Machine learning based modelling for lower heating value prediction of municipal solid waste

Abstract

Municipal solid waste (MSW) is a heterogeneous and complex fuel. Time-resolved knowledge of its physical and/or chemical properties is key to ensure stable Waste-to-Energy (WtE) plant operation. For this purpose, Gaussian processes regression (GPR) models were developed to predict the daily lower heating value of MSW using historical data from a WtE plant, together with weather and calendar data. The training dataset consisted of 730 observation points between January 2017 and December 2018, and the validation dataset had 294 observation points between January and October 2019. Both the unoptimized GPR and the hyperparameter optimized GPR models developed in this study showed better prediction accuracy than the models reported in the literature, achieving mean absolute errors (MAE) of 0.592 and 0.688 MJ/kg and mean absolute percentage errors (MAPE) of 5.23 and 6.05%. For the first time, online process data were utilized for MSW lower heating value prediction, freeing the model development from laborious ultimate and proximate analysis or waste fractionation. The GPR model proposed is not only superior according to the accuracy indicators but can also be used in online operation and learn from new data as opposed to the static models found in literature. In future work, the model could be extended to include more response variables to extract more information to be used in process optimization and control.