Leveraging Mutual Information in Algorithm-Assisted Dynamic Feature Selection for Short-term SOx Emissions Forecasting

Abstract

Traditional feature selection methods often impose fixed constraints on the size of feature subsets, creating a notable gap between academic research and industrial applications. To address this issue, this research study introduces Algorithm-Assisted Dynamic Feature Selection, referred to as AADFS, which synergizes genetic algorithms and simulated annealing with mutual information to optimize the trade-off between exploration and exploitation. This allows for the dynamic identification of the most relevant features without rigid constraints. AADFS has demonstrated improvements in prediction accuracy, especially in the short term, and reductions in computational complexity across a variety of datasets, ranging from corporate emissions data to publicly available datasets, including solar energy forecasts, meteorological patterns, electricity consumption, and financial market trends. AADFS aims to offer valuable insights for researchers and industry professionals seeking to transform theoretical advancements into practice.

Date of Award	2025
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology
Supervisor	Dit Yan YEUNG (Supervisor) & Ki On Ng (Supervisor)