This study proposes an integrated regulation model that synergizes dynamic threshold optimization with bidirectional long short-term memory (Bi-LSTM) prediction to address the volatility characteristics of renewable energy generation and grid integration requirements. The key findings demonstrate that the dynamic threshold optimization, leveraging ICEEMD-VMD decomposition for fluctuation feature extraction and dynamic programming for threshold calibration (t = 0.59), achieves 82.42% identification accuracy in power mutation events with a 5-min advance warning capability, substantially mitigating grid dispatch response delays. Simultaneously, the Bi-LSTM prediction framework exhibits superior performance in high-frequency (1 Hz) power forecasting within 120-s horizons, reducing RMSE and MAE by approximately 15% compared to conventional models while maintaining an R2 coefficient exceeding 90%, thereby validating its exceptional capability in capturing global temporal features for time-series prediction. The integrated approach effectively coordinates multi-timescale operational requirements through adaptive threshold mechanisms and bidirectional temporal dependency modeling, offering a robust solution to address the inherent intermittency and volatility challenges in renewable energy systems.