A Short-Term Prediction Method for Power Loads Based on Dynamic Clustering and Time-Frequency Decomposition

The integration of renewable energy and demand response mechanisms has intensified the uncertainty and volatility of power load patterns. Accurate short-term load forecasting (STLF) is critical for grid stability and economic dispatch. This paper proposes a hybrid prediction method combining dynamic clustering and time-frequency decomposition to address the challenges of load variability. First, an adaptive dynamic fuzzy Cmeans (DFCM) clustering algorithm is introduced to categorize users based on evolving consumption patterns. Second, the Variational Mode Decomposition (VMD) technique decomposes load sequences into intrinsic mode functions (IMFs) to capture multi-scale temporal features. These components are then reconstructed and fed into an XGBoost model for prediction. Validated on a real-world dataset from Southwest China, the proposed method reduces mean absolute percentage error (MAPE) by 18.7% compared to traditional static clustering and single-scale models. The results demonstrate superior accuracy and robustness in handling abrupt load changes and periodic fluctuations.