Water quality prediction is of great significance for water environmental protection and management. Traditional water quality prediction methods are mainly based on linear models, and they fail to extract nonlinear relationships. Recurrent neural network-based ones have shortcomings such as being difficult for long-term prediction and unable to capture spatial dependencies. To address these issues, this work proposes an improved spatio-temporal prediction model called GATWNet, which combines a -G raph AT tention network with a W aveNet model based on dilated causal convolution to predict water quality at multiple sites over a future time period. GATWNet jointly captures the spatial information of river networks and the temporal information of each water quality monitoring sensor. Two real-world datasets-based experimental results demonstrate the proposed GATWNet achieves higher prediction accuracy than several baseline models.