Integrating LLM-based agents with uncertainty-aware optimization for water-energy-carbon nexus management in irrigation districts.

Efficient irrigation water use is vital for food security, economic returns and ecological protection, yet it faces multiple uncertainties within the water-energy-carbon (WEC) nexus. Conventional optimization models are often limited by complexity, lack of interpretability, and poor alignment with routine management. This study develops an LLM-agent driven intelligent optimization framework for WEC-coupled irrigation management. The framework comprises three sequential agents: a task analysis agent that converts natural language instructions into standardized model configurations; an algorithm execution agent that runs a scenario-based multi-objective fuzzy-credibility constrained programming model (integrated with NSGA-III and AHP-TOPSIS); and a result analysis agent that interprets optimization outputs, compares alternative schemes, identifies trade-offs, and generates structured decision reports. A hybrid LLM setup uses DeepSeek-V4-Flash for task parsing and Qwen3.6-Plus for decision analysis. Applied to four instruction types (standard, punctuation-free, swapped word-order, and ambiguous scenarios), the framework achieved 100% task completion without manual intervention, demonstrating efficiency in automated parsing, execution, and reporting. In the Zhaokou Yellow River Diversion Irrigation Area Phase II, the framework quantified critical management trade-offs. Under the 75% hydrological frequency, increasing the credibility level from 0.5 to 1.0 increases water shortage by 9.43×106 m3, pollutant emissions by 0.18×103 tonnes, carbon emissions by 0.05×106 tonnes, and decreases net economic benefit by 2.56×106 CNY. By improving accessibility and interpretability, this framework offers an interactive decision-support pathway for irrigation water management under hydrological and parametric uncertainties.