Time Series Prediction Based On Deep Belief Network

With the development of the Internet of Things,the dimension of time series data has increased rapidly,which has brought various difficulties to time series prediction.Firstly,because the time series prediction tasks are time-sensitive,the prediction method is required to quickly calculate the prediction results.In addition,the increase in time series dimensions makes the data highly nonlinear,requiring predictive models with a high degree of functional fit ability.Therefore,it is of great practical significance to take into account both learning speed and prediction accuracy.In this paper,the time series prediction model based on deep belief network is studied for the problems faced by the above time series prediction:(1)Based on the characteristics of noise interference,time-effects and complex relationship functions of time series data in real-world systems,a hybrid neural network prediction model based on deep belief network and gated recurrent unit is proposed(Deep Belief Network-Gated Recurrent Unit,DBN-GRU).Firstly,the time series signal is decomposed by empirical mode decomposition to suppress the influence of noise.Then,the pre-training mechanism of deep belief network is used to train the network weight,thus accelerating the overall learning speed of the network.Finally,through the gate structure of the gated recurrent unit,the data time correlation is memorized,and the ability of the whole model to extract the time and space relationship of the data is strengthened.The performance of DBN-GRU for highly nonlinear and strong time-dependent time series prediction is verified by experiments.(2)A Deep Belief Network-glial-Gated Recurrent Unit(DBN-g-GRU)is proposed for the problems that the deep belief network can only predict a single time series factor.The introduction of glial cells to assist hidden layer neurons to learn other time series data related information enhances the ability of predictive models to deal with strongly coupled time series.And the glial is connected by the formation of a chain,and the activation pulse is transmitted in one direction,thereby accelerating the speed of network learning.Experiments show that the DBN-g-GRU algorithm can effectively achieve good results in a strongly coupled time series environment.