DATA-DRIVEN MODEL FOR SMART HOME ENERGY PREDICTION USING DEEP AND TREE-BASED LEARNING

Abstract

With residential energy consumption accounting for over 27% of global electricity usage and smart homes projected to grow to over 1.2 billion devices by 2030, optimizing energy usage through accurate prediction has become a critical need. However, current energy prediction systems often struggle with inconsistent sensor data and poor generalization across varied household patterns, leading to inefficient energy utilization and limited scalability. This work proposes a robust energy prediction framework tailored for smart home environments, addressing the shortcomings of traditional machine learning models through a hybrid deep learning approach. Initially, raw sensor data—often noisy and incomplete—is passed through a comprehensive data preprocessing pipeline involving normalization, noise filtering, and missing value imputation to ensure consistency and model reliability. As a baseline, the widelyused Decision Tree Regressor (DTR) is employed to evaluate performance limits of conventional models. Building upon this, we introduce a hybrid Convolutional Neural Network with Random Forest Regressor (CNN-RFR) architecture. The CNN component is responsible for extracting spatial and temporal energy usage patterns from multivariate timeseries data, effectively capturing localized consumption behaviors. The extracted features are then fed into a Random Forest Regressor, which brings robustness against overfitting and enhances interpretability. Experimental results demonstrate that the proposed CNN-RFR model significantly outperforms DTR and standalone models, yielding improvements in RMSE and MAE by substantial margins. This study highlights the potential of hybrid deep learning models to drive smarter, more adaptive home energy systems while preserving computational efficiency