TL;DR
This paper presents a comparative study of four time-series forecasting models—SARIMAX, Prophet, LSTM, and GRU—to predict municipal solid waste (MSW) generation. By incorporating external factors like holidays, weekends, and weather patterns into a multi-year dataset from Sri Lanka, the researchers found that LSTM and SARIMAX offer the most robust performance for urban resource planning.
What Problem Does This Paper Address?
As urban populations swell, "trash management" has shifted from a logistical chore to a critical environmental concern. Efficient waste management relies on accurate forecasting: if you know how much waste will be generated, you can optimize collection schedules, save fuel, and reduce the environmental footprint of landfills.
However, many existing forecasting models suffer from "data myopia." They often rely on small datasets or ignore the "human element"—the fact that waste generation isn't just a linear trend but is heavily influenced by holidays, weekends, and seasonal weather changes. This paper aims to bridge that gap by conducting a factor-wise analysis of how these external variables impact waste production and which modeling architectures capture these nuances best.
Methods
The authors compared four distinct modeling approaches:
- SARIMAX: A statistical model that combines seasonal autoregression and moving averages with "exogenous" (external) variables like holiday flags.
- Prophet: An additive model developed by Meta (Facebook) designed to handle outliers, missing data, and dramatic seasonal shifts.
- LSTM (Long Short-Term Memory): A type of Recurrent Neural Network (RNN) capable of learning long-term dependencies in sequential data.
- GRU (Gated Recurrent Unit): A more computationally efficient "cousin" of the LSTM that simplifies the gating mechanism while maintaining similar performance.
The researchers didn't just feed raw data into these models. They employed a robust preprocessing pipeline, using Isolation Forests for anomaly detection and linear interpolation to handle missing values. They also explicitly integrated temperature, humidity, and holiday binary flags as features.
Datasets
The study utilized a real-world MSW dataset from Boralasgamuwa, Sri Lanka, spanning seven years (2012–2018).
- Training Period: 2012–2017.
- Testing Period: 2018.
- Enhancements: The authors enriched the original dataset with historical weather data (temperature and humidity) from Weather Underground and national holiday schedules to create a more comprehensive feature set.
Key Insights and Ideas
- The "Holiday Dip": Contrary to what one might expect from consumer-heavy holidays, the data showed a decrease in waste generation during holidays and weekends. This is likely due to reduced institutional and commercial operations during these periods.
- Seasonality Trumps Direct Weather Correlation: While temperature and humidity are associated with waste generation, they don't have a simple linear relationship. Instead, they act as proxies for broader seasonal behaviors. Waste peaks during Summer and Winter but dips during the Monsoon and Post-Monsoon months.
- Architectural Strengths:
- SARIMAX excelled at capturing explicit, recurring weekly patterns because of its dedicated seasonal component.
- LSTM proved superior at capturing the complex, non-linear "noise" of daily human activity, making it the best overall predictor for the full year.
- Prophet's Struggle: Despite its popularity for business forecasting, Prophet showed the poorest performance in this specific domain, suggesting that MSW data might be too volatile for its additive approach.
Key Results
The models were evaluated using standard metrics: Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and the coefficient of determination ($R^2$).
- Overall Performance: LSTM was the clear winner for general forecasting, achieving the highest $R^2$ of 0.9329 and the lowest MAPE of 4.62%.
- Seasonal Accuracy: SARIMAX was the most accurate when looking specifically at seasonal averages, with a near-perfect $R^2$ of 0.9955 in that category.
- Deep Learning vs. Statistical: Both LSTM and GRU significantly outperformed Prophet, proving that deep learning is well-suited for the irregularities of municipal waste data.
Limitations and Future Work
The authors acknowledge that while their model is robust, it is currently limited to a single geographic area. Future research directions include:
- Hybrid Models: Combining the statistical rigor of SARIMAX with the non-linear learning capabilities of LSTMs.
- Socio-Economic Features: Integrating data on household income, employment rates, and GDP per capita to see how economic shifts influence the "trash bin."
- Granularity: Expanding the study to include more diverse administrative units to test the models' generalizability across different urban layouts.
For a deeper dive into the comparative assessment of how this work fits into the broader field, see the authors' summary of existing literature:
Summary updated