A football optimization–enhanced deep learning framework with large language model guidance for desalination systems

The prediction of desalination system performance remains a critical challenge due to the complex, nonlinear, and multivariate nature of membrane-based processes, where data quality issues, feature redundancy, and suboptimal model configurations often undermine predictive reliability. Although deep learning models exhibit strong representational power, they are highly sensitive to preprocessing, feature selection, and hyperparameter tuning, while existing AutoML and rule-based optimization approaches are neither semantically aware nor sufficiently robust. To address these limitations, this paper proposes a joint data-oriented and optimization-based framework that integrates large language model (LLM)-controlled preprocessing with metaheuristic-driven feature selection and hyperparameter optimization. Specifically, several state-of-the-art LLMs are evaluated for preprocessing and feature selection, while hyperparameter optimization of deep learning models, with a particular focus on Spatio-Temporal Graph Convolutional Networks (STGCNs), is performed using the Football Optimization Algorithm (FbOA). Extensive experiments conducted on a synthetic desalination dataset demonstrate that FbOA consistently outperforms competing optimizers, yielding lower average prediction error, significantly reduced feature subsets, and improved stability. The final FbOA-optimized STGCN achieves substantial improvements in performance, with mean squared error reaching the order of \(10⁻⁷\) and very high correlation and efficiency indicators ( \(R^2> 0.98\) , NSE \(> 0.97\) , WI \(> 0.98\) ), significantly surpassing baseline and post-feature-selection configurations. These results confirm that combining LLM-guided preprocessing with FbOA-driven optimization leads to near-optimal predictive performance and efficient feature utilization. The proposed framework offers a versatile and scalable solution for desalination process modeling, with strong potential for enabling digital twins and predictive maintenance in next-generation water treatment systems.