A Short-term Electricity Load Forecasting Model:CEEMDAN-SE-VMD+SelfAttention-TCN Fusion Model

Abstract Under the increasing electricity consumption trend and complex power consumption forms, accurate power load forecasting faces severe challenges. This paper proposes the CEEMDAN-SE-VMD + SelfAttention-TCN Fusion model. Initially, the original power load data undergo decomposition using CEEMDAN to generate Intrinsic Mode Functions (IMFs). These IMFs are then adaptively differentiated into high- and low-frequency components using Sample Entropy (SE), with high-frequency components further extracted twice via VMD. Subsequently, the original load data are reconstructed into sequences of high and low-frequency components. These sequences are input into TCN enhanced by self-attention for prediction, integrating results into final predictions. Using historical power load data from ENTSO Germany, an example achieves RMSE 24.1293, MAE 17.268, and R2 0.9838. Compared to SelfAttention-TCN without reconstruction, RMSE decreases by 25.5326, MAE 19.3769, R2 increases by 0.0515. The example illustrates that signal decomposition and reconstruction effectively capture latent deep features in load data, aiding model learning. Self-attention enhances TCN's ability to learn complex dependencies, improving prediction accuracy. This model demonstrates high tolerance to raw load data and strong performance in short-term power load forecasting tasks.