PATTERN RECOGNITION BASED SPEED FORECASTING METHODOLOGY FOR URBAN TRAFFIC NETWORK

Tettamanti, Tamás; Csikós, Alfréd; Kis, Krisztián Balázs; Viharos, Zsolt János; Varga, István · 2018 · Crossref

DOI: 10.3846/16484142.2017.1352027

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper proposes a comprehensive methodology for short-term traffic speed forecasting in signalized urban road networks, addressing a gap in existing research that primarily focuses on freeways or arterial roads. Motivated by the needs of Intelligent Transportation Systems (ITS) and shared mobility services, the study aims to predict average traffic speeds for specific road links 5, 15, and 30 minutes into the future. The authors identify two critical challenges in this domain: the sensitivity of data-driven models to incomplete or biased input data from sensors and Floating Car Data (FCD), and the difficulty of selecting optimal input parameters for neural network training. To address these issues, the authors developed a two-stage Artificial Neural Network (ANN) approach using Multi-Layer Perceptrons (MLP). The first stage employs a Euclidean-distance-based feature selection algorithm to rank and select the most relevant input features from a large pool of statistical parameters (including moments, tendencies, and speed measurements). This reduces input dimensionality and improves model efficiency. The second stage introduces a built-in incomplete data handling mechanism. Unlike traditional imputation methods that may distort data, this approach dynamically reconfigures the neural network for each training sample by temporarily disabling neurons corresponding to missing values, thereby ensuring robust performance even with intermittent data. The methodology was validated using a microscopic traffic simulation of the Oktogon square area in Budapest, Hungary, conducted in VISSIM. The simulation replicated real-world traffic dynamics and signal plans, generating 3,600 data records from 60 six-hour simulation runs. These records were split into training (2,500 records) and testing (1,100 records) sets. The study focused on predicting speeds for individual links, using data from topologically connected links as inputs. The authors compared their proposed ANN method against Support Vector Machines (SVM) and evaluated different feature selection strategies, including the proposed Euclidean-distance method, Maximum-Relevance Minimum-Redundancy (mRMR), and expert knowledge-based selection. The results demonstrated that the proposed ANN methodology, particularly when combined with the Euclidean-distance feature selection and built-in missing data handling, provided accurate short-term speed forecasts. The dynamic handling of incomplete data proved superior to standard imputation techniques, as it avoided introducing artificial distortions into the training dataset. The study concludes that this full methodology offers a robust solution for urban traffic prediction, capable of handling the complexities and data irregularities inherent in real-world urban networks, thus supporting more effective traffic management and ITS applications.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• anticipation