Traffic congestion forecasting using machine learning methods

Background. This study develops a comprehensive approach to traffic congestion forecasting using synthetic data that simulates the dynamics of urban traffic. A hybrid methodology is proposed that combines time series analysis and deep learning, which is highly relevant for modeling nonlinear dependencies and patterns in traffic data.

Purpose. The purpose of this study is to develop and test a predictive model capable of accurately forecasting traffic congestion levels while accounting for seasonal and weather-related factors.

Materials and methods. To identify patterns in the data, additive time series decomposition, spectral analysis based on the fast Fourier transform, and autocorrelation analysis were applied. The predictive model was implemented using a two-stage approach: the classical ARIMA algorithm was used for baseline forecasting, while an LSTM architecture with two recurrent layers and regularization was trained on 24-hour sequences. Additionally, to compare and validate the results, the ensemble method Random Forest was used, configured with the following hyperparameters: 200 trees, maximum depth of 12, minimum samples per leaf of 2.

Results. The results demonstrate the superiority of the LSTM model over ARIMA and Random Forest in terms of predictive accuracy, as confirmed by visual comparison of forecasts with test data and by the mean squared error metric. Key factors influencing congestion were identified, including daily traffic intensity cycles, increased load during precipitation events (up to 30% during snow and 20% during rain), as well as temperature-dependent modulation of traffic flow.