Extended Abstract
Background
In recent years, population growth, industrialization, increased urbanization, lifestyle changes, and shifts in consumption patterns have caused an increase in the quantity of municipal solid waste (MSW). Improper management of MSW constitutes a serious threat to public health and the environment, leading to pollution of water, soil, and air resources and causing various diseases in humans. Given the existing problems in managing MSW, moving toward sustainable and efficient waste management and employing novel technologies is essential. In this regard, smart technologies, and artificial intelligence in particular, have attracted attention as effective tools for transforming waste management. Accurate prediction of waste generation is one of the key components of sustainable waste management, as reliable information about waste quantities is required at various stages of reduction, collection, recycling, and disposal; such information can help prevent infrastructure failures, increase efficiency, and reduce costs. Since conventional and traditional methods are unable to accurately represent the complex, multi-factor dynamics of waste generation, the use of advanced algorithms and hybrid AI-based approaches has gained increasing attention. Considering the challenges in quantitatively assessing waste and the need to develop data-driven approaches in this field, this study aims to predict the quantity of MSW in Tabriz, to present an AI-based model, and to examine the accuracy of neural network–based algorithms.
Methods
This study was carried out in Tabriz, with a population of over 1.7 million and an area of about 324 square kilometers. Daily waste generation data were obtained from the Waste Management Organization of Tabriz Municipality for the years 2019 to 2024. First, the data were checked and corrected for missing, outlier, or invalid values to ensure their accuracy and reliability for modeling. The model used was an LSTM-based neural network with an 8-layer architecture, each layer containing 50 hidden units. To predict waste quantities, a 16-time-step input sequence from the time series was fed into the model at each step, and a learning rate of 0.001 was adopted to balance learning speed and the prevention of overfitting. The training process used the Adam optimizer and was run for 600 epochs to ensure that the model reached satisfactory convergence. The model was implemented in Python using specialized libraries, including Pandas and NumPy for data processing, Matplotlib for data visualization, and the PyTorch framework for developing and training the LSTM network. These measures made it possible to design a stable and efficient model capable of extracting hidden patterns for data-driven waste management through time-series prediction. The RMSE, MAE, and MAPE metrics were used to evaluate the model's performance and accuracy.
Results
Based on the available data, the average daily MSW generation in Tabriz was approximately 980 tons. The daily tonnage data for waste were examined, and 11 implausible records were removed; these records were replaced using a 20-day moving average based on the 10 days preceding and the 10 days following each removed record. Model performance was evaluated in two phases. In the first phase, data from 2019 to 2024 were examined, with 85% of the data used for training and the remaining 15% used for testing. The RMSE and MAE values in this phase for the training and testing data were calculated as 52.58 and 72.91 tons/day, and 33.94 and 55.01 tons/day, respectively. In the second phase, data from 2024 were also incorporated into the model; similarly, 85% of the data from 2019 to 2024 were used for training and 15% for testing. The RMSE, MAE, and MAPE values for the training and testing data were 50.62 and 70.47 tons/day, 32.61 and 53.02 tons/day, and 3.52% and 5.86%, respectively. The low MAE values indicate that the model's prediction errors were small compared to the actual data and that the model performed well in predicting the overall trend of waste generation. The RMSE values were also within a reasonable range, indicating that the model demonstrated adequate robustness against sudden fluctuations and outliers. The low MAPE values likewise indicate acceptable results.
Conclusion
This study developed an LSTM-based model to predict the quantity of MSW in Tabriz. The results show that neural networks can identify complex patterns, fluctuations, and nonlinear relationships in waste generation. Nevertheless, challenges remain in using neural networks to predict waste quantities. Such challenges include the model's dependence on the accuracy and validity of the data, limited data availability, and difficulty in determining an optimal network structure. Extending the length of the data period could improve the model's long-term predictive performance. In general, the LSTM model may serve as an effective approach for forecasting MSW generation and can assist urban managers in more precise planning for waste reduction, collection, transportation, recycling, and disposal.
Practical Implications of Research
The practical implications of this research include supporting more proactive and data-driven municipal solid waste management through the prediction of waste quantity using artificial intelligence. The application of the LSTM model can support urban decision-makers in anticipating future collection and treatment requirements, allocating resources efficiently, and improving recycling and resource recovery strategies. However, the practical implementation of the model requires the establishment of a coherent data infrastructure, continuous and standardized data collection, validation of the model under real-world conditions, and regular updating in response to changes in waste generation patterns. Overall, AI-based prediction can serve as a decision-support tool to shift municipal solid waste management from reactive decision-making toward a more data-driven, proactive, evidence-based, and sustainable approach.