🤖 machine learning

Generalising Travel Time Prediction To Varying Route Choices In Urban Networks

This paper introduces the Generalised Travel Time Predictor (GenTTP), a framework that overcomes the limitations of existing graph neural network models by accurately predicting system-wide travel times and flows across varying route choices rather than being restricted to fixed demand realizations.

Original authors: Łukasz Gorczyca, Kacper Drozd, Michał Bujak, Rafał Kucharski

Published 2026-05-11

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Łukasz Gorczyca, Kacper Drozd, Michał Bujak, Rafał Kucharski

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict how long a city-wide delivery trip will take.

The Old Way: Guessing the Pattern
Most current computer programs that predict traffic act like a weather forecaster who only looks at the sky. They look at how traffic usually behaves during rush hour (the "recurring pattern") and guess, "Okay, it's 5 PM, so it will be congested here."

These programs are great at predicting what happens when everyone drives the same way they always do. But they fail miserably if something changes. If a fleet of delivery trucks decides to take a different route, or if a new app tells drivers to avoid a certain street, these old programs get confused. They can't imagine "what if" scenarios. They just keep predicting the same old traffic jam, even if the drivers have actually changed their plans.

The New Way: The "What If" Simulator (GenTTP)
The researchers at Jagiellonian University built a new tool called GenTTP (Generalised Travel Time Predictor). Think of this not as a weather forecaster, but as a traffic architect.

Instead of just watching what happens, GenTTP asks: "If I tell 1,000 drivers to take Route A instead of Route B, how will the traffic change?"

It works by looking at two things at the same time:

The Current Traffic: How crowded the roads are right now (like looking at the current water level in a river).
The Plan: The specific route map for every single vehicle (like knowing exactly where every boat in the river is heading).

How It Works (The Creative Analogy)
Imagine the city's road network is a giant, complex plumbing system with pipes (roads) and water (cars).

The Old Models just watch the water flow and guess where it will go next based on yesterday's flow. They assume the pipes are always used the same way.
GenTTP has two special eyes:
- Eye 1 (The Flow Watcher): It looks at the water pressure and current speed in the pipes using a "Graph WaveNet" (a smart camera that sees how water moves through the whole pipe system).
- Eye 2 (The Route Planner): It looks at the blueprint of where the water is supposed to go. It uses a "LSTM" (a memory unit) to remember the plan for every single vehicle.

These two eyes feed into a "brain" (a fusion layer) that combines the current reality with the future plan. It then calculates the total time it will take for all the water to get through the system.

Why This Matters
The paper highlights a specific problem: Connected Autonomous Vehicles (CAVs). Imagine a future where a company controls a fleet of self-driving taxis. They can tell all their cars to take a specific route to avoid a jam.

The old models would say, "I don't know what you're doing, I'll just predict the usual jam."
GenTTP says, "Ah, I see you are sending 500 cars down the back streets. I can calculate exactly how that will change the traffic and how much time everyone will save (or lose)."

The Results
The team tested this on a real city network (Ingolstadt, Germany) with over 1,000 simulated vehicles.

Accuracy: When they asked the model to predict the total travel time for different route plans, it was very close to the "real" answer (only about a 10% difference).
Comparison: The old models were way off, with errors ranging from 18% to a massive 82%.
Speed: GenTTP is much faster than running a full, second-by-second simulation of every car. It's like using a smart calculator instead of building a physical model of the city to test your ideas.

In a Nutshell
GenTTP is a tool that stops traffic prediction from being a passive observation of the past. It allows city planners and fleet managers to actively test different route strategies and see exactly how those choices will impact travel times, even if those choices are unusual or never happened before. It bridges the gap between "what is happening" and "what could happen if we change the plan."

Technical Summary: Generalising Travel Time Prediction To Varying Route Choices In Urban Networks

Problem Statement
Existing traffic forecasting models, predominantly based on Graph Neural Networks (GNNs), are limited to predicting future congestion based on recurring, historical demand patterns. These models treat traffic flow as an observational time-series problem, inferring future states solely from past states. Consequently, they fail to generalize when route choices vary, such as during disruptions, unexpected events, or when managing fleets of Connected Autonomous Vehicles (CAVs). In these scenarios, the same travel demand distributed across different paths can produce substantially different network-wide congestion patterns and total travel times. Current methods cannot answer "counterfactual" questions (e.g., "What would happen to travel time if these vehicles took a different route?") because they lack an explicit mechanism to process route assignment data.

