Short-Term Load Forecasting for Residential-Level Smart Microgrids: A Comparative Evaluation of Machine Learning and Deep Learning Architectures

Abstract

Accurate short-term load forecasting (STLF) is essential for modern grid operations, enabling efficient scheduling, demand response, and renewable energy integration. This paper presents a systematic comparison of five forecasting architectures applied to a large dataset of 98 residential homes, with 1-minute and 15-minute smart meter readings spanning 2015-2023. The models include an XGBoost pipeline with extensive feature engineering, a tuned CatBoost implementation, a feedforward neural network with multi-output regression, a multi-scale convolutional Kolmogorov-Arnold network (MCKAN), and a long short-term memory (LSTM) network with a 7-day lookback. All models are trained globally, pooling data across homes while incorporating home-specific categorical variables, including an assignment to a simulated microgrid topology with nine kiosks and three phases. Hyperparameter optimization is performed using Optuna and Keras Tuner. CatBoost achieves the lowest test MAE across all horizons, from 0.4089 (15-minute) to 0.7419 (30-day), outperforming XGBoost by 4-10% and deep learning models by larger margins. The findings support Sustainable Development Goals 7, 9, and 13 and provide actionable insights for energy management, particularly in the South African context of load shedding and grid decarbonization.