A Comparative Study of Advanced Machine Learning and Deep Learning Models for Municipal Solid Waste Forecasting: A Case Study of Surat, India

Abstract

To make the city planning sustainable, especially in rapidly growing cities of the world like Surat in India, the implementation of effective waste management is crucial.  The primary factor governing this is the ability to accurately predict the quantity of municipal solid waste (MSW) likely to be generated. This study presents a comprehensive comparative analysis of various predictive models including linear regressions, kernel approaches, gradient boosting as well as the deep learning architectures. Using historical data from Surat, systematic preprocessing and feature engineering generated 419 features representing temporal, socio-economic, climatic, COVID-19, and mobility factors. The novel contribution of this study is the systematic feature engineering framework that explicitly encodes temporal structure (419 engineered features including lagged values, rolling statistics, and seasonal decomposition), enabling simple linear models to capture complex waste generation patterns. Ten distinct models, ranging from statistical approaches to machine learning and deep learning were evaluated and compared. Advanced ensemble models, including LightGBM (R² = 0.983), CatBoost (R² = 0.977), and XGBoost (R² = 0.970) demonstrated strong performance. The best-performing models (OLS and Gaussian Process Regression) achieved R² = 0.997 with Mean Absolute Percentage Error (MAPE) = 1.43%. In this study, linear models trained within a few milliseconds and achieved per-sample inference times on the order of 0.004-0.008 ms, whereas the tuned MLP and tree-based ensembles required seconds of training and millisecond-level inference, corresponding to differences of roughly two to three orders of magnitude in computational cost. Other notable performers include Lasso regression (R² = 0.979), tuned MLP (R² = 0.968), and Random Forest (R² = 0.960). The results demonstrate that feature engineering has greater influence on forecasting accuracy than model complexity, with OLS using engineered features (R² = 0.997) outperforming the MLP model (R² = 0.966) by approximately 3.1% while providing substantially faster predictions. Feature importance analysis identified lagged MSW values, rolling statistics, demographic indicators, festival effects, and COVID-19 lockdown impact as key predictors. The research finds that adoption of systematically designed feature engineering framework is a valuable tool for MSW management. The study provides comprehensive model benchmarking and practical recommendations for policymakers pursuing sustainable urban development.