Mustafa Kamal, Fandisya Rahman, Qori Afiata Fiddina, Sulthan Rafif, Riza Akhsani Setyo Prayoga
Accurate multi horizon traffic forecasting is essential for intelligent transportation systems and real time mobility management. While recent studies highlight the efficiency of lightweight linear architectures such as DLinear, their effectiveness on complex real world traffic data remains unclear. This paper presents a comprehensive comparative study between LSTM and DLinear using a 50,000 sample, 50 sensor subset of the Traffic dataset. Both models are evaluated under unified settings across four prediction horizons: 96, 192, 336, and 720 steps. Experimental results show that LSTM consistently achieves lower forecasting error, providing approximately 10-15% lower MSE and MAE compared to DLinear across all horizons. LSTM better captures nonlinear temporal dynamics and sharp congestion fluctuations, while DLinear performs competitively only in mid range horizons and degrades noticeably for short and long forecasting ranges due to its channel wise linear assumptions. Qualitative analyses, including forecast trajectories, sample wise error distributions, and sensor wise MAE heatmaps, further confirm the robustness of LSTM on heterogeneous traffic behaviors. These findings highlight the continued relevance of recurrent models and provide practical insights into selecting efficient forecasting architectures for real world traffic prediction. © 2025 IEEE.
Telkom University, Information Technology Surabaya Campus Department, Surabaya, Indonesia; Telkom University, Informatics Surabaya Campus Department, Surabaya, Indonesia; Telkom University, Information Technology (Surabaya) Dept., Surabaya, Indonesia; Telkom University, Science Data (Surabaya) Dept., Surabaya, Indonesia; Universitas Negeri Surabaya, Information Technology Education Dept., Surabaya, Indonesia
