MSW-DeepStack: Innovative Municipal Solid Waste Prediction Model for Informed Decision-Making

This section explores the latest advancements and applications of ML and DL in predicting the MSW generation rate, paving the way for more sustainable and viable solutions. The municipal solid waste generation in Thailand was predicted using the grey model (GM) by Li et al. [13]. Multivariate and univariate applications of GM were performed, with factors like population density, GDP per capita, household expenditure, and household size taken into account. The models with the smallest Mean Absolute Percentage Error (MAPE) were GM(1,1) and GM(1,3), with values of -0.1 and 2.124, respectively. The study estimates a 10–25% rise in garbage generation from 2018 to 2030. MSW was quantified by Mensah et al. [14] using data gathered from both residential and commercial locations in Ethiopia. When compared to other models, the Multi-Linear Regression model fared the best with a coefficient of determination of 0.72. The GDP per capita is inversely proportional to the number and size of municipal solid waste disposal facilities, according to research by Khanal [15]. The facility type most desired varies from region to region, with some areas preferring incineration and others preferring landfills. The study highlights the need for updated waste management policies, procedures, technology, and administration in China.

The Waste Management Output Index (WMOI), the Diversion Gross Domestic Product (DGDP) ratio and the Current spending per ton handled (CuPT) were applied to evaluate the efficiency of waste management systems in Canada [16]. To facilitate the comparison of jurisdictions, an additive weighting technique was adopted. The study observed increasing WMOI tendencies in most regions aside from Nova Scotia and decreasing DGDP ratio tendencies across all jurisdictions except for Nova Scotia. In addition, CuPT tendencies showed that Saskatchewan and Alberta were more productive than other provinces. Kathmandu Metropolitan City (KMC) in Nepal utilized linear regression analysis to examine the rising trend of the MSW generation rate [17]. There is a strong relationship between population size and the amount of trash produced. Using Artificial Neural Network Multi-Layer Perceptron (ANN-MLP), Support Vector Regression (SVR), and Random Forest (RF) with an autoencoder (AE), Islam et al. [18] proposed a hybrid ML model to forecast demolition waste management. The hybrid model, especially the AE-ANN model, significantly improved the performance of the ANN model, reducing the error rate by 27%.

A comprehensive survey analyzed ML techniques implemented for managing waste within smart cities, focusing on waste generation and disposal phases [19]. The survey exposed various challenges in this field, including the lack of real-time data availability, the absence of standardized benchmarking tests for evaluating ML models, and the need for well-defined long-term waste management plans. Islam et al. [18] conducted a study to predict yard waste generation in Winnipeg, Canada, using quarterly data [20]. They implemented the grey models GM(1,1) and GM(1, N) to predict individual and multivariate factors, respectively. Both models outperformed linear and non-linear models, achieving low MAPE values. The MAPE for in-sample data ranged from 0.06% to 10.39% for GM(1,1) and from 5.64% to 7.54% for GM (1, N). Lu et al. developed the WGMod, an advanced ML model to predict MSW generation in China [21]. WGMod identified key influencing factors such as annual precipitation, population density, and annual mean temperature, achieving an R² value of 0.939.

Bayesian-optimized ANN models with ensemble uncertainty analysis were used to predict heterogeneous MSW generation rates [22]. The study demonstrated a correlation between MSW physical composition and other indices, resulting in more reliable predictions with lower relative standard deviations. The study report suggests that 42,873 tons per day (t/d) of municipal solid waste will be produced in Malaysia by 2030, with 44% of it being food waste. Wu et al. [23] proposed an optimized ANN coupled with the Particle Swarm Optimization (PSO) algorithm to forecast waste quantities in Poland. The ANN-PSO model outperformed the conventional ANN model in terms of the coefficient of efficiency (CE)-0.11. In Vietnam, the solid waste generated from selected residential areas was evaluated using six different ML models [24]. Among them, Random Forest (RF) and k-nearest neighbors (KNN) performed better with an R² of 0.96-0.97.

