Multi-strategy cooperative prediction model for short-term electric load based on Ultra-NOA-CNN-DLKA-GRU
To address the problems of insufficient nonlinear feature extraction, imbalanced long- and short-term dependency modeling, and limited prediction accuracy caused by hyperparameter sensitivity in traditional short-term electric load forecasting, this paper proposes a hybrid forecasting method for short-term electric load based on Convolutional Neural Network (CNN), Deep Large Kernel Attention (DLKA), and Gated Recurrent Unit (GRU), and designs an improved Nutcracker Optimization Algorithm (NOA), termed Ultra-NOA, for hyperparameter optimization. First, in view of the strong nonlinearity, significant fluctuations, and multiscale variations of load series, CNN is employed to extract local patterns and high-frequency fluctuation features, thereby enhancing the perception of peak-valley switching and abrupt changes. Subsequently, a DLKA module is introduced to adaptively weight and screen key features, improving the ability to focus on informative signals while suppressing noise interference. On this basis, GRU is used to model temporal correlations and jointly characterize long-term trends and short-term disturbances. Furthermore, to reduce the uncertainty of manual parameter tuning, an improved NOA algorithm is constructed to adaptively search for key hyperparameters, thereby improving training stability and parameter matching. Experimental results show that the proposed model achieves strong predictive performance on a real load dataset, with a test-set R² of 0.922 and an RMSE of 238.855. Ablation experiments verify the effectiveness of the CNN, DLKA, and GRU modules. Comparative experiments demonstrate that Ultra-NOA-CNN-DLKA-GRU outperforms multiple baseline models in both error metrics and fitting performance, with R² improved by approximately 8.5% to 17.6% and RMSE reduced by approximately 27.9% to 39.8%. The results indicate that the proposed method can effectively improve the accuracy, robustness, and generalization ability of shortterm electric load forecasting.