Research On Short-term Load Forecasting Method Based On Robust Multidimensional Autoregression

With the gradual development of power market,power load forecasting is required to be more timely,accurate and reliable.Curren tly the load forecasting plays an important role in the power system operation and planning.At the same time,the power load has obvious randomness and uncertainty,and it is difficult to guarantee the high prediction accuracy.Therefore,it is of great significance to predict the power load and improve the prediction accuracy.Based on the analysis of power load characteristics,we propose a multidimensional autoregressive prediction method of time series in this thesis,and investigate the short-term prediction of power load from two aspects according to the problem that the accuracy of multidimensional autoregressive forecasting is affected by the abnormal value and randomness of power load.Aiming at outliers in the power load data and the sensitivity of multidimensional autoregressive prediction model to outliers,a robust prediction method based on maximum correntropy criterion and autoregression(MCC-AR)was proposed in this thesis.The maximum correntropy is introduced into the constrained least square multidimensional autoregressive prediction model,and the local similarity between data points is measured according to the Gaussian kernel width in the correntropy.For nonlinear nonconvex optimization problems,a quadratic programming method based on the properties of conjugate convex functions is proposed,and the semidefinite relaxation method is used to solve the constrained regression problem.In addition,several kinds of parameter selection methods in the prediction model are analyzed and discussed respectively,and the rationality and correctness of the selected parameters are proved.Results show that the proposed regression model with maximum correntropy optimization is more accurate and reliable than multidimensional autoregression and multidimensional differential autoregressionIn view of the randomness of power load,the variational Mode Decomposition(VMD)was used to analyze and process the data.The components of different time scales are effectively separated by non-recursive method,which avoids the problem of mode aliasing and has better noise filtering effect.To solve the problem that the number of decomposition is difficult to determine in this method,the sample entropy is used to determine the noise margin,and the average sample entropy and mutual correlation are used to choose adaptively.A load sequence analysis model of improved VMD is established,and the effectiveness of this method is proved by simulation experiments.The improved VMD method is used to decompose and process the real load data,and the MCC-AR method is used for short-term prediction.The experimental results show that the improved VMD decomposition method has better prediction accuracy in short-term load forecasting than EMD decomposition method.