AB-LSTM: Attention Bidirectional Long Short-Term Memory for Multivariate Time-Series Forecasting

Abstract

Due to its applications in numerous fields, such as weather forecasting, multivariate time series forecasting has attracted significant interest. This paper promotes AB-LSTM (Attention Bidirectional Long Short-Term Memory) as an innovative method for precise multivariate time series forecasting. The AB-LSTM model combines the strengths of bidirectional LSTM (Bi-LSTM) and attention mechanisms to capture the temporal dependencies and interdependencies among multiple time series variables. The bidirectional nature of the LSTM allows the model to incorporate past and future information, enabling more robust predictions. The attention mechanism focuses on the most relevant time steps and variables, enhancing the model's ability to extract meaningful patterns and relationships from the input data. The performance of AB-LSTM is then evaluated in the experiment to the public datasets of Beijing PM2.5. The results were compared with state-of-the-art baseline models. The results demonstrate that AB-LSTM outperforms the baseline models regarding forecasting accuracy based on RMSE 20.966, particularly for long-range predictions and complex datasets. The AB-LSTM model offers interpretability by providing attention weights, indicating the relative importance of each input variable at different time steps.