Short-Term Electricity Load Forecasting in Power Systems Using BO-VMD-STGCN

Abstract

To address the strong nonlinearity, multi-fluctuation coupling, and spatio-temporal correlation characteristics inherent in short-term power load, a forecasting model based on BO-VMD-STGCN (bayesian optimization, variational mode decomposition and spatio-temporal graph convolutional networks) is proposed. Firstly, BO is utilized for the adaptive tuning of key parameters in VMD, thereby achieving an optimal multi-scale decomposition of the original load sequence and mitigating data non-stationarity. Subsequently, the decomposed modal components along with the reconstructed sequences are fed into the STGCN, wherein a stacking mechanism with residual connections is incorporated to synergistically extract temporal dependencies and spatial correlations via spatio-temporal convolutions. Experimental results demonstrate that the proposed model achieves an MAPE, RMSE, and MAE of 2.84%, 62.68, and 38.36, respectively, along with an R2 of 0.92. Compared with LSTM, STGCN, STM-GCN, and GCN-LSTM, the R2 metric is improved by 26.39%, 19.73%, 18.42%, and 35.29%, respectively. Exhibiting marked superiority over benchmark approaches in both goodness-of-fit and predictive fidelity, the proposed model underscores its intrinsic strengths in jointly capturing spatio-temporal dependencies and adaptively disentangling non-stationary temporal dynamics. It thus establishes a reliable, accurate, and interpretable paradigm for short-term electricity load forecasting.