Scholarlink

<p>Accurate prediction of aviation-related carbon dioxide (CO₂) emissions is critical for supporting sustainable air transport and climate policy decisions. This study proposes a hybrid physics-guided machine learning framework that integrates domain knowledge with advanced deep learning models to improve both prediction accuracy and interpretability. The dataset, obtained from EUROCONTROL, covers aviation emissions across European countries between 2010 and 2025. The proposed approach combines a physics-based baseline with deep learning architectures, including Long Short-Term Memory (LSTM), one-dimensional Convolutional Neural Network (1D-CNN), and Transformer models, using a residual learning strategy. In addition, a stacked ensemble model is developed to leverage the strengths of individual learners and enhance predictive performance. Model evaluation is conducted using regression metrics such as Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE), as well as classification metrics including Area Under the Curve (AUC) for high-emission detection. Results show that the hybrid model achieves high explanatory power (R² = 0.9986) and improved local accuracy (MAE ≈ 17,536), outperforming standalone models in interpretability and variance explanation. The ensemble model achieves the lowest prediction error (RMSE ≈ 21,948), demonstrating superior overall performance. Furthermore, explainability is ensured through SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) analyses.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A hybrid physics-guided ensemble learning framework for predicting aviation co₂ emissions with explainable artificial intelligence

  • 期刊论文

  • 发表日期：

    2026年6月15日

查看全文
  • Yıldırım Özüpak

摘要

Accurate prediction of aviation-related carbon dioxide (CO₂) emissions is critical for supporting sustainable air transport and climate policy decisions. This study proposes a hybrid physics-guided machine learning framework that integrates domain knowledge with advanced deep learning models to improve both prediction accuracy and interpretability. The dataset, obtained from EUROCONTROL, covers aviation emissions across European countries between 2010 and 2025. The proposed approach combines a physics-based baseline with deep learning architectures, including Long Short-Term Memory (LSTM), one-dimensional Convolutional Neural Network (1D-CNN), and Transformer models, using a residual learning strategy. In addition, a stacked ensemble model is developed to leverage the strengths of individual learners and enhance predictive performance. Model evaluation is conducted using regression metrics such as Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE), as well as classification metrics including Area Under the Curve (AUC) for high-emission detection. Results show that the hybrid model achieves high explanatory power (R² = 0.9986) and improved local accuracy (MAE ≈ 17,536), outperforming standalone models in interpretability and variance explanation. The ensemble model achieves the lowest prediction error (RMSE ≈ 21,948), demonstrating superior overall performance. Furthermore, explainability is ensured through SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) analyses.