This paper preprocesses missing and anomalous power load data to reduce the error of power load forecasting. Loess smoothing-based time series decomposition algorithm (STL) is introduced to initially decompose the data into trend component, period component and residual component according to the inner and outer loop process. Aiming at the noise limitation of the residual component, the method of reconstruction is carried out by improving the STL-ICEEMDAN quadratic decomposition. Combined with sample entropy and maximum information number, it is quadratically divided into high-complexity component and low-complexity component to improve the prediction accuracy. The results show that the trend component autocorrelation coefficients are 0.786 and 0.729, the correlation is high, there is periodicity, and the decomposition method is effective. The root mean square error, average absolute error, average absolute percentage error, and relative absolute error of power load prediction of this paper’s model are 91.66kW, 77.91kW, 0.88%, and 0.98%, respectively, and the prediction error is smaller than the comparison model.
