AI-driven short-term load forecasting enhanced by clustering in multi-type university buildings: Insights across building types and pandemic phases

Abstract

Accurate forecasting of electricity demand is pivotal for optimizing energy management in smart buildings and propelling the advancement towards net-zero goals. However, scalable, accurate, and robust building load prediction models are scarce in educational campuses, whose diverse functions and consumption patterns, resembling large-scale urban environments, are often hindered by data scarcity and lack of cross-building heterogeneity analysis. This study presents an AI-based short-term building load forecasting framework integrating K-means clustering with BiLSTM regression, tailored for various energy consumption patterns across different building types. The model requires no auxiliary variables and is adaptable to both regular operational conditions and disruptions caused by the COVID-19 pandemic. The clustering-enhanced model adeptly identifies unique energy consumption patterns and significantly improves prediction accuracy, with a 3.65 % increase in mean R2 and a 55.19 % reduction in standard deviation under normal conditions. During the pandemic, its performance is further amplified, with a 4.90 % increase in mean R2 and a 62.41 % reduction in standard deviation, highlighting its robustness. The model shows particularly high accuracy in buildings with consistent energy profiles, such as teaching and research facilities, while it encounters greater challenges in mixed-use and office buildings due to their variable energy patterns. The pandemic underscores the model's limitations in adapting to abrupt operational shifts, signaling a pressing need for future enhancements to incorporate adaptive forecasting techniques. This research substantiates the application of AI in building energy systems, contributing to the development of nearly zero-energy buildings (NZEB) and supporting the transition towards sustainable urban energy systems.