Interpretable Deep Learning for Short-Term Electricity Load Forecasting with Calibrated Prediction Intervals

Authors: , , , , ,

Abstract: Accurate electricity load forecasting is vital for power grid management, market operations, and renewable energy integration. This paper presents a comprehensive study on multivariate short-term load forecasting using deep learning models—specifically Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks with attention variants. We introduce a data-driven pipeline encompassing extensive exploratory analysis, advanced feature engineering, and systematic model refinement guided by permutation feature importance. Furthermore, we investigate uncertainty quantification (UQ) via Deep Ensembles and Monte Carlo (MC) Dropout to derive reliable prediction intervals. The refined multivariate LSTM demonstrates consistent gains over baseline implementations and robust predictive capability across short-term horizons, including day-ahead (h=24). UQ analysis indicates superior calibration and coverage with MC Dropout relative to Deep Ensembles, offering actionable insight for risk-aware decision making in modern energy systems.

WSEAS Transactions on Power Systems, ISSN / E-ISSN: 1790-5060 / 2224-350X, Volume 20, 2025, Art. #30

PDF
DOI
XML
Cite
×
Citation Tools
WSEAS AMA
Hussein Ahmad Ahmad, Taha Benarbia, Seyyed Kasra Mortazavi, Mohamed El Bahnasawi, Mahmoud Hamed, Kyandoghere Kyamakya, "Interpretable Deep Learning for Short-Term Electricity Load Forecasting with Calibrated Prediction Intervals," WSEAS Transactions on Power Systems, vol. 20, pp. 380-393, 2025, DOI:10.37394/232016.2025.20.30
Hussein Ahmad Ahmad, Taha Benarbia, Seyyed Kasra Mortazavi, Mohamed El Bahnasawi, Mahmoud Hamed, Kyandoghere Kyamakya. Interpretable Deep Learning for Short-Term Electricity Load Forecasting with Calibrated Prediction Intervals. WSEAS Transactions on Power Systems. 2025;20:380-393. 10.37394/232016.2025.20.30
Copy to Clipboard
Citation copied to Clipboard

Certification

---