AI-Driven Micro-Load Forecasting for Sustainable Smart Homes Using Lightweight Hybrid Models

Abstract

Accurate appliance-level forecasting is needed for smart-home energy management because aggregate household forecasts do not reveal which devices create short-term peaks. Many recent approaches also depend on complex deep-learning or ensemble pipelines that are difficult to deploy in resourceconstrained environments and provide limited support for device-level interpretation. This study proposes a lightweight micro-load forecasting framework that combines a compact long short-term memory (LSTM) network with a linear regression correction layer and temporal feature engineering. A synthetic 30-day dataset was generated at 15-minute intervals for an air conditioner, refrigerator, washing machine, and lighting load using realistic daily cycles, event-based activations, and controlled noise. The proposed model achieved an overall mean absolute error of 0.12 kW across the four appliance channels. Forecast-guided scheduling was then applied to shift flexible operation away from peak periods, producing an estimated monthly electricity reduction of 6%. SHAP analysis identified the temporal and lag features that influenced each appliance forecast. The results show that a lightweight and explainable micro-load model can support practical scheduling decisions and improve energy efficiency in sustainable smart homes.