IoT-Based Residential Energy Forecasting with Explainable Feature Selection for Sustainable Building Management in the Industry 5.0 Era

Buildings are a key application domain for improving energy efficiency and sustainability, while Industry 5.0 emphasizes sustainable and human-centric digital transformation. This study presents a comparative framework for IoT-based residential energy consumption forecasting, combining time-series models with explainable artificial intelligence (XAI)-based feature analysis. Using a publicly available dataset from a residential building in Leicester, United Kingdom, the study predicts the next-hour total power consumption of six monitored appliances using a 24-h historical window and measurements available at or before the forecast time. Ridge Regression, conventional machine learning models, iTransformer, and TimeFilter are evaluated using 993 out-of-sample observations. SHapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) support feature analysis and selection. With all features, Ridge Regression achieves an RMSE of 35.869 W, an MAE of 23.261 W, and an R2 of 0.265. Feature selection has varying effects across models and evaluation metrics and does not necessarily improve predictive performance. Persistence and 24-h seasonal-naïve methods serve as baselines. The findings assess explainable IoT-based forecasting without requiring a full Digital Twin infrastructure and may offer insights relevant to Industry 5.0-oriented building management. Further validation across buildings and operating conditions is needed to assess generalizability.