Title: Advancing Time Series Forecasting: Hierarchical Methods, Probabilistic Models, and Domain Knowledge Integration from Power Systems to Retail

Date: Dec 17th, 2024

Time: 8 AM - 10 AM

Location: CODA 1215, meeting link

Hanyu Zhang 

Machine Learning PhD Student

H. Milton Stewart School of Industrial and Systems Engineering

Georgia Institute of Technology

Committee: 

Dr. Pascal Van Hentenryck (Advisor), H. Milton Stewart School of Industrial and Systems Engineering 

Dr. Yao Xie, H. Milton Stewart School of Industrial and Systems Engineering 

Dr. Siva Theja Maguluri, H. Milton Stewart School of Industrial and Systems Engineering 

Dr. B. Aditya Prakash, School of Computational Science and Engineering 

Dr. Terrence Mak, Department of Data Science & AI, Monash University

Abstract: This thesis advances time series forecasting through three novel methodological contributions addressing key challenges in modern forecasting applications. First, it introduces the Bundle-Predict-Reconcile (BPR) framework, which improves hierarchical wind power forecasting by learning optimal groupings of related time series while maintaining forecast consistency across different levels. Second, it develops a weather-informed probabilistic forecasting framework that combines Temporal Fusion Transformers with Gaussian copula methods to capture spatio-temporal dependencies in renewable energy systems. Finally, it presents LLMForecaster, an innovative approach that leverages large language models to incorporate unstructured textual information into time series forecasts, significantly improving accuracy for products with seasonal patterns. The proposed methods are extensively validated on real-world datasets from power systems and retail domains, demonstrating substantial improvements over existing approaches in forecast accuracy and uncertainty quantification.