An Adaptive Transfer Learning Framework for Data-Scarce HVAC Power Consumption Forecasting

Abstract

Heating, ventilation, and air conditioning (HVAC) systems constitute a large proportion of building energy consumption and provide considerable potential for power grid regulation. While the HVAC power consumption forecasting task is generally straightforward with sufficient historical data, it becomes challenging when dealing with scarce data. Such situation is common in cases of intermittent data collection or early system implementations, where precise forecasting is required despite limited data available. Considering accessible datasets from nearby or similar HVAC systems through energy management systems, this paper proposes an adaptive transfer learning framework to tackle this issue. Specifically, the framework leverages diverse source domains, employing model-level regularizers to quantify domain discrepancies and an adaptive parameter regulation mechanism to dynamically align source domains with the target domain. Embedded within the framework, a unique deep learning architecture with attention mechanisms is proposed, capable of identifying complex temporal patterns and hierarchical features in HVAC systems. Experiments on public HVAC datasets demonstrate the generalization, accuracy and robustness of our methodology under diverse data-scarce scenarios.