A Time-Series Model of Gated Recurrent Units Based on Attention Mechanism for Short-Term Load Forecasting

Accurate short-term load data is foundational for rigorous research and policy formulation. We propose a machine learning framework capable of precisely forecasting future loads and addressing missing data to meet this requirement. This framework integrates considerations of temporal autocorrelation and inter-feature correlations, ensuring a high degree of generalizability across diverse load datasets. Our investigation assesses the efficacy of several machine learning architectures, including convolutional neural networks and gated recurrent units, in solving this problem. We also introduce a neural network-based time-series model to extract and leverage temporal and inter-feature correlations within these datasets. The proposed models are subjected to rigorous experimentation and validation using a real-world dataset to ascertain their robustness. This study aims to offer robust and efficient methodologies for analyzing electricity consumption data in both academic and practical contexts.