Stacked ensemble learning approach for PCM-based double-pipe latent heat thermal energy storage prediction towards flexible building energy

Abstract

Dynamic performance prediction of the PCM-based double-pipe latent heat thermal energy storage (LHTES) system plays an important role for building peak load shifting. This study presents a stacked ensemble learning framework for the phase change performance prediction of LHTES systems. The proposed model has a two-layer structure involving base and meta models: Regression Tree, Support Vector Regression and Linear Regression. Sensitivity analysis has been utilized for the feature extraction considering possible geometrical and physical parameters of LHTES systems with a fixed PCM volume that can store 0.94 kWh of energy within 255 min. Case studies show that the proposed model outperforms the base models in terms of both prediction accuracy and robustness. Compared with base models, the proposed model has a maximum 7.82% (charging) enhancement in Mean Absolute Percentage Error (MAPE) over all-range data sets and a maximum 16.43% enhancement in MAPE for the initial discharging stage, and has more stable predictive results within the high accuracy zone (accuracy > 95%) with a maximum 7.3% (charging) increase in the distribution density over all-range data sets and a maximum interquartile range reduction of 25.6% (middle discharging stage) in Mean Absolute Error. Application of the proposed model achieved more building peak load reduction.