Effective operational planning in public transit systems (PTS) needs estimates of future ridership as one of the inputs. The forecasting of passenger demands is challenging in PTSs due to changes travel behavior, presence of multi-seasonality data, influence of external events, etc. and often limited availability of historical data. In this study, riderhip forecasting has been performed for urban bus transport network. A set of statistical and machine learning models were selected based on literature review, data characteristics of PTS demand data, and our past studies ins same domain. A real-life large dataset on 110 routes of an urban public bus network has been used for experimental analysis. Further, several calendar-related and operations-related new variables were derived to help better model the travel patterns. These variables include the long weekend indicator, route direction indicator along with typical, day of the week, week of the year, month, weekend indicator, holiday indicator. Using selected forecasting methods, the 1-day, 7-day, and 28-day ahead forecasts were generated and performance is evaluated using the Mean Error (ME), Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Symmetric Mean Absolute Percentage Error (SMAPE) metrics. Based on empirical analysis, it has been observed that the selected machine learning models outperform statistical models. However, there are differences in performance with respect to the different prediction horizon, such as, errors increase as the horizon becomes longer. The study will help PTS authorities to improve operational planning using daily ridership forecasts generated from the proposed forecasting methods.

Keywords

Public Bus Transportation, Ridership Forecasting, Urban Transit Planning