Prediction of Traffic Volume Based on Deep Learning Model for AADT Correction

Abstract

Accurate traffic volume data are crucial for effective traffic management, infrastructure development, and demand forecasting. This study addresses the challenges associated with traffic volume data collection, including, notably, equipment malfunctions that often result in missing data and inadequate anomaly detection. We have developed a deep-learning-based model to improve the reliability of predictions for annual average daily traffic volume. Utilizing a decade of traffic survey data (2010–2020) from the Korea Institute of Civil Engineering and Building Technology, we constructed a univariate time series prediction model across three consecutive sections. This model incorporates both raw and adjusted traffic volume data from 2017 to 2019, employing long short-term memory (LSTM) techniques to manage data discontinuities. A power function was integrated to simulate various error correction scenarios, thus enhancing the model’s resilience to prediction inaccuracies. The performance of the model was evaluated using certain metrics, such as the mean absolute error, the root mean squared error, and the coefficient of determination, thus validating the effectiveness of the deep learning approach in refining traffic volume estimations.

Keywords

• annual average daily traffic volume (AADT)
• traffic volume
• long short-term memory (LSTM)
• power function
• deep learning