A Research On Several Types Of Time Series Forecast Models Based On Deep Neural Network

With the advancement of science and technology,artificial intelligence also developed rapidly and reflecting its excellent performance and potential.One of the most eye-catching is the deep neural network,the recurrent neural network and convolutional neural network in the deep neural network are models developed around time and space,respectively.As the most important time data,how to learn the historical laws of time series and predict their future trends has always been the focus of scholar's research.Correct prediction of time series can bring a lot of convenience to people,and traditional time series analysis once reached a bottleneck due to the shortcomings of the method.However,neural network model used to deal with new problems in time series prediction.First,this paper introduces traditional time series analysis methods and some neural network methods.Traditional time series methods include moving average models(MA),autoregressive models(AR)and autoregressive moving average models(ARMA).These traditional methods are suitable for wide stationary time series.The feature of wide stationary series is that the statistical characteristics of the entire time series do not change with the time shift,it means the mean and covariance of the time series do not change with the time shift.The disadvantages of the traditional method is that for non-wide stationary time series,the effects of these methods are generally not ideal,and for the particularly complex and highly nonlinear time series like stock data,the effect of prediction is limited.Neural network methods include BP neural network and recurrent neural network(RNN).Both methods use the strong learning ability of neural network to make predictions based on the features of existing data.Compared with traditional methods,neural network methods are more accurate and can be applied to most time series.Among them,recurrent neural networks are often used to process time series.Neural network methods also has its disadvantages,BP neural network is a local search optimization method,it is easy to fall into local extremes,recurrent neural networks are prone to gradient explosion and gradient disappearance problems and it is difficult to deal with long dependence problems.The prediction results is often not ideal when the span is large.Based on the gradient explosion,disappearance of gradient and long dependence problem of recurrent neural network,this paper proposes using long short-term memory network and its variant gated recurrent unit as the basic model of time series prediction.Based on this,we proposes to combine the long short-term memory network and gated recurrent unit to build a new model called GRU-LSTM model.Compared with the gated recurrent unit,this model is more complex that it can learn more features and has a higher accuracy.Compared with long short-term memory,it has less parameters and it is easier to train.The three models mentioned above can effectively predict time series in the time dimension,but they basically do not consider the spatial dimension.In order to predict time series by combining time and space,further proposed to combine graph neural network with GRU-LSTM model to form a new model based on time and space,called GC-GLSTM model.In this way,the model first uses the graph convolution network to process the spatial information,then uses the gated recurrent unit and long short-term memory network to process the time information.This paper uses the stock data(Dalian Thermal Power 600719,Dalian Friendship 000679)and traffic data,the two types of time series use four models to make prediction.The stock data does not have a spatial structure,while the traffic data has a specific spatial structure.Finally,the performance of each model is analyzed and compared,then select the best prediction model.Experimental results show that introducing spatial information in the time series can indeed improve the final prediction result.The innovations of this article are as follows:1.In this paper,a new network(GRU-LSTM)is constructed by long short-term memory and gated recurrent unit.In order to use to spatial information in time series,this paper combines graph convolutional network and recurrent neural network to build GC-GLSTM network,then we use the constructed feature formula to find the relationship between each feature,we give a graph structure to the time series based on the connections between features.Experimental results show that the GC-GLSTM model with the spatial information is superior to other recurrent neural networks.2.This paper selects two different types of time series data to reflect the universality of the model.In order to evaluate the differences of each model better,different evaluation indexes are adopted,and the model results in this paper are compared with long-short term memory network and gated recurrent unit.Experimental results show that the model in this paper can be applied to these two different kinds of time series.