Cubillos [25] developed and optimized an ANN model to predict municipal solid waste (MSW) generation rates in mainland China, considering regional variations. The study found that regional differences had a major impact on MSW prediction and suggested that building out regional models independently might improve predictive capability. After dividing the regions into southern, northern, and western, the outcome improved from R² 0.916 and RMSE 59.3 to R² 0.968/0.946/0.943 and RMSE 6.4/9.7/17.6, respectively. In Nagpur, India, the monthly data was analyzed using non-linear autoregressive neural models [26]. A couple of NAR models were set with different hidden layers, and both models performed better with lower absolute maximum errors of 6.45% and 3.05%. A multi-site Long Short-Term Memory (LSTM) neural network was utilized to analyze historical data of weekly waste weights from households in Herning, Denmark [27]. The study demonstrated that the adoption of a multi-site approach enhances the forecasting performance of the LSTM model by an average of 28%. Moreover, the LSTM models outperformed traditional methods such as ARIMA, demonstrating an average improvement of 85%.

Kannangara et al. [28] compared six different ML models to forecast MSW in India. A hybrid model of Genetic Algorithm (GA) and ANN was found to be the best performer with an R² value of 0.87. The fuzzy logic method was utilized to investigate the socio-economic components of MSW generation in China [29]. Due to the limited dataset, the GM (1, 1), Linear Regression, and ANN models were chosen for accurate forecasting. The GM (1, 1) model is suitable for short datasets as it accounts for the lack of social and other predictor values, while Linear Regression performs well on larger historical datasets. On the other hand, ANN is effective for short datasets, capturing patterns and data relationships to generate outputs. The ANN model and Decision Tree (DT) were implied to evaluate MSW, paper, leaf, yard, and kitchen organic wastes, along with socio-economic factors [30]. The ANN model performed better with the R² of 0.72 for MSW, and the population, education and income were highly correlated with the waste generation rate. To predict the generation of MSW, a waste prognostic tool, regression, and time series analysis were employed [31]. The authors concluded that the biodegradable waste content would be the fraction with the highest percentage.

To tackle the escalating issue of MSW, it is evident that ML and DL techniques have become indispensable instruments. These approaches offer a comprehensive array of tools to enhance the precision of waste prediction and uncover fundamental influencing elements. By employing these methodologies to navigate intricate datasets, scholars are striving to unveil intricate correlations between waste generation and diverse socio-economic and environmental variables.

Despite the challenges caused by digitization and the COVID-19 pandemic, all fourteen datasets exhibited a downward trend. However, they remain a major factor in the production of MSW. Figure 1 shows the total waste generation from 2003 to 2023, while Figure 2 shows the average annual rate of recycling. Over this period, both garbage generation and recycling rates rose gradually from 2003 until 2013, after which they leveled off. In 2013, the global recycling rate peaked at 62%. By 2014, however, both waste produced and recycling rates had decreased by 4.3% and 7.4%, respectively. In the same year, there was a significant drop in the amount of construction and demolition (C&D) waste recycled, which accounted for 25% of the overall decline in recycling compared to that of 2013. This drop was the first in a decade, and it was the result of the NEA’s persistent advocacy for waste recycling programs.

1.png

Figure 1. Total waste generated in Singapore between 2003 and 2023

2.png

Figure 2. Average waste recycling rate of Singapore between 2003 and 2023

Due to extrinsic factors such as the COVID-19 pandemic, subsequent years witnessed additional fluctuations. Solid waste generation in Singapore fell for the second year in a row in 2020, down to about 5.88 Million Metric Tons (MMT) from 2019. This represented a 19% decrease from 2019. The total amount was almost identical to that which was discarded in 2008. There was a decline from 73% in 2019 to 68% in 2020 in the non-domestic sector and a decline from 17% to 13% in the domestic sector over the same time. The worldwide outbreak created an economic slowdown in 2020, which in turn affected the generation and recycling of ferrous metal scrap and construction and demolition debris.

