Lightweight Reservoir Fusion Architecture for Multi-Horizon Load Forecasting in Power Systems
Year: 2026
Type: article
Abstract: Reliable multi-horizon forecasting of electricity demand is vital for effective scheduling, planning, and operational decision-making in contemporary power systems. This paper introduces a computationally efficient framework for real-time forecasting using a Stacked Reservoir Computing Architecture (SRCA) with dynamic horizon fusion and ridge-regularized readout. The proposed framework captures nonlinear temporal interactions using hierarchically stacked reservoirs with fixed internal dynamics in order to reduce gradient-based training alongside significant pre-processing. To ensure consistency and temporal continuity across the forecast horizons, the model employs a dynamic horizon-fusion mechanism that adaptively blends the horizon-specific outputs generated by the ridge-regularized readout layer. Furthermore, the system includes raw-mode Variational Mode Decomposition (VMD) features that enhance signal diversity and representational depth without complicating the model. A comprehensive ablation study and statistical validation process further identified the optimal configuration awhich shows strong predictive performance, with MAPE values of 1.17% and 2.82% and R2 scores of 0.989 and 0.973, respectively, across two real-world electricity demand datasets. Comparative evaluation further shows that forecast-aware variational mode decomposition (FA-CVMD) reduces the average root mean square error (RMSE) by 2.13% and 37.33% compared with forecast-aware complete ensemble empirical mode decomposition (FA-CEMD) on the two datasets, respectively. The SRCA framework requires shorter training time, making it suitable for quick and adaptable forecasting. The results demonstrate that the suggested architecture is a suitable and computationally efficient multi-horizon load forecasting method for current power systems.
Source: Smart Grids
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7117147623
