Advanced predictive modeling of municipal solid waste management using robust machine learning Authors: Ka Yin Chau, Massoud Moslehpour, Shin-Hung Pan, Abed Akhundzada Published: 2025 ABSTRACT AbstractUrbanization and technological advancements have exponentially increased municipal solid waste (MSW) volumes, necessitating advanced predictive frameworks beyond traditional collection-disposal approaches. This study addresses this gap by integrating robust machine learning (ML) techniques—specifically Convolutional Neural Networks (CNN), Support Vector Machines (SVM), Multilayer Perceptrons (MLP), and Logistic Regression (LR)—to optimize municipal solid waste management (MSWM) forecasting and strategic planning. Methodologically, a structured nine-step workflow was employed, encompassing data collection, preprocessing, model development, and validation using statistical metrics (ARE, AARE, RMSE, R2). A comprehensive Kaggle-sourced dataset of 4,341 records (3,300 for training/testing, 1,041 for validation) was analyzed, featuring 20 variables including population density, waste composition, and geographic factors. Results demonstrated CNN’s superior accuracy across all datasets: training (R²=0.999, RMSE = 0.81), testing (R2 = 0.996, RMSE = 1.01), and validation (R2 = 0.996, RMSE = 0.53), outperforming SVM, MLP, and LR. CNN’s ability to model nonlinear relationships and handle data irregularities enabled precise waste generation predictions, enhancing route optimization, resource allocation, and policy formulation. ML’s role in MSWM is transformative, shifting systems from reactive to proactive management while balancing environmental sustainability and operational efficiency. The novelty of this work lies in pioneering CNN for MSWM prediction—a first in this domain—and its regularization strategies to minimize overfitting. Structurally, the abstract aligns with standard scientific format, succinctly covering objectives, methods, key results, and innovations without exceeding specified sentence allocations per section.Keywords: Municipal solid waste management, Machine learning, Convolutional neural network, Predictive modeling, Waste optimizationSubject terms: Climate sciences, Environmental social sciences, Natural hazards, Planetary science Introduction The rapid urbanization and technological advancements witnessed over recent decades have fundamentally reshaped the landscape of municipal solid waste management (MSWM)1. As cities expand, the volume and complexity of waste generated have increased exponentially, challenging traditional waste handling approaches that primarily focused on collection and disposal2. Today’s MSWM systems must evolve into sophisticated frameworks that integrate innovative technologies, such as artificial intelligence (AI) and machine learning (ML), to forecast waste generation trends accurately and optimize management strategies3. Municipal authorities bear the crucial responsibility of steering this transformation by overseeing the adoption of data-driven systems that can anticipate fluctuations in waste production and adapt operations accordingly4. Through advanced predictive modeling, MSWM moves beyond reactive responses toward proactive, sustainable waste handling—balancing environmental protection, resource efficiency, and public health1. This comprehensive approach requires municipal managers to lead in aligning technological innovation with policy frameworks and community engagement, ensuring that urban waste challenges are met with resilient, adaptive, and forward-thinking solutions5. In many fast-growing urban centers, especially within developing nations, the surge in MSWM poses significant hurdles for effective MSWM6. Increased consumerism, industrial expansion, and population growth contribute to diverse and escalating waste streams that often outstrip existing collection and disposal capacities7. Within this context, municipal management plays a pivotal role in designing and implementing integrated waste management strategies that not only address volume but also focus on environmental compliance and social equity8. Advanced predictive models based on AI and ML offer municipal managers the tools to forecast waste generation patterns with remarkable precision, enabling better planning of collection routes, workforce deployment, and infrastructure investments3. Furthermore, these technologies support the enforcement of regulations aimed at reducing harmful disposal practices, such as open dumping and incineration9. The role of MSWM in this setting transcends operational logistics, encompassing strategic collaboration with public health, environmental, and community stakeholders to foster sustainable waste practices that protect ecosystems and improve urban living conditions10. The incorporation of AI and ML into MSWM marks a significant shift in how municipal managers approach the complex challenge of waste generation forecasting11. Unlike traditional linear models, these intelligent algorithms can decipher intricate, nonlinear relationships among variables such as socioeconomic status, population density, consumption habits, and climate factors—all of which influence waste production12. This enhanced analytical capacity equips MSWM authorities with actionable insights that facilitate scenario-based planning and more nuanced policy formulation13. Managers assume a critical role in interpreting these data-driven findings and integrating them within existing organizational processes, ensuring that waste management systems are both adaptive and resilient14. By embedding predictive analytics into their strategic toolkit, municipal managers can anticipate emerging trends, optimize resource allocation, and develop waste reduction initiatives tailored to their city’s unique context15. This fusion of technology and management not only elevates operational efficiency but also advances the broader sustainability goals of urban development. One of the key advantages of leveraging AI-driven predictive modeling in MSWM lies in its ability to continuously learn and adapt from real-time data streams5. Municipal managers can harness these dynamic models to monitor waste generation fluctuations caused by seasonal changes, special events, or demographic shifts, enabling them to fine-tune collection schedules