Methodology: The GenTTP Framework
The authors propose the Generalised Travel Time Predictor (GenTTP), a learning-based surrogate model designed to predict system-wide travel time conditioned on explicit route assignments. Unlike traditional models that map past traffic states to future states, GenTTP maps a sequence of historical traffic loads and route assignment matrices to future traffic flows and total travel time.

The architecture consists of two primary branches fused via a Multi-Layer Perceptron (MLP):

Flow Processing Branch: This branch processes historical traffic load data ( $Q$ ) using a Graph WaveNet-based Graph Convolutional Network (GCN). It captures spatiotemporal correlations and existing congestion patterns. To address the issue of overestimating flow on unoccupied road segments (a common failure in MSE-trained models), the authors introduce a Sparsity-Aware Smoothing Mechanism (a low-occupancy gate) that explicitly handles nodes with null traffic.
Assignment Processing Branch: This branch processes the route assignment matrix ( $A$ ), which represents the planned paths of vehicles. It utilizes a sequence modeling approach (LSTM or GRU) to encode the directional traffic propagation and future network load implied by the specific routing strategy.
Fusion and Prediction: The representations from both branches are fused using either an attention mechanism or data concatenation. The fused features are passed through an MLP decoder with a sigmoid gate to predict the next-step traffic flow ( $\hat{Q}_t$ ). By iterating this process, the model generates a complete sequence of future traffic flows, which are then aggregated to estimate the system-wide total travel time ($TT$).

Key Contributions

Assignment-Aware Prediction: The paper introduces a traffic flow predictor conditioned on route choices, moving beyond the "assignment-agnostic" nature of state-of-the-art forecasting models.
Hybrid Architecture: Development of a multi-head neural network that integrates microscopic traffic propagation (via assignment inputs) with macroscopic flow directionality (via traffic state inputs).
Sparsity Handling: Introduction of a Sparsity-Aware Smoothing Mechanism that mitigates Mean Squared Error (MSE) overestimation on low-occupancy nodes, a typical issue in urban traffic networks.
Counterfactual Capability: The framework enables the evaluation of "what-if" scenarios for different routing strategies without running computationally expensive microscopic simulations for every candidate assignment.

Experimental Results
The method was validated on the Ingolstadt urban network (734 nodes, 1034 vehicles) using data generated from 1,000 microscopic simulations (SUMO) with varying route assignments.

Accuracy: GenTTP achieved a total travel time prediction error of 10.3% (using the attention-based fusion variant). In comparison, baseline models (STAEformer, ADGCRN, Graph WaveNet) exhibited errors ranging from 18% to 82.7%.
Flow Prediction: The model demonstrated an average discrepancy of 2.64 per node in flow prediction (MAE), outperforming baselines which struggled to capture dynamic changes in occupancy, particularly on sparsely populated nodes.
Efficiency: GenTTP inference is significantly faster than running microscopic simulations, making it a viable surrogate for rapid scenario analysis.

Significance and Claims
The paper claims that GenTTP addresses a critical gap in transportation modeling: the inability to predict network outcomes for varying route assignments. While previous models are observational (predicting the continuation of existing patterns), GenTTP is interventional, capable of analyzing the congestion consequences of specific routing policies.

The authors position this work as a vital tool for the coordination of Connected Autonomous Vehicle (CAV) fleets, where routing strategies can be actively controlled rather than merely observed. By providing accurate, dynamic estimates for non-recurring or atypical demand distributions, the framework supports rapid scenario evaluation and decision-making for fleet operators aiming to optimize system performance or prevent congestion emergence. The authors modestly note that current results are confined to a specific micro-simulator and demand variety, suggesting further validation is required to assess generalizability across different networks and conditions.

Technical Summary: Generalising Travel Time Prediction To Varying Route Choices In Urban Networks

More like this