Coupled CST-GAN Data Augmentation and A-STGCN for Predicting Nitrogen-Phosphorus Migration and Unraveling Algal Bloom Mechanisms in the Taihu River-Lake System

As a typical complex aquatic ecosystem, the Taihu Basin relies heavily on the material exchange between the lake and its connected rivers, a critical factor driving eutrophication and algal blooms. However, existing water quality monitoring data suffer from spatiotemporal sparsity and insufficient sample sizes, limiting the accuracy of deep learning models in simulating long-term nitrogen and phosphorus (N/P) migration and mining algal bloom mechanisms. To address these challenges, this study proposes a comprehensive framework integrating data augmentation with graph deep learning. Specifically, a Conditional Spatio-Temporal Generative Adversarial Network (CST-GAN) was first constructed to learn inherent distribution patterns and generate high-quality augmented data, significantly expanding the sample scale. Subsequently, the Attention-based Spatio-Temporal Graph Convolutional Network (A-STGCN) was employed to model the river-lake system as a topological graph. Crucially, leveraging the interpretability of the attention mechanism embedded within A-STGCN, this study moved beyond black-box prediction to successfully identify primary N/P input channels, quantify the response mechanisms of algal blooms to N/P migration fluxes, and pinpoint the key driving factors triggering outbreaks. This research demonstrates a closed-loop approach from data augmentation to system simulation, providing a scientific basis for precise pollution control and early warning of algal blooms in the Taihu Basin.