Data-driven heterogeneous ensemble forecasting framework incorporating multi-scale criteria

Hongmin Li

doi:10.1016/j.eswa.2025.130537

Data-driven heterogeneous ensemble forecasting framework incorporating multi-scale criteria

2025-11-29

0

PRE

AI

DOI：10.1016/j.eswa.2025.130537

Original

Original request for help

Save

Abstract

En 中文

The increasing forecasting demands of complex system pose significant challenges for individual deep learning model to effectively capture the multifaceted complexity and inherent uncertainty of data. Conventionally, the ideology of ‘decomposition and integration’ has been universally adopted to eliminate the temporal complexity, yet neglecting the validity of multi-scale decomposition and heterogeneity of multi-scale features. Based on these limitations, this study proposes a data-driven heterogeneous ensemble forecasting framework. Specifically, multi-scale decomposition is conducted based on constrained variational approach, center frequency method and Gold Rush Optimization algorithm, then decomposed subsequences are aggregated as different time-scale criteria by Lempel-Ziv algorithm. In addition, a scale-granularity adapted learning strategy is introduced in this study to extract the differentiated characteristics. Moreover, by integrating the concept of transfer learning, a graph-attention mechanism is embedded into the VAR model to forecast the uncertain trends from neighboring nodes in complex system. Empirically, experimental validations on air pollution datasets from the Harbin-Changchun agglomeration demonstrate that the proposed forecasting framework exhibits remarkable accuracy and stability. This framework offers a critical and effective approach to high-precision prediction in domains such as energy consumption and environmental monitoring.

Data-driven heterogeneous ensemble forecasting framework incorporating multi-scale criteria

Abstract

Journal

Organization