Short-Term Load Interval Prediction Based on DBO-Optimized Variational Mode Decomposition and CNN-BiLSTM-Attention

To address the challenges in short-term power load forecasting posed by the non-stationary, multi-scale, and nonlinear characteristics of load data, this study proposes a short-term load interval forecasting method based on a Convolutional Neural Network-Bidirectional Long Short-Term Memory-Attention (CNN-BiLSTM-Attention) model, where Variational Mode Decomposition (VMD) is optimized by the Dung Beetle Optimizer (DBO). First, the Pearson correlation coefficient is applied to select four input features strongly correlated with load demand: visibility, temperature, atmospheric pressure, and historical load demand. Second, with minimum envelope entropy adopted as the fitness function, the DBO adaptively searches for the optimal number of modes and penalty factor of VMD, which effectively mitigates the non-stationarity of the original load time series by separating it into components of different scales. Next, a hybrid CNN-BiLSTM-Attention forecasting model is constructed. The CNN module extracts local spatiotemporal features, the BiLSTM module captures bidirectional long-term temporal dependencies, and the attention mechanism dynamically assigns weights to key information representations, collectively enabling high-precision point and interval forecasting of load demand. The results show that the proposed model achieves an MAE of 0.0893 MW, an RMSE of 0.1156 MW and an R2 of 0.9618, which are 17.0% and 17.3% lower and 1.17% higher than those of the best-performing benchmark model, while the interval coverage reaches 96.9% with the narrowest average width of 22.5% and the lowest interval RMSE of 0.231 MW.