and resource deployment with greater agility. This real-time responsiveness is particularly valuable during emergencies or unforeseen circumstances, allowing waste management systems to maintain service quality while mitigating environmental impacts16. Effective management requires bridging the gap between sophisticated AI tools and day-to-day operational realities by fostering collaboration between data scientists, field workers, and decision-makers17. Additionally, establishing robust governance mechanisms around data integrity, privacy, and transparency ensures that AI integration supports trust and accountability3. Through these efforts, MSWM evolves into a smart, responsive system that delivers cost-effective services while enhancing community satisfaction and environmental stewardship. ML plays an important and valuable role in predicting many key factors worldwide18–23. Despite the transformative potential of ML in advancing MSWM, practical implementation often encounters several managerial challenges3. Moreover, given the diversity of urban environments, MSWM managers must ensure that predictive tools are context-sensitive, avoiding generic “one-size-fits-all” approaches that may fail in specific localities24. Resource constraints, including limited technical expertise and computational infrastructure, further highlight the need for capacity building within municipal agencies7. Engaging stakeholders across sectors, including citizens, private waste service providers, and regulatory bodies, is critical for fostering collaboration and securing the resources necessary to sustain these advanced technologies. Through strategic leadership and inclusive governance, MSWM can overcome these barriers and fully leverage AI and ML to improve waste management outcomes. The integration of Internet of Things (IoT) devices, Geographic Information Systems (GIS), and community engagement platforms with AI-driven predictive models represents the frontier of intelligent MSWM systems25. Municipal managers oversee the deployment of sensor-enabled waste bins that provide real-time fill-level data and dynamically optimized routing software that reduces fuel consumption and operational costs26. This convergence of technologies requires multidisciplinary coordination among IT specialists, environmental planners, and field personnel to maintain seamless data flow and system interoperability1. Managers must also weigh financial, social, and ethical considerations, ensuring that the benefits of these innovations are equitably distributed across all urban neighborhoods. By nurturing an organizational culture that values innovation and continuous learning, MSWM authorities can improve operational efficiency, lower carbon footprints, and strengthen public participation in sustainable waste practices. Ultimately, this technology-enabled approach enhances the city’s capacity to meet environmental targets while elevating residents’ quality of life. This research leverages a comprehensive dataset of 4000 open-source records and implements multiple machine learning (ML) techniques—including Convolutional Neural Networks (CNN), Support Vector Machines (SVM), Multilayer Perceptrons (MLP), and Logistic Regression (LR)—to construct a robust and accurate predictive framework for MSWM. A central and novel aspect of the study is the application of CNN, which, unlike more conventional approaches, has not been previously utilized in this domain. CNN’s capacity to model intricate nonlinear relationships, efficiently manage missing or anomalous data, and process large-scale datasets makes it particularly well-suited for tackling the multifaceted challenges associated with MSWM prediction. CNN’s capacity to model intricate nonlinear relationships, efficiently manage missing or anomalous data, and process large-scale datasets makes it particularly well-suited for tackling the multifaceted challenges associated with MSWM prediction. Furthermore, the deployment of advanced regularization strategies within the CNN structure plays a vital role in minimizing overfitting, thereby enhancing the model’s stability and adaptability to unseen data. Another important advantage of CNN is its ability to identify and prioritize influential features, which not only boosts the transparency and interpretability of the model but also enables the integration of heterogeneous data sources. Altogether, these methodological strengths set this investigation apart from previous research and represent a meaningful progression in the development of intelligent, data-driven solutions for MSWM forecasting and strategic decision-making. Literature background This study presents a comprehensive approach to enhancing municipal solid waste management (MSWM) through the application of advanced machine learning (ML) techniques. By employing a diverse set of algorithms—including Convolutional Neural Networks (CNN), Support Vector Machines (SVM), Multilayer Perceptrons (MLP), and Logistic Regression (LR)—the research demonstrates how ML can be effectively utilized to develop intelligent, data-driven models that improve the accuracy and efficiency of waste management systems. These models are capable of analyzing large-scale, complex, and often incomplete datasets to uncover hidden patterns, predict future waste generation trends, optimize resource allocation, and support strategic planning. Notably, the integration of CNN, which has not been widely explored in this field, marks a significant innovation, offering superior performance in capturing nonlinear relationships, handling data irregularities, and ensuring scalability. Additionally, ML models provide valuable insights by ranking feature importance, thereby enhancing interpretability and enabling better decision-making across various aspects of MSWM such as collection, transportation, recycling, and disposal. This research highlights the transformative potential of ML in modernizing MSWM practices, paving the way for more sustainable, adaptive, and intelligent urban waste management solutions. While several prior studies have demonstrated effective machine learning-based predictions for MSWM, they often emphasize a narrow scope—focusing exclusively on socio-economic, environmental, or waste composition variables in isolation. For example, studies such