Time Series Forecasting Model Based on the Adapted Transformer Neural Network and FFT-Based Features Extraction.

Area of Science:

• Data Science
• Machine Learning
• Signal Processing

Background:

• Accurate time series forecasting is vital across finance, climatology, and engineering.
• Neural networks struggle with sensor data due to volume, noise, and long-term dependencies.
• Existing models face challenges with diverse sensor data characteristics.

Purpose of the Study:

• To enhance time series forecasting accuracy for sensor-collected data.
• To address limitations of current neural network models in handling complex time series.
• To improve predictive performance by incorporating frequency-domain information.

Main Methods:

• Proposed an adapted transformer architecture for time series prediction.
• Introduced a data preprocessing method using fast Fourier transform (FFT) to convert time-domain to frequency-domain.
• Enriched data with complex-valued frequency-domain features to boost informational content.

Main Results:

• The proposed model demonstrated superior performance across three diverse sensor datasets.
• Achieved higher accuracy compared to state-of-the-art models like LSTM, DeepAR, and Transformer.
• Consistently outperformed existing methods across five distinct performance metrics.

Conclusions:

• The adapted transformer model with FFT preprocessing significantly improves time series forecasting accuracy.
• The method effectively handles challenges posed by large volumes, noise, and long-term dependencies in sensor data.
• This approach offers a robust solution for accurate forecasting in data-driven applications.