Research On Time Series Forecasting And Application Based On Long Short-term Memory

With the rapid development of network and information technology,the amount of data is increasing explosively.Fully mining the valuable information existing in the huge data resources is helpful for scientific decision-making.Among these massive data,there is an important type of data called time series data.However,time series usually exhibit high-dimensional,non-stationary,and non-linear characteristics.Especially in the era of big data,new data forms and new complex relationships between variables in the data bring unprecedented challenges to time series forecasting.Long short-term memory(LSTM)has an excellent time series learning ability.Thus it is a research hot spot in the field of time series prediction.This thesis studies the time series forecasting problem in the specific application problems of various industries using the LSTM.The main content includes the following three aspects:Firstly,a prediction model named DE-LSTM is designed based on differential evolution algorithm(DE)and LSTM.This model effectively utilizes the optimization ability of DE algorithm and the time series learning ability of LSTM,and is used to solve the problem of electricity price prediction.As there is a problem that the parameters of LSTM and the lag length of electricity price are difficult to determine,the prediction model effectively uses the optimization ability of the DE to find the important parameters of LSTM and the lag length of electricity price.Three electricity price prediction cases are used to verify the performance of the DE-LSTM prediction model.The sensitivity of the selection of fitness value selection,data partitioning method,and related parameters of DE are analyzed.Secondly,an RF-DE-LSTM combined prediction model is proposed.This model combines the feature selection ability of random forest(RF),the optimization ability of DE,and the nonlinear sequence fitting ability of LSTM.Many factors influence high-dimensional time series,and the data features are different.It is difficult for LSTM to learn sequence features directly from high-dimensional space.The RF-DE-LSTM combination prediction model first uses the RF algorithm to select a small number of keywords from many web search index keywords to reduce the data dimension and then uses the DE algorithm to optimize the lag length of the selected search index keywords and the historical tourist arrivals.Finally,LSTM is used to perform regression modeling on the input features and the number of tourists to get the final prediction result.The results of two actual tourist arrival cases show that the performance of the proposed RF-DE-LSTM prediction model is better than some common prediction models.Finally,an EWT-attention-LSTM prediction model is proposed.This model combines the sequence decomposition ability of the empirical wavelet transform(EWT)model,the key information capture ability of the attention mechanism,and the time series learning ability of LSTM.To evaluate the relationship between the input and output of LSTM,the original energy consumption data is decomposed through EWT,and then the attention mechanism is used to assign different weights to the input influencing factors of energy consumption,which can highlight more critical influencing factors.Then LSTM is utilized to forecast each component decomposed by EWT.The performance of EWT-attention-LSTM prediction model is evaluated through three monthly energy consumption cases.