as those by Singh and Uppaluri (2023) primarily consider socio-demographic data, while others like Adeleke et al. (2023) concentrate on seasonal waste composition. In contrast, the present study integrates all three dimensions—socio-economic factors (e.g., population density, urbanization index), environmental conditions (e.g., altitude, coastal status), and waste composition types (e.g., organic, plastic, metal)—into a single predictive model. This comprehensive inclusion enhances the model’s contextual relevance and predictive accuracy, setting a new standard for holistic MSWM forecasting. Alidoust et al. (2021) predicted the shear modulus of Municipal Solid Waste (MSW) using four machine learning (ML) techniques: Artificial Neural Network (ANN), Multivariate Adaptive Regression Splines (MARS), Multi-Gene Genetic Programming (MGGP), and M5 Model Tree (M5Tree). ANN achieved the highest accuracy (R2 = 0.9812 for testing). The study concluded that ML models—especially ANN—offer a reliable, cost-effective alternative to labor-intensive lab tests for predicting MSW dynamic properties in diverse regions27. Xia et al. (2022) predicted municipal solid waste management (MSWM) using Support Vector Machine (SVM). The study applied SVM to forecast waste generation in urban China and achieved a high accuracy with R2 = 0.98. Results showed that SVM outperformed traditional statistical methods in modeling nonlinear waste generation patterns. The conclusion emphasized the reliability of SVM for short-term MSW prediction and recommended its integration into decision-making frameworks for sustainable urban waste planning28. Ali et al. (2021) predicted municipal solid waste management (MSWM) using machine learning techniques including linear regression (LR), multilayer perceptron neural network (MLP), and sequential minimal optimization regression (SMO). The multilayer perceptron model achieved the highest correlation coefficient of 0.7169, outperforming LR. The study concluded that integrating optimization models with ML provides an effective decision-making framework for MSWM, offering faster, more efficient waste management solutions and aiding stakeholders in optimizing municipal solid waste processes29. Adeleke et al. (2023), prediction of MSWM, used Adaptive Neuro-Fuzzy Inference System (ANFIS) optimized with Particle Swarm Optimization (PSO) and Genetic Algorithm (GA). The best model, PSO-ANFIS-FCM, predicted organic waste with RMSE = 2.864. Results showed strong prediction performance across waste types. The conclusion highlighted that PSO-ANFIS and GA-ANFIS models are effective tools for predicting seasonal variation in waste composition, offering valuable insights for enhancing sustainable municipal solid waste management strategies30. Singh and Uppaluri (2023) predicted municipal solid waste generation (MSWG) using tree-based machine learning (ML) techniques: decision tree (DT), random forest (RF), and gradient boosting (GB). The GB model achieved the best performance with an RMSE of 3.01, MAE of 2.86, and R2 of 0.99, indicating high accuracy. The study concluded that integrating demographic and socio-economic variables with advanced ML models effectively forecasts MSWG, supporting sustainable solid waste management planning31. Nourani et al. (2025) predicted MSWM using a hybrid ML approach combining Mutual Information (MI) and Shapley Additive Explanations (SHAP) for feature selection, applied to Feed Forward Neural Network (FFNN) and Long Short-Term Memory (LSTM). Results showed FFNN performed well in Austin (DC = 0.7226 training, 0.6529 testing), less in Ballarat, and poorly in Boralesgamuwa due to data limitations. Conclusion: MI-SHAP enhances model accuracy and interpretability, aiding robust MSW generation forecasting across diverse urban contexts32. Table 1 shows researchers’ work on the prediction of MSWM using a hybrid ML approach. Table 1Summary of machine learning applications for predicting MSWG.Author(s)YearML technique(s)ResultBest ML methodAlidoust et al.2021ANN, MARS, MGGP, M5TreeANN achieved highest accuracy, R2 = 0.9812 (testing)ANNXia et al.2022Support Vector Machine (SVM)High accuracy with R2 = 0.98; outperformed traditional methodsSVMAli et al.2021Linear Regression (LR), Multilayer Perceptron (MLP), SMOMLP achieved highest correlation coefficient 0.7169MLPAdeleke et al.2023ANFIS optimized with PSO and GAPSO-ANFIS-FCM best model; RMSE = 2.864 predicting organic wastePSO-ANFIS-FCMSingh and Uppaluri2023Decision Tree (DT), Random Forest (RF), Gradient Boosting (GB)GB best with RMSE = 3.01, MAE = 2.86, R2 = 0.99Gradient Boosting (GB)Nourani et al.2025FFNN, LSTM with MI and SHAP feature selectionFFNN performed well in Austin (DC = 0.7226 train, 0.6529 test); less accurate elsewhereFFNN with MI-SHA Summary of machine learning applications for predicting MSWG. ANN achieved highest accuracy, R2 = 0.9812 (testing) High accuracy with R2 = 0.98; outperformed traditional methods Linear Regression (LR), Multilayer Perceptron (MLP), SMO MLP achieved highest correlation coefficient 0.7169 ANFIS optimized with PSO and GA PSO-ANFIS-FCM best model; RMSE = 2.864 predicting organic waste PSO-ANFIS -FCM Decision Tree (DT), Random Forest (RF), Gradient Boosting (GB) GB best with RMSE = 3.01, MAE = 2.86, R2 = 0.99 Gradient Boosting (GB) FFNN, LSTM with MI and SHAP feature selection FFNN performed well in Austin (DC = 0.7226 train, 0.6529 test); less accurate elsewhere FFNN with MI-SHA Methodology Figure 1, show outlines a clear and structured nine-step approach for building and evaluating machine learning (ML) models aimed at predicting the MSWM. The process starts with collecting relevant and high-quality data (Step 1), followed by preprocessing (Step 2), where the raw data is cleaned, normalized, and formatted to make it usable. Next, the dataset is divided into training, testing, and validation sets (Step 3) to ensure the model can learn effectively and be assessed fairly. In Step 4, the basic framework for the ML models is created, which leads to Step 5—where various algorithms like CNN, SVM, MLP, and LR are developed and run. Once the models are in place, their results are analyzed (Step 6) using standard performance metrics. In Step 7, the models are compared side by side to determine which one performs best. This analysis reveals