A hybrid forecasting framework based on support vector regression with a modified genetic algorithm and a random forest for traffic flow prediction

Zhang, Lizong; Alharbe, Nawaf; Luo, Guangchun; Yao, Zhiyuan; Li, Ying · 2018 · OpenAlex-citations

DOI: 10.26599/tst.2018.9010045

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 addresses the challenge of accurate short-term traffic flow forecasting, a critical component of intelligent transportation systems for managing congestion and guiding travelers. The authors identify that existing methods often struggle with the complexity and variability of real-world traffic data. Specifically, Support Vector Regression (SVR) models, while effective for nonlinear problems, suffer from two main issues: the inclusion of redundant or noisy features reduces efficiency, and poor parameter selection leads to suboptimal performance. To resolve these issues, the study proposes a novel hybrid framework, termed RF-CGASVR, which integrates Random Forest (RF) for feature selection and a modified Genetic Algorithm (GA) with chaotic characteristics for parameter optimization. The proposed framework operates in three stages. First, the RF algorithm evaluates feature importance by generating multiple decision trees and assessing the impact of each feature on prediction accuracy, providing a ranked list of potential inputs. Second, an enhanced Genetic Algorithm (CGA) optimizes the SVR parameters (kernel width, insensitive loss, and regularization constant). Unlike standard GAs, which are prone to premature convergence and local optima, the CGA employs a tent map method for chaotic population initialization and a chaotic mutation strategy to balance randomness and ergodicity. The fitness of each chromosome in the GA is determined by the Root Mean Square Error (RMSE) of the SVR model using the current feature subset and parameters. Third, the framework iteratively removes low-effect features and refines parameters to identify the optimal combination of the most informative feature subset and SVR settings. The framework was evaluated using real-world traffic data collected from eight sensors near the I-605 interstate highway in California, sourced from the California Department of Transportation’s Performance Measurement System. The experiments tested the model on two road layouts: straight roads and crossroads, to assess performance in both simple and complex spatiotemporal scenarios. The RF-CGASVR model was compared against several baseline methods, including standard SVR with grid search, Back-Propagation Neural Networks (BPNN), and Autoregressive Integrated Moving Average (ARIMA). Performance was measured using RMSE and Mean Absolute Percentage Error (MAPE). The results demonstrate that the proposed RF-CGASVR model achieves superior forecasting accuracy compared to the benchmark methods. The hybrid approach effectively reduces the number of required features while improving prediction reliability, addressing the limitations of traditional SVR and other machine learning techniques. The study concludes that integrating feature selection within the learning process and using chaotic optimization for parameters significantly enhances the robustness and precision of short-term traffic flow predictions. This framework offers a more efficient and accurate tool for traffic management systems, particularly in handling the nonlinear and irregular patterns inherent in real-world traffic data.

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.

• telematics crash prediction