Title: Probabilistic Forecasting in Energy Markets

Date: Thursday, November 19, 2026

Time: 12:00 PM – 2:00 PM ET

Location: George Tower, Room 607, and virtual via Microsoft Teams

Meeting link: https://teams.microsoft.com/meet/251426656237229?p=Rtcf8xozrWXtD8QpJP

(Meeting ID: 251 426 656 237 229, Passcode: X7PR3MY3)

Moyi Guo

Ph.D. Candidate

H. Milton Stewart School of Industrial and Systems Engineering

Georgia Institute of Technology

Committee

1. Dr. Shijie Deng (Advisor), H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology
2. Dr. David Goldsman, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology
3. Dr. Valerie Thomas, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology
4. Dr. Shihao Yang, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology
5. Dr. Daniel Molzahn, School of Electrical and Computer Engineering, Georgia Institute of Technology
6. Dr. Aram Sogomonian, COO and Partner, GenBright LLC, an ENGIE company

Abstract

Electricity load and price forecasts underpin unit commitment, market bidding, and risk management, yet the series they target are strongly seasonal, heteroskedastic, and prone to extreme events, and operational decisions increasingly require not a single number but a calibrated description of uncertainty. This dissertation develops probabilistic forecasting methods for deregulated electricity markets in three research threads, progressing from point accuracy to calibrated uncertainty to joint scenarios across locations. The first thread proposes GARCHFormer, a Transformer-style encoder–decoder that maintains conditional-mean, conditional-variance, and fused latent paths at every layer, linked by a GARCH-inspired attention mechanism and a FiLM-based variance-to-mean fusion. Evaluated on public load and price datasets under both standard and capacity-controlled protocols, it delivers consistently strong point accuracy while learning time-varying volatility directly. The second thread extends this architecture to probabilistic forecasting with quantile, sigma-scaled, and Gaussian output modes and benchmarks it against a roster of deep learning models. Because raw prediction intervals from different models reach different coverage levels, it introduces a calibrate-then-compare protocol built on conformalized quantile regression, refined with per-horizon corrections and temporal copulas, and ranks forecasters by minimax regret across markets; under this rule the proposed model is the most robust choice. The third thread turns to multi-location renewable generation, developing a copula-based framework that produces spatially coherent solar and wind scenarios, improves on an industrial baseline, and stitches together scenario sets for wind speed and wind generation that come from separate forecast providers. Together, the results indicate that explicitly modeling volatility improves both point and probabilistic forecasts, that probabilistic forecasters should be compared only after they are calibrated, and that copula-based scenario generation turns marginal forecasts into the joint scenarios that market operations require.