Short term power load forecasting plays a crucial role in maintaining stable operation of the power system and optimizing resource allocation. A new short-term power load forecasting model was proposed to address the complexity and randomness of power load data, as well as the low accuracy of existing forecasting models. This model included an improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) and permutation entropy (PE) reconstruction part, as well as a bidirectional long short-term memory (BiLSTM) and kernel extreme learning machine (KELM) prediction part. Firstly, ICEEMDAN was used to decompose complex load data into n relatively stable subsequences, thereby reducing the randomness of the data. Permutation entropy was introduced to calculate the PE value of each subsequence for reconstruction, effectively reducing the computational scale of the model. Secondly, the BiLSTM model was used to explore the intrinsic connections between data, and to learn and predict various reconstructed sequences. Finally, KELM was used to perform non-linear fitting on the predicted values of the reconstructed sequence, further improve prediction accuracy. The experimental results show that the ICEEMDAN-PE-BiLSTM-KELM model reduces root mean square error (RMSE) by 106.05 MW, mean absolute error (MAE) by 62.34 MW, and mean absolute percentage error (MAPE) by 0.877% compared to traditional long short-term memory (LSTM) models. It can be seen that this model can better solve the complexity and randomness of data while improving prediction accuracy.