Short-term Load Forecasting Based On Variable Weighted Synthesis Of Multiple Space Support Vector Machines

As an important part of power system,power load forecasting provides a basis for power grid planning and dispatching.Accurate short-term power load forecasting can not only save resources,reduce economic costs,but also ensure the safe operation of the power grid.Aiming at the problems of low accuracy and low stability in short-term load forecasting,this paper proposes four new methods.First,short-term load forecasting method based on similar historical situation screening improved SVM.Aiming at the shortage that the similarity-day method is too large to fully excavate and utilize the historical data,and the universality is not high,a more general concept of similar historical situation is introduced,and the forecasting method of improved support vector machine based on similar historical situation screening is designed.Firstly,the correlation analysis method is used to analyze the power load at the historical time to determine the candidate set of forecasting features.Furthermore,the forecasting phase space is constructed by means of the input-output relational expansion of the historical data of time series.Then,with the help of situation similarity analysis,the historical situation of similarity was screened and the structure of training data was optimized.Finally,the SVM with Gaussian kernel function was used to conduct learning modeling on the training data after structural optimization,so as to realize the power load forecasting based on the historical similar situation.The results of the example analysis show that compared with the traditional SVM method,the improved SVM method of similar historical situation screening improves the prediction accuracy and stability by 0.493 and 0.268 percentage points respectively.In particular,the training data organization method of similar historical situation screening significantly improved the forecasting performance of decision tree,partial least squares and other forecasting methods,with an increase of about 0.1 to 2 percentage points.Second,short-term load forecasting model based on variable weight synthesis kernel function improving similar situation SVM.Aiming at the problem that the prediction accuracy and stability of single kernel function SVM can not meet the application demand,the variable weight synthetic kernel function is designed,and the forecasting method that the synthetic kernel function improving similar situation SVM is given.This method first analyzes the features of Gaussian,polynomial and other kernel functions,and then uses similar scenario sets to train SVM with different kernel functions.Secondly,the forecast validity index is introduced,and the simulation experiment is designed to test the prediction performance of the model.Finally,according to the test results,the kernel function variable weight is constructed,and with the help of the kernel function variable weight synthesis,the SVM with higher prediction performance and stronger stability is constructed.The case study showed that the method was more accurate and stable than similar historical situation screening improved SVM.Third,short-term load forecasting method based on variable weighted synthesis multiple space Gaussian kernel support vector machines.Aiming at the problem that the prediction performance of single forecasting method is easy to fluctuate and its accuracy and stability are difficult to guarantee,the concept of forecasting phase space is introduced,and the forecasting method based on variable weight synthesis multiple space Gaussian kernel SVM is presented.The method firstly selected the prediction features by weighted correlation degree,and then randomly selected some historical loads with high correlation degree as input to establish a series of forecasting phase space.Secondly,SVM was used to build the forecast models in each forecasting phase space.Then,by using re-sampling technology to design the simulation experiments to test the forecast performance of the sub-models,and the weight of each model were determined based on the performance test results.Finally,load forecasting was obtained by variable weighted synthesis of multiple phase space forecasting models.The analysis of load forecasting example showed that the ensemble forecasting method based on multiple forecasting phase space modeling and variable weight synthesis can significantly improve the prediction stability while ensuring the prediction accuracy.Fourth,short-term load forecasting model based on variable weight synthesis multiple space similar situation improved SVM.This method tries to improve the prediction performance by applying the similar situation screening and variable weight synthesis multiple space forecasting model.Firstly,multiple forecasting phase spaces were constructed,and similar historical scenarios were screened in each forecasting phase space.Using similar situation improving SVM method to learn a series of spatial sub-models,and designing simulation experiments to test their performance.Secondly,the similarity between the projection of the situation to be predicted in different forecasting phase space and the similar historical situation used in modeling in this space is calculated,and the spatial sub-model’s constant weight is designed based on the similarity,and the state adjustment weight is designed based on the spatial sub-model’s forecasting performance state.Finally,the variable weight weight of the spatial sub-model is obtained by superimposing the similarity constant weight and the predicted performance state adjustment weight,and the synthesized short-term load forecasting is realized by means of multiple space variable weight synthesis.Two independent simulation experiments are designed to test the prediction performance of the model,and compared with the current methods.The experimental results show that this method has higher accuracy and stronger stability than other commonly used forecasting methods,such as variable weight synthesis kernel function improving similar situation SVM,variable weight synthesis multiple space Gaussian kernel SVM,similar historical situation screening improving SVM,random forest,ensemble decision tree,decision tree and partial least square.In summary,similar situation screening and variable weight synthesis multiple spaceforecasting models are both effective methods to improve model prediction performance and have a certain universality.Among them,the integration of multiple models significantly improves the stability of prediction.However,this is not endless.As the number of models increases,their ability to continue to improve gradually declines,and the complexity of models requires more debugging and optimization parameters,and the difficulty of model training becomes greater.More often,we need to make a compromise between performance and complexity.In addition,the parameters of the model in this paper are mostly selected and set through variable parameter debugging.If the optimization methods such as particle swarm optimization are used to optimize the model parameters,the prediction performance is expected to be further improved.