A Multi-objective transfer learning framework for time series forecasting with Concept Echo State Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Time Series Analysis

Background:

• Time series forecasting faces challenges in feature extraction and knowledge transfer, especially with heterogeneous data.
• Existing methods often require extensive hyperparameter tuning and struggle with diverse, limited, or missing data.

Purpose of the Study:

• Introduce a novel transfer learning framework for time series forecasting.
• Address challenges in feature extraction and knowledge transfer across heterogeneous datasets.
• Enhance forecasting performance and efficiency.

Main Methods:

• Utilized Concept Echo State Network (CESN) with a multi-objective optimization strategy.
• Optimized CESN for individual data sources to extract targeted features.
• Implemented a multi-network architecture for knowledge sharing among Echo State Networks (ESNs).

Main Results:

• Successfully extracted targeted features capturing unique dataset characteristics.
• Achieved improved forecasting performance through effective knowledge sharing.
• Reduced the need for extensive hyperparameter tuning by optimizing specific CESN parameters.

Conclusions:

• The proposed CESN-based transfer learning framework offers a promising solution for time series forecasting.
• The framework is particularly effective for datasets that are diverse, limited, or contain missing values.
• This approach enhances both the performance and efficiency of time series forecasting models.