Novel Machine Learning-Based Time Series forecasting Model By combining short-term and long-term time patterns

Abstract

Abstract Time series forecasting is widely applied in various domains. One of the disadvantages of long-term series is not using the characteristics of the frequency domain. These features generally exist in short-term patterns of time series and indicate rapid fluctuations in a short period of time. In this article, a new model of long-term forecasting of multi-dimensional time series is proposed, in which by using local and global neural network design, in addition to extracting long-term patterns, using Fourier transform and frequency domain characteristics of short-term patterns as well has been extracted. The proposed model obtains time and frequency range by combining long-term patterns with short-term time patterns and seasonal forecasting. The output of the forecasts is used to predict the trend and by combining it with the time domain and frequency, the final output of the time series is obtained. The simulation results show that the proposed model has better prediction accuracy than conventional models.