A Faster Time Series Data Prediction Method Based on LSTM
Xuanpeng Song · 2022
The complex structure of LSTM increases the number of parameters and leads to an increase in training time. We propose an improved prediction method for time series data based on LSTM, which can significantly reduce the training time while ensuring a certain prediction accuracy. Our method first uses wavelet decomposition to decompose the data into low-frequency data and high-frequency data and then uses LSTM to learn the characteristics of low-frequency data, use Random Forest to learn the characteristics of high-frequency data, and finally uses wavelet reconstruction to reconstruct the predictions of LSTM and Random Forest for different frequency data into prediction data. Test results on datasets in three different domains show that our method can predict the overall trend of time series data well, but the prediction results for local details are slightly worse. Compared with using LSTM directly, our method increases the average mae by 15.52% and the average mse by 31.10% on the three datasets but reduces the average training time by 69.66%.