Research On Combined Forecasting Model Of Power Load Based On Deep Learning

Load forecasting is an important part of power system planning,which directly affects the safety and reliability of power grid operation.The real-time and high-precision prediction of load is the key measure to improve the efficiency of the entire power grid and to maximize the operation efficiency of the power related enterprises.The traditional simple load forecasting model is good at dealing with the linear relationship between the sequences.When facing the highly nonlinear characteristics of the actual data,its forecasting ability will be significantly reduced.Nowadays,in the context of the rise of big data analysis and the continuous upgrading of computing tools,deep learning models are widely used in the field of power load forecasting because they can efficiently mine potential non-linear relationships between data.Aiming at the problem of load forecasting,this paper relies on the powerful feature learning capabilities of deep learning to focus on data preprocessing,input variable feature extraction,and the construction of combined prediction models.The specific work is as follows:Based on the theoretical basis,the characteristic analysis and data preprocessing of the historical load have been completed.Firstly,the daily,monthly and annual load characteristics of the power load are analyzed by load characteristic index parameters.After entering the data preprocessing phase,that is to use the box plot method to detect and correct the outliers of the real-time data information of the England region from December 1,2013 to November 30,2014,which included six types of influencing factors.In view of the problem that too many input variables are easy to lead to the decline of model stability,an effective feature selection and evaluation mechanism of the main influencing factors of power load is established.After using the method of Maximal Information Coefficient(MIC),Grey Relation Analysis(GRA)and the method of Mean Decrease Impurity(MDI)based on random forest to complete the screening experiment of the factors affecting the power load,the improved MIV-BPNN algorithm is used to determine the effectiveness of the above three feature selection methods.The final results show that the dominant influencing factors screened by the GRA correlation analysis method are the most accurate.Based on the analysis of the internal variation law of power load,a combined algorithm VMD-DBN,which combines Variational Mode Decomposition(VMD)technology and Deep Belief Network(DBN),is proposed and obtained higher prediction accuracy in the load prediction experiment.Relying on VMD technology,the original load is decomposed into a series of modal functions with more obvious characteristics in the time domain and frequency domain,and then the input variables are determined according to the specific characteristics of each modal function.Then the DBN depth learning algorithm is used to model and analyze them respectively.Finally,the prediction data of each modal function at the same time point is accumulated to obtain the final prediction result.By comparing the prediction results of VMD-DBN algorithm with DBN and classic algorithms BPNN and RBFNN on two different data sets,the superiority of the proposed algorithm was confirmed.Based on the potential linear and non-linear characteristics of the power load sequence,a combined forecasting algorithm is proposed,which combines the linear model Auto Regressive Moving Average(ARMA)and the non-linear model Long Short Term Memory Network(LSTM).Compared with a variety of single algorithms,the forecasting accuracy is significantly improved.By using the ability of deep learning algorithm LSTM in mining data non-linear relationship,the residual sequence obtained after ARMA model prediction is modified to realize the complementary advantages between models,and then effectively improve the accuracy of load forecasting.After comparing the prediction results with single ARMA,LSTM and SVM on two different data sets,the effectiveness of the proposed algorithm in short-term power load forecasting is verified.