<p>Pond aquaculture is the predominant mode of freshwater aquaculture, where accurate assessment and early warning of future water quality are essential for effective management and the healthy growth of cultured species. Most current studies focus on water quality prediction but face challenges, including the limitations of single-factor models, the high cost of multi-factor models, and insufficient attention to dynamic early warning. This study proposes a dynamic early warning approach combining a parallel CNN-GRU fusion model with an Improved Comprehensive Water Quality Identification Index (ICWQII). The method utilizes the parallel CNN-GRU model to simultaneously predict multiple water quality parameters. A game theory-based algorithm optimizes empirical, objective, and prediction-error weights to derive dynamic comprehensive weights, which are then integrated with the ICWQII to evaluate future water quality levels. Notably, the model features a lightweight architecture containing only 0.15 million parameters, with a single-sample forward inference computational cost of approximately 0.23 MFLOPs. The proposed method achieved an accuracy of 0.8241 and a missing alarm rate (MAR) of 0.1104 on the constructed dataset. The results demonstrate that the proposed method balances lightweight characteristics with accuracy, enabling proactive water quality warning. This lays a foundation for extension to diverse aquaculture environments and species, providing technical support for intelligent water quality regulation.</p>