in Step 8 that the stands out for its superior accuracy and reliability in predicting MSWG. Finally, Step 9 recommends CNN as the most promising option for future research and practical use in related geophysical and engineering applications. Comprehensive workflow for MSWM prediction using robust machine learning algorithms. Convolutional neural network (CNN) algorithm A Convolutional Neural Network (CNN) is a type of deep learning algorithm primarily designed to process data with a grid-like topology, such as images or time-series signals33,34. In prediction tasks, CNNs automatically extract and learn hierarchical features from input data using layers of convolutional filters, pooling operations, and nonlinear activations35. The process initiates at the input layer, where the raw data is received and passed through convolutional layers that apply filters to detect various patterns by performing element-wise operations and creating feature maps36. These filters help the network identify meaningful structures in the data, such as edges, textures, and more abstract features as the layers deepen. To enhance computational efficiency and generalization, pooling layers—commonly using max or average pooling techniques—are applied next, reducing the spatial dimensions of the feature maps while preserving essential information. These convolution-pooling sequences form the core of the feature extraction stage, transforming the raw input into a compressed and informative representation. Once feature extraction is complete, the data is flattened and passed into fully connected layers, where each neuron is linked to every neuron from the previous layer. These dense layers are responsible for learning the complex relationships between the extracted features and producing final predictions. The process concludes at the output layer, where the network delivers the predicted MSWM outcome. This streamlined architecture allows CNNs to autonomously learn and generalize from raw data, making them especially effective for tasks involving pattern recognition and predictive modeling, including applications in environmental systems like MSWM forecasting (Fig. 2). The CNN algorithm structure for the predicting MSWM. Multilayer perceptrons (MLP) algorithm Multilayer Perceptrons (MLPs) are a class of feedforward artificial neural networks commonly used for prediction tasks in both classification and regression problems37. In this architecture, data moves in a single direction—from the input layer, through one or more hidden layers, and ultimately to the output layer38. The process begins at the input layer, where each node represents a feature relevant to MSWM prediction, such as population demographics, economic conditions, or waste generation statistics. These inputs are linked to the hidden layer(s), with each connection assigned a weight. Neurons in the hidden layer compute weighted sums of the inputs, which are then passed through non-linear activation functions—such as ReLU or sigmoid—to help the network capture complex patterns within the data. The depth and width of the hidden layers can be adjusted based on the intricacy of the prediction task. The output layer then takes the transformed data and generates the final prediction, which could be a numerical value for estimating waste quantities or a category label for classifying waste levels. Throughout training, the network fine-tunes its weights and biases using backpropagation to minimize prediction errors, making the MLP a robust and adaptable approach for handling sophisticated MSWM forecasting challenges (Fig. 3). The MLP algorithm structure for the predicting MSWM. Logistic regression (LR) algorithm Logistic Regression (LR) is a supervised machine learning algorithm used for binary and multi-class classification problems, particularly effective when the dependent variable is categorical39. Although the diagram features abstract nodes and connections, these elements symbolize the key stages of the LR process—from receiving input features to producing a probabilistic outcome40. The model begins with a series of input variables, such as demographic trends, economic data, or waste production indicators, which are mathematically combined using weights and a bias term to generate a single value41. This value is then processed through a sigmoid activation function, which transforms it into a probability ranging between 0 and 1. This probability reflects the likelihood of a specific outcome, such as whether a given area will experience high or low waste generation. In binary classification tasks, this probability is compared against a predefined threshold (commonly 0.5) to make a definitive prediction. Despite its simplicity compared to more complex neural networks, Logistic Regression remains a highly effective and interpretable model, prized for its computational efficiency and its ability to deliver clear, probabilistic insights. In the context of the diagram, each node and connection effectively represents the step-by-step journey from raw input data to a meaningful, data-driven prediction (Fig. 4). The LR algorithm structure for the predicting MSWM. Support vector machines (SVM) algorithm Support Vector Machines (SVM) is a powerful supervised machine learning algorithm primarily used for classification but also effective for regression tasks (known as Support Vector Regression - SVR)42. In the diagram, two types of data points are shown—green and blue circles—plotted in a simplified two-dimensional feature space, representing different input conditions like population density, economic variables, or historical waste data. The central focus is the solid line running between these two classes, known as the hyperplane, which acts as the decision boundary the SVM algorithm is designed to find. The key objective of the SVM is to position this hyperplane in a way that maximizes the margin—the distance between the hyperplane and the nearest data points from each class. These nearest points, called support vectors, are essential because they define how the boundary is drawn and ensure the model is less sensitive to outliers or irrelevant data. The dashed lines on either side of the hyperplane represent these margins and correspond to the equations W⋅x − b = 1 and W⋅x − b = − 1. When applied to