In contrast, the production of waste reached 6.94 MMT in 2021, an 18% increase over 2020. The rise was attributed to the increase in economic activity. While the aggregate recycling rate for waste generated in 2021 increased to 55% from 52% in 2020, it did not reach the 2019 level before the pandemic. In the year-on-year analysis, the non-domestic sector envisioned its recycling rate rising from 68% to 70%, while the domestic sector remained at 13%. More waste was generated in Singapore in 2022, the second consecutive year of increase as economic activity continued to pick up. Both non-domestic and residential waste production rose from 2021 to 2022, from 5.12 Mt to 5.53 Mt and 1.82 MMT to 1.86 MMT, correspondingly. In 2022, the recycling rate increased to 57% from 55% the previous year. This rise is attributable to the greater amounts of construction and demolition waste generated by an increase in demolition projects. Thus, these inferences highlight that the waste generation rate would increase in tandem with the economy and population.

3.png

Figure 3. Total waste generated based on the category

4.png

Figure 4. Average waste recycling rate based on the category

Upon analyzing Figure 3 and Figure 4, it becomes apparent that Paper & Cardboard (P&C) waste constitutes a significant portion, amounting to approximately 25 MMT. However, only half of this waste is recycled, while the rest is either sent to landfills or waste-to-energy convertible plants. Similarly, the quantities of ferrous metals and C&D waste are almost equal, each contributing around 22 MMT. Subsequently, their recycling rates are also similar, at approximately 96% and 98%, respectively. On the other hand, used slag demonstrates a high recycling rate of 97%. Out of a total waste generation of 6.54 MMT, a substantial amount (6.3 MMT) is being recycled, showcasing efficient recycling practices for this waste material. Additionally, plastic waste accounts for an estimated 16.75 MMT of total waste production, of which only 1.67% was recycled during the 20 years.

Despite the issue of food insecurity affecting a segment of the population, it is noteworthy that the overall food waste in Singapore remains significant at approximately 14 MMT. However, only 13% of the food waste was recycled, indicating that a considerable amount of food is being discarded and wasted. The recycling rate of wood and timber waste is 63%, with a total waste production of 6.73 MMT, almost 4.67 MMT of which is recycled. The horticultural waste production rate was 6 MMT, of which 2.4 MMT of waste was sent to landfills. Similarly, stones, ceramics, and rubber wastes were produced in a lesser amount of 5.66 MMT, of which only 4% were recycled. This waste type has the lowest recycling rate among both domestic and non-domestic waste, as it includes a mix of miscellaneous waste categories that require distinct management and recycling approaches.

Among the waste types, ash and sludge waste have the second lowest recycling rate, with only 6% of the total production of 2.65 MMT being recycled. The overall waste production from the textile and leather sector is 3.06 MMT, and although the amount of waste is less, it has been growing significantly in the modern era. Non-ferrous metal waste is produced in a relatively small amount of 2.19 MMT, but due to its higher recycling rate, approximately 89% of the total waste produced is recycled. Meanwhile, glass waste has a recycling rate of 16% and a total production of 1.48 MMT, making it the second least produced waste type. Scrap tires are the least produced waste type, with only 0.52 MMT of waste generated, of which almost 84% is recycled.

The insight produced by the descriptive study indicates that addressing the challenges in MSW is necessary. The suboptimal recycling rates of various waste types, such as stones, ceramics and rubber, ash and sludge, and textile waste emphasize the demand for novel forecasting techniques to direct the waste management strategies. Hence, to effectively deal with the escalating problems in handling waste, research on waste generation forecasting is vital.

The attribute "total waste generated" for each data type was measured in tons. The dataset was cleansed by converting duplicate names into unique ones and eliminating redundant entries. To facilitate easier access and computation, this attribute was normalized by dividing the column with all its values by one million. The MSW dataset was then prepared in a suitable format for time series forecasting, where the data is divided into input sequences and target values using a rolling window approach. When the lookback is specified, the function forms input sequences with 1 past time step, adding a time step dimension to each sample. As a result, the original 2D arrays representing the input data are transformed into 3D arrays with the additional time step dimension, which is required for certain forecasting models like GRU. The MSW dataset was then split into two segments: 14 years (2003-2017) allocated for training, and 6 years (2018-2023) designated for testing, following a 75:25 ratio. Since the rolling window approach was utilized, the last data point of the MSW dataset was omitted.

