Seasonal Pattern Recognition-based Advanced Hybrid Machine Learning Methods for Residential Energy Consumption Forecasting in Smart Grid Networks

Abstract

This research explores machine learning (ML) techniques, with an emphasis on neural networks (NNs), to predict household energy consumption in the context of smart grids. The study evaluates algorithms such as CNN, LSTM, RF, AdaBoost, and CatBoost with a primary focus on hybrid models to enhance forecast precision. The research adopts a comprehensive data collection strategy while dividing the time series data into training (80%) and testing (20%) segments for rigorous model evaluation. Hybrid approaches that combine several algorithms significantly outperform single models, and LSTM-CatBoost is at the forefront by exhibiting lower error rates and higher R² values. To avoid overfitting, the research employs cross-validation in conjunction with early stopping, thus, producing robust and trustworthy models. The paper provides a wealth of information on energy forecasting that paves the way for the energy efficiency and sustainability of smart grids, which are essentially energy management systems based on the latest technology and consumer feedback mechanisms.