Forecasting Traffic on a Junction using Gated Recurrent Unit (GRU)

Cindy Prilia Ariestanti, Muhammad Idham Ananta Timur, S.T., M.Kom.

2025 | Skripsi | ELEKTRONIKA DAN INSTRUMENTASI

Kemacetan lalu lintas telah menjadi masalah besar di kota-kota di seluruh dunia, dan model linier tradisional seringkali tidak mampu mengelola pola lalu lintas harian yang rumit dan non-linier. Untuk mengatasi masalah ini, studi ini mengusulkan penggunaan deep learning berbasis arsitektur Gated Recurrent Unit (GRU) untuk prediksi volume lalu lintas di area tertentu yang dipantau. Data set Metro Interstate Traffic Volume, dari tahun 2012 hingga 2018 dengan gap pengumpulan data antara tahun 2014 hingga 2015, digunakan. Proses pra-pemrosesan data dan seleksi fitur berbasis korelasi dilakukan, serta pengujian konfigurasi berbeda dengan satu, tiga, dan lima layer GRU. Hasil penelitian menunjukkan bahwa konfigurasi terbaik adalah GRU tiga layer dengan fitur-fitur paling informatif, mencapai RMSE 0,0525 dan MAPE 17,03%. Penelitian ini menyimpulkan bahwa jaringan GRU mampu mendeteksi pola musiman secara efektif dan memiliki low-footprint issue (kompleksitas terhadap kecepatan) dibandingkan dengan jenis recurrent network lainnya.

Traffic congestion has become a huge problem in cities all around the world, and traditional linear models are often unable to manage the intricate, non-linear patterns of daily traffic. In an effort to overcome this issue, this study proposes the use of deep learning based on the Gated Recurrent Unit (GRU) architecture for traffic volume prediction in a specific area where monitoring is done. The Metro Interstate Traffic Volume dataset, from 2012 to 2018 with a data collection gap between 2014 to 2015, was used. Data preprocessing and correlation-based feature selection, and tested different configurations with one, three, and five layers of GRUs were done. Our findings indicate that the best-performing configuration was the three-layer GRU with the most informative features, reaching an RMSE of 0.0525 and MAPE of 17.03%. The research implies that GRU networks are capable of effectively detecting the seasonal patterns and have a low-footprint issue (complexity to speed) in comparison to other types of recurrent networks.

Kata Kunci : Traffic Volume Forecasting, Gated Recurrent Unit, Deep Learning, Time Series Analysis, Feature Selection, Intelligent Transportation Systems