The architecture MSW-DeepStack model analyzes the most prevalent categories of MSW, and employs a fusion of Machine Learning (ML) and Deep Learning (DL) techniques. These methods were employed to extract meaningful patterns and insights from the data. The architecture of the MSW-DeepStack model is structured as a two-tiered framework. In particular, the performance of the model was significantly elevated by integrating the Grid Search Optimized XGBoost model (GSO-XGBoost) as opposed to the conventional XGBoost model [37-40]. In light of this enhancement, the initial phase entailed applying the GSO-XGBoost model to all fourteen MSW datasets. Thus, the GSO-XGBoost and GRU models were coupled for optimal performance in the first tier. The key output of this combination was the column stacking of GSO-XGBoost prediction and GRU prediction. This stacked columnar data was meticulously constructed to be used as the input for the subsequent tier. A meta-regressor, in this case a Random Forest (RF), is used in the second phase of the model. This RF model operates on the column-stacked predictions obtained from the first tier. The role of the RF model in this architecture is to utilize the combined predictive power of the initial GSO-XGBoost and GRU models, thereby refining the overall predictive accuracy of the MSW-DeepStack model. Overall, the innovative architecture of the MSW-DeepStack model capitalizes on the strengths of GSO-XGBoost, GRU, and RF models, resulting in a comprehensive solution for addressing the challenges of MSW quantity prediction. Each of the fourteen unique MSW datasets was run through the MSW-DeepStack model to improve the quantity of waste forecasting.

6.1 Evaluation metrics

The key performance indicators such as Mean Absolute Percentage Error (MAPE), Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Maximum Error (ME), and R-squared error (R²) were utilized to mathematically represent the above-mentioned KPIs (refer to Eqs. (6)-(10)).

$M A P E=\left[\frac{1}{N} \sum_{i=1}^N\left|\frac{y_a, \hat{y}_a}{y_a}\right|\right] * 100$            (6)

$M A E=\frac{1}{N} \sum_{i=1}^N\left|y_a, \hat{y}_a\right|$            (7)

$R M S E=\sqrt{\frac{1}{N} \sum_{i=1}^N\left(y_a, \hat{y}_a\right)^2}$            (8)

$M E=\operatorname{Max}_a\left|y_a, \hat{y}_a\right|$            (9)

$R^2=1-\frac{\sum_{i=1}^N\left(y_a, \hat{y}_a\right)^2}{\sum_{i=1}^N\left(y_a, \bar{y}\right)^2}$            (10)

where, ${{y}_{a}}$ is the actual value at time a, ${{\hat{y}}_{a}}$ is the predicted value at time a, $\bar{y}$ is the Mean target value, and N is the number of data points.

The observed pattern of ash and sludge waste reached a high in 2016, pointing to a considerable rise in trash creation that could most likely be linked to industrial activity or development. Since then, the pattern has stabilized, suggesting a constant output of trash (refer to Figure 6). Significantly, the GRU model had the highest MAPE, at 13%, of all models. The MAPE values of 3.75% and 3.45% were recorded by the Random Forest and GSO-XGBoost models, respectively, indicating moderate performance. The proposed MSW-DeepStack model performed exceptionally well, with a MAPE of 0.27%.

As shown in Figure 7, C&D debris increased over time, though this trend was slowed by the pandemic. RMSE values of 0.5070, 0.4255, and 0.5117 were attained by the GSO-XGBoost, Random Forest, and GRU models, respectively, indicating inadequate performance. In contrast, with an RMSE of 0.1225, the MSW-DeepStack model substantially outscored the other models. A similar upward trend was seen in Figure 8 which displays the ferrous metal wasted over time. ME values of 0.34 and 0.44 were seen in the GRU, Random Forest and GSO-XGBoost models, while the ME value of 0.1 was observed in the proposed model, which is a significant improvement. Tables 2 (a)-(d) represent results obtained from the employed models.

