The Time Series Prediction Method Based On Improved Deep Belief Network And The Application Of Load Forecasting

High quality prediction results have important guiding significance for people's production and life,which is helpful for people to make more reliable decisions.Prediction is a theory to estimate the future development trend of a thing according to the law of its development in the past.In recent decades,experts and scholars in various fields have put forward tens of thousands of forecasting methods,and the discipline of forecasting has developed by leaps and bounds.With the advent of the era of big data,mining the massive information contained in unstructured and semi-structured data poses new challenges to the prediction methods.At the same time,how to improve the accuracy of prediction methods in specific fields has become an important research direction of prediction research.As a new deep learning method,deep belief network can learn the inherent characteristics of sample data set compared with traditional prediction method.It can be widely used in many fields and achieve good results.This paper mainly examines the prediction performance of the improved deep belief network in the time series data set,which also applies the improved model to the short-term load prediction to investigate the generalization ability of the improved model.The main work of this paper is as follows:(1)The optimal configuration strategy of the basic model of deep belief network is designed.Deep belief network is a special deep neural network.According to the improvement direction of artificial neural network,the influences of different activation functions and network parameter optimization methods on the basic model of deep belief network were investigated,and the optimal configuration of the deep belief network model was found,which laid a theoretical foundation for the later research.(2)The point prediction model based on restricted Boltzmann machine and recurrent neural network was constructed.Time series is a dynamic system.According to the belief network is a kind of static mapping relationship between input and output,feedback mechanism of the recursive neural network is introduced into the belief network model.A deep learning framework based on restricted boltzmann machine and the recursive neural network is constructed,which is applied to point prediction model of time series.The comparison results show that performance of the proposed new framework has higher quality.(3)The lower and upper bounds estimation method based on deep belief network is proposed to construct the interval prediction model.The method of upper and lower bound estimation,which is often used to construct interval prediction,is based on neural network.Considering that the neural network have the disadvantage of easily trapped into local minima and slow convergence speed,the upper and lower bound estimation method based on deep belief network is improved,and the upper and lower bound estimation method based on deep belief network is given.(4)The improved deep belief network is applied to short-term load power forecasting.Considering the importance of short-term load power prediction to power system and the generalization ability of the improved model,the improved deep belief network was respectively applied to the point prediction and interval prediction of short-term load power to improve the accuracy of short-term load power prediction.Based on the characteristics of load power data changing periodically with time and the concept of similar days,the improved deep belief network was used to analyze the predicted performance of load data at different times,different dates and in different seasons.