A Hybrid Frequency-Disentangled Framework for Household Energy Consumption Forecasting via VMD-Trend Decomposition and Attentive CNN-BiLSTM
IEEE Access · 2026 · 14 · 69766-69777
DOI: 10.1109/access.2026.3687392 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Muge Li und 3 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Accurate household energy forecasting is critically challenged by the inherent stochasticity of occupant behavior, which manifests as irregular high-frequency fluctuations superimposed on deterministic consumption patterns. Conventional end-to-end deep learning models struggle with such non-stationary signals, as they force neural networks to simultaneously capture multi-scale dynamics ranging from stable seasonal trends to abrupt behavioral spikes within a unified learning objective. This research proposes a frequency-disentangled hybrid framework that systematically decouples signal components through a two-stage decomposition strategy: Trend-Residual Decomposition (TRD) followed by Variational Mode Decomposition (VMD), integrated with an attentive CNN-BiLSTM backbone. The TRD stage first isolates the long-term consumption trend via sliding-window filtering (168-hour window), separating it from complex residuals. Subsequently, VMD resolves the residual into $K=6$ quasi-orthogonal intrinsic mode functions (IMFs) representing distinct frequency bands, effectively neutralizing non-stationarity while avoiding the mode-mixing artifacts inherent in empirical decomposition methods. Each IMF, enriched with engineered temporal features (hour-of-day, day-of-week, weekend indicator), is modeled by a dedicated CNN-BiLSTM-Attention expert that learns component-specific temporal dependencies. The final forecast is reconstructed by aggregating IMF-level predictions with the deterministic trend. Empirical evaluations on the Individual Household Electric Power Consumption (IHEPC) dataset demonstrate that the proposed framework achieves a mean absolute percentage error (MAPE) of 11.95%, compared to 32.42% for the state-of-the-art decomposition-agnostic CNN-BiLSTM-Attention baseline, representing a 63.1% relative error reduction. Ablation studies substantiate the indispensability of the integrated architecture: the end-to-end variant (lacking TRD-VMD decomposition) suffers a catastrophic performance collapse with MAPE surging to 45.99%. Furthermore, excluding the attention mechanism or CNN layers increases the error to 13.67% and 12.80%, respectively, confirming that each module is essential for mitigating the “frequency-mixing” effect in volatile residential energy signals. These results establish that frequency-domain disentanglement is not merely an enhancement but a prerequisite for modeling volatile residential energy consumption patterns in next-generation building energy management systems.
Zitieren
APA
Li, M., Tian, Z., Zhang, S., & Jin, C. (2026). A Hybrid Frequency-Disentangled Framework for Household Energy Consumption Forecasting via VMD-Trend Decomposition and Attentive CNN-BiLSTM. IEEE Access, 14, 69766-69777. https://doi.org/10.1109/access.2026.3687392
Vancouver
Li M, Tian Z, Zhang S, Jin C. A Hybrid Frequency-Disentangled Framework for Household Energy Consumption Forecasting via VMD-Trend Decomposition and Attentive CNN-BiLSTM. IEEE Access. 2026;14:69766-69777. doi:10.1109/access.2026.3687392
BibTeX
@article{Li2026,
author = {Li, Muge and Tian, Zhongyu and Zhang, Sunyu and Jin, Changjie},
title = {A Hybrid Frequency-Disentangled Framework for Household Energy Consumption Forecasting via VMD-Trend Decomposition and Attentive CNN-BiLSTM},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {69766--69777},
doi = {10.1109/access.2026.3687392}
}
EUVIMED