Deep Clustering and Adaptive Mode Decomposition of Industrial Loads for Temporal Complexity

Abstract

Accurate extraction of underlying features from electrical loads is crucial for intelligent microgrid management.To address the heterogeneous characteristics and microscopic nonstationary properties of industrial loads, a shunt-cooperative feature extraction framework relying on MSE is proposed in this work. Different from traditional methods that use a single feature extraction or clustering model to process the entire sequence, MSE is introduced to quantify the dynamic complexity of load sequences, and effective classification is performed for loads with strong and weak time-series characteristics. For strong time-series loads, a multi-view Variational AutoEncoder (VAE) deep clustering model with adaptive dual-distance metrics fused with Euclidean distance and LB_Keogh distance is constructed to accurately extract macroscopic typical power consumption patterns. For weak time-series loads, a “coarse screening-fine classification” hybrid decomposition architecture based on CEEMDAN-NGO-VMD is proposed, where the complexity index (LZC) is introduced to adaptively strip low-frequency trends and Northern Goshawk Optimization (NGO) is used to optimize VMD and suppress high-frequency mode aliasing. Simulation results show that, compared with the CEEMDAN-only method and the NGO-VMD-only algorithm, the proposed method achieves accurate extraction of multi-scale time-frequency features and effectively overcomes the mode aliasing problem of traditional decomposition methods in terms of clustering performance. Finally, complex loads are reconstructed into multi-modal feature components with clear physical meanings, providing a reliable feature basis for subsequent identification of multi-modal power consumption behaviors in industrial microgrids.