DRCNet : Decoupled residual correction network for multivariate time series forecasting

Yao, Bin and Wang, Tie and Bilal, Muhammad (2026) DRCNet : Decoupled residual correction network for multivariate time series forecasting. International Journal of Intelligent Networks, 7. pp. 19-30. ISSN 2666-6030

Abstract

Accurate prediction of solar radiation energy can precisely assess the power generation capacity of photovoltaic systems, thereby optimizing grid dispatch and ensuring the stability of power supply. The prevailing methodology employs temporal models or spatiotemporal graph neural networks, which are used to extract the meteorological dependencies for solar radiation or conduct site-specific modeling. However, these studies overlearn the geographical features of the training domain, and lack the extrapolation continuity of unknown coordinates in space. To resolve these problems, this study constructs a decoupled residual correction network, which separates the region-invariant time series rules from geographically specific deviations through a dual-path parallel network. The first path uses a frequency-embedded encoder for general time series prediction based on multivariate variables. The second path uses an independent network to predict amplitude and vertical offsets from geographical features and residual features. Cross-regional zero-shot extrapolation verification utilized data from 12 sites (11 source domains and 1 target domain). DRCNet achieved the best performance on all four datasets, with the average absolute error (MAE) reduced to below 0.1 (specifically ranging from 0.062 to 0.095), significantly outperforming the second-best model Crossformer (0.36). Our codes will be released at https://github.com/Phoenix-Snow/lab-explore-DRCNet.