A Comprehensive Survey of Deep Learning for Short-Term Load Forecasting in Smart Grids

Short-term load forecasting (STLF) is a critical function for the efficient, reliable, and economic operation of modern smart grids. This survey provides a comprehensive and critical review of the application of deep learning (DL)techniques to STLF. We analyze the strengths, weaknesses, and suitability of various DL approaches, including recurrent neural networks (RNNs), convolutional neural networks (CNNs), hybrid models, attention mechanisms, physics-informed neural networks (PINNs), and quantum machine learning (QML) methods. The review goes beyond simple summarization, offering a comparative analysis of different architectures and methodologies, and identifying key research trends and gaps. We delve into the specific challenges of STLF, such as handling non-linearity, incorporating exogenous variables, dealing with limited or noisy data, and achieving scalability and interpretability. Finally, we propose concrete directions for future research, emphasizing the need for robust, explainable, computationally efficient, and uncertainty-aware DL models for STLF.