Zero-shot traffic flow prediction with foundation models: a comparison with deep learning approaches

Area of Science:

• Artificial Intelligence
• Transportation Engineering
• Data Science

Background:

• Traffic flow prediction is vital for urban mobility, congestion reduction, and road safety.
• Deep learning models offer accuracy but demand large datasets and extensive training.
• Foundation models demonstrate strong performance in time series prediction tasks.

Purpose of the Study:

• To evaluate the efficacy of foundation models (Lag-Llama, Chronos) for zero-shot traffic flow prediction.
• To compare the predictive accuracy of foundation models against traditional deep learning models.
• To analyze the impact of model size, context length, and training data on foundation model performance.

Main Methods:

• Application of two foundation models, Lag-Llama and Chronos, for zero-shot traffic flow prediction.
• Comparative analysis of foundation model performance against established deep learning models.
• Investigation of factors influencing prediction accuracy, including context length and model size.

Main Results:

• Foundation models significantly outperform deep learning models in traffic flow prediction.
• Foundation models demonstrate effectiveness under both normal conditions and disruptive events.
• Larger models and longer context lengths improve accuracy but increase inference time.

Conclusions:

• Foundation models offer a practical and efficient alternative for traffic flow prediction, requiring less data and training time.
• Model selection based on comprehensive training data is crucial for optimal zero-shot performance.
• Foundation models represent a promising advancement for real-world intelligent transportation systems.