A Computer Vision and Transformer-Based Framework for Short-Term Energy Consumption Forecasting in Smart Grids

Abstract

Accurate short-term energy consumption forecasting is critical for the efficient operation of smart grids to balance the loads in advance, implement demand response on the consumer side, and reduce distribution costs. However, most of the existing deep learning methods, including LSTMs and CNN-based models, are unable to capture both long-range temporal patterns and the fine-grained local patterns within high-resolution smart meter data simultaneously. In this paper, we propose a novel hierarchical transformer-based framework called EF-Transformer for short-term energy forecasting, which consists of a multi-scale patch embedding module, a dual-branch temporal attention encoder, and an adaptive decomposition gate that decomposes and handles non-stationary trend and seasonal components. We evaluate the proposed model on three publicly available benchmark datasets, including the UCI Individual Household Electric Power Consumption dataset, the Electricity Load Diagrams 2011-2014 dataset, and the Smart Meter Energy Consumption dataset, and on the UCI household dataset, EF-Transformer achieves the best MAE of 0.091 kWh, RMSE of 0.134 kWh, and MAPE of 2.87% among the state-of-the-art models, including LSTM, Auto former, Informer, and Patch-TST. We conduct ablation experiments to demonstrate that all the components of the architecture contribute to the state-of-the-art performance, and overall, EF-Transformer provides a scalable and interpretable solution for smart grid operators who require accurate short-term consumption forecasts at both household and grid levels.