Abstract: Water quality prediction is essential for ecological water management. Water quality exhibits complex non-stationary dynamics and multi-dimensional nonlinear relationships due to temporal evolution and environmental changes. In multi-parameter prediction of river water quality, the spatial-temporal dependence is complex, and traditional models struggle to integrate dynamic topology with? long-period features effectively. Therefore, we propose a spatial-temporal graph convolutional network model. In the temporal dimension, a masked sub-series transformer module extracts long-term trends through self-supervised pre-training. Combined with dilated causal convolution, it captures cumulative water quality effects. It solves the lag problem of traditional models in responding to sudden changes. In the spatial dimension, a dynamic graph learning module fuses a predefined station-distance adjacency matrix with a dynamic residual map to generate adaptive graph structures. Experimental results demonstrate that our model outperforms existing methods for predicting water quality. The prediction R2 for all water quality indicators exceeds 93%.