Enhanced Long Short-Term Memory for Landfill Area Estimation Based on Domestic Solid Waste Prediction

Abstract

An innovative approach tackles the challenge of domestic solid waste generation prediction through machine learning techniques. To overcome limitations in capturing complex temporal patterns faced by conventional Long Short- Term Memory (LSTM) models for time series forecasting, an enhanced version called e-LSTM is introduced. The e-LSTM model incorporates key improvements to address standard LSTM shortcomings. A hybrid activation function named SigmoRelu is introduced to boost the model’s ability to capture intricate time series patterns. Additionally, the Radam optimizer is adopted to optimize the learning process and enhance convergence. Dropout layers are integrated within the LSTM architecture to counter overfitting and ensure robust generalization to new data. Extensive experiments compare the e-LSTM model’s performance against standard LSTM and GRU models, highlighting its significant advancement. The e-LSTM model achieves superior predictive accuracy in waste generation forecasting compared to standard LSTM and GRU models.  Overall, the proposed e-LSTM model marks a substantial step forward in domestic solid waste prediction, effectively addressing the limitations of traditional LSTM models. The combination of SigmoRelu activation, Radam optimization, and dropout mechanisms yields a robust and accurate predictive framework. Empirical results confirm the model’s superiority, positioning it as a valuable tool for waste management applications and decision-making.