Developing a multivariate time series forecasting framework based on stacked autoencoders and multi-phase feature.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Multivariate time series forecasting is crucial for intelligent decision-making.
• Deep learning models face challenges with non-linear patterns and data randomness.
• Existing models struggle with simultaneous effectiveness and robustness.

Purpose of the Study:

• To propose a novel prediction framework for enhanced time series forecasting.
• To address limitations in current deep learning models for complex datasets.
• To improve accuracy and robustness in multivariate time series prediction.

Main Methods:

• A multi-phase feature selection technique for optimal feature and lag window selection.
• A stacked autoencoder strategy using Long Short-Term Memory (LSTM) and temporal convolution networks.
• Two autoencoders: one for random weight initialization, another for temporal relations.

Main Results:

• The proposed framework significantly outperforms existing models on real-world datasets (Energy Appliances, PM2.5, Solar Radiation).
• Achieved approximately 40% improvement in Mean Absolute Error (MAE) for PM2.5 data.
• Demonstrated substantial improvements in Mean Squared Error (MSE) and MAE for Solar Radiation data.

Conclusions:

• The novel framework offers superior generalization and predictive accuracy.
• It effectively models intricate temporal dynamics in multivariate time series.
• The approach provides a robust solution for complex forecasting tasks across domains.