Time-Series Prediction Research Based on Combined Prophet-LSTM Models

EasyChair Preprint 9843

Abstract

Time series forecasting models are an important model for practical applications, but are some relatively complex class of modelling and forecasting problems due to their reliance on the sequence of events. The advent of Facebook's open source framework Prophet is a breakthrough in traditional time series forecasting models, which is simple, efficient, flexible and highly robust. However, any single prediction model for non-linear time series prediction still suffers from low accuracy and inability to extract the composite features of time series well. To this end, we propose an innovative approach to time series prediction based on the Prophet model and adding the long-short memory network model LSTM to form a combined Prophet-LSTM model. Firstly, the origin of time series forecasting is introduced, and several classical time series forecasting models are listed and their shortcomings are analyzed; secondly, the principles and advantages of the combined Prophet-LSTM model are elaborated; finally, the trend change of temperature is predicted using the Shanghai temperature data set as a sample, and the good prediction results confirm that the combined model is an excellent forecasting tool, which is worth studying and promoting application.

Keyphrases: Gradient disappearance, Gradient explosion, Overfitting, Prophet, Regressionmodels, robustness

@booklet{EasyChair:9843,
  author    = {Gao Jun-Gang},
  title     = {Time-Series Prediction Research Based on Combined Prophet-LSTM Models},
  howpublished = {EasyChair Preprint 9843},
  year      = {EasyChair, 2023}}