Mesoscale Traffic Forecasting for Real-Time Bottleneck and Shockwave Prediction

TL;DR

The paper tackles real-time mesoscale traffic forecasting on highways by integrating spatial self-attention with LSTM (SA-LSTM) and a Laplacian Pyramid loss to emphasize fine-scale spatio-temporal details. It formulates the problem as a structured data-series task using INRIX speeds across 21 segments on I-24 Nashville and introduces an efficient n-step SA-LSTM that mitigates accumulation error for multi-minute forecasts while maintaining sub-millisecond inference. Empirical results show SA-LSTM outperforms baselines, especially under high congestion, and the n-step variant provides the best overall trade-off between short-term accuracy and long-horizon reliability. The approach has practical implications for real-time traffic management and CAV-in-the-loop control, enabling timely mitigation of bottlenecks and shockwaves in operational networks.

Abstract

Accurate real-time traffic state forecasting plays a pivotal role in traffic control research. In particular, the CIRCLES consortium project necessitates predictive techniques to mitigate the impact of data source delays. After the success of the MegaVanderTest experiment, this paper aims at overcoming the current system limitations and develop a more suited approach to improve the real-time traffic state estimation for the next iterations of the experiment. In this paper, we introduce the SA-LSTM, a deep forecasting method integrating Self-Attention (SA) on the spatial dimension with Long Short-Term Memory (LSTM) yielding state-of-the-art results in real-time mesoscale traffic forecasting. We extend this approach to multi-step forecasting with the n-step SA-LSTM, which outperforms traditional multi-step forecasting methods in the trade-off between short-term and long-term predictions, all while operating in real-time.

Figures (12)

• Figure 1: Representation of an LSTM cell. The Cell State ($C_t$, in orange) runs through the entire sequence. It stores and transmits information across time steps while selectively modifying or forgetting parts of it. The Hidden State ($h_t$, in purple) is the output of the LSTM cell at a specific time step. It carries information that is relevant to the current time step's prediction or output. It is also influenced by the cell state and the input at that time step. LSTMs employ three gate types (forget in brown, input in blue, and output in red) to regulate how information is managed within the cell state and the hidden state.
• Figure 2: The Target Road Segment of CIRCLES: I-24 Westbound in Nashville, Tennessee, seen within the highlighted region.
• Figure 3: In the red contour of the figure, one observes the chronological progression of congestion on the specified segments. A notable persistent bottleneck is evident at Exit 59. This congestion initiates at approximately 6:00 a.m., likely attributable to the augmented commuting demand upstream, and it fully resolves by around 9:00 a.m.
• Figure 4: Illustrative representation of the validation datasets. Top Row: Three representative snapshots from the Easy Validation set, showcasing common traffic patterns with periodic congestion and the prominence of temporal dependencies. Bottom Row: Three exemplar visuals from the Hard Validation set, highlighting moments of intense congestion, significant vehicle interactions, and the emphasis on spatial dependencies.
• Figure 5: A single cell from an SA-LSTM network. The SA-LSTM is an LSTM in which the output gate, in red, is augmented with self-attention.