Large language model assisted hyper-heuristic evolutionary algorithm for groundwater level prediction.

Area of Science:

• Environmental Science
• Artificial Intelligence
• Computational Science

Background:

• Traditional models and ANNs struggle with groundwater level prediction accuracy.
• Metaheuristic optimization methods are time-consuming and require expert input.

Purpose of the Study:

• To propose a novel Large Language Model-assisted hyper-heuristic evolutionary algorithm framework (LLMHHEA) for enhanced groundwater level (GWL) prediction accuracy.
• To integrate LLM generative intelligence with hyper-heuristic search via a co-evolutionary mechanism.

Main Methods:

• Developed LLMHHEA with dynamic mutation strategies (metaheuristic, ANN, LLM-evolved), mutation-type-constrained crossover, and adaptive selection.
• Employed a hybrid encoding scheme for co-evolutionary mechanism.
• Validated on two GWL datasets and one temperature time series dataset.

Main Results:

• LLMHHEA demonstrated stable convergence and superior predictive results compared to traditional metaheuristic-ANN combinations.
• Identified optimal LLM-improved algorithm-ANN pairings for different datasets (e.g., LLM-KOA-ANFIS, LLM-GA-BP).
• LLM-evolved mutations significantly improved generalization and prediction accuracy over original algorithms.

Conclusions:

• LLMHHEA offers a new intelligent paradigm for complex optimization and prediction tasks.
• The framework shows generalizability across diverse data types and temporal characteristics.
• Provides robust technical support for sustainable water resource management.