The short-term prediction of daily traffic volume for rural roads using shallow and deep learning networks: ANN and LSTM

Predicting daily traffic volume in the short term is of great importance for rural roads since it assists in relieving congestion, trip planning, and improving the level of service (LOS). Benchmark parametric methods like seasonal autoregressive integrated moving average (SARIMA) show less accurate predictions when traffic flow sustains irregularities. In addition, the SARIMA is not sophisticated enough to properly employ big data. Shallow learning techniques like the artificial neural network (ANN) cannot capture short-term and long-term time dependencies of daily traffic volume. Therefore, long short-term memory (LSTM) has been suggested to estimate the daily traffic volume of rural roads. LSTM offers a proper estimate of the daily traffic volume using a unit structure, with no need to create a separate model for each piece of a road. The daily traffic volume for three types of roads, i.e., high-volume roads, international roads for transit of goods, and recreational roads leading to the city of Mashhad, Iran, was estimated using LSTM. The research results demonstrated that SARIMA displayed constant sinusoidal variations around the annual average of daily traffic (AADT) and was very weak in capturing the irregularities of daily volume. Unlike SARIMA, ANN was generally better at following the volume trends. However, unrealistic spikes and drops in the forecasts of ANN occurred on days close to national holidays. LSTM resulted in the highest percentage of estimation accuracy (95%) while having no overfitting issue. Better estimates were obtained for the international freight transit roads where the volume variations were lower.