Power consumption prediction in warehouses using variational autoencoders and tree-based regression models

• A novel deep learning and tree-based hybrid model for warehouse power prediction. • Variational Autoencoders for dimensionality reduction and latent feature extraction. • ExtraTree model achieving peak accuracy in warehouse power consumption forecasting. • Enhanced transparency and interpretability to foster human-machine trust. • Precision power forecasting for cost savings and C O 2 emissions reduction. Precise power consumption prediction is essential for efficient energy management, resource allocation, infrastructure planning, and cost optimization. In this paper, power consumption prediction is addressed using machine learning (ML) techniques. The CU-BEMS dataset is employed to represent power consumption and variations in ambient conditions within a seven-floor building. The dataset is explored, preprocessed, and analyzed. A theoretical framework linking power consumption and floor features is developed using a variational autoencoder (VAE) to compress all features from all zones on all floors into a latent space of only 15 features. This latent space is then combined with a specific zone (Zone 2) on a particular floor (Floor 6) and fed into a tree-based regressor layer to predict all features in this zone. A windowing function is used to create lagged versions of past data to predict future feature values in a multi-output scenario (6 outputs). It is found that for 1-hour-ahead forecasting, the ExtraTree regressor achieves the highest accuracy with a coefficient of determination ( R 2 ) of 97.4% and Mean Absolute Error ( M A E ) of 0.46. The prediction of power consumption features reaches 99.2%, specifically for the air conditioning unit’s consumed power. The proposed solution outperforms previous works, with LightGBM as the best regressor for 10-minute-ahead forecasting ( M A E = 0.218 , R 2 = 98.8 ) and 20-minute-ahead forecasting ( M A E = 0.317 , R 2 = 97.4 ). The proposed framework demonstrates its power in linking all floors and consumption zones, making it generalizable, accurate, transferable, and efficient for predicting power consumption within industrial facilities and habitats.