A Novel IoT-based deep learning framework for real-time waste forecasting: Optimizing multi-waste categories using AutoML

Abstract

Municipal solid waste (MSW) management faces significant challenges in waste generation forecasting due to insufficient data and limitations of traditional predictive methods, such as ARMA and ARIMA, which struggle with nonlinear, high-frequency data. This study introduces an IoT-based deep learning framework to enhance forecasting accuracy by utilizing real-time data from 3052 smart recycling bins across three Chinese cities. A Bi-LSTM model was applied to predict multiple waste streams daily, with AutoML techniques optimizing performance through automated hyperparameter tuning and model selection. Analysis 14,039,838 waste records, the Bi-LSTM achieved a mean absolute percentage error (MAPE) of 18.7 %, successfully predicting six waste categories at a daily frequency. By leveraging large-volume, granular, and frequently updated IoT data, this approach enables dynamic waste management, optimizing collection routes and reducing costs. The results highlight the potential of IoT and AI integration for advanced waste management, providing substantial support for operational efficiency and informed policy decision-making.