6.png

Figure 6. Ash and sludge waste

7.png

Figure 7. Construction and debris waste

8.png

Figure 8. Ferrous metal waste

Table 2. (a) Results of GSO-XGBoost

Table 2. (b) Results of GRU

Table 2. (c) Results of RF

Table 2. (d) Results of MSW-DeepStack

9.png

Figure 9. Food waste

10.png

Figure 10. Glass waste

11.png

Figure 11. Horticultural waste

Regarding food waste (see Figure 9), the MAPE values for the optimized XGBoost and Random Forest models hover around 9.3%, while the MAPE for the GRU model is closer to 7.5%. The proposed MSW-DeepStack model performed exceptionally well, with a much lower MAPE than the state-of-the-art model. As consumer and production practices shift, so do the seasonal patterns of glass debris which are represented in Figure 10. The MSW-DeepStack fared best of all used models, with an MAE of 0.0163, an RMSE of 0.0014, and a ME of 0.0022. While the GSO-XGBoost model showed a ME of 0.011, the RF model only dropped to 0.0109. The GRU model outperformed GSO-XGBoost and RF, however, it was still inferior to the proposed model as its ME was only 0.0044. In 2017, there was an uptick in the rate of discarded garden materials, but the trend has since leveled off (refer to Figure 11). The RMSE values for the GSO-XGBoost, RF, and GRU models ranged from 0.1384, 0.0448, and 0.0971, which is comparable but not remarkable. The proposed MSW-DeepStack model, on the other hand, was superior, with an RMSE of 0.0245. This further demonstrates its superiority in reducing discrepancies between forecasted and observed values.

The statistics shown in Figure 12 represent a seemingly random trend for non-ferrous metal waste. Compared to other models, both GSO-XGBoost and RF showed moderately higher ME values (0.06 and 0.069, respectively). With an ME of 0.029, the GRU model outperformed GSO-XGBoost and RF in terms of prediction accuracy. Meanwhile, the MSW-DeepStack model emerged as the most favourable choice, achieving a negligible ME of 0.0115. P&C waste illustrated in Figure 13 denotes a complex pattern in numbers. The proposed MSW-DeepStack model achieved the highest MAPE (4.2%) among the models appraised. The MAPE values for the GSO-XGBoost, GRU and RF models, on the other hand, were 6.5%, 6% and 7.8%, respectively.

12.png

Figure 12. Non-ferrous metal waste

13.png

Figure 13. Paper and cardboard

14.png

Figure 14. Plastics

15.png

Figure 15. Scrap tyres

Over the past decade, there have been distinct trends in the data on plastic trash (refer to Figure 14). The GSO-XGBoost and RF models performed moderately, with MAPE values of 12.7% and 11.4%, respectively, while the GRU model performed better, with a MAPE value of 6.2%. Particularly, the proposed MSW-DeepStack model performed the best of all models tested, with a MAPE of 1.7%. Although scrap tyres are discarded in modest numbers, the associated data structure is intricate (as illustrated in Figure 15). Compared to the other models, MSW-DeepStack performed the best, with a ME of 0.0033. GSO-XGBoost, RF, and GRU all produced ME values of 0.0114, 0.0117, and 0.0118, respectively.

16.png

Figure 16. Stones, ceramics & rubber

17.png

Figure 17. Textile and leather

18.png

Figure 18. Used slag

19.png

Figure 19. Wood and timber

Waste quantities of stones, ceramics, and rubber showed a linear trend until 2017, then declined, as displayed in Figure 16. When compared to the GSO-XGBoost, RF, and GRU models, the proposed model’s predictions for this type of waste had the lowest MAPE (8.6%), MAE (0.0398), RMSE (0.0094), and ME (0.0133). Waste materials from the textile and leather industries, as well as used slag and timber and wood, exhibited seasonal patterns (refer to Figures 17-19). The MSW-DeepStack model worked quite well, with lower RMSE values of 0.0246, 0.0152, and 0.0355, even for waste with such seasonal data patterns. Comparatively higher RMSE, MAPE, MAE, and ME values indicated that the performance of the remaining models was merely average.

