Research On Hybrid Model Of Short-term Power Load Forecasting Based On Ceemdan Decomposition

Power load forecasting is an important part of power system decision-making and planning.Accurate and reliable load forecasting results can ensure the rationality of power system regulation and power supply stability,realize effective dispatch of power supply and demand,promote energy conservation and emission reduction,and improve economic benefits,has important practical significance for the sustainable and stable development of the power system and society.In this paper,a complete set of adaptive noise empirical mode decomposition(Complete Ensemble Empirical Mode Decomposition with Adaptive Noise,CEEMDAN),Long Short Term Memory(Long Short Term Memory,LSTM)and Multiple Linear Regression(MLR)as basically,considering the influencing factors of load,the research on short-term power load forecasting is carried out.The main work is as follows:(1)Aiming at the problem that the prediction accuracy of the LSTM network is easily affected by the randomness of the initialization parameters,particle swarm optimization(PSO)is used to optimize the network parameters of the LSTM model.Combining the Pearson correlation coefficient to analyze the correlation between the load and the impact factor,the impact factor with strong correlation is selected as the input of the prediction model,which reduces the input dimension of the prediction model.Experimental results show that the proposed prediction method significantly improves the prediction accuracy of the LSTM neural network.(2)Aiming at the volatility and randomness of the power load sequence,the CEEMDAN algorithm is used to decompose the nonlinear and non-stationary load sequence into several sub-components containing local features,combined with the advantages of LSTM to deal with the long-term non-stationary component prediction problem and MLR processing the stationary component realizes the characteristics of rapid prediction while ensuring accuracy.This paper proposes a hybrid prediction model based on CEEMDAN-LSTM-MLR.The zero-crossing rate is used to divide each sub-component into stationary and non-stationary components,which are respectively predicted by MLR and LSTM models.Aiming at the different characteristics of the feature trends of the sub-components,the Pearson correlation coefficient is used again to analyze the correlation between the selected influence factor and each sub-component,and the feature input with high correlation is constructed.The experimental results show that the maximum daily average absolute percentage error of the hybrid model CEEMDAN-LSTM-MLR within a week does not exceed 2%;compared to the combined model in which all sub-components are predicted by the LSTM network,the hybrid model is ensuring the prediction accuracy at the same time,it speeds up the prediction speed of the model.(3)On the basis of(2),a CCEEMDAN-PSO-LSTM-MLR hybrid prediction model based on component recombination is proposed(component recombination is denoted as CCEEMDAN).Combining similar sub-components according to the sample entropy value effectively reduces the complexity of the prediction model.The experimental results show that the model has a significant improvement in both the prediction accuracy and the prediction speed.