Abstract: In recent years,the research enthusiasm for time series data in academic and industrial fields has been increasing,but the frequency information contained in it still lacks effective modeling.The studies found that time series forecasting relies on different frequency patterns,providing useful clues for future trend forecasting:short-term series forecasting relies more on high-frequency components,while long-term forecasting focuses more on low-frequency data.In order to better mine the multi-frequency mode of time series,this paper proposes a multi-feature adaptive frequency domain prediction model (MAFD).MAFD is composed of two stages.In the first stage,it uses XGBoost algorithm to measure the importance of the input vector and selects high-importance features.In the second stage,the model integrates the frequency feature extraction of the time series and the frequency domain modeling of the target sequence,and builds an end-to-end prediction network based on the dependence of the time series on the frequency mode.The innovation of MAFD is reflected in the predictive network's ability to automatically focus on diffe-rent frequency components according to the dynamic evolution of the input sequence,thereby revealing the multi-frequency pattern of the time series and strengthening the learning ability of the model.This work uses 4 datasets from different fields to verify the performance of the model.The experimental results show that compared with the existing classic prediction models,MAFD has higher accuracy and smaller lag.

Key words: Adaptive modeling, Deep learning, Multi-frequency pattern, Time series prediction, XGBoost