Comparative evaluation of a hybrid decomposition-based neural network model for multi-horizon hourly traffic forecasting

Accurate short- and long-term forecasting of hourly traffic volumes is essential for operational planning in toll station environments, where demand fluctuations affect staffing, queue formation, and resource allocation. This study proposes a hybrid decomposition–forecast–reconstruction framework based on Singular Spectrum Analysis, Autoregressive Modeling, and Artificial Neural Networks (SSA–AR–ANN) for multi-horizon traffic forecasting. The methodology first decomposes the original hourly traffic series into dominant reconstructed components that represent trend and periodic patterns while isolating noise. Future values of each reconstructed component are then predicted using component-wise autoregressive models, and a multilayer perceptron is trained to reconstruct the final traffic forecasts from the predicted components. The proposed approach is evaluated using real-world hourly vehicle arrival data from the Iasmos toll station on the Egnatia Motorway (Greece), covering the period January 2014 to October 2018. Forecasting performance is assessed across multiple horizons ranging from 6 to 336 h and compared against seasonal autoregressive integrated moving average (SARIMA) and long short-term memory (LSTM) networks. Results demonstrate that the hybrid SSA–AR–ANN model provides competitive accuracy across the examined horizons and substantially lower computational cost than the LSTM baseline, supporting its suitability for near real-time operational deployment.