Globalization for Scalable Short-term Forecasting of Heterogeneous Loads

Power systems are structured hierarchically, emphasizing the need for multi-step load forecasting not only at the system level but also at primary and secondary substations, area levels, and individual Points of Delivery (PoDs). This necessitates scalable forecasting techniques that can handle numerous measuring points with diverse characteristics while ensuring timely and accurate predictions. While intuitive and locally accurate, traditional local forecasting models (LFMs) struggle with scalability, becoming computationally expensive and less efficient as network size and data volume grow. In contrast, global forecasting models (GFMs) enhance generalizability, scalability, and robustness through globalization and cross-learning. However, GFMs assume that input time series are inherently related, overlooking potential spatiotemporal data heterogeneity. Heterogeneity in time series refers to diverse characteristics across different series or within a single series over time, which can bias forecasts and reduce generalizability if unaddressed. Data heterogeneity can be broadly categorized into spatial and temporal heterogeneity. Spatial heterogeneity arises from structural and behavioral differences across geographical areas, customer types (e.g., residential, commercial, industrial), and grid levels (e.g., system-level, substations, PoDs). Temporal heterogeneity, on the other hand, refers to changes within a single time series over time due to evolving user behavior, technological upgrades, or external disruptions such as extreme weather, wildfires, or policy interventions, which can manifest as data drift or concept drift. This thesis investigates globalization in the presence of data heterogeneity in power systems. First, a method is proposed to identify and segment price-responsive customers as an expression of temporal heterogeneity. It then examines global load forecasting under data drifts, comparing feature-transforming and target-transforming models to reveal how globalization, heterogeneity, and drift influence performance. The role of globalization in peak load and zero-shot forecasting is also explored. To address spatial heterogeneity and balance globality with locality, we propose time series clustering (TSC) methods, model-based TSC for feature-transforming models, and weighted instance-based TSC for target-transforming ones. Finally, we study globalization in multistep and probabilistic settings, emphasizing the importance of globalization and tackling temporal heterogeneity and sample indistinguishability. We propose a regime-aware global forecasting framework with temporal embeddings to address the challenges posed by temporal heterogeneity, while introducing a temporal globalization strategy to capture both long-term trends and short-term dynamics. Extensive experiments on real-world datasets validate the effectiveness of the proposed approaches.