An Improved SVR Combination Model Based On VMD Research On Short-term Grid Load Forecdiction

The rapid economic development and the load fluctuations caused by the epidemic have brought great challenges to power load forecasting.Scientific and accurate power load prediction is very important to the stable operation of the power system,and accurate short-term power load prediction provides a reliable basis for power grid dispatching,operation and maintenance of equipment,start-stop and maintenance of generator sets.However,the change law of the power load itself is very complex,and it will be affected by the weather,temperature and humidity,holidays and other factors,which makes it difficult to predict the power load.In view of the above problems,this paper does the following research:In view of the problem of missing values and outliers of the power load itself,KNN was used to conduct missing value processing,and the Mahallidean distance was used to determine and correct the outliers,and the missing values and outlier values were effectively tested and corrected.In view of the influence of multiple factors,the MIC maximum information coefficient is used to quantitatively analyze the relationship between the various factors,determining the temperature as the key factor affecting the power load,and taking the periodic influence into account in the load prediction system,and constructing the index system of the influence factors of load prediction.To bring historical load data into load prediction,we use the GRU model with a historical memory function and a modified SVR model.first,Here,we construct both the unifactorial and multifactorial GRU models,For the parameters of the GRU model(learning rate,regularization coefficient,and number of hidden layer nodes),Using the seagull algorithm and the particle swarm algorithm for optimization,The SOA-GRU and PSO-GRU models were constructed,By example,the SOA-GRU model prediction effect is good;next,Where traditional SVR models are insensitive to historical time series,It is susceptible to the kernel function and the parameters(c and g),In this paper,by improving the traditional SVR model,Scroll the historical load backward,To study the influence of historical load input dimensions on prediction accuracy,For the model input parameters,Building the SSA-improved SVR and HPO-improved SVR models using the sparrow and hunter prey algorithms,respectively,After experimental verification,When the historical power load input dimension is 8,The improved SVR model can effectively improve the prediction accuracy;The rbf kernel function predicts better results for both the traditional SVR model and the improved SVR model;last,Compared with the SOA-GRU,PSO-GRU,SSA-improved SVR,HPO-improved SVR and other models,HPO-The improved SVR model has the best prediction performance.For the nonlinear and instability problems of the power load itself,we introduce the variational mode decomposition(VMD)algorithm into the short-term load prediction.VMD was used to decompose short-term loads at different frequencies,and to reduce the influence of internal differentiation,different components were predicted by HPO-improved SVR model and SOA-GRU model,respectively,and the predicted subsequences were reconstructed to obtain the predicted structure.The results show that the VMD-clustering-HPO-improved SVR model is compared to the HPO-improved SVR model without VMD-clustering Figure [46] Table [26] Reference [81]...