CL-Informer: Long time series prediction model based on continuous wavelet transform.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Time series prediction is crucial but challenging.
• Existing deep learning models like Informer have limitations in capturing multi-scale data characteristics and dependencies.
• Accurate long sequence prediction remains an open problem.

Purpose of the Study:

• To propose a novel deep learning model, CL-Informer, for improved time series prediction accuracy.
• To enhance the Informer model by integrating continuous wavelet transform and LSTM layers.
• To evaluate the performance of CL-Informer against other state-of-the-art forecasting models.

Main Methods:

• Developed the CL-Informer model by adding a continuous wavelet transform (CWT) embedding layer to the Informer architecture.
• Utilized a Long Short-Term Memory (LSTM) layer to further capture data dependencies and process CWT-derived information.
• Compared CL-Informer with Informer, Informer+, and Reformer on five diverse datasets for time series forecasting.

Main Results:

• CL-Informer achieved an average Mean Squared Error (MSE) reduction of 30.64% for univariate prediction horizons.
• CL-Informer demonstrated an average MSE reduction of 10.70% for multivariate prediction horizons.
• The proposed model showed superior performance in long sequence prediction compared to the original Informer model.

Conclusions:

• The CL-Informer model significantly improves the accuracy of time series prediction.
• The integration of CWT and LSTM effectively captures multi-scale data characteristics and dependencies.
• CL-Informer offers a more precise and reliable approach for long sequence forecasting.