Retrofitting Energy IoT Devices in the Residential Sector: Efficient Machine Learning for Day-Ahead Forecasting

Electricity forecasting at the household level is vital for optimizing residential energy consumption and empowering grid flexibility provision. At the same time, many residents are reluctant to share their energy data, encouraging local and privacy-preserving model deployment. The growing availability of IoT (Internet of Things) devices in modern residences provides a unique opportunity to tap into the computational potential of these devices and enable advanced data-driven energy optimization. Achieving this, however, required deploying effective yet computationally efficient forecasting pipelines, given the computational constraints faced by HEMS (Home Energy Management System) that control residential devices. These systems typically operate with modest CPUs and minimal RAM, lacking high-performance GPUs or modern multi-core processors, thus presenting a computationally constrained environment. Therefore this study evaluates efficient machine learning algorithms, within computationally restricted environments for key household-level time series applications namely, forecasting heat pump (HP) and household consumption and photovoltaic (PV) power generation. The employed models in this study are XGBoost, CatBoost, LightGBM, Random Forest, Stochastic Gradient Descent (SGD) Regressor, and Kolmogorov-Arnold networks (KAN), for day-ahead forecasting. The models were evaluated based on overall accuracy, computational efficiency, and their ability to predict peak periods accurately, creating a Pareto front of models excelling in precision and runtime speed. Although performance can vary by case, the analysis suggests that the SGD Regressor consistently offers superior speed, KAN generally achieves lower RMSE scores, and tree-based models demonstrate strong performance in capturing peak demand. This study provides practical insights for deploying machine learning models in computationally limited settings, supporting efficient forecasting in residential energy systems with the broader objective of enhancing grid flexibility.