Model for non-periodic time series prediction

Abstract

In practical applications, purely periodic data is relatively rare, and most data often exhibit non-periodic characteristics, making it challenging to predict or describe through simple periodic changes. Single neural network frequently encounters issues such as overfitting, difficulties in capturing long-term dependencies, and limited ability to capture nonlinear relationships when dealing with non-periodic time series. To effectively forecast non-periodic time series, the ILTNet model was proposed based on the Informer model. The ILTNet model integrated linear prediction (AR model) and nonlinear prediction (Informer model and recurrent skip components), enabling effective capture of long-term dependencies. Experiments show that ILTNet model demonstrates significant advantages in non-periodic time series forecasting compared to the LSTNet, Informer, AR, and GRU models. For example, on the Exchange Rate dataset, ILTNet model reduces RSE by 0.0333 and 0.0277 compared to LSTNet model at horizons of 96 and 128, respectively.Compared to Informer model, ILTNet model achieves significant RSE reductions at all horizons, particularly reducing RSE by 0.2877 at a horizon of 96.

Keywords

• time series forecasting
• convolutional neural network
• LSTNet
• Informer