Scholarlink

<p>This study evaluates three decision tree-based regression models, namely CART, CHAID, and Exhaustive CHAID, for predicting the gross calorific value of biomass and municipal solid waste from ultimate analysis data. To achieve the best predictive accuracy and outcomes, multiple critical parameters such as tree depth, node size, and complexity parameters were systematically tuned in the optimization of the model. Out of the three models, the regression CART model outperformed the rest, with the lowest MAE (mean absolute error) of 0.663 (training) and 1.071 (testing), MSE (mean squared error) of 1.166 (training) and 2.517 (testing), and highest coefficient of determination (R<sup>2</sup> = 0.950 for training and 0.865 for testing). Exhaustive CHAID model came at second rank with R<sup>2</sup> value of 0.880 (training) and 0.832 (testing) while CHAID model was slightly less accurate. The predicted GCV is most strongly influenced by carbon (C) and oxygen (O), according to partial dependence plots, with C exerting a positive influence and O a negative one. Both the SHAP and perturbation-based sensitivity analyses established O and C as the two most dominant predictors, with O showing the highest perturbation range (≈ 16) followed by C (≈ 9.5). Overall, the CART model exhibited robustness, interpretability, and data efficiency, resulting in accurate GCV estimation for a wide and range of biomass and MSW samples.</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

Explainable tree-based machine learning models for predicting gross calorific value of various waste from ultimate analysis

  • 期刊论文

  • 发表日期：

    2026年3月9日

查看全文
  • Chaitanya Nidhi,
  • Prabhat Kumar Singh

摘要

This study evaluates three decision tree-based regression models, namely CART, CHAID, and Exhaustive CHAID, for predicting the gross calorific value of biomass and municipal solid waste from ultimate analysis data. To achieve the best predictive accuracy and outcomes, multiple critical parameters such as tree depth, node size, and complexity parameters were systematically tuned in the optimization of the model. Out of the three models, the regression CART model outperformed the rest, with the lowest MAE (mean absolute error) of 0.663 (training) and 1.071 (testing), MSE (mean squared error) of 1.166 (training) and 2.517 (testing), and highest coefficient of determination (R2 = 0.950 for training and 0.865 for testing). Exhaustive CHAID model came at second rank with R2 value of 0.880 (training) and 0.832 (testing) while CHAID model was slightly less accurate. The predicted GCV is most strongly influenced by carbon (C) and oxygen (O), according to partial dependence plots, with C exerting a positive influence and O a negative one. Both the SHAP and perturbation-based sensitivity analyses established O and C as the two most dominant predictors, with O showing the highest perturbation range (≈ 16) followed by C (≈ 9.5). Overall, the CART model exhibited robustness, interpretability, and data efficiency, resulting in accurate GCV estimation for a wide and range of biomass and MSW samples.