A Hybrid Framework Integrating Traditional Models and Deep Learning for Multi-Scale Time Series Forecasting.

Area of Science:

• Data Science
• Machine Learning
• Statistical Modeling

Background:

• Accurate time series forecasting is vital for decision-making across various fields.
• Existing methods struggle to achieve high accuracy for both short-term and long-term predictions.
• Integrating traditional statistical models with deep learning offers potential for improved forecasting.

Purpose of the Study:

• To propose a general hybrid forecasting framework integrating statistical and deep learning models.
• To enhance time series prediction accuracy by capturing both short-range patterns and long-range dependencies.
• To offer a robust and interpretable forecasting solution.

Main Methods:

• Developed a hybrid forecasting framework combining ARIMA with deep learning models (LSTM, Transformer).
• Implemented a multi-scale prediction mechanism and a dual-stage forecasting process.
• Fused outputs from statistical and deep learning components using an adaptive mechanism.

Main Results:

• The hybrid framework consistently outperformed stand-alone ARIMA, LSTM, Transformer, Informer, and Autoformer models.
• Achieved state-of-the-art accuracy on eight diverse public datasets for both short-horizon and long-horizon forecasts.
• Ablation studies confirmed the significant contribution of each module within the framework.

Conclusions:

• The proposed hybrid approach offers a promising direction for combining statistical and deep learning techniques in time series forecasting.
• The framework demonstrates superior accuracy, interpretability, and robustness compared to existing methods.
• This integrated approach effectively addresses the challenges of accurate forecasting across different time scales.