Regression based prediction of higher heating value for refuse-derived fuel using convolutional neural networks predicted elemental data and spectrographic measurements

Abstract

Higher heating value (HHV) is the key parameter for replacing Refuse-Derived Fuel (RDF) with fossil fuels in the cement industry. HHV can be measured with a bomb calorimeter or predicted from direct elemental data by using regression models. Both methods require the continuous use of special laboratory equipment and are time consuming. To overcome these limitations, this study aims to predict the HHV value of RDF from predicted elemental data by using regression models. Therefore, once the predicted elemental data are generated, there will be no need to have continuous elemental data to predict HHV. Predicted elemental data were generated from direct elemental data and Near Infrared (NIR) camera-based spectrometric data by using a deep learning model. A convolutional neural networks (CNN) model was used for deep learning and was trained with 10,500 NIR image samples, each of which was 28×28×1. Different regression models (Linear, Tree, Support-Vector Machine, Ensemble and Gaussian process) were applied for HHV prediction. According to these results, higher R² values (>0.85) were obtained with Gaussian process models (except for the Rational Quadratic model) for the predicted elemental data. Among the Gaussian models, the highest R² (0.95) but the lowest Root Mean Square Error (RMSE) (0.0563), Mean Squared Error (MSE) (0.0317) and Mean Absolute Error (MAE) (0.0431) were obtained with the Mattern 5/2 model. The results of predictions from predicted elemental data were compared to predictions from direct elemental data. The results show that the regression from predicted elemental data has an adequate prediction (R²=0.95) compared to the prediction from the direct elemental data (R²=0.99).

Graphical abstract