Closed-loop autonomous scheduling of multi-level irrigation canal-gate systems using an LLM-agent framework: prompting strategies and model heterogeneity

This is the authors' abstract. We don't add key points for this paper.

To address real-time control challenges caused by nonlinear hydrodynamic feedback and multi-objective scheduling coupling in complex multi-level irrigation systems, this study developed an LLM-based autonomous irrigation scheduling framework, using the Storm Water Management Model (SWMM) as a physics-driven hydraulic simulation environment for feedback generation, scheduling validation, and low-cost iterative debugging prior to field deployment. Through MCP-based workflows on the Dify platform, bidirectional interaction between the LLM decision-making module and the hydraulic simulation system was achieved. An MQTT-based communication abstraction layer was further introduced to support rapid integration with heterogeneous field devices through configurable topic mapping, enabling flexible replacement of control targets without redesigning scheduling logic. A four-hour continuous scheduling experiment was conducted on a tree-topology irrigation network consisting of one main canal, three branch canals, and 60 irrigation fields. The performances of GPT-5, DeepSeek-R1, Qwen3, and Kimi-k2 were evaluated under Direct Instruction (DI), Chain-of-Thought (CoT), and Constraint-Augmented (CA) prompting strategies. Results show that the Constraint-Augmented (CA) strategy effectively mitigates numerical insensitivity through embedded physical constraints, achieving reliable hydraulic regulation. Under CA prompting, Qwen3-32B achieved a Stability Index (SI) of 99.7%, a Goal Achievement Rate (GAR) of 95%, zero constraint violations (CVF = 0), and an irrigation uniformity coefficient of variation of 6%, while incurring a single-call API cost of 0.009 CNY. Comparative evaluation revealed a practical trade-off among regulation performance, inference efficiency, and deployment cost, identifying Qwen3-32B as the most balanced configuration for the current 63-gate irrigation scheduling task. The integration of LLM reasoning, hydraulic simulation, and MQTT-based communication further provides a pathway toward deployable cyber-physical irrigation scheduling systems.