Abstract
In its simplest form, the traffic flow prediction problem is restricted to predicting a single time-step into the future. Multi-step traffic flow prediction extends this set-up to the case where predicting multiple time-steps into the future based on some finite history is of interest. This problem is significantly more difficult than its single-step variant and is known to suffer from degradation in predictions as the time step increases. In this paper, two approaches to improve multi-step traffic flow prediction performance in recursive and multi-output settings are introduced. In particular, a model that allows recursive prediction approaches to take into account the temporal context in term of time-step index when making predictions is introduced. In addition, a conditional generative adversarial network-based data augmentation method is proposed to improve prediction performance in the multi-output setting. The experiments on a real-world traffic flow dataset show that the two methods improve on multi-step traffic flow prediction in recursive and multi-output settings, respectively.
Introduction
Predicting traffic flow in the immediate future is a well-studied task, but extending these predictions to longer horizons—multi-step forecasting—presents significant challenges. Traditional strategies like recursive, direct, and multi-output modeling each have inherent drawbacks. Recursive models suffer from "drift" as errors accumulate at each step. Direct models are computationally expensive and ignore temporal dependencies. Multi-output models, while capturing dependencies, are prone to overfitting and struggle with far-future uncertainty. This paper addresses these issues by introducing two novel strategies: Conditional-Data as Demonstrator (C-DaD) and C-GAN-based data augmentation.
Methodology
Conditional-Data as Demonstrator (C-DaD)
The C-DaD approach enhances the "Data as Demonstrator" (DaD) concept, where a model is retrained on its own predicted trajectories to learn error correction. The authors identify that the nature of error correction varies by time step. To address this, they introduce a time-step index ($n$) into the input vector. By augmenting the input with this index, the model becomes sensitive to its position in the prediction trajectory, allowing for context-aware error correction.
C-GAN Data Augmentation
For multi-output models, the authors propose using a Conditional Generative Adversarial Network (C-GAN) for data augmentation. Unlike standard Gaussian noise, the C-GAN learns to synthesize realistic historical data conditioned on future labels. This provides a robust source of training data that mimics actual traffic patterns, helping multi-output models generalize better and resist overfitting.
Experimental Results
The methods were evaluated using real-world traffic data from the Caltrans Performance Measurement System (PeMS), aggregated into 15-minute intervals.
Performance Gains
The C-DaD approach demonstrated substantial improvements over standard recursive methods, achieving a 22.92% reduction in Mean Squared Error (MSE) and a 27.89% reduction in Mean Absolute Error (MAE). It also outperformed the original DaD algorithm, validating the importance of the time-step index.
For multi-output models, C-GAN augmentation resulted in an 18.47% improvement in MSE compared to baseline multi-output models. Notably, C-DaD is computationally efficient as it requires training only a single model, unlike direct or hybrid methods that scale with the prediction horizon.
Discussion and Conclusion
The core insight of this work is that providing temporal context (the "clock") allows recursive models to understand their own state of degradation. Furthermore, "data-aware" augmentation via C-GANs proves superior to simple noise injection. While GANs introduce training complexity, the performance gains in multi-step forecasting are significant. Future work may explore applying these techniques to other sequential domains, such as video prediction, where recursive drift remains a primary hurdle.
Summary updated