Research on dynamic cooling load prediction method of cascaded-CPCMs building based on machine learning

Phase change buildings utilize reversible phase change for thermal energy storage, offering significant energy-saving potential. However, accurate load prediction in such buildings remains challenging due to complex heat transfer dynamics. This study constructs three test rooms—Room 1# (conventional), Room 2# (single-CPCMs), and Room 3# (cascaded-CPCMs with cascaded phase change temperatures: 19.18 °C, 23.05 °C, 26.29 °C)—to investigate thermal performance and develop a novel dynamic cooling load prediction framework. The cascaded-CPCMs design integrates vertically layered composite phase change materials to sequentially absorb/release heat, while a hybrid SSA-VMD-PCA methodology optimizes data quality for load forecasting. Specifically, the Sparrow Search Algorithm (SSA) optimizes Variational Mode Decomposition (VMD) parameters, decomposing original load data into Intrinsic Mode Functions (IMFs). Principal Component Analysis (PCA) then reduces dimensionality by extracting key features from IMFs. The processed data is fed into machine learning models (MLR, SVM, Bo-XG Boost) combined with a sliding window technique for dynamic predictions. Results show cascaded-CPCMs reduce summer temperature swings by 15 % and cooling loads by 16.52 % compared to conventional buildings. The SSA-VMD-PCA framework enhances prediction accuracy by 59.88 %, with the MLR model achieving the highest precision (R² ≥ 0.9998, MAE ≤ 0.1067). This study provides a validated methodology for scalable energy-efficient building design and adaptive thermal management.