This　research　focuses　on　the　fluctuation　recognition　and　short-term　forecasting　of　new　energy　power　generation,　specifically　targeting　wind　and　solar　farms.　We　propose　a　novel　two-layer　decomposition　model　based　on　ICEEMD-VMD　to　preprocess　the　non-stationary　and　volatile　output　power　data,　effectively　reducing　noise　and　improving　data　quality.　The　study　explores　the　fluctuation　patterns　of　wind　and　solar　power　generation,　establishing　a　dynamic　thresholdbased　fluctuation　recognition　model　to　detect　significant　power　changes.　For　short-term　forecasting,　a　Bi-LSTM　model　is　employed　to　predict　the　output　power　components　obtained　from　the　ICEEMD-VMD　decomposition.　The　model＇s　performance　is　evaluated　using　RMSE,　MAE,　and　goodness　of　fit　(R2),　demonstrating　high　accuracy　and　reliability.　Our　findings　indicate　that　the　proposed　methods　can　effectively　capture　and　predict　power　generation　fluctuations,　providing　valuable　insights　for　grid　operation　and　energy　management.　©　2025　IEEE.