To address the problems that the water level of tidal river networks exhibits nonlinear fluctuations under the coupling effects of tides, runoff, and sluice-pump regulation, and that topographic differences of water systems result in low accuracy of holistic unified simulation and high cost and complexity of individual station modeling, an intelligent water level prediction method for tidal river networks was proposed based on a multi-scale fusion spatiotemporal attention network (MSTANet) with spatiotemporal clustering. By integrating multi-scale temporal features and spatial dependencies, this method enhanced the simulation capability for water level evolution processes under multiple disturbances, such as tides and sluice-pump regulation; meanwhile, spatiotemporal clustering was introduced to achieve watershed subdivision modeling, effectively mitigating the performance degradation caused by spatial heterogeneity. The validation in Wennan Water Control Area of Shanghai shows that this method has good simulation accuracy and computational efficiency and possesses engineering practicability and application prospects.