Electrical load forecasting is a cornerstone of smart grid management, relying heavily on Artificial Intelligence (AI) to predict power demand accurately. Deep learning models, particularly Artificial Neural Networks (ANN), are widely employed for this purpose; however, they typically require massive data aggregation and significant computational time. This paper provides an extensive survey of deep learning-based load forecasting techniques published between 2015 and 2020. The survey categorizes existing literature into Distributed Deep Learning (DDL), Back Propagation (BP), and non-BP based methodologies. A key finding of the survey suggests that shifting away from centralized data dependency can significantly reduce computational overhead. Consequently, the authors present a novel conceptual DDL model for smart grids. This model incorporates the Hilbert-Schmidt Independence Criterion (HSIC) Bottleneck technique, which aims to improve forecasting accuracy while maintaining computational efficiency. This research contributes a valuable framework for future smart grid implementations, emphasizing the transition toward more decentralized and intelligent energy management systems.

Keywords: Smart Grid, Load Forecasting, Deep Learning, Distributed Deep Learning (DDL), HSIC Bottleneck, Artificial Neural Network (ANN), Energy Systems.