<p>The safe operation of large-scale irrigation districts is threatened by dynamic risks that exhibit complex spatiotemporal dependencies. Conventional prediction models often fail to capture these coupled dynamics and, critically, lack the ability to diagnose risk causality, limiting their practical utility. To address these limitations, this study proposes an enhanced spatiotemporal attention model (ESTAM) for traceable risk evolution forecasting. The model introduces a parallel attention architecture to simultaneously decode temporal dependencies and spatial correlations within the canal network. This structure, combined with a multitask learning paradigm, enables the model to perform causal diagnosis of primary risk drivers, moving beyond conventional risk warnings. The experimental results demonstrate that the ESTAM achieves a 91.88% prediction accuracy, substantially outperforming the baseline models. Further validation through case studies confirms the model’s robustness in tracing risk propagation and attributing causes, highlighting its potential for advancing proactive and fine-grained risk management in irrigation districts.</p>