An AI-driven hybrid approach for intelligent load forecasting

Abstract

Accurate load forecasting is essential for the intelligent operation of modern power systems, where efficient energy management directly impacts grid stability and sustainability. However, the inherent challenges posed by load series, including nonlinearity and complex temporal dependencies, often compromise the performance of conventional forecasting methods. To address these issues within the context of AI-driven design, this study introduces a hybrid forecasting approach under a decomposition-and-ensemble framework that integrates variational mode decomposition (VMD), gated recurrent unit network (GRU), and Harris Hawks optimization (HHO). First, the VMD first decomposes the complex raw time series into multiple subsequences with constrained bandwidths and distinct center frequencies, simplifying the underlying evolutionary patterns to be learned. Second, the GRU, as an effective variant of recurrent neural networks, independently captures both short- and long-term temporal dependencies within each subsequence via its gating mechanism. Finally, the HHO estimates optimal weights for nonlinear integration of individual forecasting results. The proposed method is evaluated on four kinds of electricity load datasets from Ireland, using four quantitative metrics along with the complexity-invariant distance (CID). Experimental results demonstrate superior performance over multiple single and hybrid models in one-step-ahead forecasting. Overall, this work presents an AI-driven solution for intelligent load forecasting, thereby contributing to the advancement of smart energy management in power systems.