Research On Hybrid Load Forecasting Methods Of Fused Magnesium Based On Decomposition And Reconstruction

As a kind of refractory materials,electro-fused magnesium is widely used in metallurgy,building materials,aerospace,military industry and other fields.However,the power consumption of electro-fused magnesium is huge and the load fluctuates sharply.Accurately predicting the changes of load can implement peak cut,which can not only reduce the cost of electricity for enterprises and improve economic efficiency,but also save power supply capacity and realize energy saving and consumption reduction,so it is of great importance of research and application.Affected by the dynamic shift of operating modes in the electro-fused magnesium smelting process,the fluctuation of electricity load has the characteristics of strong non-linearity,nonstationary,small time granularity,etc.,which increases the difficulty of load prediction.To address this problem,this paper studies a hybrid prediction method for electro-fused magnesium load based on decomposition-reconstruction,which improves the robustness and accuracy of prediction by sequence decomposition and heterogeneous reconstruction,and further extends to multi-step hybrid prediction method for electro-fused magnesium load by combining with error correction strategy.The specific research work includes:First,the fluctuation characteristics of electro-fused magnesium load are studied,and a single model prediction method of fused magnesium load is designed.Combined with the mechanism analysis,the electro-fused magnesium load data is preprocessed,and the abnormal data and redundant information are removed.The third-order lag features of 57 attributes affecting the load are extracted,and the feature engineering is constructed using Pearson correlation coefficient.Six representative models are selected to predict loading for electro-fused magnesium,such as AutoRegressive Moving Average Model（ARMA）,Random Forest（RF）,Support Vector Machine（SVM）,e Xtreme Gradient Boosting（XGBoost）,Light Gradient Boosting Machine（Light GBM）and Long Short-Term Memory Network（LSTM）.A comparative experimental study is carried out using industrial data,and it is found that SVM and ARMA have the worst prediction among all models.In the case of simple features and small sample size,RF performs the best metrics compared with other models,while XGBoost and Light GBM have limited advantages and average performance.Secondly,for the strong non-linearity,non-stationary and small time granularity of load fluctuation,a hybrid predicting method based on decomposition-reconstruction for electro-fused magnesium load is designed.The original sequence is decomposed into high-frequency and lowfrequency components using Discrete Wavelet Transform（DWT）based on energy entropy optimization.Through the experiment,it is found that the decomposition-reconstruction greatly improves the prediction precision.Taking DWT-RF and DWT-XGBoost as examples,RMSE,MAE and MSE are reduced by about one-third compared with RF and XGBoost,and MAPE is reduced from 4.41% and 4.55% to 1.432% and 1.501%,and the ",EVS,and PCCs are all increased greatly to more than 0.93.In addition,a sequence approximate feature mining model based on RF is designed for low-frequency components,and a sequence detail feature mining model based on LSTM for high-frequency components,and the heterogeneous hybrid model DWT-RFLSTM is reconstructed to predict the electricity load change of electro-fused magnesium.Through experimental comparison,it can be found that the heterogeneous hybrid model DWT-RF-LSTM has the best robustness and prediction effect compared with DWT-RF,DWT-XGBoost and DWTLGBM.Thirdly,a hybrid prediction method of electro-fused magnesium load based on Complete Ensemble Empirical Mode Decomposition with Adaptive Noise（CEEMDAN）and Variational Mode Decomposition（VMD）is designed.For the problem of mode mixing in CEEMDAN decomposition,the subsequence complexity is analyzed by sample entropy,and the sequences are recombined with similar entropy values.The experimental results show that the recombined CEEMDAN decomposition based on sample entropy not only reduces the computational cost but also the error significantly.In order to improve the VMD decomposition effect,a self-adaptive penalty factor optimization algorithm based on similarity is designed to set the VMD parameters,and the Sparrow Search Algorithm（SSA）is introduced to optimize the LSTM parameters.Through experimental comparison,the designed model VMD-RF-SSA-LSTM has the best prediction precision than other hybrid models in electro-fused magnesium load prediction.Fourth,combined with decomposition prediction and error correction strategy,a multi-step method for electro-fused magnesium load prediction is designed.This method first mines load features from the predicted load subsequence,then uses Principal Component Analysis（PCA）to extract the principal components,and uses the principal components without interference and prediction errors as training datasets to construct RF models for error correction.Taking the 5-step backward prediction of electro-fused magnesium load as an example,the experimental comparison shows that after the error correction strategy’s improvement,the multi-step prediction precision of12 hybrid models based on three decomposition strategies has been greatly improved.Among them,VMD-RF-LSTM-PCA-E has the best performance,and the MAPE of 1st step,2nd step,3rd step,4th step and 5th step during 5-step backward prediction is 1.4%,1.67%,1.98%,2.24% and 2.26%,respectively,which verifies the effectiveness of the multi-step prediction method for electro-fused magnesium load.