An Experimental Comparison of Machine Learning and Deep Learning Models for Energy Consumption Prediction in Smart Buildings

An Experimental Comparison of Machine Learning and Deep Learning Models for Energy Consumption Prediction in Smart Buildings

Mina Rajabi1 Mehdi Yaghoubi2

1) Master s student, Department of Computer Engineering, Faculty of Engineering, Golestan University, Gorgan, Iran, Email:
2) Assistant Professor, Department of Computer Engineering, Faculty of Engineering, Golestan University, Gorgan, Golestan, Iran, Email: m.yaghoubi(att)gu.ac.ir

IOI Code : AAHD-ABKGFK
Publication : National Conference on Information Technology, Nanotechnology, Artificial Intelligence and Technological Futures Studies(itnaf.ir)
Abstract :
    Accurate prediction of energy consumption in smart buildings is essential for optimizing energy efficiency, reducing costs, and enabling effective demand response strategies. This study presents a controlled experimental comparison of classical machine learning and deep learning approaches for predicting appliance energy consumption in a low-energy building. We evaluate four models: Random Forest, XGBoost, Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) networks, using the publicly available UCI Appliances Energy Prediction dataset. The experimental design employs rigorous temporal train-validationtest splitting to prevent data leakage, with evaluation across five random seeds for statistical robustness. Our results demonstrate that the LSTM model achieves the best predictive performance with a Mean Absolute Error (MAE) of 55.74±7.08 Wh, achieving lower MAE values than classical machine learning approaches (Random Forest: 158.89±2.55 Wh, XGBoost: 192.41±15.25 Wh). However, statistical testing revealed no significant differences between models at α =0.05, likely due to limited statistical power from five random seeds. All models exhibited negative R2 values, indicating they perform worse than simply predicting the mean, which suggests the dataset presents significant challenges for prediction. The findings provide practical guidance for selecting appropriate models for smart building energy prediction tasks and highlight the trade-offs between predictive accuracy and computational cost.

Keywords : energy consumption prediction smart buildings machine learning deep learning lstm gru random forest xgboost