CNN-LSTM-Based Forecasting of Peak Ground Acceleration from Early Seismic Waveforms

Accurate and rapid peak ground acceleration (PGA) forecasting is vital for effective earthquake early warning (EEW) systems. This study presents a deep learning approach that combines convolutional and bidirectional recurrent layers with an attention mechanism to estimate PGA using the first three seconds of three-component seismic waveforms. The model is trained on strong-motion records from Japan’s K-NET network, covering events from March 2022 to March 2025. We compare three architectures: a fully connected neural network (ANN), a plain LSTM model, and our final CNN–BiLSTM model with attention mechanism. The proposed model outperforms both baselines, achieving a validation Mean Absolute Error of 2.61 gal and R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> of 0.714. These results highlight the effectiveness of combining convolutional and sequential modeling with an attention mechanism for real-time seismic forecasting.