<p>Power load forecasting is a core task of power system management, and its accuracy directly affects energy dispatch efficiency, operation cost optimization, and grid stability maintenance. In order to improve the performance of load forecasting, this paper proposes a hybrid model (CEEMDAN-SE-xLSTM-Transformer), which integrates a fully integrated empirical modal decomposition with the adaptive noise algorithm (CEEMDAN) and the sample entropy method (SE) to improve the xLSTM-Transformer architecture. Specifically, CEEMDAN-SE is first utilized to decompose the load sequence into multiple power load subsequences with significant differences in complexity. Then, to address the limitations of the traditional transformer in long-distance dependency capture and local temporal dynamics modeling, the xLSTM-Transformer module is designed to enhance the local feature interactions through a gating mechanism with memory units and capture the global dependencies using multi-head attention. Finally, the forecast results of each IMF component are integrated to reconstruct the final load forecast value. The study is conducted based on empirical evidence from a regional grid in Australia, and compared with benchmark models such as LSTM, XGBoost, standard transformer, and CEEMDAN-Transformer. The results show that the MAPE and RMSE of the model in this paper are 0.84% and 86.37, respectively, which are better than the selected other models. The proposed method provides a more robust solution for load forecasting in high-volatility scenarios through the co-optimization of signal decomposition and hybrid architecture.</p>