• Propose a hybrid VMD–SingEn–CNN–BO for short-term load forecasting. • VMD denoises; SingEn recalibrates; CNN learns; BO optimizes hyperparameters. • Achieves lowest MAE, MSE, RMSE, MAPE on VI and NSW datasets. • Outperforms LSTM, BI-LSTM, CNN, VMD-CNN baselines in accuracy. • Discusses accuracy–computation trade-off and offers tuning guidelines. Short-term load forecasting remains challenging because electricity demand is often noisy and nonstationary. This study proposes a hybrid VMD–SingEn–CNN–BO framework, in which Variational Mode Decomposition decomposes the load signal, Singular Entropy removes uninformative modes, and a compact convolutional neural network predicts the retained components. Bayesian optimization jointly tunes the decomposition, selection, and network hyperparameters. Experiments on half-hourly electricity load data from New South Wales and Victoria, Australia, from May 2009 to May 2014 show that the proposed model achieves the lowest MAPE among the compared baselines, reaching 1.17% and 2.31%, respectively. The statistical significance of the obtained improvements is further verified using the Wilcoxon signed-rank and the Friedman tests. The accuracy improvement comes at the expense of longer offline optimization time, while online forecasting remains efficient.
Nguyen et al. (Fri,) studied this question.