A Time Series Prediction Model Based on Long Short-Term Memory Networks

Abstract

Since real time series are usually divided into cyclical and non-cyclical features, the traditional Auto Regressive Integrated Moving Average (ARIMA) model often shows certain limitations in the modeling of cyclical time series. In this regard, Long Short-Term Memory (LSTM) networks are introduced to predict periodic time series, and missing values are filled by linear interpolation to improve data integrity. The time series are divided into sliding windows according to the period, and the data of the number of people getting on and off the bus in Sino-Singapore Tianjin Eco-City is used as a sample to train and test the model. The test results show that the average absolute error (MAE) of the LSTM model is 80.246, and the root mean square error (RMSE) is 108.52, which is 91.24 and 89.75 lower than the Gated Recurrent Unit (GRU) model, and 83.03 and 104.56 lower than the ARIMA model. The goodness of fit of the LSTM model is 0.902, which is 0.281 and 0.168 higher than that of the ARIMA model and the GRU model, respectively.