Related Experiment Videos
Time Series Forecasting and Classification Models Based on Recurrent with Attention Mechanism and Generative
Kun Zhou1,2, Wenyong Wang1, Teng Hu1,2
1School of Computer Science and Engineering, University of Electronic Science and Technology of China, Chengdu 611731, China.
Sensors (Basel, Switzerland)
|December 19, 2020
Summary
Deep learning models, including temporal convolutional networks (TCNs), show promise for time series classification and forecasting. TCNs offer significant speed improvements over traditional methods while maintaining accuracy.
Area of Science:
- Machine Learning
- Deep Learning
- Time Series Analysis
Background:
- Traditional statistical methods have long been used for time series classification and forecasting.
- Deep learning has achieved success in various domains but is less explored in time series analysis.
Purpose of the Study:
- To propose and evaluate state-of-the-art neural network models for time series tasks.
- To compare the performance of deep learning models against classical methods.
Main Methods:
- Review of Long Short-Term Memory (LSTM), Temporal Convolutional Network (TCN), and Generative Adversarial Network (GAN).
- Application of LSTM with autoencoder and attention, TCN, and GAN to time series classification and forecasting.
- Introduction of Gaussian sliding window weights to accelerate training.
Main Results:
- TCN demonstrated superiority over LSTM in sequence modeling.
- Proposed TCN reduced training time by approximately 80% with comparable accuracy.
- GAN training instability was addressed through hyperparameter tuning and the Adam optimizer, achieving competitive forecasting accuracy.
Conclusions:
- Deep learning, particularly TCNs, offers effective and efficient solutions for time series classification and forecasting.
- TCNs provide a significant advantage in terms of reduced computational time.
- GANs can be effectively utilized for time series forecasting with careful implementation.
