Short Term Load Forecasting Based On Data Preprocessing And Deep Belief Network

The power industry is the basic industry of our country's national economy and it provides the basic driving force for industry and other sectors of the national economy.With the rapid economic development,China will also usher in a new upsurge of power grid construction.As the production and consumption of electric energy are simultaneous,the theoretical research on electric power load forecasting has become more and more important.Precise and timely power load forecasting is of great importance in formulating a reasonable plan for power generation and distribution,planning and construction of the power system,surveying the demand of the electricity market,bidding for the Internet and ensuring safe and efficient electricity consumption in production and life.In this thesis,a county in Zhejiang Province as a research area,analyze the data preprocessing,load characteristics,load factors,selection of similar days in training set,optimization of network structure and establishment of a number of load forecasting models based on the intelligent algorithm of intelligent mining knowledge and neural network.Combined with specific cases,the model predictive results were compared and verified,and also design and develop of the corresponding load forecasting software.Data preprocessing is carried out on the Collected load data,the interpolationis and curve fitting method are used to deal with the null value,the Raidah criterion is used to find the singular value.Combined with the survey of geographical,climate,political,economic and other conditions of the surveyed area,study and analyze the power load characteristics and various factors that affect the load.In order to increase the efficiency of model sample set training,this thesis starts with two aspects: Firstly,selects the appropriate training samples,that is,adopts the Adaptive K-means clustering of optimal K value of BWP indicator to select similar days,and secondly,simplifys the Neural network,the KPCA nonlinear dimension reduction method is used to reduce nodes in network input layer.Then,aiming at the shortcomings of traditional BP neural network,such as slow convergence rate andeasy fall into local optimum,this thesis proposes a BP neural network optimized by adaptive genetic algorithm(AGABP)and a prediction algorithm based on deep belief network(DBN),uses sliding window dynamic prediction model,analyzes phase error index by concrete examples.The DM test method is also used to verify that the short-term power load forecasting model based on deep belief network has better forecasting ability.Based on Visual C ++ and Matlab,the related functions and software of load forecasting software are preliminarily implemented in this thesis.