Multivariate Time Series Forecasting Based on Sliding Attention Correction LSTM Network
Jincheng Li, Linli Zhou, Liangtu Song · 2024
Time series signals are a common type of data, containing audio data, weather information, electricity consumption, and consumption indices. At the same time, these time series signals are also intertwined with multiple related factors, making it difficult to predict their future trends from a single time series signal. For accurate multivariate time series forecasting, this paper proposes a multivariate prediction model, the Sliding Attention Correction Long Short-Term Memory Network (SAC-LSTM). The model uses LSTM to predict the time series signal, and adopts the sliding attention module to capture the valid information in the relevant variables to correct the predictive results. According to the experimental results, the proposed prediction model in this paper is the best in both of efficiency and effect.