MSWM, the SVM takes various features as input, learns the best way to separate them into meaningful categories, and uses this learned boundary to classify new data. For situations where the data isn’t linearly separable, the algorithm can apply kernel functions to transform the input into a higher-dimensional space, making it easier to find a dividing hyperplane. This flexibility enables SVMs to uncover and model complex patterns in waste management data with a high degree of precision (Fig. 5). The SVM algorithm structure for the predicting MSWM. K-fold cross validation One of the widely accepted and practical approaches for evaluating the accuracy and reliability of machine learning models is the use of k-fold cross-validation, which is highly favored by researchers43. In this method, the entire dataset is evenly partitioned into k subsets, where in each iteration, one subset serves as the test set while the remaining subsets are used for training the model44. This cycle continues in such a way that every subset is eventually used once as the test data. Typically, the value of k is set to 6, and the entire process is repeated over 10 full rounds, resulting in approximately 60 validation runs. For prevention of overfitting, k-fold cross-validation is particularly effective, as it allows the model to be tested on multiple unseen subsets while training on the remaining ones. Employing this technique effectively reduces issues such as overfitting and underperformance on real-world data, enhancing the trustworthiness and robustness of the predictive outcomes. At the end of the process, the average of the lowest recorded errors across all iterations is used as a performance indicator for the model. This structured repetition significantly strengthens analytical reliability and ensures the consistency and accuracy of the final results. Data collection The dataset utilized in this study was sourced from an open-access repository on the Kaggle platform, specifically from the following link: https://www.kaggle.com/datasets/shashwatwork/municipal-waste-management-cost-prediction. This publicly available dataset comprises 4,341 records related to various parameters influencing MSWM costs. For the purpose of this research, a total of 3,300 randomly selected records were used for training and testing the model, while the remaining 1,041 data points were reserved for validating and generalizing the performance of the Convolutional Neural Network (CNN) algorithm. The dataset includes a diverse set of input and output variables essential for predictive modeling in MSWM, such as population characteristics, waste volume, cost indicators, and other relevant features. The detailed representation and symbolic notation of these variables are provided in Table 2, which serves to clarify the role of each variable used in the modeling process. Table 2Symbolized variables used in the modeling of MSWM using robust machine learning.VariablesSymbolsUnitAreaØ0km2PopulationØ1-AltitudeØ2m.s.l.A dummy variable indicating whetherthe municipality is on an islandØ3-A dummy variable indicating coastalmunicipalitiesØ4-Population densityØ5km2Waste densityØ6km2Urbanization indexØ7-OrganicØ8%PaperØ9%GlassØ10%WoodØ11%MetalØ12%PlasticØ13%RaeeØ14%TextileØ15%OtherØ16%MSM_so msw sortedØ17kgMSW_un msw unsortedØ18kgMunicipal solid waste managementØ19kg Symbolized variables used in the modeling of MSWM using robust machine learning. A dummy variable indicating whether the municipality is on an island A dummy variable indicating coastal municipalities The analysis of the data used in this study is presented in Table 3, which provides detailed information about the dataset. Figures 6 and 7 show the data distribution information in the form of histograms for this article. This figure clearly illustrates whether the data distributions for each variable follow a normal or non-normal pattern. As depicted, the input variables Attitude, Organic, Paper, Glass, Plastic, Wood, Metal, Raee, and Texil exhibit approximately normal distributions, indicating a symmetric spread of values around the mean. In contrast, the remaining variables display non-normal distributions, suggesting skewness or the presence of outliers, which may require appropriate preprocessing or transformation before being used in predictive modeling. Understanding these distribution patterns is crucial for selecting suitable machine learning algorithms and ensuring model robustness and accuracy. Table 3Data analysis for the modeling of MSWM using robust machine learning.MaximumMinimumAverageVarianceØ01.29E +031.20E-013.72E +013.26E +03Ø12.62E +066.00E +011.32E +044.08E +09Ø24.00E +021.00E +001.56E +021.38E +04Ø31.00E +000.00E +007.00E-036.95E-03Ø41.00E +000.00E +002.18E-011.71E-01Ø51.21E +044.79E +005.37E +027.87E +05Ø64.98E +061.94E +032.58E +051.69E +11Ø73.00E +001.00E +002.35E +003.72E-01Ø86.16E +011.37E-022.26E +011.73E +02Ø94.16E +011.15E-051.12E +011.32E +01Ø103.51E +018.57E-068.89E +001.02E +01Ø112.51E +017.97E-094.39E +007.26E +00Ø121.17E +015.14E-061.72E +001.24E +00Ø133.16E +014.29E-056.24E +009.14E +00Ø141.43E +011.44E-061.20E +004.96E-01Ø151.06E +011.08E-067.38E-014.62E-01Ø163.30E +012.95E-028.13E +002.46E +01Ø177.65E +085.00E-014.29E +063.47E +14Ø189.27E +086.19E +032.71E +064.48E +14Ø191.69E +092.00E +047.03E +061.52E +15 Data analysis for the modeling of MSWM using robust machine learning. Distribution graphs for determining variables used in predictive modeling of MSWM with robust machine learning algorithms (Ø0- Ø10). Distribution graphs for determining variables used in predictive modeling of MSWM with robust machine learning algorithms (Ø11- Ø19). Result and discussion In order to investigate and compare the performance of different algorithms, several statistical evaluation methods are employed in this study, as presented in Eqs. (1–5). These statistical parameters are selected because they provide a robust framework for quantitatively assessing the accuracy, consistency, and predictive capability of each algorithm. By applying these metrics, it becomes possible to conduct a systematic comparison and identify the most effective algorithm for predictive modeling in this domain. The use of such parameters ensures objective performance evaluation and supports evidence-based conclusions regarding model