Deep Learning Models Benchmark for Smart Meter Energy Forecasting

Accurate energy consumption forecasting is crucial for the efficient operation of power systems, directly impacting costs, energy management, and system maintenance. With the increasing availability of high-resolution data from smart meters, data-driven methods have become prevalent for both short-term and long-term forecasting. However, a comprehensive comparative analysis of modern deep learning models on real-world smart meter data has been lacking.This research presents an empirical benchmark of nine contemporary deep learning models, encompassing linear, MLP-based, convolutional, and Transformer architectures, for time-series forecasting. The models were evaluated on two public smart meter datasets, focusing on three key factors: the length of historical input, the prediction horizon, and the chosen model architecture. The findings indicate that extending historical context improves accuracy up to a certain point, beyond which benefits diminish, while accuracy consistently decreases with longer prediction horizons.The study also examined the trade-off between prediction accuracy and computational complexity. Deep learning models consistently surpassed classical baselines, with lightweight architectures achieving comparable performance at significantly reduced computational costs. Architectural differences proved more impactful for longer forecasting horizons and more heterogeneous datasets. Furthermore, a subgroup analysis across various geodemographic and household categories revealed that model choice had limited impact for most population segments, suggesting robust performance across diverse user groups.