A Translation-Invariant Neural Network Architecture for Time Series Forecasting

Targeting the problem that the whole series morphological similarity measurement method usually cannot mine the morphological trend change between time series as a whole, an extended layer with rich output representation is proposed, and the auto-encoder network is combined to automatically learn the global similarity with translation invariant from time series data, to realize the global feature extraction of time series data and the improvement of time series prediction effect. Experimental results show that the proposed structure performs excellently in all cases in the forecasting task of multiple real-world time series datasets. © 2024 Beijing University of Posts and Telecommunications. All rights reserved.