Electricity demand forecasting based on hybrid BiGRU-BiLSTM model prediction framework with dual time domain decomposition

In the context of the ongoing energy transition, there is a growing need for enhanced short-term flexibility in electricity demand. To address the complexity of nonlinear electricity demand data, this study introduces a hybrid prediction framework, namely the CEEMDAN-VMD-BiGRU-BiLSTM model. Initially, the CEEMDAN (Complete Ensemble Empirical Mode Decomposition with Adaptive Noise) method is employed to decompose the original data, characterized by significant fluctuations. Following this, the sample entropy of each decomposed component is computed and categorized into high-frequency, low-frequency, and trend sequences using K-means clustering based on the sample entropy values. For high-frequency sequences exhibiting substantial randomness, a further decomposition is performed using VMD (Variational Mode Decomposition). Subsequently, each component resulting from the dual decomposition process is predicted using a BiGRU-BiLSTM (Bidirectional Gated Recurrent Unit & Bidirectional Long Short-Term Memory) hybrid model. The final prediction outcomes for each sequence are synthesized through the BiGRU-BiLSTM model. Comparative experiments were conducted with baseline models, including BPNN (Back Propagation Neural Network), LSTM (Long Short-Term Memory), GRU (Gated Recurrent Unit), BiGRU (bidirectional gated recurrent unit), and the single decomposition CEEMDAN-BiGRU-BiLSTM framework. Evaluations based on six metrics— $\mathrm{R}^{{2}}$, MSE, MAE, RMSE, MAPE, and SMAPE—demonstrate the superior performance of the proposed hybrid model, achieving an ${\mathrm {R}}^{2}$ of ${99.22\%}$, significantly surpassing other models.