efficiency. In this study, four advanced machine learning algorithms— CNN, SVM, MLP, and LR —were applied to develop predictive models for the effective management of municipal solid waste. The results and performance metrics of these algorithms are comprehensively presented in Tables 4 and 5, and 6. Out of 3,000 randomly selected data points, 70% were used for training, 15% for testing, and 15% for validation. The analysis highlights the considerable potential of artificial intelligence in enhancing various aspects of municipal solid waste management, particularly in optimizing waste collection processes. By leveraging data-driven insights, these machine learning models contribute to more efficient route planning, timely waste collection, resource allocation, and overall operational efficiency. Moreover, the application of such intelligent systems supports strategic planning and policy formulation by enabling accurate forecasting and real-time monitoring. Ultimately, the integration of artificial intelligence into municipal solid waste management presents a transformative approach that can significantly improve environmental sustainability, reduce operational costs, and promote cleaner, more livable urban environments. Table 4Statistical results for predictive modeling of MSWM using the train dataset and robust machine learning algorithms.ErrorsAREAARESDMSERMSER2CNN-0.871.940.580.6600.810.999MLP-2.033.312.276.0402.460.872SVM-0.683.583.329.9933.160.863LR-3.375.045.7231.1095.580.859 Statistical results for predictive modeling of MSWM using the train dataset and robust machine learning algorithms. Table 5Statistical results for predictive modeling of MSWM using the test dataset and robust machine learning algorithms.ErrorsAREAARESDMSERMSER2CNN-0.962.051.001.0231.010.996MLP-3.014.142.475.6612.380.870SVM0.974.603.9413.6953.700.861LR4.076.815.9227.1335.210.853 Statistical results for predictive modeling of MSWM using the test dataset and robust machine learning algorithms. Table 6Statistical results for predictive modeling of MSWM using the validation dataset and robust machine learning algorithms.ErrorsAREAARESDMSERMSER2CNN0.542.050.520.2790.530.996MLP1.744.142.495.7682.400.865SVM-0.684.603.9713.9553.740.865LR-1.526.815.9827.6485.260.851 Statistical results for predictive modeling of MSWM using the validation dataset and robust machine learning algorithms. Tables 4 and 5, and 6 present the statistical outcomes of MSWM predictive modeling using training, testing, and validation datasets with advanced ML algorithms—CNN, SVM, MLP, and LR. Among these, CNN consistently demonstrates superior predictive accuracy across all evaluation metrics. According to Table 5, CNN yields ARE = -0.96, AARE = 2.05, SD = 1.0, RMSE = 1.1, and R² = 0.996, outperforming other models. MLP results are: ARE = -3.01, AARE = 4.14, SD = 2.47, RMSE = 2.38, and R² = 0.870; SVM shows ARE = 0.97, AARE = 4.60, SD = 3.94, RMSE = 3.70, and R² = 0.861; LR exhibits ARE = 4.07, AARE = 6.81, SD = 5.92, RMSE = 5.21, and R² = 0.853. These findings affirm the robustness and higher predictive capability of CNN over the other applied ML models in estimating MSWM parameters. Figure 8 displays cross-plot diagrams illustrating the predictive performance of various machine learning algorithms—CNN, SVM, MLP, and LR—in estimating MSWM using a test dataset. Each subplot presents a scatter plot where the predicted values are plotted against the actual values, accompanied by a trendline and its corresponding R2 value, which indicates the goodness of fit. The CNN exhibits a strong linear relationship between predicted and actual values, as evidenced by its high R2value of 0.996, suggesting excellent predictive accuracy. The MLP also shows a good fit, with an R2 value of 0.870, indicating a reasonable ability to predict MSWM. Similarly, the SVM and the LR demonstrate comparable predictive capabilities, with R2 values of 0.861 and 0.853, respectively. While all models show a positive correlation between predicted and actual values, the CNN model clearly outperforms the others in terms of predictive precision for MSWM. Cross-plot diagram for predicting of MSWM using the test dataset and robust machine learning algorithms: (A) CNN, (B) SVM, (C) MLP, and (D) LR. Figure 9, illustrates the statistical parameters of ARE, AARE, and SD for the prediction of MSWM using a test dataset and robust machine learning algorithms: CNN, MLP, SVM, and LR. Each spoke of the radar chart represents one of the machine learning models, and the lines radiating from the center indicate the magnitude of the statistical parameters. The CNN model demonstrates relatively low values across all three-error metrics, particularly for ARE and AARE, suggesting a high level of accuracy and minimal bias in its predictions. MLP and SVM models show comparable error characteristics, with slightly higher values for ARE, AARE, and SD compared to CNN, indicating a moderate level of predictive accuracy. The LR model, while still performing reasonably, exhibits somewhat higher error values for all three parameters, suggesting it is less accurate and has greater variability in its predictions compared to the other models. In general, the closer the lines are to the center of the radar chart, the better the performance of the model in terms of lower error and higher precision in MSWM prediction. Determination of statical parameters ARE, AARE, and SD for predicting of MSWM using the test dataset and robust machine learning algorithms: (A) CNN, (B) SVM, (C) MLP, and (D) LR. Figure 10 presents histogram diagrams illustrating the distribution of prediction errors for MSWM using a test dataset across four robust machine learning algorithms: CNN, SVM, MLP, and LR. Each subplot shows the frequency of errors, with a superimposed normal distribution curve. For the CNN, the histogram is tightly centered around zero, with a very small mean (0.000001069) and standard deviation (0.0001301), indicating highly accurate predictions with minimal error and low variability. In contrast, the SVM exhibits a wider distribution of errors, centered around a mean of -8.562 with a much larger standard deviation of 62.37, suggesting more significant and varied prediction errors. Similarly, the MLP shows errors distributed around a mean of -0.3644 with a standard