A Cooling Load Prediction Method for Multifunctional Buildings Based on NMF-Kalman Filter Rolling Decomposition

Cooling load prediction is essential for energy-efficient HVAC operation and building energy management. In multifunctional buildings, heterogeneous load characteristics across functional zones increase the complexity of total cooling load prediction. Although decomposition–prediction methods can alleviate this problem, conventional full-sequence decomposition may introduce future information, leading to overly optimistic performance evaluation, whereas strictly causal rolling decomposition often suffers from unstable decomposition results. This study proposes a non-negative matrix factorization and Kalman-filter-based rolling decomposition method (NMF-KFRD). Constrained NMF extracts functional-zone load templates, while a rolling decomposition framework is constructed in which Kalman filtering recursively estimates contribution weights to improve decomposition continuity and stability. Prediction models are subsequently developed for the decomposed functional-zone loads, and the total cooling load is reconstructed from their predicted values. A case study using hourly cooling load data from a multifunctional building in Guangzhou, China, is conducted. With a back-propagation neural network (BP), NMF-KFRD achieves an RMSE of 227.49 kW, an MAE of 127.83 kW, and a MAPE of 9.59%, reducing these metrics by 23.2%, 31.5%, and 41.8% versus direct prediction, and by 13.9%, 20.0%, and 22.0% versus conventional rolling decomposition prediction. Overall, NMF-KFRD improves prediction accuracy while avoiding future-information leakage and maintaining decomposition stability, providing a reference for HVAC system optimization.