MACHINE LEARNING-DRIVEN LOAD PREDICTION AND REAL-TIME ENERGY OPTIMIZATION IN SMART MICROGRIDS

Abstract

To solve the dual problem of load forecasting and real-time energy optimization in smart microgrids, the use of machine learning (ML) methods offers a game-changer. In this paper the authors offer a hybrid system composed of Long Short-Term Memory (LSTM) networks to predict daily loads in the short-term, accompanied by Deep Q-Learning to execute the dynamic energy management. The LSTM model exhibited superior forecasting capabilities with high-resolution campus-based microgrid data and obtained a mean absolute percentage error (MAPE) as low as 3.25%, especially when considering low-variability periods. Simultaneously, the reinforcement learning (RL) agent, which was trained on Deep Q-Networks (DQN), succeeded in minimizing dispersal expenses, minimizing grid-dependency, and maximizing battery and renewable resource use by adapting optimal dispatch based on a simulation environment. A comparative review of rule-based and baseline approaches indicated a 22.8 percent less total energy cost and 26.7 percent lower peak load demand. Its flexibility, active management, and use of local renewable power generation favour the increasing need of sustainable and intelligent systems of energy. The study highlights the benefits of ML-powered microgrids to energy independence, resiliency, and carbon emissions minimization.

Keywords

Smart Microgrids, Load Forecasting, Long Short-Term Memory (LSTM), Reinforcement Learning, Deep Q-Learning, Energy Optimization, Renewable Energy, Battery Management, Demand Prediction, Machine Learning in Power Systems