ICEEMDAN- and LSTM-Enhanced Hybrid Cosine-Attention iTransformer for Ultra-Short-Term Load Forecasting

Accurate power system load forecasting is the core prerequisite for guaranteeing the safe and stable operation of power grids and supporting the efficient scheduling of power systems. To improve the accuracy of load forecasting and portray the non-stationarity and multi-scale characteristics of the load sequence, this paper proposes a short-term load forecasting method based on ICEEMDAN decomposition-LSTM feature extraction-hybrid cosine attention mechanism iTransformer. Firstly, the original load sequence is decomposed using the Improved Complete Ensemble Empirical Modal Decomposition (ICEEMDAN) to extract the intrinsic modal function (IMF) and residuals (Res), and the multidimensional input feature set is constructed by combining exogenous variables such as meteorology. Secondly, multi-source features were extracted using the Long Short-Term Memory (LSTM) network to capture the complex nonlinear correlations and long-term dependencies. Finally, the extracted features are input into the iTransformer model that introduces the hybrid cosine attention mechanism. The hidden feature representation is obtained through the encoder layer modeling, and the linear mapping in the output layer generates the load forecast value. The results show that the prediction method proposed in this paper achieves better performance and can effectively improve the accuracy of short-term load prediction, which provides an effective technical support for the short-term scheduling and flexible operation of the power system.