Short-Term Load Forecasting with Attentive Neural Processes: Adaptivity and Uncertainty Estimation

Short-term load forecasting (STLF) is essential for the efficient operation and stability of modern power systems, particularly in smart grids with high penetration of renewable energy sources. In this study, we propose an enhanced Attentive Neural Process (ANP) framework for STLF, incorporating a customized loss function with an additional term to improve predictive performance. The ANP framework enables personalized and accurate forecasting by leveraging its latent variable representation and attention mechanism to adapt predictions based on newly observed consumption patterns. Its ability to encode an arbitrary number of observations allows for efficient real-time updates without requiring extensive retraining, allowing for rapid and continuous adaptation to evolving data. Through probabilistic modeling, ANP can quantify uncertainty in predictions, which can play a crucial role in risk-aware decision-making in power systems. Experimental results show that our ANP based approach achieves the highest prediction accuracy, reducing the Mean Average Error (MAE) by between 14.95% and 60.97% compared to a range of alternative learning approaches.