利用集成學習模式與不同時間特徵預測短期公車服務運量

Using Ensemble Learning Model and Time-Based Feature Extraction for Short-Term Bus Service Ridership Forecasting

Urban bus service has the flexibility of high adjustment. In the planning process of timetable or headway period, operators or public authorities can construct the forecasting model with simple historical data to the face of short-term forecasting, except that the middle or long-term forecasting needs to have enough time to collect related complex data. This study uses the ensemble learning model of machine learning to forecast short-term city bus passenger volume (BPV) demand, and it uses the volume collection of e-ticket data as input data, so as the feature selection of different temporal dimensions to enhance the forecast effect. From the forecast results, it can be found that all three temporal features contribute to the improvement of forecast performance, and only using the feature of ''BPV at the same time compared to yesterday'' leads the forecast to improve dramatically. The other case only uses ''the total passenger volume compared to yesterday''; the Final case only uses ''the passenger volume compared to an hour ago''. If all three characteristics are added to the model, MAPE will be refined from 50.247% to 8.313%. In addition, compared with other models, the Stacking model used in this study has the best prediction effect.