Research on Low Correlation sequence data prediction based on data fusion

In the field of multidimensional time series data prediction, correlation is an important influencing factor. However, in practical applications, it is often difficult to find other-dimensional variables that are highly correlated with the target variable, or these variables are hard to obtain. In this study, we address the challenge of predicting low-correlation multidimensional time series data and propose a new method based on data fusion for low-correlation sequence data prediction. Firstly, considering the characteristics of low-correlation time series data, which exhibit time-varying correlations, we introduce batch processing sliding windows to more accurately capture the correlations between dimensions. Secondly, we incorporate an Attention Turn Network (ATT-TURN-NN) to enhance the credibility of the prediction results. For low-correlation data, the overall correlation among the data is low, but the correlations between the data collected by different sensors and the target variable differ. By separately predicting different time series data, we achieve dimensionality reduction and improve the efficiency of the model. Additionally, we adopt a data fusion strategy that calculates weights based on covariance to fuse the predicted results of the multidimensional data. This strategy assigns different weights to different time series data, thereby reducing the bias impact of a single data source on the prediction and enhancing the robustness and stability of the model. Experimental results validate the superiority of the proposed method. Compared to other models, our method demonstrates better accuracy and efficiency, with a coefficient of determination reaching 0.96.