Deep LSTM-Based Transfer Learning Approach for Coherent Forecasts in Hierarchical Time Series

Area of Science:

• Data Science
• Machine Learning
• Time Series Analysis

Background:

• Hierarchical time series data are prevalent in real-world applications, posing challenges for accurate and consistent forecasting.
• Ensuring forecast consistency across different aggregation levels in hierarchical structures is complex and computationally intensive.

Purpose of the Study:

• To develop an effective Deep Long Short-Term Memory (DLSTM) auto-encoder (AE) model for hierarchical time series forecasting.
• To leverage transfer learning to mitigate the computational burden and data requirements of training DLSTM models in hierarchical architectures.

Main Methods:

• Developed a DLSTM model in an auto-encoder (AE) configuration tailored for hierarchical time series.
• Implemented a transfer learning strategy, initially training on bottom-level series and then transferring features to upper levels.
• Evaluated the DLSTM-AE approach against traditional and machine learning methods using energy and tourism datasets.

Main Results:

• The proposed DLSTM-AE approach achieved superior forecasting accuracy compared to all benchmark methods in both case studies.
• The model demonstrated an enhanced ability to produce coherent forecasts across all levels of the hierarchy.
• Transfer learning effectively reduced training time and data dependency for upper-level series.

Conclusions:

• The DLSTM-AE model with transfer learning offers a powerful and efficient solution for hierarchical time series forecasting.
• This method significantly improves both accuracy and coherence in forecasting complex hierarchical data.
• The approach provides a viable alternative to existing methods, particularly for large-scale hierarchical forecasting tasks.