A Time Series Transformer Attention Model for Enhancing Bicyclist Volume Estimation Using Data Fusion and Feature Selection Techniques

This paper evaluates transformer-based time series models' effectiveness for enhancing bicycle traffic prediction by integrating crowdsourced data sets, such as Strava, with traditional count data. By leveraging the transformer's ability to capture long-term dependencies in time series data, we aim to improve the accuracy and robustness of bicycle volume estimation across different regions. To further enhance model performance, we incorporate advanced feature selection techniques, including Correlation Analysis and Random Forest feature importance, to identify the most influential predictors. This optimal feature selection significantly improves model accuracy and robustness compared to using all available features. Our results demonstrate that feature-selected transformers outperform models utilizing all features, showing improvements across metrics such as RMSE, PRMSE, MAE, MAPE, MPE, and R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>. Using Annual Average Daily Bicycle Travel (AADBT) as the target variable, the feature-selected transformer models demonstrate enhanced generalization and transferability. This study highlights the potential of combining advanced feature selection techniques with transformer-based models to provide more accurate and reliable bicycle volume estimates, which are crucial for urban planning and safety analysis. The fusion of crowdsourced and traditional data, along with feature selection, not only improved forecast accuracy but also highlighted the complementary strengths of both data types, particularly in low-volume conditions, enhancing model performance and adaptability.