Machine Learning to Forecast the Quantities of Construction and Demolition Wastes: A Case Study on the City of Austin

Abstract

This thesis presents a machine-learning based framework to forecast construction and demolition waste (C&DW) quantities in the City of Austin using building- permit data from 2016 to 2024. The study was motivated by the need for more reliable forecasting tools to support long-term planning of recycling infrastructure, landfill capacity, and waste-management logistics. After data cleaning and preprocessing, multiple predictive models were evaluated, including Linear Regression, Neural Network, Random Forest, and Gradient Boosting Machine (GBM). The results showed that outlier treatment using the interquartile range (IQR) method substantially improved predictive performance and model stability. Among the tested methods, tree-based ensemble models consistently performed best, with Random Forest emerging as the strongest overall model and GBM as a close alternative. The framework was then enhanced by integrating contextual demographic and economic variables, including tract-level population density, population change, diversity index, consumer price index, and consumer confidence index. This enrichment improved model performance and showed that C&DW generation depends not only on project-level characteristics but also on broader neighborhood and market conditions. Finally, the predicted permit-level quantities were translated into a baseline origin-destination (OD) allocation, showing that predicted waste flows can be assigned efficiently while also revealing strong concentration in both waste generating zones and receiving facilities.