The history of disasters caused by heavy rains in Peru, characterized by mudslides, floods with devastating intensities that have resulted in significant loss of life and material damage, underscores the urgency of effective early warning systems. In particular, events such as those of 2016–2017 and 2024 at the site of our research highlight the region’s vulnerability to rainfall intensity activity. In this context, this work presents a novel approach for heavy rainfall early warning systems using deep learning architectures, specifically Long Short-Term Memory (LSTM) networks and Temporal Fusion Transformer (TFT) networks. The study focuses on predicting hourly maximum rainfall, a crucial parameter for issuing timely warnings and mitigating the dangers of heavy rains that can cause material damage and loss of human life. Using a comprehensive dataset comprising 2590 records with 4 parameters (maximum temperature, minimum temperature, relative humidity and precipitation) hydrometeorological data from the Chalhuanca station in Peru, the study employs LSTM and TFT networks to predict daily maximum rainfall values. Data preprocessing involves homogenizing hydrometeorological data and dividing it into training, validation, and testing sets. The results indicate that both LSTM and TFT networks show promising performance in different time windows (6, 12, and 24 h). For LSTM networks, the 12-h time window produces the most accurate predictions, with a balanced accuracy of 99.85% and an accuracy in F1-Score per class: (a) 87.71%, (b) 99.77%, and (c) 100%. Meanwhile, TFT networks achieve competitive results, especially with 12-h time windows, showing a balanced accuracy of 99.72% and an accuracy in F1-Score per class: (a) 88%, (b) 100% and (c) 99.58%. The findings underscore the potential of deep learning architectures to improve heavy rainfall warning systems, contributing to more effective disaster preparedness and response strategies in disaster-prone regions.