[Paper Review] Traffic Volume Prediction using Memory-Based Recurrent Neural Networks: A comparative analysis of LSTM and GRU

Lokesh Chandra Das|arXiv (Cornell University)|Mar 22, 2023

Traffic Prediction and Management Techniques4 citations

TL;DR

This paper proposes and evaluates Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks for real-time traffic volume prediction using the Metro Interstate Traffic Volume dataset. Results show GRU outperforms LSTM with shorter historical sequences, while LSTM excels with longer sequences, demonstrating both models' effectiveness in complex, dynamic traffic environments.

ABSTRACT

Predicting traffic volume in real-time can improve both traffic flow and road safety. A precise traffic volume forecast helps alert drivers to the flow of traffic along their preferred routes, preventing potential deadlock situations. Existing parametric models cannot reliably forecast traffic volume in dynamic and complex traffic conditions. Therefore, in order to evaluate and forecast the traffic volume for every given time step in a real-time manner, we develop non-linear memory-based deep neural network models. Our extensive experiments run on the Metro Interstate Traffic Volume dataset demonstrate the effectiveness of the proposed models in predicting traffic volume in highly dynamic and heterogeneous traffic environments.

Motivation & Objective

• Address the challenge of accurate real-time traffic volume forecasting in dynamic, complex urban environments.
• Overcome limitations of parametric models like ARIMA, which fail under irregular or non-stationary traffic patterns.
• Evaluate the performance of non-linear memory-based deep learning models—LSTM and GRU—for short-term traffic prediction.
• Investigate the impact of varying historical sequence lengths and feature sets on prediction accuracy.
• Provide a comparative analysis of computational efficiency and convergence behavior between LSTM and GRU models.

Proposed method

• Utilize Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks as memory-based RNNs to model temporal dependencies in traffic time-series data.
• Structure the prediction task as a multi-variate time-series problem, using past observations from $t-l$ to $t$ to predict future traffic volumes at $t+1$ to $t+f$.
• Apply MinMaxScaler for feature normalization to stabilize training and improve model convergence.
• Train models using a 24-hour and 6-hour historical window to assess performance across different sequence lengths.
• Evaluate models using standard regression metrics: Mean Squared Error (MSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE).
• Implement end-to-end training and inference pipelines with hyperparameter tuning across multiple network architectures and feature subsets.

Figure 3: Statistics of the dataset

Experimental results

Research questions

• RQ1 How do LSTM and GRU models compare in predicting traffic volume across different historical sequence lengths?
• RQ2 What is the impact of reducing the number of input features on the prediction accuracy of LSTM and GRU models?
• RQ3 Which model—LSTM or GRU—achieves better performance in terms of MSE, MAE, and MAPE under varying network configurations?
• RQ4 How does the computational efficiency and convergence speed differ between LSTM and GRU during training?
• RQ5 Can GRU outperform LSTM in scenarios with limited historical data, and under what conditions does LSTM surpass GRU?

Key findings

• GRU outperforms LSTM when the historical sequence length is short (6 hours), achieving lower MSE, MAE, and MAPE across all tested network settings.
• LSTM achieves better performance than GRU when longer historical sequences (24 hours) are used, indicating superior capability in capturing long-term dependencies.
• Both LSTM and GRU produce predictions very close to actual traffic volumes, as shown in visual comparisons across multiple time points.
• GRU exhibits slower convergence than LSTM, requiring more training time per epoch despite having fewer parameters.
• The choice of feature set significantly affects performance, with reduced features still yielding strong results, though full feature sets generally improve accuracy.
• Normalization and inverse transformation of predictions using MinMaxScaler were critical but computationally intensive steps in the pipeline.

Figure 6: MSE, MAE, and MAPE errors of LSTM and GRU by varying neural network settings. The network uses 24 hours of observations to predict the next hour’s traffic volume.