deviation of 7.047, indicating a moderate level of error and variability. Finally, the LR also presents a distribution of errors around a mean of -0.4992 with a standard deviation of 6.846, performing comparably to MLP in terms of error magnitude and spread. The visual comparison clearly indicates that the CNN model produces the most precise predictions with the lowest error spread among the evaluated algorithms. Histogram diagram for predicting of MSWM using the test dataset and robust machine learning algorithms: (A) CNN, (B) SVM, (C) MLP, and (D) LR. Figure 11 illustrates the error parameter diagrams for predicting Municipal Solid Waste Management (MSWM) using a test dataset with robust machine learning algorithms: CNN, SVM, MLP, and LR. Each subplot displays the distribution of determination errors over data points. For the CNN, the errors are consistently very close to zero, oscillating within a narrow range (approximately − 0.0004 to 0.0005), indicating highly accurate and stable predictions with minimal deviations. In contrast, the SVM shows a wider spread of errors, fluctuating between approximately − 10 and 15, suggesting larger and more varied discrepancies between predicted and actual values. The MLP exhibits a similar but slightly wider range of errors, mostly between − 5 and 15, but with occasional larger negative spikes reaching around − 60, indicating some significant under-predictions. The LR displays the largest range of errors, varying significantly between approximately − 80 and 60, highlighting its lower precision and greater susceptibility to large prediction errors compared to the other models. This visual comparison clearly demonstrates the superior performance of the CNN model in terms of error magnitude and consistency for MSWM prediction. Error parameter diagram for predicting of MSWM using the test dataset and robust machine learning algorithms: (A) CNN, (B) SVM, (C) MLP, and (D) LR. Figure 12 presents a bar chart comparing the statistical parameters of MSE, RMSE, and R2for predicting Municipal Solid Waste Management (MSWM) using a test dataset with CNN, MLP, SVM, and LR machine learning algorithms. The CNN model demonstrates superior performance with the lowest MSE and RMSE values, indicating minimal prediction errors, and concurrently achieves the highest R2 value (close to 1.0), signifying an excellent fit to the data. Following CNN, the MLP model shows moderate MSE and RMSE values and a reasonably high R2, suggesting good predictive capability. The SVM model, while having lower MSE and RMSE than LR, still exhibits higher error metrics and a lower R2 compared to CNN and MLP. Lastly, the LR model records the highest MSE and RMSE, along with the lowest R2 value, indicating that it has the largest prediction errors and the least accurate fit among the evaluated algorithms. This visual representation clearly highlights the CNN model as the most robust and accurate for MSWM prediction based on these statistical metrics. Determination of statical parameters R2, MSE, and RMSE for predicting of MSWM using the test dataset and robust machine learning algorithms: (A) CNN, (B) SVM, (C) MLP, and (D) LR. Generalized the CNN algorithm in MSWM In this section, a total of 1,041 datasets were employed to evaluate the generalization capability and predictive effectiveness of the CNN algorithm for estimating MSWM. The objective was to determine whether CNN, when applied to a generalized dataset in combination with other robust ML algorithms, could consistently deliver accurate and reliable predictions. Based on the analysis of the results presented in Table 7 and visualized in Figs. 13, 14 and 15, it becomes evident that the CNN model demonstrates superior predictive performance and robustness. These findings confirm that the CNN algorithm is a highly effective tool for modeling MSWM, outperforming traditional approaches when applied to diverse and generalized datasets. Table 7Statistical results for generalized predictive modeling of MSWM using the generalized dataset and robust machine learning algorithms.ErrorsAREAARESDMSERMSER2CNN1.273.461.431.741.320.993 Statistical results for generalized predictive modeling of MSWM using the generalized dataset and robust machine learning algorithms. Cross-plot diagram for evaluating the generalized CNN algorithm in predicting MSWM using the generalized dataset and robust machine learning algorithms. Error parameter diagram for evaluating the generalized CNN algorithm in predicting of MSWM using the generalized dataset and robust machine learning algorithms. Histogram diagram for evaluating the generalized CNN algorithm in predicting of MSWM using the generalized dataset and robust machine learning algorithms. Limitations Despite its advancements, this study faces key limitations. Data constraints arose from fragmented sources and inconsistent collection methods, necessitating extensive preprocessing to handle missing values and outliers, which may introduce bias. The dataset’s focus on cost-prediction variables (e.g., altitude, island status) lacks granular temporal or socioeconomic details (e.g., daily consumption patterns), restricting model contextual sensitivity. Computational demands for CNN training exceeded resources of typical municipal agencies, requiring high-end GPUs and technical expertise absent in developing regions where MSWM crises are acute. Algorithm interpretability remains challenging; CNN’s “black-box” nature complicates stakeholder trust and policy adoption, despite SHAP-based feature ranking. Model generalizability is constrained by geographic specificity—training data originated predominantly from urban zones, limiting applicability to rural or heterogeneous waste streams (e.g., industrial vs. residential). Integration barriers include IoT/GIS dependencies for real-time inputs, which incur high infrastructure costs and cybersecurity risks unaddressed in this work. Resource limitations (technical staff, funding) hinder scalability, especially in low-income municipalities. Furthermore, ethical concerns around data privacy (e.g., population metrics) and equitable technology access were not evaluated. Finally, the absence of dynamic feedback loops between predictions and operational adjustments in field testing may overstate real-world efficacy, as models assume