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

Time series forecasting plays a critical role in modern power systems and supply chains, where accurate predictions enable efficient operations and resource allocation. This thesis advances the field through three novel methodological contributions. First, the thesis introduces the Bundle-Predict-Reconcile (BPR) framework for hierar- chical wind power forecasting. BPR integrates asset bundling with machine learning and forecast reconciliation techniques to improve prediction accuracy at both individual and system-wide levels. This approach demonstrates significant improvements over traditional methods when tested on large-scale industry data. Second, the dissertation develops a weather-informed probabilistic forecasting framework that combines Temporal Fusion Transformers with Gaussian copula methods. This approach effectively captures spatio-temporal dependencies in renewable energy systems while pro- viding robust uncertainty quantification. The framework shows superior performance in predicting load, wind, and solar power generation across the MISO system. Finally, this work presents LLMForecaster, a novel approach that leverages large language models to incorporate unstructured textual information into time series forecasts. Applied to retail demand prediction, LLMForecaster significantly improves forecast accuracy for prod- ucts with seasonal patterns by capturing domain knowledge from product descriptions and attributes. Together, these contributions advance the state-of-the-art in time series forecasting by effectively integrating hierarchical structures, probabilistic methods, and domain knowledge across different applications.

Date

2025-01-13

Resource Type

Text

Resource Subtype

Dissertation

Full item page