Enhanced Proximal Policy Optimization for dynamic cross-regional agricultural machinery dispatching under demand uncertainty

Abstract

This paper studies dynamic cross-regional agricultural machinery dispatching in which heterogeneous machines must be reassigned after new plot demands arrive during execution. The scientific challenge is to make fast residual rescheduling decisions from the true field state, while respecting typed service compatibility, time windows, cross-cooperative coordination costs, and mandatory return-to-home operations. This paper formulates a route-based mixed-integer linear model that unifies static dispatching and event-driven residual rescheduling through machine-state interpolation and closed-loop makespan minimization. Enhanced Proximal Policy Optimization is then developed (Enhanced PPO), which aligns training with residual deployment states, mitigates delayed reward propagation in terminal-makespan objectives, and combines learned dispatch preferences with feasibility-aware online repair. On a 20-case benchmark covering four disturbance archetypes, Enhanced PPO obtains the best static and dynamic makespan in all 40 case–scale–scenario comparisons. Across competing algorithms, the median static makespan gaps range from 6.10% to 9.51%, while the median dynamic gaps range from 1.19% to 3.49%; statistical tests further show significant and more stable performance across all disturbance scenarios.