Decomposition–Ensemble Learning for Power Load Forecasting: Analysis of Noise Sensitivity and Algorithmic Stability

Precise short-term load forecasting is critical for smart grid reliability, yet the resilience of predictive models against ambient noise and sensor degradation remains insufficiently explored. This study systematically evaluates the noise sensitivity and operational stability of three forecasting paradigms (baseline LSTM, EEMD-LSTM, and CEEMD-LSTM) across varying signal-to-noise ratios (0–25 dB SNR) and sampling densities. Empirical analyses reveal a critical operational paradox: despite its theoretical sophistication, the CEEMD-LSTM architecture exhibits severe algorithmic instability and vulnerability to residual noise artifacts, yielding an RMSE standard deviation of ±1918 MW under high-fidelity conditions. Conversely, the EEMD-LSTM framework demonstrates superior robustness. Acting as an effective noise-mitigating filter, it sustains a Mean Absolute Percentage Error (MAPE) below 3.5% across all noise regimes, notably achieving sub-3.0% errors under severe interference (5 dB SNR). Furthermore, evaluating memory-optimized sparse sampling exposes a predictive trade-off, whereby reduced sampling density substantially exacerbates forecasting errors. These findings indicate that the EEMD-LSTM model, underpinned by dense sampling, provides a highly stable, noise-resilient predictive mechanism for smart grids subjected to uncertain measurements.