Multivariate Time Series Prediction Based on Temporal Change Information Learning Method.

Area of Science:

• Machine Learning
• Time Series Analysis
• Deep Learning

Background:

• Multivariate time series prediction faces challenges in extracting inter-series dependencies across time.
• Attention-based deep learning models show promise but neglect temporal dynamics in objective functions and optimization.
• Existing methods struggle to adaptively capture abrupt and slow changes in time series data.

Purpose of the Study:

• To propose a novel Change Information Learning (CIL) method for enhanced multivariate time series prediction.
• To address the limitation of ignoring temporal change information in current deep learning prediction models.
• To improve the extraction of impact information from non-predictive series on the target series at different time stages.

Main Methods:

• The CIL method incorporates Mean Absolute Error (MAE) and Mean Squared Error (MSE) into the objective function to assess varying amplitude errors.
• Second-order difference technology is employed within the objective function to adaptively capture abrupt and slow change impacts.
• A Long Short-Term Memory (LSTM) network with a transformation mechanism is utilized for comprehensive temporal dependence extraction, avoiding saturation.
• An optimization algorithm adaptively memorizes current and historical moment estimations, enhancing temporal change information acquisition in error gradients without extra hyperparameters.

Main Results:

• The proposed CIL method demonstrates significant advantages in computational overhead compared to existing approaches.
• Experimental results on three diverse datasets confirm the superior prediction effect of the CIL method.
• The method effectively captures temporal change information, leading to improved prediction accuracy.

Conclusions:

• The CIL method offers a robust solution for multivariate time series prediction by integrating temporal change dynamics.
• The approach enhances the adaptability and accuracy of deep learning models in complex time series tasks.
• CIL provides a computationally efficient and effective strategy for time series forecasting.