IMVMD-MADNet: A Hybrid Framework for Multi-Scale Prediction of Chiller Energy Consumption

Abstract

Accurate prediction of chiller energy consumption is crucial for the efficient operation and intelligent management of heating, ventilation, and air conditioning (HVAC) systems in large buildings. However, such prediction remains challenging due to the multi-scale temporal coupling and nonstationary dynamics of chiller systems. Therefore, an IMVMD-MADNet hybrid framework integrating Improved Multivariate Variational Mode Decomposition (IMVMD) and a Multi-scale Aggregation Decomposition Network (MADNet) is proposed for chiller energy consumption prediction. To avoid information leakage and capture multi-scale features, a rolling local decomposition strategy with adaptive mode selection is employed. First, IMVMD performs stepwise decomposition within a sliding window, and the optimal number of modes is determined using envelope entropy. Then, sample entropy is used to reconstruct the multivariate modes into high-, medium-, and low-frequency components. Subsequently, a dual-branch MADNet combining wavelet-domain time-frequency modeling (WDP) and time-domain causal dependency modeling (TDP) predicts each component, and the results are aggregated to generate the final prediction. Bayesian optimization is employed to optimize the key hyperparameters. One year of real industrial chiller data from a plant in Huizhou, China, is used to evaluate the proposed model against 11 forecasting models. Results show that the dual-branch architecture outperforms single-branch models. The proposed model achieves the best performance across all forecasting horizons, with its advantage becoming more pronounced as the forecasting horizon increases, demonstrating stable predictive performance under the same-plant setting.