Long sequence time-series forecasting method based on multi-scale segmentation

Abstract

Addressing challenges posed by large volumes of historical data, high computational complexity and stringent prediction accuracy requirements in long sequence time-series forecasting, we propose a Transformer model incorporating the concept of multi-scale segmentation. The model enhances the Transformer architecture by employing multi-scale segmentation to slice the time series into multiple time periods for training and prediction, thereby reducing the complexity of long time series and improving prediction accuracy. Experimental results on the real-world power transformer dataset, encompassing variables like electricity transformer temperature, electricity consumption load, and weather demonstrate that the proposed Transformer model based on the multi-scale segmentation approach outperforms traditional benchmark models such as Transformer, Informer, gated recurrent unit, temporal convolutional network and long short term memory in terms of mean absolute error (MAE) and mean squared error (MSE). The proposed Transformer model achieves an MSE of 0. 367 and an MAE of 0. 407 in experiments with a prediction length of 192 on the Weather dataset, consistently surpassing other models. By leveraging the advantages of Transformer model and incorporating the multi-scale segmentation approach, the proposed model achieves faster computational speed and superior predictive performance.