OPTIMIZATION MODEL FOR URBAN PUBLIC TRANSPORT OPERATIONS USING MACHINE LEARNING-BASED DEMAND PREDICTION

Abstract

Angkot, a route-based urban paratransit or public minibus service, remains an important component of daily mobility in Manado City. However, its predominantly supply-based operating pattern does not systematically adjust vehicle deployment to hourly passenger demand, which can produce low occupancy during off-peak periods and insufficient capacity during demand peaks. This study integrates machine-learning-based hourly passenger-demand prediction with an integer-constrained fleet-sizing model for the Paal Dua-Pasar 45 route. Direct observations were collected for 21 days, from 25 May to 14 June 2026, through an on-board survey of one angkot and a separate vehicle-headway survey; 1,538 passenger boardings were directly recorded. Random Forest and XGBoost were compared using a chronological 80:20 holdout and Leave-One-Day-Out Cross-Validation (LODO). Random Forest remained the better-performing model after tuning (LODO R² = 0.462; MAE = 0.76) compared with XGBoost (R² = 0.421; MAE = 0.83). The selected predictions were transformed from vehicle-level boardings to estimated corridor demand using observed hourly vehicle frequency and were then entered into the fleet-sizing model. The model-derived corridor demand averaged 57 passengers per hour, and the resulting fleet requirement averaged six vehicles, ranging from five to nine vehicles. Because corridor demand was inferred from one directly surveyed vehicle and predictive performance was moderate, these fleet values should be interpreted as route-specific planning estimates rather than externally validated operating prescriptions. The study demonstrates a practical framework for linking demand prediction with adaptive fleet allocation while highlighting the need for multi-vehicle validation and uncertainty-aware deployment.