Uncertainty-Aware Short-Term Load Forecasting using Autoformer with Split Conformal Prediction

Abstract

Short-term load forecasting (STLF) plays a critical role in modern smart grid operation by enabling reliable dispatch, reserve allocation, and demand-side management. Although Transformer-based architectures have demonstrated strong point forecasting performance, most existing approaches remain deterministic and do not provide calibrated uncertainty estimates required for risk-aware operational decisions. This paper proposes an uncertainty-aware STLF framework that integrates a multi-horizon Transformer forecaster with split conformal prediction to construct distribution-free and empirically calibrated prediction intervals. The Transformer captures long-range temporal dependencies in load dynamics, while conformal calibration ensures finite-sample coverage guarantees without requiring distributional assumptions. Experiments conducted on Zone 1 of the Tetouan power consumption dataset demonstrate that the proposed framework achieves competitive point forecasting accuracy, obtaining an RMSE of 1089.73 and MAE of 860.09, corresponding to a MAPE of 3.17%. Furthermore, the conformal module provides reliable uncertainty quantification, achieving an empirical prediction interval coverage probability (PICP) of 0.891 at a nominal 90% confidence level, with a PINAW of 0.114. These results indicate that the proposed approach effectively balances predictive accuracy and uncertainty reliability, making it suitable for practical uncertainty-aware short-term load forecasting in smart grid applications.