A hybrid machine-learning model for predicting the waste generation rate of building demolition projects

Abstract

Information on waste generation rate (WGR) is useful for waste management. Recently, several studies have been conducted to predict WGR using artificial intelligence (AI) to the end of realizing smart waste management. Additionally, to improve the performance of machine learning (ML) predictive models, several strategies have also been tested of recent. This study aimed to develop a hybrid ML predictive model to enhance prediction performance for small datasets consisting mainly of categorical variables. Artificial neural network (multi-layer perceptron) (ANN (MLP)) and support vector machine regression (SVMR) algorithms were selected, and categorical principal components analysis (CATPCA) was applied. Accordingly, four predictive models—ANN (MLP), SVMR, CATPCA–ANN (MLP), and CATPCA–SVMR—were developed. The CATPCA–ANN (MLP) model showed some improvements in statistical metrics as compared to the ANN (MLP) model, and the CATPCA–SVMR model showed a far better performance across all statistical metrics than the SVMR model. The best prediction performance was found in the CATPCA–SVMR model (R2 = 0.594, R = 0.770), which was thus considered the best model of the four developed. Here, the mean DWGR was 1165.04 kg/m2 for the observed values, and that for the predicted values was 1161.52 kg/m2. Thus, a novel method was proposed for developing a hybrid ML model to enhance prediction performance for small datasets consisting of categorical variables. The results of this study enable the use of ML algorithms, which are disadvantageous with respect to use of categorical variables, by using CATPCA, and we suggest a new AI approach to develop a predictive DWGR model with excellent predictive performance.