Аңдатпа
Real-time interpretation of seismic waveform data is essential for monitoring signal dynamics and detecting abnormal activity. However, forecasting seismic amplitudes remains challenging because seismic signals are highly variable, noisy, and non-stationary. This study presents an automated, gap-aware real-time forecasting pipeline for minute-level seismic signal amplitudes built on IRIS FDSN waveform data, together with a rigorous, statistically grounded comparison of seven forecasting models: Naive persistence, ARIMA, Ridge regression, HistGradientBoosting, a standard Gated Recurrent Unit (GRU), an LSTM, and a physics-inspired regularized GRU (PI-GRU) that adds temporal smoothness and amplitude-growth constraints to the training loss. Models were evaluated on 81 independently sampled, non-overlapping forecasting windows drawn from a continuous 91-day observation period across four horizons (5, 15, 30, and 60 minutes), using paired non-parametric significance testing, a quantitative ablation study, and noise- and missing-data robustness testing, and the main findings were additionally verified under an alternative peak-priority signal aggregation. Classical models, particularly Ridge regression and HistGradientBoosting, achieved significantly lower forecasting error than every neural model tested (Friedman and Holm-corrected Wilcoxon tests, p<0.0001) and were robust to synthetic noise and missing data, while the proposed PI-GRU was, contrary to expectation, the weakest-performing model overall and did not produce smoother forecast trajectories than an unregularized baseline GRU. We report this as an honest negative result: as currently configured, the physics-inspired regularization examined here does not demonstrate an accuracy or stability advantage over classical machine learning baselines for this task. The principal contributions of this work are the automated real-time forecasting pipeline itself and a reusable, rigorous evaluation methodology, together with specific, testable hypotheses - concerning hyperparameter tuning, training budget, and calibration of the physics-inspired constraint - for why the neural models underperformed and how this might be addressed in future work.