<p>The intensification of short-term heavy rainfall events has led to a notable increase in water-related disasters, especially inland flood inundation in lowland regions (such as deltas), where conventional drainage systems encounter challenges. To resolve this issue, it would be practical to develop a real-time forecasting method for inland flood inundation based solely on rainfall. As a first step, a method was developed to reproduce the computational results of conventional numerical simulation models using machine learning with LSTM, a long short-term memory (LSTM) algorithm. To allow for variations in the pattern of hyetographs used in the learning process, this method involved segmenting the temporal dimension into ten equally spaced intervals determined by the accumulation of rainfall at 50%. Subsequently, the model was optimized to leverage its optimal parameters, including the ideal number of training data, maximum epoch efficiency, and the most practical combination of hyperparameters, which was achieved through sensitivity analysis. Upon implementing these optimized parameters and solely utilizing rainfall data for prediction, the results indicated that for both flood inundation volume and depth, all of the statistical measures R², Nash–Sutcliffe efficiency, and percent bias fall within the ‘very good’ category. Additionally, computational resource assessments demonstrated a shorter average rendering time. This study concluded that LSTM-based inland flood modeling offers a faster and more resource-efficient alternative to traditional methods. It also shows promise for real-time predictions of inland flood inundation and improving disaster management in lowland areas.</p>