Combining kohonen maps with arima time series models to forecast traffic flow

van der Voort, Mascha C.; Dougherty, Mark; Watson, Susan · 1996 · OpenAlex-citations

DOI: 10.1016/s0968-090x(97)82903-8

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 introduces KARIMA, a hybrid method for short-term traffic flow forecasting that combines Kohonen self-organizing maps (SOMs) with ARIMA time series models. The research was motivated by the limitations of existing forecasting techniques. While layered statistical approaches like ATHENA demonstrated superior performance compared to single ARIMA models or backpropagation neural networks, ATHENA required a vast number of clusters (up to 192), making it a "brute force" approach that was difficult to retrain or transfer. The authors sought a generalized solution that maintained high accuracy while significantly reducing the number of required classes. The study utilized historical traffic flow data from detectors on French motorways, specifically focusing on the convergence point at Beaune and three upstream sites. Data were aggregated into half-hourly and hourly intervals for July and August across multiple years. The methodology involved using a Kohonen map as an unsupervised classifier to group traffic patterns, with a distinct ARIMA model tuned for each resulting cluster. Initial experiments using rectangular SOMs yielded poor results, often dominated by simple temporal variables like the day of the week. The authors subsequently adopted hexagonal SOMs, which facilitated better interpretation and clustering. Clusters were defined manually based on neuron activity levels and spatial relationships on the map, typically resulting in only two to four classes. The results demonstrated that the KARIMA method achieved forecasting performance comparable to the ATHENA model for both half-hour and one-hour horizons. Validation on a separate 1990 test dataset showed that the model generalized well, with error distributions similar to or slightly better than those obtained on training data. Notably, the hexagonal map structure proved superior to rectangular layouts, and clustering based on neuron activity levels significantly improved performance over unclustered ARIMA models. The method required far fewer classes than ATHENA, simplifying the model structure. The significance of this work lies in the balance between accuracy and practicality. By reducing the number of classes to a small, manageable set, KARIMA addresses the retraining and transferability issues associated with brute-force layered models. The authors conclude that the automatic retraining capability of the Kohonen map, combined with the simplicity of fitting few ARIMA models, makes the approach robust for tracking long-term changes in traffic flow. The study confirms that traffic data is highly non-linear and benefits from "divide-and-conquer" layered models, offering a more efficient alternative to complex statistical clustering methods.

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.

• mental model of traffic