Application of Artificial Neural Network for Solid Waste Projection and Energy Content Evaluation

Abstract

Forecasts of the heterogeneous municipal solid waste (MSW) generation are vital for sustainable MSW management. Artificial neural network (ANN) models have been successfully demonstrated to predict complex MSW trends, but the negligence of the MSW's heterogeneous characteristics hinders the further application of the predictions. This study aims to adopt robust ANN models coupling Bayesian hyperparameter optimization and uncertainty analysis to forecast the heterogeneous MSW generation in a country with further application in energy content evaluation. The impact of the hyperparameter optimization is illustrated by comparing the forecast uncertainties of ANN models using the Bayesian-optimized hyperparameters and default hyperparameters using ensemble forecasts. The relative standard deviations of the ensemble forecasts show that overfitting is more susceptible in the default models (11.1% -44,400%) than in the Bayesian-optimized models (3.64% -27.7%). By using the Bayesian-optimized models, Malaysia's heterogeneous MSW generation is projected individually based on its physical composition. The total MSW generation in Malaysia is expected to grow by 12% from 2020 to 2030, with food waste as the dominant composition (44%). The energy content of the overall MSW is evaluated based on the lower calorific value of the forecasted MSW. The forecasted increase in food waste generation reduces the lower caloric value of Malaysia's MSW below the lower limit to burn without supporting fuel (i.e., 6.5 MJ/kg). An alternative scenario where food waste is segregated from the overall MSW shows an increase in the lower calorific value from 5.874 MJ/kg to 10.31 MJ/kg. Without food waste segregation, the development of organic waste treatment facilities in Malaysia should be emphasized instead of incinerators. This study not only offers useful information for capacity planning of MSW treatment facilities using a data-driven approach, but it also broadens the scope for future related research by examining the application of the forecasted outcomes using ANN models.