Diminishing returns in feature selection for short-term electrical load forecasting

Department of Power Systems and Management, Technical University of Cluj-Napoca, Cluj-Napoca, Romania

^* Corresponding author: This email address is being protected from spambots. You need JavaScript enabled to view it.

Published online: 11 June 2026

Abstract

Short-term electrical load forecasting (STLF) is essential for reliable and cost-efficient power system operation. Although modern data-driven forecasting models can incorporate many contextual predictors, the marginal benefit of increasing input dimensionality remains insufficiently quantified, potentially leading to unnecessary sensing and integration complexity. This paper investigates the diminishing returns behavior associated with contextual feature expansion within a time-series forecasting framework. A controlled experimental study is conducted using a reproducible synthetic dataset designed to emulate realistic load dynamics, including temporal autocorrelation, weather dependence, and operational regime variability. Temporal memory is explicitly enforced through lagged and rolling load features, while contextual variables are incrementally selected using fold-wise filter-based ranking under time-aware cross- validation. Results obtained across linear, kernel-based, and ensemble regression models show that temporal features account for most of the predictive power, while a limited number of contextual variables provide complementary improvements. Beyond moderate subset size, relative performance gains consistently fall below 1% in RMSE and may degrade due to redundancy effects, empirically indicating diminishing returns. An ablation study further confirms that controlled feature selection improves robustness and prevents performance instability in higher-dimensional settings. The findings suggest that accurate short-term load forecasting can be achieved using a compact and interpretable feature set, supporting cost-efficient sensor deployment strategies in smart grid applications.