Priority-Guided Action-Masked Proximal Policy Optimization for Dynamic Scheduling of Electric Power Material Verification Tasks Under Multi-Source Disturbances

Abstract

The metering verification center of Yunnan Power Grid processes dynamically arriving batches of single-phase and three-phase smart meters, low-voltage current transformers, and collection terminals through multi-operation test chains on six automated verification lines under batch arrivals, urgent re-verification orders, device efficiency drift, and stochastic breakdowns. This paper proposes PPO-TS, a priority-guided action-masked proximal policy optimization framework for online verification scheduling. The problem is formulated as a Markov decision process in which infeasible task–device assignments are removed by action masking, a fused static–dynamic priority score ranks feasible candidates, and only the top-ranked fraction is passed to the PPO policy for final selection. A simulation environment parameterized with reference to the center’s line structure, lot-scale characteristics, and disturbance profile was constructed using Brandimarte flexible job-shop instances. Across fifteen evaluation cells and six baselines with ten seeds per cell, PPO-TS achieved the highest weighted composite score in every cell. Compared with the strongest rule-based baseline, it reduced average completion time by 12.9%, tardiness by 8.4 percentage points, and interruption recovery cost by 33.3%, while increasing throughput by 5.8%, with a 0.024 decrease in equipment utilization. Ablation and sensitivity analyses indicate that action masking and priority filtering contribute most of the gain.