Road infrastructure manager: travel time forecasting
Overview
The road operator intends to adopt a tool based on artificial intelligence algorithms and machine-learning models to assess travel times across the network in near real time.
INDUSTRIES
Mobility & TransportationBackground
The requirement was to test a solution capable of replacing the current decision-making system with a machine-learning model that combines near real-time data from fixed sensors (inductive loops and cameras), floating-car data, weather conditions, and road geometry. Leveraging traffic-flow theory, the model identifies mobility patterns and forecasts travel times and congestion levels at 5, 10, and 15 minutes. The system powers an interactive dashboard.
Challenge
Identify the most suitable machine-learning model to capture correlations and mobility patterns across the numerous heterogeneous data sources available, in order to implement an interactive dashboard that enables near real-time monitoring of the network status.
Approach
As a first step, an assessment of the available data was carried out to evaluate their quality, reliability, and gaps for integration into the travel-time prediction model. Subsequently, the data-engineering phase made it possible to identify the features and correlations that best represent mobility patterns. Finally, after selecting and testing multiple models, the most effective AI model was identified to power a dynamic and interactive dashboard for traffic forecasting over time.
Results
Results Achieved
Developed a tool based on data-driven logic and AI technologies that analyses speeds and travel times and generates 5-, 10- and 15-minute forecasts. Tested within a defined operational perimeter, the system provides a dynamic view of mobility and makes it possible to monitor and optimise the performance of the predictive model.
17.6%
MAPE - Mean Absolute Percentage Error – measures the average difference between predicted and actual values
0.45
RMSE - Root Mean Squared Error – measures the average error between predicted and observed values