[Paper Review] Neural Networks Model for Travel Time Prediction Based on ODTravel Time Matrix

Ayobami Adewale, Amnir Hadachi|arXiv (Cornell University)|Apr 8, 2020

Human Mobility and Location-Based Analysis11 references4 citations

TL;DR

This study proposes a neural network-based travel time prediction model using an Origin-Destination (O-D) travel time matrix derived from GPS data. It compares Multi-Layer Perceptron (MLP) and Long Short-Term Memory (LSTM) networks, finding that while both models achieve accurate predictions, LSTM outperforms MLP in fitting observed travel times but becomes increasingly sensitive to noise as prediction horizon increases (e.g., two-step prediction).

ABSTRACT

Public transportation system commuters are often interested in getting accurate travel time information to plan their daily activities. However, this information is often difficult to predict accurately due to the irregularities of road traffic, caused by factors such as weather conditions, road accidents, and traffic jams. In this study, two neural network models namely multi-layer(MLP) perceptron and long short-term model(LSTM) are developed for predicting link travel time of a busy route with input generated using Origin-Destination travel time matrix derived from a historical GPS dataset. The experiment result showed that both models can make near-accurate predictions however, LSTM is more susceptible to noise as time step increases.

Motivation & Objective

• To develop a neural network model for public transport link travel time prediction using an O-D travel time matrix derived from historical GPS data.
• To compare the prediction accuracy of Multi-Layer Perceptron (MLP) and Long Short-Term Memory (LSTM) networks in travel time forecasting.
• To evaluate model performance across different travel time durations (short vs. long jumps) and time periods (peak vs. non-peak).
• To investigate the impact of increasing time steps on LSTM model robustness and noise sensitivity.
• To explore the potential of using deep learning models for improving real-time travel time estimation in intelligent transportation systems.

Proposed method

• Constructed an Origin-Destination (O-D) travel time matrix from historical GPS trajectory data of public bus routes.
• Split the dataset into 70% training, 15% validation, and 15% testing sets to ensure model generalization.
• Designed a Multi-Layer Perceptron (MLP) with 5 input neurons, two hidden layers (12 and 50 neurons), and a single output neuron.
• Designed an LSTM model with 3 input neurons, two hidden layers (80 and 100 neurons), and a single output neuron, using Adam optimization and a batch size of 3.
• Evaluated model performance using Root Mean Square Error (RMSE), defined as $ \text{RMSE} = \sqrt{\frac{1}{z} \sum (tt^{j}_{a,b} - TT^{j}_{a,b})^2} $, where $ z $ is the number of test samples.
• Trained the MLP for 1,000 epochs and the LSTM for 2,000 epochs using the adaGrad and Adam optimizers, respectively.

Experimental results

Research questions

• RQ1 How accurately can MLP and LSTM models predict link travel times when trained on an O-D travel time matrix from GPS data?
• RQ2 Does the LSTM model outperform the MLP model in capturing temporal dependencies in travel time sequences?
• RQ3 How does prediction accuracy vary between short-jump (0–35 min) and long-jump (>35 min) trips?
• RQ4 How do peak and non-peak hours affect the prediction performance of both models?
• RQ5 How does increasing the prediction time step (e.g., from one to two steps) impact the robustness and noise sensitivity of the LSTM model?

Key findings

• The MLP model achieved an RMSE of 1.17–1.20 minutes for short jumps and 2.48–2.61 minutes for long jumps, with minimal variation across peak and non-peak hours.
• The LSTM model demonstrated superior fitting to observed travel times compared to MLP, particularly in capturing temporal patterns across peak and non-peak periods.
• For two-step predictions, the LSTM model's performance degraded significantly due to error accumulation, indicating increased sensitivity to noise over longer horizons.
• Both models maintained prediction errors below 10% for trips lasting over 30 minutes, indicating strong performance on longer-duration journeys.
• The LSTM model achieved better overall prediction accuracy than MLP, but its performance was compromised when predicting beyond one time step due to noise propagation.
• The study confirms that while deep learning models like LSTM are effective for travel time prediction, their sensitivity to noise over extended prediction horizons remains a critical limitation.