Top

Random-Feature DeepONet–FNO Surrogate for Bus Bunching Recovery under Variable Boarding Demand

Authors

Rodríguez González
School of Computer Science, École Polytechnique, Palaiseau, 91128, France
Álvarez Torres
School of Computer Science, École Polytechnique, Palaiseau, 91128, France
Marcos Rubio
Faculty of Computer Science, Université Côte d’Azur, Nice, 06103, France

Files

Abstract

Passenger boarding can amplify irregular headways and propagate delays through a bus fleet. This study develops a boarding-conditioned random-feature surrogate and a bounded holding controller for a simplified circular corridor. A DeepONet-inspired branch–trunk bank and fixed Fourier reconstructions represent local boarding nonlinearities and interactions across the ordered fleet. Ridge fitting determines the common linear readout, and fleet centering enforces a zero-sum predicted increment. The experiment uses eight buses, five demand levels, thirty paired disturbance seeds, and forty-eight control transitions per episode. Independent training, validation, and test sets contain 1200, 200, and 400 fleet fields, respectively. The hybrid test increment root mean square error is 3.47 seconds, compared with 26.30 seconds for the branch model and 20.10 seconds for the spectral model. Under nominal demand, the hybrid controller achieves a mean episode root mean square headway coefficient of variation of 0.073. The headway-derived random-arrival waiting indicators are 120.63 seconds for hybrid control and 197.38 seconds for feedback holding. Prediction errors and closed- loop intervention outcomes are evaluated separately under a shared action bound. The results support the combined representation within the evaluated corridor, with interpretation conditioned on unequal feature counts, a capped boarding response, and exact demand observations. Capacity-matched comparisons and explicit passenger queues remain requirements for assessing broader operational transfer.

Details