Cross-Domain Reformer-XL Model for Transferable Irrigation Canal Gate Control

Irrigation districts increasingly require gate-control algorithms that can operate across canals with different cross sections, actuator delays, inflow disturbances, and sensor densities. This paper proposes a cross-domain Reformer-XL model for canal gate control, in which locality-sensitive hashing attention is coupled with segment-level memory, a domain- conditioned hydraulic embedding, and a constrained control head. The model was trained with digital-twin rollouts and historical operating records, then evaluated on four simulated and field-calibrated irrigation domains containing 38 reaches, 112 gate structures, and 16,800 daily operation sequences. Compared with nonlinear model predictive control, gated recurrent networks, and a standard Transformer, the proposed model reduced downstream water-level root-mean-square error from 7.8 cm to 4.9 cm in the most shifted target domain, lowered cumulative gate travel by 18.6%, and maintained a 96.4% constraint-satisfaction rate during unmeasured turnout disturbances. Ablation tests showed that segment memory contributed 1.4 cm of tracking improvement, sparse hashing kept 288-step inference below 0.7 s, and domain adversarial alignment reduced transfer degradation by 31.2%. A sensitivity test also showed that seven days of target-domain commissioning data were sufficient to reach 5.4 cm tracking error after adaptation. The results indicate that long-context attention with explicit hydraulic-domain adaptation can provide accurate, smooth, and computationally practical gate scheduling for modern irrigation canal automation while reducing the amount of target-domain commissioning data required for deployment.