ATFSAD: Enhancing Long Sequence Time-Series Forecasting on Air Temperature Prediction

Abstract

Long sequence time-series forecasting models pay excessive attention to irrelevant information and noise when predicting air temperature, leading to prediction drift and poor generalization ability. We propose an Air Temperature Forecasting Model Based on RenyiSparse Attention and Adaptive Time Series Decomposition(ATFSAD). The ATFSAD model follows the encoder-decoder structure, embedding RenyiSparse Attention to efficiently capture long-term dependencies from time series. Additionally, to reduce the redundancy mixed into the encoded information, an information distilling technique is designed to filter it out prior to decoding. Lastly, the multi-layer decoder is combined with adaptive time series decomposition to progressively refine the seasonal and trend components of the prediction. The experiments on the Jena meteorological dataset show that ATFSAD outperforms the state-of-the-art models Informer and FEDformer by 9.89% and 5.97% respectively, in predicting air temperature at 10-minute intervals for a full day(144 points), with a mean absolute error(MAE) of 1.73. Therefore, ATFSAD is of practical value for air temperature prediction.

Keywords

• Artificial intelligence
• deep learning
• forecasting
• time series analysis
• weather forecasting