Evaluation of Load Forecasting in Intelligent Grid Systems Through Machine Learning Techniques

Smart grid (SG) devices depend heavily on accurate load prediction for efficient management, load shedding, and optimal processing dispatch, among other applications. However, achieving precise predictions while minimizing forecast errors is challenging due to the complexity and uncertainty of smart grid data. This paper provides a comprehensive analysis and practical overview of the latest advancements in probabilistic deep learning, focusing on prediction techniques for SG systems. Deep learning (DL) and machine learning (ML) methods have been thoroughly investigated for their potential in energy forecasting. Additionally, data and hybrid approaches are examined for their ability to enhance forecasting accuracy. The Victorian electricity consumption and American Electric Power (AEP) datasets serve as the basis for a case comparison between point and probability-based forecasting techniques. The study reveals that long short-term memory (LSTM) models with suitable hyperparameter tuning outperform point prediction models when dealing with larger samples and nonlinear patterns over extended periods. Moreover, the Bayesian bidirectional RNN (BRNN) probabilistic approach demonstrates the highest accuracy, as evidenced by the lowest root mean square error (RMSE).