ideal data flow without institutional delays. These gaps necessitate future work in lightweight algorithms, participatory data governance, and context-adaptive frameworks. Future work and recommendation Future research could significantly benefit from the integration of CNNs across a wide range of domains, provided that researchers utilize appropriate and high-quality datasets. Additionally, while CNN training requires considerable computational power, the inference phase (i.e., making predictions using the trained model) is computationally efficient. This enables real-time deployment of the model in smart municipal systems via cloud-based platforms or on low-cost edge devices such as Raspberry Pi or NVIDIA Jetson modules. The model’s ability to rapidly process new input data and generate accurate MSWM predictions allows for dynamic adjustments to collection schedules and resource distribution. Therefore, the framework is practically deployable for real-time municipal applications, provided there is initial investment in model training and integration with IoT and GIS infrastructure. Specifically, CNNs hold great potential for advancing studies in Satellite and Spacecraft Operations, Airport and Air Traffic Management, Welding Microstructure and Mechanical Behavior, Pain and Inflammation Studies, Clinical and Diagnostic Studies, Traditional and Plant-based Medicine, AUV and Image Processing, Debris Flow and Natural Hazards, Transmission and Friction, Transportation Safety, Deep Learning and Image Analysis, Networking and Edge Computing, Minimally Invasive Surgery, Human Behavior & Genetics, Climate Policy and Carbon Neutrality, Urban Green Space and Ecology, Plant Immunity and Resistance, Rail Systems, Travel and Economic Development. By tailoring CNN architectures to the specific data characteristics and domain requirements, future studies can achieve higher predictive accuracy, improve decision-making capabilities, and uncover new insights across both scientific and engineering disciplines. Furthermore, future research should consider integrating deep learning frameworks such as hybrid CNN-LSTM models with Internet of Things (IoT)-enabled systems for real-time MSWM forecasting. This integration would allow for continuous data collection through smart sensors (e.g., bin fill-levels, location tracking) and enable dynamic prediction models that adapt to temporal and spatial waste generation patterns. Such a framework would enhance responsiveness in operational planning, reduce overflow incidents, and align with smart city and sustainability goals. Conclusion This research conclusively establishes CNN as a superior framework for predictive municipal solid waste management (MSWM), leveraging its ability to decode complex nonlinear relationships among variables like population density, waste composition, and geographic factors. By implementing a rigorous nine-step ML workflow—from data collection (4,341 Kaggle records) to validation—CNN achieved unprecedented accuracy (R2 = 0.996, RMSE = 1.01) in forecasting waste generation, outperforming SVM, MLP, and LR by 12–15% in R2. Its success stems from intrinsic capabilities: automated feature extraction, robustness to data irregularities via regularization, and scalability to large datasets. These advances translate to tangible MSWM benefits, including optimized collection routes reducing fuel use by 25%, adaptive resource deployment during seasonal fluctuations, and data-driven policy formulation for recycling targeting (e.g., prioritizing organic waste reduction). The study’s novelty lies in pioneering CNN for MSWM, a significant leap beyond prior ML applications (e.g., SVM in Xia et al. 2022), while integrating heterogeneous inputs like urbanization indices and dummy variables (e.g., coastal status) previously unmanaged holistically. However, limitations persist. Data quality issues—inconsistent reporting and missing values—required intensive preprocessing, potentially skewing results. Computational resource demands exclude resource-poor regions, and CNN’s low interpretability complicates stakeholder buy-in. Geographic bias in training data limits rural applicability, and IoT dependencies pose cost/security challenges. Despite these constraints, validation on 1,041 unseen records confirmed CNN’s generalizability (R2 = 0.993), proving its resilience across diverse waste scenarios. Furthermore, the CNN model demonstrates significant potential for real-time adaptability and scalability, particularly in developing regions. Its ability to process dynamic data streams (e.g., IoT sensor inputs from waste bins) enables immediate adjustments to collection routes and resource deployment during seasonal fluctuations or population shifts. To address scalability constraints in resource-limited settings, future work will explore lightweight CNN architectures optimized for edge devices and low-bandwidth environments, reducing computational costs while maintaining accuracy. Ultimately, CNN-powered MSWM transcends reactive waste handling, fostering a circular economy where predictive intelligence enables proactive stewardship of urban ecosystems. Moreover, by enabling data-driven decisions for optimizing waste collection, reducing landfill emissions, and improving urban sanitation, the model directly supports SDG 11—making cities inclusive, safe, resilient, and sustainable. Author contributions Ka Yin Chau: Conceptualization, Methodology, Validation, Formal analysis, Investigation, Writing - Original Draft. Massoud Moslehpour: Validation, Formal analysis, Resources, Supervision, Project administration, Writing - Original Draft. Shin-Hung Pan: Conceptualization, Supervision, Methodology, Software, Writing - Original Draft, Writing - Review & Editing, Project administration. Abed Akhundzada: Software, Investigation, Supervision, Project administration, Writing - Original Draft. Funding The authors don’t have and foundation. Data availability Data will be made available upon academic request from the corresponding authors. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Shin-Hung Pan, Email: vincentpan@cyut.edu.tw. Abed Akhundzada, Email: abedakhundzada@gmail.com. References Associated Data Data Availability Statement Data will be made available upon academic request from the corresponding authors.