关键词: 较短的长序列时间序列预测, 序列到序列, 长短期记忆, 自注意力机制, 特征重分配

Abstract: Aiming at the problem that most of the existing researches have studied short-sequence time series forecasting and long-sequence time series forecasting separately, which leads to the poor forecasting accuracy of the model in the shorter long-sequence time series, a Shorter Long-sequence Time Series Forecasting Model (SLTSFM) was proposed. Firstly, a Sequence-To-Sequence (S2S) structure was constructed using Convolutional Neural Network (CNN) and a new PBUSM (Probsparse Based on Uniform Selection Mechanism) self-attention mechanism, which was used to extract the features of the long-sequence input. Secondly, the newly designed "far light, near heavy" strategy reallocated the features of each time period extracted from multiple Long Short-Term Memory (LSTM) modules, which were more capable of short-sequence input feature extraction. Finally, the reassigned feature was used to enhance the extracted long-sequence input feature to improve the forecasting accuracy and realized the time series forecasting. Four publicly available time series datasets were utilized to verify the effectiveness of the model. When compared with the best-performing comparative model, Gated Recurrent Unit (GRU), the experimental results show that the Mean Absolute Error (MAE) metrics of SLTSFM are reduced by 61.54%, 13.48%, 0.92% and 19.58% for univariate time series forecasting on the four datasets. Reducing by 17.01%, 18.13%, 3.24% and 6.73% for multivariate time series forecasting. Demonstrating that the model is effective in improving the accuracy of shorter long-sequence time series forecasting.

Key words: shorter long-sequence time series forecasting, Sequence-To-Sequence (S2S), Long Short-Term Memory (LSTM), self-attention mechanism, feature reallocation