Accurate water quality prediction is essential for effective ecological water management. However, water quality exhibits complex non-stationary dynamics and multi-dimensional nonlinear relationships driven by temporal evolution and environmental variability. In multi-parameter river water quality prediction, intricate spatial-temporal dependencies make it difficult for traditional models to effectively integrate dynamic topology and long-period features. To address this challenge, we propose a spatial-temporal graph convolutional network (STGCN) model. In the temporal dimension, a masked sub-series transformer module is employed to extract long-term trends through self-supervised pretraining. Combined with dilated causal convolution, it captures cumulative water quality effects and alleviates the response lag common in traditional models when facing abrupt changes. In the spatial dimension, a dynamic graph learning module integrates a predefined station-distance adjacency matrix with a dynamic residual map to generate adaptive graph structures. Experimental results demonstrate that the proposed model outperforms existing methods in water quality prediction, achieving an R² greater than 0.93 across all water quality indicators.