Predicting building-related carbon emissions: A test of machine learning models

This chapter evaluates and compares the performance of six machine-learning (ML) algorithms in predicting China’s building-related carbon emissions. The models took into account five input parameters influencing building-related CO₂ emissions: urbanisation, R&D, population size, GDP, and energy use. The study used quarterly data throughout 1971Q1–2014Q4 to develop, calibrate, and validate the models. Each model was developed using 140 observations and validated on 36 observations. In tuning each ML model for comparative purposes, 10-fold with cross-validation approach was used in selecting the optimal hyperparameters and their associated arguments. The results indicate that the random forest (RF) model attained the highest coefficient of determination (R²) of 99.88%, followed by the k-nearest neighbour (KNN) (99.87%), extreme gradient boosting (XGBoost) (99.77%), decision tree (DT) (99.63%), adaptive boosting (AdaBoost) (99.56%), and the support vector regression (SVR) model (97.67%). Overall, the RF algorithm is the best performing ML algorithm in accurately predicting building-related CO₂ emissions, whereas the best algorithm in terms of time efficiency is the DT algorithm. The KNN model is highly recommended when practitioners want to have accurate predictions in a timely manner. RF, KNN, and DT models could be added to the toolkits of environmental policymakers to provide high-quality forecasts and patterns of building-related CO₂ emissions in an accurate and real-time manner.