@Article{ee.2026.088401,
AUTHOR = {Bowen Yuan, Huwei Liu, Jingyun Liu, Yuchen Lin},
TITLE = {Data-Driven Differential Energy Forecasting for Controlled Environmental Chambers},
JOURNAL = {Energy Engineering},
VOLUME = {},
YEAR = {},
NUMBER = {},
PAGES = {{pages}},
URL = {http://www.techscience.com/energy/online/detail/28258},
ISSN = {1546-0118},
ABSTRACT = {Controlled environmental chambers require continuous cooling, heating, humidification, ventilation and lighting to maintain stable indoor conditions, resulting in intensive and fluctuating energy demand. This study proposes a data-driven workflow for short-term differential energy consumption prediction in a BAE20-series controlled environmental chamber manufactured by Beijing Chuangyi Xintong Technology Co., Ltd. (Beijing, China) to support low-carbon operation and operational diagnostics. Chamber monitoring data were preprocessed to construct differential energy consumption, with negative differences set to zero and 99.5th-percentile (P99.5) truncation applied to reduce extreme spike effects. The resulting target series showed high sparsity, local spikes, short-term inertia and periodicity. Ridge regression, Random Forest, baseline Extreme Gradient Boosting (XGBoost), an improved XGBoost workflow, recurrent neural network baselines, autoregressive integrated moving average (ARIMA) and simple statistical baselines were evaluated using five-fold time-series validation and unified metrics. The improved XGBoost workflow achieved the lowest errors in the principal model comparison, with mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R<sup>2</sup>) values of 0.0001 ± 0.0001, 0.0008 ± 0.0005 and 0.9981 ± 0.0025, respectively. Under a separate expanding-window ablation with chronological data partitioning, eleven sensor-derived physical features reduced mean MAE by 17.00% and RMSE by 8.82% across all five folds. Unified-feature and ablation experiments indicate that the improvement mainly arises from robust target processing, rolling statistical features, time-series feature representation and regularized training, rather than model type alone. These findings suggest that energy prediction for controlled laboratory environments should align target construction and feature representation with sparse differential energy dynamics, offering a practical basis for low-carbon operation, anomaly detection and future control-oriented energy management.},
DOI = {10.32604/ee.2026.088401}
}