The power load series is characterized by volatility, complexity, and fuzziness, which brings challenges to the forecast. To fully mine the important information in the sequence and reduce the complexity and volatility of the sequence, a two-stage prediction model and method of quadratic CEEMDAN-SE-Transformer-BiLSTM combined with error correction is proposed in this paper. In the first stage, Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) is used to decompose the power load data. To reduce sequence complexity and volatility. The component is reconstructed according to the sample entropy (SE) value. The high-frequency components after reconstruction are decomposed and reconstructed again to fully reduce the sequence complexity. The Transformer model is used to forecast each component, and the predicted values are combined to obtain the final forecast data. In the second stage, error distribution data is obtained based on the error between the predicted value and the actual value, and the Bidirectional Long Short Term Memory (BiLSTM) network model is used to train and predict the error distribution data. The error-corrected load value is obtained by combining the error-predicted value with the load-predicted value. Compared with several traditional models, the results show that the prediction index R2 of the secondary CEEMDAN-SE-Transformer-BiLSTM load forecasting model established in this paper is 85.91%, MAPE is 0.19%, and RMSE is 42.1861 kW, which is more accurate and effective than other traditional forecasting models.