Research On Short-term Load Forecasting Methods Based On Deep Learning

Accurate short-term power load forecasting is extremely important for ensuring stable operation of the power system,reducing power generation costs,and avoiding waste of resources.Many factors affect short-term electric load,and the superposition of these factors leads to it being non-linear and non-stationary.Decomposing different load components from the original load series for analysis by the signal decomposition method will help improve the accuracy of load prediction,and based on this,this paper mainly combines the signal decomposition method with the gated recurrent unit network(GRU)in deep learning to conduct related research on short-term power load forecasting.The main work of the paper is as follows:Firstly,the EMD-GRU short-term load forecasting model integrating external influence factors is proposed.The model can decompose a complex original load series into several sub-series containing different characteristics of the original load series through empirical mode decomposition(EMD).Then,for each sub-series decomposed,the correlation with external related factors,including temperature,date type,etc.,is analyzed,and the factors that are highly correlated with each sub-series are selected as input features for GRU prediction models for each sub-series separately.Finally the prediction results of each prediction model are superimposed to obtain the final prediction result.Then,the EMD-GRU load forecasting model with feature selection is proposed.Based on EMD decomposition of the original load series,Pearson correlation coefficient method of filter feature selection method is used to analyze the correlation between the decomposed sub-series and the original load series.Some sub-series with high correlation with the original load series are selected as features and input into the GRU network together with the original load series to establish the prediction model.By combining the feature selection method with the EMD method to construct a prediction model,multiple random errors introduced by directly modeling and predicting multiple sub-series obtained from the decomposition are avoided,which effectively reduces the overall complexity of the model and improves the overall prediction accuracy and speed.Finally,the short-term load forecasting model based on second decomposition strategy and time series clustering is proposed.The model first uses the method of Complete Ensemble Empirical Mode Decomposition with Adaptive Noise(CEEMDAN)to decompose the original load series into several subseries,and then for the problem of large prediction error of the high-frequency sub-series,calculate the volatility of each sub-series,selects the sub-series with high volatility to use the variational modal decomposition(VMD)method for secondary decomposition,which avoids the characteristic loss caused by the direct removal of the high frequency sub-series,and makes full use of the characteristics of the original load series.Secondly,the DTW-based AP clustering method is used to cluster all the subseries,so that the subseries with similar change trends are clustered together for reconstruction.Finally,to further improve the prediction accuracy,a double Gated Recurrent Unit(GRU)model with error correction is introduced to model and predict the series after cluster reconstruction.