Forecasting waste accumulation at the Discharge Point (DP) level remains a challenging task due to the high variability, sparsity and irregularity of filling patterns in Municipal Solid Waste (MSW) systems. This paper investigates the feasibility of applying Artificial Neural Network (ANN) based models, including feedforward ANN, Long Short-Term Memory (LSTM) and Convolutional Neural Networks (CNN) architectures, to predict DP level waste filling behaviour using a real-world dataset comprising 3.85 million records collected from 200 DPs over a four-year period. While existing studies on MSW forecasting predominantly focus on aggregated spatial scales, such as facilities or districts, this work explicitly adopts a DP-centred perspective to capture the pronounced variability and irregularity inherent to individual waste disposal units. A systematic experimental framework is proposed consisting of five scenarios designed to evaluate the impact of different preprocessing strategies, including sliding window transformation, normalisation, interpolation, outliers¿ detection, and both temporal and structural aggregation. The findings demonstrate that LSTM and CNN-based models, combined with temporal aggregation intervals of 3 and 4 hours, significantly improve predictive performance, while structural aggregation across multiple DPs further enhances model stability achieving R² values up to 0.89. These findings highlight the critical role of data preprocessing and aggregation strategies in improving forecasting reliability for highly irregular DP level time series, providing practical insights for the development of decision support systems in MSW management.
