Oceanic internal waves (IWs) significantly impact marine engineering safety and submarine navigation stability, with amplitude serving as a core parameter characterizing their intensity. However, directly achieving rapid time-series forecasting of IW amplitudes remains a significant challenge, particularly for large-amplitude IWs. To date, the application of in-situ data for such forecasting tasks has not been reported. This study proposes a forecasting framework that integrates IW generation mechanisms with machine learning. We develop a CEEMDAN-based dual-track model featuring an innovative decomposition-frequency division predictionreconstruction strategy. Multiple algorithms are benchmarked separately for trend and high-frequency components to identify optimal configurations. Using Sea Level Anomaly (SLA) and tidal range data as fundamental features, the model is trained to predict the IW amplitude at the mooring location. We achieve rapid forecasting of semidiurnal IW amplitudes. The test results show excellent performance. On the test set, the RMSE is 1.19 m, and the MAE is 0.93 m. The R² exceeds 0.98. Extreme amplitude events (>40 m) further validate model reliability. This study successfully forecasts IW amplitude. It establishes a basic framework for future predictions that combine physical mechanisms with machine learning.