A high-temporal-resolution interpolation framework for time-series downscaling based on time-domain decomposition and deep learning: A case study of building occupancy data

Abstract

With the rapid development of renewable energy, the coordinated operation of demand response between buildings and the power grid has become increasingly important. This requires the support of high-temporal-resolution building-side data. However, basic building-side data are usually collected or simulated at hourly intervals, whereas grid-side demand response often requires data at 15-min intervals. This leads to a temporal mismatch between building data and grid response strategies. Therefore, this research proposes a time-series downscaling framework based on discrete wavelet transform (DWT), convolutional neural network (CNN), and TimesNet, referred to as DWT-TimesNet, which aims to generate 15-min high-temporal-resolution instantaneous data from 1-h low-temporal-resolution instantaneous observations. In the proposed framework, DWT is first used to decompose the original time series at different time scales. CNN and TimesNet models are then used to generate the corresponding components, which are then reconstructed into 15-min high-temporal-resolution data. Building occupancy data are used as a case study. The cross-validation for repeated random splits results show that, under the premise of reliable peak timing characterization, the proposed DWT-TimesNet framework significantly improves the distributional consistency of peak magnitude and volatility factor compared with linear interpolation. The median Kolmogorov–Smirnov (KS) statistic of peak magnitude is reduced from 0.0714 to 0.0275, and the median KS statistic of volatility factor is reduced from 0.6016 to 0.0934. These results show that the DWT-TimesNet framework can better preserve peak-related and volatility-related characteristics in high-temporal-resolution building data. The single-zone cooling-load case also suggests that the generated occupancy series may be useful as an input for future building load and demand response analysis.