The MSW-DeepStack model exhibits better accuracy and reduced error rates in several metrics, such as RMSE, MAPE, MAE, and ME. Reliability in capturing actual waste generation numbers is improved by a reduced MAE, which indicates that the model regularly makes smaller average errors. The decreased ME emphasizes forecasting stability by showing fewer extreme prediction deviations. Furthermore, a lower RMSE indicates less error variation, which aids in precisely simulating data oscillations. In the same way, the suggested model’s exceptionally low MAPE demonstrates its great accuracy in identifying genuine waste generation trends with little departure from actual values. Together, these metrics illustrate the MSW-DeepStack model’s robustness, making it a valuable tool for precise municipal solid waste management predictions.

Table 3. R² Value obtained by MSW-DeepStack model

Further, the R² values for the MSW-DeepStack model were superior for all used datasets, falling between 0.6 to 0.9 as represented in Table 3. The best R-squared value was 0.9, and it was found in the glass waste dataset, which exhibits a linear trend. The second-highest R² value of 0.86 was achieved for food waste, indicating strong predictive accuracy for this dataset. Both horticultural waste and plastic waste datasets secured the third-highest R² value of 0.84. In contrast, lower R² values were recorded for certain complex datasets. For example, paper and cardboard waste and textile and leather waste had R² values of 0.64 and 0.61, respectively, reflecting the complexity of their data patterns. Used slag and ferrous metal waste datasets achieved reasonably good R² values of 0.74 and 0.81, respectively. Similarly, scrap tyres and wood and timber datasets showed moderate performance with R² values of 0.7 and 0.72, respectively. Other datasets such as stones, ceramics, and rubber, and non-ferrous metals achieved R² values of 0.8 and 0.81, respectively, demonstrating stable performance. The ash and sludge and construction and debris datasets also recorded an R² value of 0.8, signifying good predictive results. Across a wide range of datasets with varying data patterns, the proposed approach consistently outperforms the state-of-the-art methods. As the next step, MSW-DeepStack model showed the least amount of error when applied to the MSW dataset. Subsequently, it was used to forecast the rate of waste production between 2023 and 2030.

20.png

Figure 20. Historical (2003-2023) and predicted data (2024-2030)

Figure 20 illustrates the historical and projected trends in garbage generation from 2003 to 2030. After a period of reduction during the pandemic, waste production has resumed its upward trend and shows no signs of slowing down until either consumption or production patterns undergo radical changes. C&D would surpass ferrous metal waste and P&C waste as the most common types of waste in the country by 2030. Notably, despite considerable efforts, it is anticipated that plastic trash will increase by almost 7 MMT by the year 2030. One further interesting conclusion drawn from MSW-DeepStack is that waste generation will grow between 2026 and 2028. Overall, the rate of waste formation is rising across the board, except for used slag.

Considering the essential roles played by waste reduction and responsible consumption in the Sustainable Development Goals (SDGs) proposed by the United Nations, attaining sustainable waste management is consistent with the SDGs. The results stress the need for comprehensive strategies that address waste minimization, recycling, circular economy, and specific regulatory changes. Collectively pursuing these measures is crucial in changing the course, marking a concrete step toward a future with sustainable waste practices and the actualization of the broader SDGs. Though Singapore utilizes incineration, waste-to-energy techniques and recyclability, the waste generation rate must be minimized to maintain a sustainable environment. Waste shipped to Semakau should be reduced by at least 30 percent daily until 2030 since the landfill is expected to reach capacity by 2035. It would aid in keeping Semakau Landfill in operation past the year 2035 [41]. Hence, these findings would aid the authorities in making the appropriate decisions with the predicted waste generation rate for the various categories of municipal solid waste.