Research On Short-Term Power Load Forecasting Algorithm Based On Neural Network Hybrid Model

The stable and safe operation of the overall structure of the power system is an important prerequisite for the country’s rapid economic development.With the scientific and orderly expansion of power consumption areas in my country,power demand and grid transformation have placed higher requirements on grid planning and dispatching.Short-term power load forecasting is Load forecasting is a very important part.It provides scientific and reasonable planning basis for the maintenance and operation of various power generation,transmission,and distribution equipment in the power system.However,the power load of my country’s power grid faces more uncertain factors,and short-term power load forecasting is also one of the urgent research topics.Based on short-term power load forecasting as the research foundation,this paper analyzes and discusses in actual cases,puts forward and proves its own arguments;the innovations and main research work of this paper are as follows:(1)This paper uses the whole society electricity load in Yixing City,Wuxi City-Jiangyin City,Jiangsu Province in 2018 as the experimental data set,in which the daily load periodicity,weekly load periodicity,seasonal load characteristics,and geographic regions of the electricity load characteristics and weather factors affecting electricity load are analyzed in detail,and the influence and connection of various potential properties on electricity load are fully discussed.Moreover,the phase space method in feature engineering is used to further optimize the pretreatment process.(2)According to the excessive dependence of wavelet transform(WT)on prior knowledge,the selection of parameters and the existence of many shortcomings of modal aliasing and end effects in empirical mode decomposition(EMD),the paper proposes an empirical wavelet transform(EWT)Combined model of short-term load forecasting combined with Deep Belief Network(DBN).EWT is good at mining potential features in uncertain multi-factor time series,and is fully prepared for the construction of the training matrix;DBN adopts the two-stage training mode of forward pre-training and reverse fine-tuning to obtain better weights and Bias matrix,the neuron uses the neuron activation method in the form of statistical probability,which makes the entire network have a very good ability to handle multi-dimensional features.(3)A new multi-layer combination model is proposed on the basis of EWT-DBN.The prediction results of EWT-DBN and the original training matrix are combined and imported into the width learning system(BLS).The multiple sets of feature mapping nodes in BLS can be better to learn higher dimensional data.Through the verification of examples,the prediction performance of EWT-DBN-BLS is better than many models.