Making large time series models work for urban load forecasting: A dual-stream large-lightweight hybrid framework with trend–fluctuation decomposition

Accurate short-term urban aggregate load forecasting ( SULF ) supports power system operation. At 15-minute resolution, urban load combines lower-variation temporal structure with rapid fluctuations associated with weather, calendar effects, and recurring operating schedules. Large time-series models (LTMs) provide reusable long-context representations. However, their rigid architectures make it difficult to model the joint effects of exogenous covariates and load composition heterogeneity in SULF, limiting cross-regional deployment. We propose CRD-DLF , a dual-stream framework that combines an LTM with a lightweight multivariate MLP-Mixer through trend–fluctuation decomposition. The LTM forecasts the lower-variation stream, while the MLP-Mixer models the rapidly varying stream. Within the latter, the Industrial Load Separator (IndSep) extracts a schedule-dependent, covariate-insensitive forecasting component associated with recurring industrial operating patterns in aggregate observations. The remaining fluctuations are modeled using meteorological and calendar covariates. Adaptive fusion combines the branch forecasts, and the Patch Structural Loss encourages structural consistency during training. Tests on the Lishui citywide and county-level datasets show that CRD-DLF improves forecasting accuracy and stability relative to the plain LTM and other state-of-the-art methods.