Peru’s seismic history, characterized by devastating earthquakes resulting in significant loss of life and property damage, underscores the urgency of effective early warning systems. Notably, events like the 1746 and 2007 Pisco earthquakes highlight the vulnerability of the region to seismic activity. In this context, this work presents a novel approach to earthquake 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 Peak Ground Acceleration (PGA), a crucial parameter for issuing timely alerts to mitigate earthquake hazards. Using a comprehensive dataset comprising 5045 seismic records from various locations in Peru, the study employs LSTM and TFT networks to predict PGA values. Data preprocessing involves homogenizing acceleration records and dividing them into training, validation, and testing sets. Results indicate that both LSTM and TFT networks demonstrate promising performance across different time windows (5, 30, and 60 s). For LSTM networks, the 60-s time window yields the most accurate predictions, with validation accuracy reaching 98.015% and testing accuracy at 88.89%. Meanwhile, TFT networks achieve competitive results, particularly with 30-s time windows, showing validation accuracy of 96.03% and testing accuracy of 91.32%. The findings underscore the potential of deep learning architectures in enhancing early warning systems, contributing to more effective disaster preparedness and response strategies in earthquake-prone regions.