SATNet: Upgraded LSTM Network For Mining Time Series Correlation Utilizing The Self-Attention Mechanism

Multivariable time series prediction is a crucial machine learning application that spans many domains, such as the forecast of users' electricity consumption and exchange rate adjustment of various currencies. In these real-world datasets, there are remarkable similarities between the data of the time series itself, and these similarities can be used for prediction. Still, the traditional methods do not take full advantage of this feature. In this paper, we continue to tap the potential of the Self-Attention Mechanism and propose a new deep learning framework, Self-Attention Time Series Network (SATNet), to improve prediction accuracy. Time series data is preprocessed by SATNet using CNN and RNN, and the Self-Attention Mechanism is used to calculate the correlation between data to improve the prediction accuracy. Then, the time features and data correlation extracted from different structures are fused through the Learnable Addition Mechanism. Finally, the traditional linear model is added to extract the mutation factors in the time series to correct the prediction results. In comparison to several of the most sophisticated baseline approaches, SATNet significantly outperformed them in our examination of real-world data with complex time series, with an average of 7% higher. Especially in the dataset with an obvious regular pattern, it has significant advantages compared with other advanced models. Online resources for data and experimental codes are accessible.