ISSN :2582-9793

dCor-Transformer: Feature-Independent Attention for Load Forecasting

Original Research (Published On: 17-Nov-2026 )
DOI : https://doi.org/10.54364/AAIML.2026.66359

Sabra ElFerchichi

Adv. Artif. Intell. Mach. Learn., - (-):-

1. Sabra ElFerchichi: University of Jeddah

Download PDF Here

DOI: 10.54364/AAIML.2026.66359

Article History: Received on: 01-Aug-26, Accepted on: 10-Nov-26, Published on: 17-Nov-26

Corresponding Author: Sabra ElFerchichi

Email: seelferchichi@uj.edu.sa

Citation: Sabra ElFerchichi. dCor-Transformer: Feature-Independent Attention for Load Forecasting. Advances in Artificial Intelligence and Machine Learning. 2026. [Ahead of Print]. https://dx.doi.org/10.54364/AAIML.2026.66359


Abstract

Sustainable resource planning and modern power grids depend on accurate electricity demand

forecasts. The underlying data present a modeling challenge due to: weather conditions,

time-of-day patterns, consumption history, and market prices. They are simultaneously

correlated and non-linear in their interactions. This study introduces the dCor-

Transformer, a multivariate-to-univariate forecasting framework that combines a statistical

feature preprocessing stage with a standard Transformer encoder. The preprocessing step

draws on distance correlation theory to structure the input space prior to attention, with

the goal of improving both computational efficiency and predictive focus. The framework

is evaluated on two complementary benchmarks — the AEMO Victoria dataset for shortterm

load forecasting and the UCI Electricity dataset for long-term horizons — to assess

its behavior across different forecasting regimes. Component-level analysis and multi-seed

validation further reinforce the reliability and interpretability of the results. Findings show

that dCor preprocessing yields the largest prediction gains on high-dimensional unstructured

input like UCI Dataset, while in the case of well engineered feature sets like AEMO it

achieves slightly lower than the Vanilla transformer accuracy while reducing training by

41%.


Statistics

Article Views: 15
PDF Downloads: 0