An Attention-Enhanced Transformer Framework for Intelligent Energy Management and Load Forecasting in U.S. Power Grids

Abstract

Accurate short-term and multi-horizon electricity load forecasting is a fundamental requirement for intelligent energy management in modern power grid systems, particularly under increasing demand variability and weather-driven consumption patterns. Conventional statistical, machine learning, and recurrent neural network models often exhibit limited capability in modelling complex non-linear relationships and long-range temporal dependencies inherent in large-scale power system data. To address these challenges, this paper proposes an Attention-Enhanced Transformer-based Multi-Horizon Weather-aware Network (ATMH-WNet) for efficient and accurate load forecasting in U.S. power grids. The proposed framework employs linear feature embedding and sinusoidal positional encoding to construct temporally informed latent representations, which are processed through a multi-layer Transformer encoder with multi-head self-attention. This design enables the model to jointly capture short-term dynamics and long-range dependencies while producing direct multi-step forecasts in a single forward pass,

thereby avoiding recursive error accumulation. The proposed model is evaluated on the PJM Interconnection hourly electricity consumption dataset spanning 2002-2018 and is compared against persistence, SARIMA, Prophet, XGBoost, and LSTM benchmarks. Experimental results demonstrate that ATMH-WNet consistently outperforms all baseline models, achieving a mean absolute error of 1325 MW, a root mean squared error of 1873 MW, a mean absolute percentage error of 2.9%, and an ????2 score of 0.97 on the held-out test set. Compared to the strongest deep learning baseline, the proposed framework reduces forecasting errors by more than 20% across major accuracy metrics. Additional qualitative analyses, including load profile alignment, residual diagnostics, and error distribution assessment, further confirm the robustness, stability, and generalization capability of the proposed approach. These results establish ATMH-WNet as an effective and scalable solution for real-world intelligent energy management and multi-horizon load forecasting applications.