A Hybrid CNN-GRU-Attention Network with Adaptive Data Decomposition for Electricity Demand Forecasting

Electricity demand forecasting is critical for efficient power system planning and operation. Nevertheless, power data frequently demonstrate pronounced non-stationarity owing to meteorological variability, holiday effects and economic fluctuations. To address multi-dimensional data challenges, this paper proposes a decomposition-integration framework based on Variational Mode Decomposition (VMD) and a hybrid neural network. The Zebra Optimization Algorithm (ZOA) adaptively tunes VMD parameters, decomposing data into clear sub-sequences. A Convolutional Neural Network-Gated Recurrent Unit model then predicts each sub-sequence in parallel, with the Convolutional Neural Network (CNN) capturing spatial features (e.g., weather indicators) and Gated Recurrent Unit (GRU) handling long-term dependencies. A multi-channel attention mechanism dynamically weights critical features, enhancing both forecasting accuracy and interpretability. Experimental results demonstrate significant error reduction and improved data decoupling, enabling optimal power resource allocation.