Predicting municipal solid waste generation using artificial intelligence: A hybrid approach of entropy analysis and SHAP for optimal feature selection

Abstract

The management of municipal solid waste (MSW) is one of the primary challenges in urban areas. To improve the accuracy of waste generation predictions, this study employed a hybrid approach that integrates Mutual Information (MI) with Shapley Additive Explanations (SHAP) for effective feature selection in Artificial Intelligence (AI) modeling. The Feed Forward Neural Network (FFNN) and Long Short-Term Memory (LSTM) models were utilized. The FFNN, a shallow learning model, is simpler and effective for capturing general patterns in data, while the LSTM, a deep learning model, is more suitable for autoregressive tasks such as predicting MSW generation. The proposed hybrid approach facilitated more precise identification of the key factors influencing MSW generation and improved the prediction models. The methodology was applied to meteorological and socio-economic data from three cities: Austin in the United States, Ballarat in Australia, and Boralesgamuwa in Sri Lanka, to examine the methodology under different conditions. The dominant factors identified included population, income, the Consumer Price Index (CPI), and lagged MSW variables with lags of 5, 10, and 20 days. The modeling performance was evaluated using the Determination Coefficient (DC) and Root Mean Square Error (RMSE). In Austin, the FFNN achieved a DC of 0.7226 during training and 0.6529 during testing. In Ballarat, the FFNN achieved training and testing DC values of 0.7037 and 0.6941, respectively. In Boralesgamuwa, due to severe data limitations, the model did not train well and showed poor performance in predictions (DC and RMSE values were significantly lower). The better performance of the model in Austin could be attributed to the longer temporal coverage of the data and greater stability in socio-economic patterns, while higher variability in socio-economic factors in Ballarat may have slightly reduced the model’s accuracy. The results from Boralesgamuwa also highlight the importance of access to quality and consistent data for developing accurate models. These findings demonstrate that the MI-SHAP method can enhance prediction accuracy by identifying both linear and nonlinear relationships among variables and provide deeper insights into the dynamics governing waste generation. This methodology can aid in developing sustainable MSW management policies across various regions.