Scholarlink

Reliability and sustainability of public transport are crucial for efficient urban mobility. This paper presents a machine learning-based approach for predicting bus delays and estimating CO \(_2\) emissions per trip. Two XGBoost regression models were trained on 124,877 real-world indian data points collected over a two-hour window, which contains GPS and bus operational data, and validated using an independent data set comprising 108,634 data points collected on a different day. The first model forecasts bus arrival times using features such as speed, previous trip delay, and temporal-spatial attributes. The second model estimates CO \(_2\) emissions per trip based on speed, trip delay, and derived variables such as fuel consumption. Both models utilize real-world GPS and operational bus data. Performance evaluation demonstrates results that are competitive with prior work on Indian city bus data, with the delay prediction model achieving a Mean Absolute Error (MAE) of 49 s and an R \(^2\) of 0.9533, while the emissions model achieves an MAE of 1.903 kg per trip and R \(^2\) of 0.99615. The results underscore the effectiveness of machine learning in improving transit operations and supporting environmental impact analysis.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Machine Learning-Based Prediction of Bus Delay and CO \(_2\) Emission Estimation Using XGBoost

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • B. V. Amogha,
  • Pranamya R. Bairy,
  • Sheba Selvam,
  • Joel D. Joy

摘要

Reliability and sustainability of public transport are crucial for efficient urban mobility. This paper presents a machine learning-based approach for predicting bus delays and estimating CO \(_2\) emissions per trip. Two XGBoost regression models were trained on 124,877 real-world indian data points collected over a two-hour window, which contains GPS and bus operational data, and validated using an independent data set comprising 108,634 data points collected on a different day. The first model forecasts bus arrival times using features such as speed, previous trip delay, and temporal-spatial attributes. The second model estimates CO \(_2\) emissions per trip based on speed, trip delay, and derived variables such as fuel consumption. Both models utilize real-world GPS and operational bus data. Performance evaluation demonstrates results that are competitive with prior work on Indian city bus data, with the delay prediction model achieving a Mean Absolute Error (MAE) of 49 s and an R \(^2\) of 0.9533, while the emissions model achieves an MAE of 1.903 kg per trip and R \(^2\) of 0.99615. The results underscore the effectiveness of machine learning in improving transit operations and supporting environmental impact analysis.