Image-based time series forecasting: A deep convolutional neural network approach.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Time Series Analysis

Background:

• Deep learning has shown significant success in computer vision tasks.
• Traditional time series forecasting relies on numerical data representations.
• There is a need for novel approaches to improve forecasting accuracy.

Purpose of the Study:

• To introduce ForCNN, a new deep learning method for univariate time series forecasting.
• To explore the use of visual representations (images) of time series data for forecasting.
• To evaluate the performance of ForCNN against established forecasting models.

Main Methods:

• Developed ForCNN, a deep neural network combining convolutional and dense layers.
• Utilized image-based input for time series data, converting numerical data into visual representations.
• Examined three deep convolutional neural network architectures: VGG-19, ResNet-50, and a custom design.

Main Results:

• ForCNN successfully generated point forecasts using image-based time series data.
• Performance was evaluated on benchmark datasets from the M3 and M4 forecasting competitions.
• The proposed image-based method demonstrated superior performance compared to standard and state-of-the-art models.

Conclusions:

• Image-based time series forecasting using deep learning is a viable and effective approach.
• ForCNN offers a novel perspective on time series analysis by leveraging visual data processing.
• This method has the potential to advance the field of forecasting accuracy.