Research On Time Series Prediction Model Based On Improved Recurrent Neural Network

With the rapid development of information technology,social life has entered the era of big data.Different types of massive data are produced in various fields all the time.A class of data is generated sequentially and continuously at different time intervals.This type of data often contains rich information,such as financial transaction data,weather data,network traffic data.This type of data is collectively called "time series data".Time series data may imply the future development trend of some events,so it is of great practical significance to analyze and predict them.With the rapid development of machine learning,it is possible to predict time series with high precision.This paper proposes an improved recurrent neural network model—new gate control unit(NGCU).NGCU is mainly applied to the prediction of time series data,which aims to improve the prediction accuracy of the model and reduce the training time of the model under the premise that the model can alleviate the problems of "gradient disappearance" and "gradient explosion" of the recurrent neural network.Compared with long short-term memory(LSTM),NGCU eliminates the single "output gate" to reduce the computational complexity of the model,and introduces an anti-oversaturation transformation information(Tri)module to make the control unit more sensitive to data feature learning.Compared with gated recurrent unit(GRU),NGCU retains the data information from the beginning to the previous point,which makes the control unit learn the data characteristics more fully.To verify the accuracy,efficiency,and feasibility of NGCU time series prediction,support vector regression(SVR),random forest regression(RFR),recurrent neural network(RNN),LSTM,and GRU are used as the comparison models in the experiment.To analyze the extensibility of NGCU in the integration models,convolutional neural networks(CNN)and attention mechanism(AM)are introduced in this paper.Thus,the comparison models of CNN-LSTM,CNN-GRU,CNN-NGCU,CNN-AMLSTM,CNN-AM-GRU,and CNN-AM-NGCU are constructed.To verify the generalization ability of the models,three different data sets are selected in the experiment: the air quality data set,the Hang Seng Index data set,and the gold futures data set.The three datasets differ in data volume,stability,and eigenvalue dimension.In the three datasets,the air quality index,the Hang Seng Index closing price,and the gold closing price are predicted respectively.Experimental results show that the single model NGCU is better than the other five single models,and the model training time is lower than LSTM and GRU.In the extension experiment of the integrated models,the prediction results and training time of CNN-NGCU are better than those of CNN-LSTM and CNN-GRU,and the prediction results and training time of CNN-AM-NGCU are better than those of CNN-AM-LSTM and CNN-AM-GRU.In this experiment,the time series prediction model based on CNN-AM-NGCU has the highest prediction accuracy.Therefore,the time series prediction model based on NGCU can achieve high-accuracy prediction while ensuring low training time.