A bicycle traffic prediction framework using pretrained large language models

Abstract

Bicycle traffic, as a sustainable mode of transport, helps reduce urban emissions and promote livable cities. Accurate short-term prediction of bicycle traffic flow can support real-time and near-real-time traffic management decisions in urban cycling systems. However, this task is challenging due to limited high-quality bicycle data, sparse sensor coverage, and high sensitivity to external factors. In this work, we propose a Bicycle Spatial-Temporal Large Language Model (BiSTLLM) for bicycle traffic prediction, inspired by the strong time series modeling capabilities of large language models (LLMs). In BiSTLLM, we define the historical traffic sequence at each location as tokens, allowing the model to learn complex spatial-temporal patterns. Specifically, we first introduce a weather lagging processor using causal temporal convolution to capture the delayed effects of weather on bicycle traffic. We also incorporate semantic embeddings of Point-of-Interest (POIs) and bicycle lane accessibility to represent land use and infrastructure, enhancing the model’s ability to learn both localized activity patterns and broader spatial-temporal dynamics. We then further introduce a fine-tuning strategy to better tailor the general knowledge of a pre-trained LLM to the specific characteristics of bicycle traffic. Experiments on real-world bicycle traffic flow datasets demonstrate that BiSTLLM outperforms existing state-of-the-art models. Notably, BiSTLLM exhibits strong performance even in few-shot settings, highlighting its potential for accurate bicycle traffic prediction in data-sparse environments.