • EP4401013B1 patent drawingFigure 1~2
  • EP4401013B1 patent drawingFigure 3
  • EP4401013B1 patent drawingFigure 4~6

AI summary

A congestion prediction device (120) includes a correction-model generating unit (152) that generates a correction model that is a learning model for predicting, from values detected by one or more sensors, congestion-related information of a time point at which the one or more sensors perform detection, where the first model is generated by using values indicated by past sensor data indicating the values detected in the past by the one or more sensors installed in n areas out of m areas (where m is an integer of two or more) as input data and using the congestion-related information indicated by past congestion-area data indicating the congestion-related information including a past congestion level of each of the m areas as correct data; and a model learning unit (162) that generates a prediction model that is a learning model for predicting a future congestion level from the congestion-related information, by using the congestion-related information of a first time point indicated by the past congestion-area data as input data and using the congestion level in the congestion-related information of a second time point indicated by the past congestion-area data as correct data, the second time point being a time point after the first time point.