In this study, a graph guided spatiotemporal correlation prediction model (GSCPM) is proposed to solve the problem of spatiotemporal relationship modeling and lag impact in watershed flood prediction. The model encodes the time correlation of the historical attributes of each monitoring point through multiple long and short-term memory networks (LSTM), then uses graph convolution neural network (GCN) to find the geographic and spatial dependencies between detection points. In addition, we propose rainfall delay features, flood discharge delay features and upstream water level delay features to find the lag effects between variables, and perform extensive experiments on real-world watershed dataset.The experimental results demonstrate that the GSCPM model is superior to LSTM, RNN and other models, which is suitable to be widely used in watershed flood prediction.