Evaluation of the security of the water–energy–food system in the Yellow River Basin using explainable machine learning models

ABSTRACT Workflow constructing a WEF security indicator system, defining thresholds, preprocessing data, training nine machine learning models, applying random forest regression, and conducting SHAP-based driving factor analysis. The security of the water–energy–food (WEF) system is an important guarantee of sustainable regional development. Assessing WEF security and its drivers helps tackle resource and environmental challenges. Traditional assessment methods like system dynamics, logistic curve, and fuzzy logic rely on subjective inputs or specific assumptions, causing model inaccuracies that distort security evaluations. Based on 2012–2021 Yellow River Basin data, this study constructed a WEF security evaluation indicator system based on the pressure–state–response framework and established security threshold standards. Nine machine learning models were compared to identify the optimal approach for WEF security assessment, and Shapley additive explanations explained each factor's contribution and driving mechanism. The results indicated that random forest regression outperformed other models in evaluating the WEF security of the Yellow River Basin. The WEF security index plummeted in 2017, while changes in other years were relatively minor. Spatially, high-security areas were concentrated in the eastern and southern regions of the basin. Analysis of influencing factors revealed that most factors had similar levels of contribution to WEF security, with no single dominant factor, and their impacts were nonlinear and phased. This study provided a novel framework for WEF security evaluation and offered scientific support for policy formulation.