Time Series Forecasting Based On LSTM Hybrid Model

Time series data often reflects the law of the development of an event,and contains a wealth of potential information.Long Short-Term Memory(LSTM)can be used for time series prediction because of the uniqueness of their network structure,which can store information and solve long-term dependence problems.The research on LSTM is conducive to efficient mining of important information in time series data,and has an important impact on the development of social and economic aspects.In this paper,LSTM is taken as the research object,and the XGB-LSTM hybrid model based on Xgboost feature selection and LSTM neural network is proposed.An empirical analysis was conducted on three different time series data sets: a bank stock data,SML2010 and Beijing PM2.5.The research in this paper concentrated in the following three aspects:Firstly,an exploratory analysis of LSTM neural network based on a bank stock dataset is carried out to explore the effects of parameter time step,number of neurons and batch size on the performance and efficiency of LSTM neural network.It provides a reference value for how to improve the training precision and speed of LSTM in subsequent research.The research shows that when the remaining parameters are the same,LSTM training duration increases with time step and number of neurons,and different parameters will have a certain impact on model performance.Therefore,using the empirical information combined with the actual situation of the data to determine the approximate range of parameters firstly,and then using grid search,random search and other methods to adjust parameters reasonably,which is beneficial to improve the prediction accuracy and training speed of LSTM neural network.Secondly,the XGB-LSTM hybrid model,LSTM and PCA-LSTM hybrid models are respectively established on the SML2010 dataset for comparative analysis,demonstrating the effectiveness of the XGB-LSTM hybrid model.The results show that the XGB-LSTM hybrid model not only saves time overhead,but also reduces noise and improves the accuracy and stability of the model.Finally,the XGB-LSTM hybrid model is applied to the prediction of PM2.5 concentration in Beijing,and compared with Xgboost,LSTM,GRU and BP neural network,further demonstrating the effectiveness and application value of the XGB-LSTM hybrid model.It is proved that the LSTM neural network and its hybrid model have higher prediction accuracy than other models in time series prediction.The XGB-LSTM hybrid model is more suitable for high-dimensional large sample data sets.