Hybrid Time-Series Forecasting and Cost Optimization for Smart Homes Using Decomposition and Machine Learning

This paper proposes a hybrid framework for short-term residential energy forecasting and cost
optimization in smart home environments. The study leverages a synthetically generated
dataset comprising appliance-level energy consumption across 10 households during 2022,
sampled at 15-minute intervals and enriched with environmental and behavioral features. The
forecasting methodology integrates Seasonal-Trend decomposition using LOESS (STL) to
isolate trend and seasonal components, with a Random Forest regressor trained on the residuals
using lagged load features, rolling statistics, indoor and outdoor conditions, occupancy
indicators, and electricity pricing as exogenous predictors. Forecast generation is formulated
as the additive combination of the decomposed trend, seasonal components, and residual
predictions. The model is evaluated using a time-based train-test split, with forecasting
accuracy assessed through Mean Absolute Error (MAE), Root Mean Squared Error (RMSE),
and Mean Absolute Percentage Error (MAPE). In parallel, a cost optimization module is
developed under a time-of-use (TOU) pricing scheme, where washer operation sessions are
detected at 15-minute resolution and reallocated using a greedy heuristic to minimize daily
costs while respecting operational constraints. Experimental results demonstrate that the hybrid
model improves prediction accuracy relative to baseline methods, while the optimization
strategy yields tangible reductions in electricity expenditure. This work highlights the potential
of combining decomposition-driven forecasting with appliance-level scheduling to support
energy-efficient smart home management.