Extreme weather events, combined with human-induced factors, such as expanding impervious surfaces and inadequate drainage infrastructure, are driving escalating urban flood risks worldwide. In this study, we present a novel spatiotemporal Long Short-Term Memory (LSTM)-based surrogate of the U.S. Environmental Protection Agency (EPA)’s Storm Water Management Model (SWMM) to predict maximum water depth and inflow at the asset level within urban drainage networks. The high-resolution SWMM model, encompassing the full network of conduits and manholes, was first calibrated and validated using U.S. Geological Survey (USGS) observations. The LSTM surrogate was then trained on data from 5,000 rainfall events across seven Annual Recurrence Intervals (ARIs) ranging from 1 to 100 years. The SWMM-LSTM surrogate model consistently achieves high predictive performance for both water depth and inflow, highlighting its robustness across diverse storm scenarios and ARI conditions. Hyperparameter optimization via grid search revealed task-specific configurations: larger hidden layers with moderate dropout improved water depth predictions, while deeper network architectures with minimal dropout optimized inflow forecasts. By providing rapid, computationally efficient predictions without compromising accuracy, the SWMM-LSTM surrogate offers a practical tool for real-time flood risk assessment, scenario evaluation and actionable decision-making in complex urban drainage systems.