Neural ODE-Guided Imputation Computing for Building IoT Time Series with High Missing Gaps

Imputing missing data in Building Internet of Things (BIoT) poses a significant challenge due to the complex multi-dimensional variations and the inevitable occurrence of long-span missing gaps during maintenance or repair periods. In this paper, we introduce ODE-MI, a novel neural ODE-guided Multiple Imputation framework tailored for BIoT time series with high missing gaps. ODE-MI integrates shapelet-wise convolution, node embeddings, and dynamic graph embeddings with transfer delay to capture compound spatio-temporal (ST) volatilities, trends, and high-order impact transfer. Our contributions include the design of a multi-faceted backbone network capable of capturing long-distance ST characteristics in BIoT, and the development of the ODE-MI framework to handle long-term missing data within networked dynamic attributes. Empirical assessments using open datasets demonstrate that ODE-MI outperforms existing imputation methods, especially in scenarios involving continuous long-span missing gaps.