Scholarlink

Accurate traffic speed prediction has become essential for mitigating congestion and improving mobility, thereby contributing to safety, public health, and economic benefits. This study presents a comprehensive evaluation of deep learning models for short-term traffic speed forecasting using a time-series dataset from the National Performance Management Research Data Set (NPMRDS), focused on a selected road segment or Traffic Message Channel (TMC) in Connecticut state, USA, over seven years (2017–2024). We pre-process the data to ensure consistent hourly intervals and evaluate nine deep learning architectures: Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Encoder-Decoder LSTM, attention-based sequence to sequence, transformer, Convolutional Neural Network (CNN), CNN-LSTM, and Temporal Convolutional Network (TCN) under various input feature (univariate, bivariate, and multivariate) combinations. Additionally, we implement a Seasonal AutoRegressive Integrated Moving Average with eXogenous factors (SARIMAX) model as a statistical baseline. Experimental results demonstrate that all deep learning models outperform SARIMAX, with GRU achieving the best univariate Mean Absolute Error (MAE) of 1.74 miles per hour (mph) and CNN achieving the best overall performance (MAE = 1.798 ± 0.026 mph).

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep Learning-Based Time-Series Prediction of Traffic Speed Using NPMRDS Dataset

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Saeka Rahman,
  • Md Motiur Rahman,
  • Asheka Rahman,
  • Sangho Park,
  • Shuju Wu,
  • Miad Faezipour

摘要

Accurate traffic speed prediction has become essential for mitigating congestion and improving mobility, thereby contributing to safety, public health, and economic benefits. This study presents a comprehensive evaluation of deep learning models for short-term traffic speed forecasting using a time-series dataset from the National Performance Management Research Data Set (NPMRDS), focused on a selected road segment or Traffic Message Channel (TMC) in Connecticut state, USA, over seven years (2017–2024). We pre-process the data to ensure consistent hourly intervals and evaluate nine deep learning architectures: Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Encoder-Decoder LSTM, attention-based sequence to sequence, transformer, Convolutional Neural Network (CNN), CNN-LSTM, and Temporal Convolutional Network (TCN) under various input feature (univariate, bivariate, and multivariate) combinations. Additionally, we implement a Seasonal AutoRegressive Integrated Moving Average with eXogenous factors (SARIMAX) model as a statistical baseline. Experimental results demonstrate that all deep learning models outperform SARIMAX, with GRU achieving the best univariate Mean Absolute Error (MAE) of 1.74 miles per hour (mph) and CNN achieving the best overall performance (MAE = 1.798 ± 0.026 mph).