RESEARCH OF FAMILIES OF METHODS FOR FORECASTING TRANSPORT NETWORK CONGESTION

Abstract

The paper conducts a comprehensive investigation and systematization of method families for predicting traffic network congestion, a critical component of modern Intelligent Transportation Systems (ITS). The relevance of this topic is driven by rapid urbanization and the need for efficient traffic flow management to reduce congestion and improve safety.

The study performs a comparative analysis of three main classes of approaches: parametric statistical models, Machine Learning (ML), and Deep Learning (DL). It demonstrates that traditional methods like ARIMA and Kalman Filtering, while effective for short-term predictions on isolated segments, are limited in modeling nonlinear spatiotemporal dependencies of complex networks. ML methods, such as Support Vector Machines (SVM) and Random Forests, show greater flexibility but often require complex feature engineering.

Special attention is given to the deep learning family, particularly Recurrent (RNN, LSTM) and Convolutional Neural Networks (CNN), and their evolution into Graph Neural Networks (GNN). The paper details architectures like DCRNN (Diffusion Convolutional Recurrent Neural Network) and STGCN, which model traffic flow as a diffusion process on a directed graph, effectively capturing road network topology and spatial correlations.

The research also highlights emerging hybrid and ensemble approaches. It reviews the H-STGCN model, which integrates navigation data regarding future driver intentions via a specialized domain transformer, improving the prediction of non-recurring congestion. The effectiveness of ensemble methods (Stacking), which combine diverse base learners (e.g., MLP and SVC) to enhance robustness and accuracy, is also analyzed.

It is concluded that the most promising direction involves developing hybrid models that combine the spatial modeling capabilities of GNNs with attention mechanisms and the integration of external contextual data.